refactor: simplify inference engine and backend dispatch

- merge _generate_streaming/_generate_non_streaming into single _generate() with stream flag
- delete dead GenerationRequest class and generate_with_request method
- inline _next_token helper into generate_async
- replace flash-attn double-checked locking with functools.lru_cache
- extract _write_and_gather_kv helper shared by TorchNative/FlashAttn backends
- inline _kv_cache_is_contiguous into its sole call site in FlashAttnBackend
- change default backend priority from flash>cuda>torch to cuda>flash>torch
- add ASTR_BACKEND env var to override default backend at resolve time
- add supports_graph() static method to AttentionBackend ABC, override in CudaBackend
- replace isinstance(get_backend(), CudaBackend) with get_backend().supports_graph() in executor
- add torch.cuda.is_available() guard to CudaBackend.supports()

.dockerignore

@@ -4,6 +4,8 @@
 # Allow necessary files
 !astrai/
 !scripts/
-!assets/
+!docs/
+!csrc/
+!setup.py
 !pyproject.toml
 !README.md

.github/workflows/release.yml

@@ -0,0 +1,100 @@
+name: Release
+
+on:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  build-pure:
+    name: Build pure-Python wheel
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Build wheel (no CUDA)
+        run: |
+          pip wheel . --no-deps -w dist/
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: pure-wheel
+          path: dist/*.whl
+          if-no-files-found: error
+
+  build-cuda-linux:
+    name: Build CUDA wheel (Linux, ${{ matrix.cuda_tag }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda_tag: "cu128"
+            cuda_ver: "12.8.0"
+          - cuda_tag: "cu130"
+            cuda_ver: "13.0.0"
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install torch (${{ matrix.cuda_tag }})
+        run: |
+          pip install torch --index-url https://download.pytorch.org/whl/${{ matrix.cuda_tag }}
+
+      - name: Setup CUDA (${{ matrix.cuda_ver }})
+        uses: Jimver/cuda-toolkit@v0.2.35
+        with:
+          cuda: "${{ matrix.cuda_ver }}"
+
+      - name: Build wheel (with CUDA kernels)
+        run: |
+          CSRC_KERNELS=true pip wheel . --no-deps --no-build-isolation -w dist/
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: cuda-wheel-linux-${{ matrix.cuda_tag }}
+          path: dist/*.whl
+          if-no-files-found: error
+
+  release:
+    name: Attach wheels to release
+    needs: [build-pure, build-cuda-linux]
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Download pure-Python wheel
+        uses: actions/download-artifact@v4
+        with:
+          name: pure-wheel
+          path: release-assets/pure
+
+      - name: Download CUDA wheels (all variants)
+        uses: actions/download-artifact@v4
+        with:
+          pattern: cuda-wheel-linux-*
+          merge-multiple: true
+          path: release-assets/cuda
+
+      - name: Verify release assets
+        shell: bash
+        run: |
+          set -euo pipefail
+          pure_wheels=(release-assets/pure/*.whl)
+          cuda_wheels=(release-assets/cuda/*.whl)
+          test "${#pure_wheels[@]}" -eq 1
+          test "${#cuda_wheels[@]}" -ge 1
+
+      - name: Create release & upload assets
+        uses: softprops/action-gh-release@v2
+        with:
+          files: |
+            release-assets/pure/*.whl
+            release-assets/cuda/*.whl
+          tag_name: ${{ github.ref_name }}
+          generate_release_notes: true

.gitignore

@@ -7,8 +7,15 @@
 # Allow specific file types and root files
 !astrai/**/*.py
 !scripts/**/*.py
-!scripts/**/*.sh
 !tests/**/*.py
+!csrc/**/*.py
+!csrc/CMakeLists.txt
+
+!csrc/**/*.cu
+!csrc/**/*.h
+!csrc/**/*.cuh
+
+!scripts/**/*.sh
 
 # Allow GitHub files
 !/.github/**
@@ -18,8 +25,13 @@
 !/.dockerignore
 !/Dockerfile
 !/docker-compose.yml
-!/assets/**
+!/docs/**
 !/CONTRIBUTING.md
 !/LICENSE
 !/pyproject.toml
 !/README.md
+# Allow extension modules (only source .py)
+!/astrai/extension/**/*.py
+
+# Allow build files
+!/setup.py

CONTRIBUTING.md

@@ -20,9 +20,6 @@ Run the following checks **in order** — CI will reject if any fail.
 ruff format .
 ```
 
-> **Note**: `ruff format` may rename parameters (e.g. `mask` → `attn_mask`).  
-> Always review the diff after formatting.
-
 ### 2. Import sorting
 
 ```bash
@@ -44,7 +41,7 @@ python -u -m pytest tests/ -v
 
 > Failed tests may leave orphan tempdirs under `%TEMP%`. Clean them manually if needed.
 
-### 4. (Optional) Full pre-commit check
+### 4. (Optional) Full pre-commit check script
 
 If you have Git Bash available:
 
@@ -52,12 +49,17 @@ If you have Git Bash available:
 bash scripts/pre_commit.sh
 ```
 
-This runs format check, import sort check, and tests in one go.
+The script installs development dependencies by default, then runs the format
+check, import sort check, and tests. If dependencies are already installed, use:
+
+```bash
+bash scripts/pre_commit.sh --skip-deps
+```
 
 ## Commit Style
 
 ```
-fix/feat/chore/docs/refactor/perf/test/style/ci/build/revert : short description (~50 chars)
+type: short description (~50 chars)
 
 - bullet point body (each ~60 chars)
 ```
@@ -73,7 +75,7 @@ fix/feat/chore/docs/refactor/perf/test/style/ci/build/revert : short description
 |---------|-------|-----|
 | `ruff check --select I` fails | Wrong import order | `ruff check . --select I --fix .` then `ruff format .` |
 | `ruff format` changed many files | Not formatted before commit | Review diff carefully before staging |
-| Pre-commit hook rejects | Tests or lint failed | Fix individually, do not `--no-verify` |
+| Pre-commit check script fails | Dependency install, tests, or lint failed | Fix the failing step; use `--skip-deps` only when dependencies are already installed |
 | Tests fail with tempdir left | Test crash | Clean `%TEMP%` manually |
 
 ## Submitting Changes
@@ -93,7 +95,7 @@ fix/feat/chore/docs/refactor/perf/test/style/ci/build/revert : short description
 
 ## License
 
-By contributing, you agree that your contributions will be licensed under the [GPL-3.0 License](LICENSE).
+By contributing, you agree that your contributions will be licensed under the [Apache-2.0 License](LICENSE).
 
 ---
 

Dockerfile

@@ -1,8 +1,16 @@
 # AstrAI Dockerfile - Multi-stage Build (Optimized)
+#
+# CUDA version selection:
+#   docker build -t astrai .
+#   docker build -t astrai --build-arg CUDA_TAG=cu128 .
+#   docker build -t astrai --build-arg CUDA_TAG=cu130 .
+# Default: cu128
 
 # Build stage - use base image with minimal build tools
 FROM ubuntu:24.04 AS builder
 
+ARG CUDA_TAG=cu128
+
 WORKDIR /app
 
 # Install Python 3.12 and minimal build dependencies
@@ -20,10 +28,12 @@ ENV PATH="/opt/venv/bin:$PATH"
 
 # Copy source code and install (deps read from pyproject.toml)
 COPY astrai/ ./astrai/
+COPY csrc/ ./csrc/
+COPY setup.py .
 COPY pyproject.toml .
 RUN pip install --no-cache-dir --upgrade pip \
     && pip install --no-cache-dir . \
-    --extra-index-url https://download.pytorch.org/whl/cu128
+    --extra-index-url "https://download.pytorch.org/whl/${CUDA_TAG}"
 
 # Production stage
 FROM ubuntu:24.04 AS production
@@ -43,7 +53,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 # Copy application code
 COPY astrai/ ./astrai/
 COPY scripts/ ./scripts/
-COPY assets/ ./assets/
+COPY docs/ ./docs/
 COPY pyproject.toml .
 COPY README.md .
 

LICENSE

@@ -1,674 +1,201 @@
-                    GNU GENERAL PUBLIC LICENSE
-                       Version 3, 29 June 2007
-
- Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
- Everyone is permitted to copy and distribute verbatim copies
- of this license document, but changing it is not allowed.
-
-                            Preamble
-
-  The GNU General Public License is a free, copyleft license for
-software and other kinds of works.
-
-  The licenses for most software and other practical works are designed
-to take away your freedom to share and change the works.  By contrast,
-the GNU General Public License is intended to guarantee your freedom to
-share and change all versions of a program--to make sure it remains free
-software for all its users.  We, the Free Software Foundation, use the
-GNU General Public License for most of our software; it applies also to
-any other work released this way by its authors.  You can apply it to
-your programs, too.
-
-  When we speak of free software, we are referring to freedom, not
-price.  Our General Public Licenses are designed to make sure that you
-have the freedom to distribute copies of free software (and charge for
-them if you wish), that you receive source code or can get it if you
-want it, that you can change the software or use pieces of it in new
-free programs, and that you know you can do these things.
-
-  To protect your rights, we need to prevent others from denying you
-these rights or asking you to surrender the rights.  Therefore, you have
-certain responsibilities if you distribute copies of the software, or if
-you modify it: responsibilities to respect the freedom of others.
-
-  For example, if you distribute copies of such a program, whether
-gratis or for a fee, you must pass on to the recipients the same
-freedoms that you received.  You must make sure that they, too, receive
-or can get the source code.  And you must show them these terms so they
-know their rights.
-
-  Developers that use the GNU GPL protect your rights with two steps:
-(1) assert copyright on the software, and (2) offer you this License
-giving you legal permission to copy, distribute and/or modify it.
-
-  For the developers' and authors' protection, the GPL clearly explains
-that there is no warranty for this free software.  For both users' and
-authors' sake, the GPL requires that modified versions be marked as
-changed, so that their problems will not be attributed erroneously to
-authors of previous versions.
-
-  Some devices are designed to deny users access to install or run
-modified versions of the software inside them, although the manufacturer
-can do so.  This is fundamentally incompatible with the aim of
-protecting users' freedom to change the software.  The systematic
-pattern of such abuse occurs in the area of products for individuals to
-use, which is precisely where it is most unacceptable.  Therefore, we
-have designed this version of the GPL to prohibit the practice for those
-products.  If such problems arise substantially in other domains, we
-stand ready to extend this provision to those domains in future versions
-of the GPL, as needed to protect the freedom of users.
-
-  Finally, every program is threatened constantly by software patents.
-States should not allow patents to restrict development and use of
-software on general-purpose computers, but in those that do, we wish to
-avoid the special danger that patents applied to a free program could
-make it effectively proprietary.  To prevent this, the GPL assures that
-patents cannot be used to render the program non-free.
-
-  The precise terms and conditions for copying, distribution and
-modification follow.
-
-                       TERMS AND CONDITIONS
-
-  0. Definitions.
-
-  "This License" refers to version 3 of the GNU General Public License.
-
-  "Copyright" also means copyright-like laws that apply to other kinds of
-works, such as semiconductor masks.
-
-  "The Program" refers to any copyrightable work licensed under this
-License.  Each licensee is addressed as "you".  "Licensees" and
-"recipients" may be individuals or organizations.
-
-  To "modify" a work means to copy from or adapt all or part of the work
-in a fashion requiring copyright permission, other than the making of an
-exact copy.  The resulting work is called a "modified version" of the
-earlier work or a work "based on" the earlier work.
-
-  A "covered work" means either the unmodified Program or a work based
-on the Program.
-
-  To "propagate" a work means to do anything with it that, without
-permission, would make you directly or secondarily liable for
-infringement under applicable copyright law, except executing it on a
-computer or modifying a private copy.  Propagation includes copying,
-distribution (with or without modification), making available to the
-public, and in some countries other activities as well.
-
-  To "convey" a work means any kind of propagation that enables other
-parties to make or receive copies.  Mere interaction with a user through
-a computer network, with no transfer of a copy, is not conveying.
-
-  An interactive user interface displays "Appropriate Legal Notices"
-to the extent that it includes a convenient and prominently visible
-feature that (1) displays an appropriate copyright notice, and (2)
-tells the user that there is no warranty for the work (except to the
-extent that warranties are provided), that licensees may convey the
-work under this License, and how to view a copy of this License.  If
-the interface presents a list of user commands or options, such as a
-menu, a prominent item in the list meets this criterion.
-
-  1. Source Code.
-
-  The "source code" for a work means the preferred form of the work
-for making modifications to it.  "Object code" means any non-source
-form of a work.
-
-  A "Standard Interface" means an interface that either is an official
-standard defined by a recognized standards body, or, in the case of
-interfaces specified for a particular programming language, one that
-is widely used among developers working in that language.
-
-  The "System Libraries" of an executable work include anything, other
-than the work as a whole, that (a) is included in the normal form of
-packaging a Major Component, but which is not part of that Major
-Component, and (b) serves only to enable use of the work with that
-Major Component, or to implement a Standard Interface for which an
-implementation is available to the public in source code form.  A
-"Major Component", in this context, means a major essential component
-(kernel, window system, and so on) of the specific operating system
-(if any) on which the executable work runs, or a compiler used to
-produce the work, or an object code interpreter used to run it.
-
-  The "Corresponding Source" for a work in object code form means all
-the source code needed to generate, install, and (for an executable
-work) run the object code and to modify the work, including scripts to
-control those activities.  However, it does not include the work's
-System Libraries, or general-purpose tools or generally available free
-programs which are used unmodified in performing those activities but
-which are not part of the work.  For example, Corresponding Source
-includes interface definition files associated with source files for
-the work, and the source code for shared libraries and dynamically
-linked subprograms that the work is specifically designed to require,
-such as by intimate data communication or control flow between those
-subprograms and other parts of the work.
-
-  The Corresponding Source need not include anything that users
-can regenerate automatically from other parts of the Corresponding
-Source.
-
-  The Corresponding Source for a work in source code form is that
-same work.
-
-  2. Basic Permissions.
-
-  All rights granted under this License are granted for the term of
-copyright on the Program, and are irrevocable provided the stated
-conditions are met.  This License explicitly affirms your unlimited
-permission to run the unmodified Program.  The output from running a
-covered work is covered by this License only if the output, given its
-content, constitutes a covered work.  This License acknowledges your
-rights of fair use or other equivalent, as provided by copyright law.
-
-  You may make, run and propagate covered works that you do not
-convey, without conditions so long as your license otherwise remains
-in force.  You may convey covered works to others for the sole purpose
-of having them make modifications exclusively for you, or provide you
-with facilities for running those works, provided that you comply with
-the terms of this License in conveying all material for which you do
-not control copyright.  Those thus making or running the covered works
-for you must do so exclusively on your behalf, under your direction
-and control, on terms that prohibit them from making any copies of
-your copyrighted material outside their relationship with you.
-
-  Conveying under any other circumstances is permitted solely under
-the conditions stated below.  Sublicensing is not allowed; section 10
-makes it unnecessary.
-
-  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
-
-  No covered work shall be deemed part of an effective technological
-measure under any applicable law fulfilling obligations under article
-11 of the WIPO copyright treaty adopted on 20 December 1996, or
-similar laws prohibiting or restricting circumvention of such
-measures.
-
-  When you convey a covered work, you waive any legal power to forbid
-circumvention of technological measures to the extent such circumvention
-is effected by exercising rights under this License with respect to
-the covered work, and you disclaim any intention to limit operation or
-modification of the work as a means of enforcing, against the work's
-users, your or third parties' legal rights to forbid circumvention of
-technological measures.
-
-  4. Conveying Verbatim Copies.
-
-  You may convey verbatim copies of the Program's source code as you
-receive it, in any medium, provided that you conspicuously and
-appropriately publish on each copy an appropriate copyright notice;
-keep intact all notices stating that this License and any
-non-permissive terms added in accord with section 7 apply to the code;
-keep intact all notices of the absence of any warranty; and give all
-recipients a copy of this License along with the Program.
-
-  You may charge any price or no price for each copy that you convey,
-and you may offer support or warranty protection for a fee.
-
-  5. Conveying Modified Source Versions.
-
-  You may convey a work based on the Program, or the modifications to
-produce it from the Program, in the form of source code under the
-terms of section 4, provided that you also meet all of these conditions:
-
-    a) The work must carry prominent notices stating that you modified
-    it, and giving a relevant date.
-
-    b) The work must carry prominent notices stating that it is
-    released under this License and any conditions added under section
-    7.  This requirement modifies the requirement in section 4 to
-    "keep intact all notices".
-
-    c) You must license the entire work, as a whole, under this
-    License to anyone who comes into possession of a copy.  This
-    License will therefore apply, along with any applicable section 7
-    additional terms, to the whole of the work, and all its parts,
-    regardless of how they are packaged.  This License gives no
-    permission to license the work in any other way, but it does not
-    invalidate such permission if you have separately received it.
-
-    d) If the work has interactive user interfaces, each must display
-    Appropriate Legal Notices; however, if the Program has interactive
-    interfaces that do not display Appropriate Legal Notices, your
-    work need not make them do so.
-
-  A compilation of a covered work with other separate and independent
-works, which are not by their nature extensions of the covered work,
-and which are not combined with it such as to form a larger program,
-in or on a volume of a storage or distribution medium, is called an
-"aggregate" if the compilation and its resulting copyright are not
-used to limit the access or legal rights of the compilation's users
-beyond what the individual works permit.  Inclusion of a covered work
-in an aggregate does not cause this License to apply to the other
-parts of the aggregate.
-
-  6. Conveying Non-Source Forms.
-
-  You may convey a covered work in object code form under the terms
-of sections 4 and 5, provided that you also convey the
-machine-readable Corresponding Source under the terms of this License,
-in one of these ways:
-
-    a) Convey the object code in, or embodied in, a physical product
-    (including a physical distribution medium), accompanied by the
-    Corresponding Source fixed on a durable physical medium
-    customarily used for software interchange.
-
-    b) Convey the object code in, or embodied in, a physical product
-    (including a physical distribution medium), accompanied by a
-    written offer, valid for at least three years and valid for as
-    long as you offer spare parts or customer support for that product
-    model, to give anyone who possesses the object code either (1) a
-    copy of the Corresponding Source for all the software in the
-    product that is covered by this License, on a durable physical
-    medium customarily used for software interchange, for a price no
-    more than your reasonable cost of physically performing this
-    conveying of source, or (2) access to copy the
-    Corresponding Source from a network server at no charge.
-
-    c) Convey individual copies of the object code with a copy of the
-    written offer to provide the Corresponding Source.  This
-    alternative is allowed only occasionally and noncommercially, and
-    only if you received the object code with such an offer, in accord
-    with subsection 6b.
-
-    d) Convey the object code by offering access from a designated
-    place (gratis or for a charge), and offer equivalent access to the
-    Corresponding Source in the same way through the same place at no
-    further charge.  You need not require recipients to copy the
-    Corresponding Source along with the object code.  If the place to
-    copy the object code is a network server, the Corresponding Source
-    may be on a different server (operated by you or a third party)
-    that supports equivalent copying facilities, provided you maintain
-    clear directions next to the object code saying where to find the
-    Corresponding Source.  Regardless of what server hosts the
-    Corresponding Source, you remain obligated to ensure that it is
-    available for as long as needed to satisfy these requirements.
-
-    e) Convey the object code using peer-to-peer transmission, provided
-    you inform other peers where the object code and Corresponding
-    Source of the work are being offered to the general public at no
-    charge under subsection 6d.
-
-  A separable portion of the object code, whose source code is excluded
-from the Corresponding Source as a System Library, need not be
-included in conveying the object code work.
-
-  A "User Product" is either (1) a "consumer product", which means any
-tangible personal property which is normally used for personal, family,
-or household purposes, or (2) anything designed or sold for incorporation
-into a dwelling.  In determining whether a product is a consumer product,
-doubtful cases shall be resolved in favor of coverage.  For a particular
-product received by a particular user, "normally used" refers to a
-typical or common use of that class of product, regardless of the status
-of the particular user or of the way in which the particular user
-actually uses, or expects or is expected to use, the product.  A product
-is a consumer product regardless of whether the product has substantial
-commercial, industrial or non-consumer uses, unless such uses represent
-the only significant mode of use of the product.
-
-  "Installation Information" for a User Product means any methods,
-procedures, authorization keys, or other information required to install
-and execute modified versions of a covered work in that User Product from
-a modified version of its Corresponding Source.  The information must
-suffice to ensure that the continued functioning of the modified object
-code is in no case prevented or interfered with solely because
-modification has been made.
-
-  If you convey an object code work under this section in, or with, or
-specifically for use in, a User Product, and the conveying occurs as
-part of a transaction in which the right of possession and use of the
-User Product is transferred to the recipient in perpetuity or for a
-fixed term (regardless of how the transaction is characterized), the
-Corresponding Source conveyed under this section must be accompanied
-by the Installation Information.  But this requirement does not apply
-if neither you nor any third party retains the ability to install
-modified object code on the User Product (for example, the work has
-been installed in ROM).
-
-  The requirement to provide Installation Information does not include a
-requirement to continue to provide support service, warranty, or updates
-for a work that has been modified or installed by the recipient, or for
-the User Product in which it has been modified or installed.  Access to a
-network may be denied when the modification itself materially and
-adversely affects the operation of the network or violates the rules and
-protocols for communication across the network.
-
-  Corresponding Source conveyed, and Installation Information provided,
-in accord with this section must be in a format that is publicly
-documented (and with an implementation available to the public in
-source code form), and must require no special password or key for
-unpacking, reading or copying.
-
-  7. Additional Terms.
-
-  "Additional permissions" are terms that supplement the terms of this
-License by making exceptions from one or more of its conditions.
-Additional permissions that are applicable to the entire Program shall
-be treated as though they were included in this License, to the extent
-that they are valid under applicable law.  If additional permissions
-apply only to part of the Program, that part may be used separately
-under those permissions, but the entire Program remains governed by
-this License without regard to the additional permissions.
-
-  When you convey a copy of a covered work, you may at your option
-remove any additional permissions from that copy, or from any part of
-it.  (Additional permissions may be written to require their own
-removal in certain cases when you modify the work.)  You may place
-additional permissions on material, added by you to a covered work,
-for which you have or can give appropriate copyright permission.
-
-  Notwithstanding any other provision of this License, for material you
-add to a covered work, you may (if authorized by the copyright holders of
-that material) supplement the terms of this License with terms:
-
-    a) Disclaiming warranty or limiting liability differently from the
-    terms of sections 15 and 16 of this License; or
-
-    b) Requiring preservation of specified reasonable legal notices or
-    author attributions in that material or in the Appropriate Legal
-    Notices displayed by works containing it; or
-
-    c) Prohibiting misrepresentation of the origin of that material, or
-    requiring that modified versions of such material be marked in
-    reasonable ways as different from the original version; or
-
-    d) Limiting the use for publicity purposes of names of licensors or
-    authors of the material; or
-
-    e) Declining to grant rights under trademark law for use of some
-    trade names, trademarks, or service marks; or
-
-    f) Requiring indemnification of licensors and authors of that
-    material by anyone who conveys the material (or modified versions of
-    it) with contractual assumptions of liability to the recipient, for
-    any liability that these contractual assumptions directly impose on
-    those licensors and authors.
-
-  All other non-permissive additional terms are considered "further
-restrictions" within the meaning of section 10.  If the Program as you
-received it, or any part of it, contains a notice stating that it is
-governed by this License along with a term that is a further
-restriction, you may remove that term.  If a license document contains
-a further restriction but permits relicensing or conveying under this
-License, you may add to a covered work material governed by the terms
-of that license document, provided that the further restriction does
-not survive such relicensing or conveying.
-
-  If you add terms to a covered work in accord with this section, you
-must place, in the relevant source files, a statement of the
-additional terms that apply to those files, or a notice indicating
-where to find the applicable terms.
-
-  Additional terms, permissive or non-permissive, may be stated in the
-form of a separately written license, or stated as exceptions;
-the above requirements apply either way.
-
-  8. Termination.
-
-  You may not propagate or modify a covered work except as expressly
-provided under this License.  Any attempt otherwise to propagate or
-modify it is void, and will automatically terminate your rights under
-this License (including any patent licenses granted under the third
-paragraph of section 11).
-
-  However, if you cease all violation of this License, then your
-license from a particular copyright holder is reinstated (a)
-provisionally, unless and until the copyright holder explicitly and
-finally terminates your license, and (b) permanently, if the copyright
-holder fails to notify you of the violation by some reasonable means
-prior to 60 days after the cessation.
-
-  Moreover, your license from a particular copyright holder is
-reinstated permanently if the copyright holder notifies you of the
-violation by some reasonable means, this is the first time you have
-received notice of violation of this License (for any work) from that
-copyright holder, and you cure the violation prior to 30 days after
-your receipt of the notice.
-
-  Termination of your rights under this section does not terminate the
-licenses of parties who have received copies or rights from you under
-this License.  If your rights have been terminated and not permanently
-reinstated, you do not qualify to receive new licenses for the same
-material under section 10.
-
-  9. Acceptance Not Required for Having Copies.
-
-  You are not required to accept this License in order to receive or
-run a copy of the Program.  Ancillary propagation of a covered work
-occurring solely as a consequence of using peer-to-peer transmission
-to receive a copy likewise does not require acceptance.  However,
-nothing other than this License grants you permission to propagate or
-modify any covered work.  These actions infringe copyright if you do
-not accept this License.  Therefore, by modifying or propagating a
-covered work, you indicate your acceptance of this License to do so.
-
-  10. Automatic Licensing of Downstream Recipients.
-
-  Each time you convey a covered work, the recipient automatically
-receives a license from the original licensors, to run, modify and
-propagate that work, subject to this License.  You are not responsible
-for enforcing compliance by third parties with this License.
-
-  An "entity transaction" is a transaction transferring control of an
-organization, or substantially all assets of one, or subdividing an
-organization, or merging organizations.  If propagation of a covered
-work results from an entity transaction, each party to that
-transaction who receives a copy of the work also receives whatever
-licenses to the work the party's predecessor in interest had or could
-give under the previous paragraph, plus a right to possession of the
-Corresponding Source of the work from the predecessor in interest, if
-the predecessor has it or can get it with reasonable efforts.
-
-  You may not impose any further restrictions on the exercise of the
-rights granted or affirmed under this License.  For example, you may
-not impose a license fee, royalty, or other charge for exercise of
-rights granted under this License, and you may not initiate litigation
-(including a cross-claim or counterclaim in a lawsuit) alleging that
-any patent claim is infringed by making, using, selling, offering for
-sale, or importing the Program or any portion of it.
-
-  11. Patents.
-
-  A "contributor" is a copyright holder who authorizes use under this
-License of the Program or a work on which the Program is based.  The
-work thus licensed is called the contributor's "contributor version".
-
-  A contributor's "essential patent claims" are all patent claims
-owned or controlled by the contributor, whether already acquired or
-hereafter acquired, that would be infringed by some manner, permitted
-by this License, of making, using, or selling its contributor version,
-but do not include claims that would be infringed only as a
-consequence of further modification of the contributor version.  For
-purposes of this definition, "control" includes the right to grant
-patent sublicenses in a manner consistent with the requirements of
-this License.
-
-  Each contributor grants you a non-exclusive, worldwide, royalty-free
-patent license under the contributor's essential patent claims, to
-make, use, sell, offer for sale, import and otherwise run, modify and
-propagate the contents of its contributor version.
-
-  In the following three paragraphs, a "patent license" is any express
-agreement or commitment, however denominated, not to enforce a patent
-(such as an express permission to practice a patent or covenant not to
-sue for patent infringement).  To "grant" such a patent license to a
-party means to make such an agreement or commitment not to enforce a
-patent against the party.
-
-  If you convey a covered work, knowingly relying on a patent license,
-and the Corresponding Source of the work is not available for anyone
-to copy, free of charge and under the terms of this License, through a
-publicly available network server or other readily accessible means,
-then you must either (1) cause the Corresponding Source to be so
-available, or (2) arrange to deprive yourself of the benefit of the
-patent license for this particular work, or (3) arrange, in a manner
-consistent with the requirements of this License, to extend the patent
-license to downstream recipients.  "Knowingly relying" means you have
-actual knowledge that, but for the patent license, your conveying the
-covered work in a country, or your recipient's use of the covered work
-in a country, would infringe one or more identifiable patents in that
-country that you have reason to believe are valid.
-
-  If, pursuant to or in connection with a single transaction or
-arrangement, you convey, or propagate by procuring conveyance of, a
-covered work, and grant a patent license to some of the parties
-receiving the covered work authorizing them to use, propagate, modify
-or convey a specific copy of the covered work, then the patent license
-you grant is automatically extended to all recipients of the covered
-work and works based on it.
-
-  A patent license is "discriminatory" if it does not include within
-the scope of its coverage, prohibits the exercise of, or is
-conditioned on the non-exercise of one or more of the rights that are
-specifically granted under this License.  You may not convey a covered
-work if you are a party to an arrangement with a third party that is
-in the business of distributing software, under which you make payment
-to the third party based on the extent of your activity of conveying
-the work, and under which the third party grants, to any of the
-parties who would receive the covered work from you, a discriminatory
-patent license (a) in connection with copies of the covered work
-conveyed by you (or copies made from those copies), or (b) primarily
-for and in connection with specific products or compilations that
-contain the covered work, unless you entered into that arrangement,
-or that patent license was granted, prior to 28 March 2007.
-
-  Nothing in this License shall be construed as excluding or limiting
-any implied license or other defenses to infringement that may
-otherwise be available to you under applicable patent law.
-
-  12. No Surrender of Others' Freedom.
-
-  If conditions are imposed on you (whether by court order, agreement or
-otherwise) that contradict the conditions of this License, they do not
-excuse you from the conditions of this License.  If you cannot convey a
-covered work so as to satisfy simultaneously your obligations under this
-License and any other pertinent obligations, then as a consequence you may
-not convey it at all.  For example, if you agree to terms that obligate you
-to collect a royalty for further conveying from those to whom you convey
-the Program, the only way you could satisfy both those terms and this
-License would be to refrain entirely from conveying the Program.
-
-  13. Use with the GNU Affero General Public License.
-
-  Notwithstanding any other provision of this License, you have
-permission to link or combine any covered work with a work licensed
-under version 3 of the GNU Affero General Public License into a single
-combined work, and to convey the resulting work.  The terms of this
-License will continue to apply to the part which is the covered work,
-but the special requirements of the GNU Affero General Public License,
-section 13, concerning interaction through a network will apply to the
-combination as such.
-
-  14. Revised Versions of this License.
-
-  The Free Software Foundation may publish revised and/or new versions of
-the GNU General Public License from time to time.  Such new versions will
-be similar in spirit to the present version, but may differ in detail to
-address new problems or concerns.
-
-  Each version is given a distinguishing version number.  If the
-Program specifies that a certain numbered version of the GNU General
-Public License "or any later version" applies to it, you have the
-option of following the terms and conditions either of that numbered
-version or of any later version published by the Free Software
-Foundation.  If the Program does not specify a version number of the
-GNU General Public License, you may choose any version ever published
-by the Free Software Foundation.
-
-  If the Program specifies that a proxy can decide which future
-versions of the GNU General Public License can be used, that proxy's
-public statement of acceptance of a version permanently authorizes you
-to choose that version for the Program.
-
-  Later license versions may give you additional or different
-permissions.  However, no additional obligations are imposed on any
-author or copyright holder as a result of your choosing to follow a
-later version.
-
-  15. Disclaimer of Warranty.
-
-  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
-APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
-HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
-OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
-THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
-PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
-IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
-ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
-
-  16. Limitation of Liability.
-
-  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
-WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
-THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
-GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
-USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
-DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
-PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
-EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
-SUCH DAMAGES.
-
-  17. Interpretation of Sections 15 and 16.
-
-  If the disclaimer of warranty and limitation of liability provided
-above cannot be given local legal effect according to their terms,
-reviewing courts shall apply local law that most closely approximates
-an absolute waiver of all civil liability in connection with the
-Program, unless a warranty or assumption of liability accompanies a
-copy of the Program in return for a fee.
-
-                     END OF TERMS AND CONDITIONS
-
-            How to Apply These Terms to Your New Programs
-
-  If you develop a new program, and you want it to be of the greatest
-possible use to the public, the best way to achieve this is to make it
-free software which everyone can redistribute and change under these terms.
-
-  To do so, attach the following notices to the program.  It is safest
-to attach them to the start of each source file to most effectively
-state the exclusion of warranty; and each file should have at least
-the "copyright" line and a pointer to where the full notice is found.
-
-    <one line to give the program's name and a brief idea of what it does.>
-    Copyright (C) <year>  <name of author>
-
-    This program is free software: you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation, either version 3 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License
-    along with this program.  If not, see <https://www.gnu.org/licenses/>.
-
-Also add information on how to contact you by electronic and paper mail.
-
-  If the program does terminal interaction, make it output a short
-notice like this when it starts in an interactive mode:
-
-    <program>  Copyright (C) <year>  <name of author>
-    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
-    This is free software, and you are welcome to redistribute it
-    under certain conditions; type `show c' for details.
-
-The hypothetical commands `show w' and `show c' should show the appropriate
-parts of the General Public License.  Of course, your program's commands
-might be different; for a GUI interface, you would use an "about box".
-
-  You should also get your employer (if you work as a programmer) or school,
-if any, to sign a "copyright disclaimer" for the program, if necessary.
-For more information on this, and how to apply and follow the GNU GPL, see
-<https://www.gnu.org/licenses/>.
-
-  The GNU General Public License does not permit incorporating your program
-into proprietary programs.  If your program is a subroutine library, you
-may consider it more useful to permit linking proprietary applications with
-the library.  If this is what you want to do, use the GNU Lesser General
-Public License instead of this License.  But first, please read
-<https://www.gnu.org/licenses/why-not-lgpl.html>.
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -1,6 +1,6 @@
 <div align="center">
   
-  <img src="assets/images/logo.png" width="auto" alt="Logo">
+  <img src="docs/images/logo.png" width="auto" alt="Logo">
   <p>
     <strong>A lightweight Transformer training & inference framework</strong>
   </p>
@@ -8,8 +8,8 @@
 
 <div align="center">
   <img src="https://img.shields.io/badge/python-3.12+-blue.svg" alt="python">
-  <img src="https://img.shields.io/badge/license-GPL--3.0-blue.svg" alt="license">
-  <img src="https://img.shields.io/github/v/release/ViperEkura/AstrAI?label=Release&color=76bad9" alt="release">
+  <img src="https://img.shields.io/badge/license-Apache--2.0-blue.svg" alt="license">
+  <img src="https://img.shields.io/github/v/tag/ViperEkura/AstrAI?label=Release&color=76bad9" alt="release">
   <img src="https://img.shields.io/github/stars/ViperEkura/AstrAI?style=flat&label=Stars&color=76bad9" alt="stars">
   <img src="https://img.shields.io/github/forks/ViperEkura/AstrAI?style=flat&label=Forks&color=76bad9" alt="forks">
 </div>
@@ -17,17 +17,17 @@
 
 <div align="center">
   <a href="#english">English</a> •
-  <a href="assets/docs/README-zh-CN.md">中文</a> •
+  <a href="docs/README-zh-CN.md">中文</a> •
   <a href="https://github.com/ViperEkura/AstrAI/issues">Issue Tracker</a> •
   <a href="https://github.com/ViperEkura/AstrAI/discussions">Discussions</a> •
-  <a href="https://huggingface.co/ViperEk/">HuggingFace</a>
+  <a href="https://huggingface.co/ViperEkura">HuggingFace</a>
 </div>
 
 <br>
 
 ## 📖 Table of Contents
 
-- [Features](#features)
+- [Overview](#overview)
 - [Getting Started](#getting-started)
 - [Demo](#demo)
 - [Documentation](#documentation)
@@ -40,15 +40,19 @@
 <a id="english"></a>
 ## English
 
-### Features
+### Overview
 
-- 🚀 **High Performance**: Optimized for both training and inference with efficient parallelization.
-- 🔧 **Flexible**: Support for seq/sft/dpo/grpo training, customizable model architectures.
-- 💡 **Easy to Use**: Simple API with comprehensive examples and demos.
-- 📦 **Lightweight**: Minimal dependencies, easy to deploy.
-- 🔬 **Research‑Friendly**: Modular design, easy to experiment with new ideas.
-- 🤗 **HuggingFace-Style API**: AutoModel/AutoTokenizer APIs inspired by HuggingFace for easy model and tokenizer loading.
-- 🔌 **Dual API Compatibility**: Supports both OpenAI and Anthropic chat completion APIs out of the box.
+AstrAI is an end-to-end Transformer framework for building, training, evaluating, and serving models. It provides a compact PyTorch codebase for the complete model lifecycle, from declarative data preprocessing and distributed training to continuous-batching inference and OpenAI/Anthropic-compatible APIs.
+
+| Area | Capabilities |
+|---|---|
+| **Models** | Autoregressive language models and embedding models with GQA, MLA, MoE, RoPE, and extensible attention/FFN components |
+| **Training** | Pre-training (`seq`), supervised fine-tuning (`sft`), DPO, and GRPO with gradient accumulation, checkpointing, DDP, and FSDP |
+| **Data** | Declarative JSON preprocessing, configurable masking and packing, binary/JSONL storage, and streaming datasets |
+| **Inference** | Continuous batching, paged KV cache, radix prefix caching, streaming generation, and Torch/CUDA/FlashAttention backends |
+| **Serving** | FastAPI server with OpenAI and Anthropic chat completion protocols, including SSE streaming and tool calls |
+| **Evaluation** | Perplexity, MMLU, HumanEval, IFEval, IFD, and ROUGE evaluation tools |
+| **Extensibility** | Factory and registry architecture for models, datasets, training strategies, callbacks, kernels, and protocol components |
 
 ### Getting Started
 
@@ -56,11 +60,14 @@ End-to-end walkthrough in 5 steps:
 
 **1. Install**
 
+AstrAI requires Python 3.12+ and pins PyTorch exactly to `2.11.0`. Training, `scripts/tools/generate.py`, generation evaluations, and the generation demos require CUDA; CPU support is limited to components with an explicit CPU device path, such as the HTTP server and direct-scoring evaluations.
+
 ```bash
 git clone https://github.com/ViperEkura/AstrAI.git
 cd AstrAI
-pip install -e .
-# pip install -e ".[dev]"    # optional: dev dependencies (pytest, ruff)
+pip install -e .                                          # pure PyTorch (no CUDA kernels)
+# CSRC_KERNELS=true pip install -e . --no-build-isolation  # optional: fused CUDA kernels
+# pip install -e ".[dev]"                                  # dev dependencies (pytest, ruff)
 ```
 
 **2. Download model**
@@ -102,9 +109,7 @@ nohup python scripts/tools/train.py \
     --warmup_ratio=0.05 \
     --max_lr=1e-4 \
     --max_grad_norm=1.0 \
-    --adamw_beta1=0.9 \
-    --adamw_beta2=0.95 \
-    --adamw_weight_decay=0.01 \
+    --weight_decay=0.1 \
     --window_size=2048 \
     --ckpt_interval=10000 \
     --ckpt_dir=./checkpoint \
@@ -133,7 +138,7 @@ Check out the demos in the `scripts/demo/` folder:
 # Download model weights (required before running demos)
 python scripts/demo/download.py                      # model → params/
 
-# Interactive streaming chat (multi-turn, maintains history)
+# Single-turn interactive streaming prompt loop (no conversation history)
 python scripts/demo/stream_chat.py
 # Type your message after >>, type !exit to quit
 
@@ -184,7 +189,7 @@ docker run --gpus all -v /path/to/data:/data -it astrai:latest
 # Docker Compose (GPU, default)
 docker compose up -d
 
-# Docker Compose (CPU only)
+# Docker Compose CPU server profile (CUDA-only generation scripts/demos are unavailable)
 docker compose --profile cpu up -d
 ```
 
@@ -214,18 +219,23 @@ curl -X POST http://localhost:8000/v1/messages \
 curl http://localhost:8000/health
 ```
 
-See [Inference Guide](assets/docs/inference.md) for SSE streaming format, error codes, and stats endpoint.
+See [Inference Guide](docs/guides/inference.md) for SSE streaming format, error codes, and stats endpoint.
 
 ### Documentation
 
 | Document | Description |
 |----------|-------------|
-| [CLI Reference](./assets/docs/params.md) | Parameters for all CLI tools (train, server, generate, preprocess) |
-| [Architecture](./assets/docs/architecture.md) | System architecture, class diagram & design patterns |
-| [Training](./assets/docs/training.md) | Training loop, strategies & formulas |
-| [Inference](./assets/docs/inference.md) | KVCache, continuous batching, sampling & HTTP API |
-| [Data Flow](./assets/docs/dataflow.md) | Data pipeline, storage backends & dataset architecture |
-| [Preprocessing](./assets/docs/preprocessing.md) | Declarative JSON-driven data preprocessing |
+| [Get Started](./docs/get-started.md) | Installation and quickstart |
+| [CLI Reference](./docs/guides/params.md) | Parameters for all CLI tools (train, server, generate, preprocess) |
+| [Preprocessing](./docs/guides/preprocessing.md) | Declarative JSON-driven data preprocessing |
+| [Training](./docs/guides/training.md) | Training loop, strategies & formulas |
+| [Inference](./docs/guides/inference.md) | KVCache, continuous batching, sampling & HTTP API |
+| [Evaluation](./docs/guides/evaluation.md) | HumanEval, MMLU, PPL, ROUGE, IFD, IFEval |
+| [Distributed](./docs/guides/distributed.md) | Multi-GPU DDP / FSDP training |
+| [Architecture](./docs/developer/architecture.md) | System architecture, class diagram & design patterns |
+| [Data Flow](./docs/developer/dataflow.md) | Data pipeline, storage backends & dataset architecture |
+| [Internals](./docs/developer/internals.md) | Training internals: loss formulas, callback lifecycle, KV cache |
+| [CUDA Kernels](./docs/developer/cuda_kernels.md) | Custom CUDA attention kernels & benchmarks |
 
 ### Contributing
 
@@ -242,11 +252,11 @@ For major changes, please open an issue first to discuss what you would like to
 
 - **GitHub Issues**: [Issue Tracker](https://github.com/ViperEkura/AstrAI/issues)
 - **Discussions**: [GitHub Discussions](https://github.com/ViperEkura/AstrAI/discussions)
-- **HuggingFace**: [Model Hub](https://huggingface.co/ViperEk)
+- **HuggingFace**: [Model Hub](https://huggingface.co/ViperEkura)
 
 ### License
 
-This project is licensed under the [GPL-3.0 License](LICENSE).
+This project is licensed under the [Apache-2.0 License](LICENSE).
 
 ---
 

assets/docs/dataflow.md

@@ -1,109 +0,0 @@
-# Data Flow
-
-This document describes the data pipeline: from raw text to model input tensors. For creating preprocessing configs, see [Preprocessing Guide](preprocessing.md).
-
-## Contents
-
-- [Overview](#overview)
-- [Data Preparation](#data-preparation) — tokenization, format detection, backends
-- [Data Keys by Training Type](#data-keys-by-training-type)
-- [Dataset Architecture](#dataset-architecture)
-- [Sampler](#sampler)
-- [DataLoader](#dataloader)
-
-## Overview
-
-```
-JSONL Lines → Pipeline (mask builder) → Tokenized Tensors
-                                              ↓
-                                      .h5 or .bin storage
-                                              ↓
-                                      Store.load()
-                                              ↓
-                                      Store.fetch(begin, end, keys)
-                                              ↓
-                                      BaseDataset.__getitem__(idx)
-                                              ↓
-                                      Sampler → DataLoader → Training / Inference
-```
-
-## Data Preparation
-
-Raw text is tokenized via `AutoTokenizer.encode()` and saved as HDF5 (`.h5`) or binary (`.bin` + `meta.json`) files with keyed tensor groups.
-
-### Tokenization
-
-The `Pipeline` reads JSONL lines, applies the mask builder (see [Preprocessing](preprocessing.md)), and produces flat token sequences:
-
-```python
-# Per JSONL line: messages → chat template → token IDs + loss mask
-tokens = tokenizer.encode(rendered_text)        # List[int]
-loss_mask = [0, 0, 0, 1, 1, 1, 1, 1, 1]        # 0=masked, 1=train
-# Stored as flat tensors, packed with other lines by packing strategy
-```
-
-The output `meta.json` records the storage format, key names, dtype, total token count, and tensor shapes for each shard.
-
-### Format Detection
-
-`detect_format(load_path)` inspects the path:
-
-- If `load_path` is a file: checks suffix — `.h5`/`.hdf5` → `"h5"`, unknown suffix raises `ValueError`
-- If `load_path` is a directory: recursively globs for `*.h5`/`*.hdf5` files → `"h5"`, or `*.bin` + `**/meta.json` → `"bin"`
-
-### Store Backends
-
-Storage format is auto-detected by `detect_format()`; backends are dispatched via registry:
-
-```
-StoreFactory.create("h5")  → H5Store
-StoreFactory.create("bin") → MmapStore
-```
-
-**H5Store**: Reads HDF5 files, supports `share_memory_()` for multi-process DataLoader workers (copies tensors to shared memory).
-
-**MmapStore**: Memory-maps `.bin` files. OS page cache sharing is native — no explicit `share_memory_()` needed. Uses `torch.from_numpy(np.memmap(...))`.
-
-Both backends normalise tensors into `Store._data[Dict[str, List[Tensor]]]` + `Store._cum[Dict[str, List[int]]]` (cumulative lengths for bisect-based indexing).
-
-## Data Keys by Training Type
-
-| Type | Storage Keys |
-|------|-------------|
-| `seq` | `sequence` (→ input_ids, target_ids via offset-by-1) |
-| `sft` | `sequence`, `loss_mask`, `position_ids` |
-| `dpo` | `chosen`, `rejected`, `chosen_mask`, `rejected_mask` |
-| `grpo` | `prompts`, `responses`, `masks`, `rewards` |
-
-## Dataset Architecture
-
-```
-DatasetFactory.load(train_type, load_path, window_size, stride=None, storage_type=None)
-  → BaseDataset.load(load_path, storage_type=None)
-    → detect_format(load_path)
-    → StoreFactory.create(storage_type)
-    → Store.load(load_path)
-      → _normalize(raw)  # base Store, shared by both backends
-        → Store._data[Dict[str, List[Tensor]]] + _cum[Dict[str, List[int]]]
-          → BaseDataset.__getitem__(idx)
-            → get_index(idx) → [begin, end)
-            → Store.fetch(begin, end, keys) → Tensor / Dict[str, Tensor]
-```
-
-`window_size` = max input length, `stride` = step between consecutive samples (defaults to `window_size`, optional). `storage_type` defaults to `None` (auto-detect via `detect_format`).
-
-`Store.fetch(begin, end, keys)` accepts a single key (`str`) returning a `Tensor`, or a list of keys returning `Dict[str, Tensor]`. Internally uses `bisect` across multi-segment tensors. Raises `RuntimeError("Store not loaded")` if called before `load()`.
-
-## Sampler
-
-`ResumableDistributedSampler` supports checkpoint-aware distributed sampling:
-
-- Tracks `start_epoch` / `start_iter` for resume
-- Shuffle via `torch.Generator(seed + epoch)`
-- Per-replica index slicing for DDP
-
-## DataLoader
-
-Standard PyTorch `DataLoader` with configurable `batch_size`, `num_workers`, `pin_memory`, `prefetch_factor`. Sampler produces indices; dataloader fetches tensor batches via `__getitem__`.
-
-> Document Update Time: 2026-06-19

assets/docs/inference.md

@@ -1,241 +0,0 @@
-# Inference
-
-## Contents
-
-- [KV Cache](#kv-cache)
-- [KVCache System](#kvcache-system)
-- [Continuous Batching](#continuous-batching)
-- [Sampling](#sampling-strategy-pattern)
-- [Protocol Handlers](#protocol-handlers-strategy-pattern)
-- [Engine & GenerateResult](#engine--generateresult)
-- [HTTP API](#http-api) — endpoints, SSE, errors, stats
-- [Engine API](#engine-api)
-
-## KV Cache
-
-At decode time, only the last query token matters. All previous K/V are cached to avoid recomputation:
-
-$$
-o_n = \sum_j \text{softmax}\left(\frac{q_n k_j}{\sqrt{d_k}}\right) v_j
-$$
-
-RoPE is applied **before** KV cache write, not after — otherwise position encoding drift occurs.
-
-## KVCache System
-
-Seven classes working together:
-
-```
-KVCache (facade)
-  ├── PagePool         orchestrates page allocation + prefix matching
-  │     ├── Allocator   bitmask-based page allocator + ref-count + LRU eviction (inside PagePool)
-  │     └── PrefixCache hash-based prefix matching (page_hash via polynomial hash) (inside PagePool)
-  ├── TaskTable        maps task_id → page_table + cached token count
-  ├── Storage          k_cache / v_cache tensors (n_layers × n_pages × page_size × n_kv_heads × head_dim)
-  └── KvcacheView      bundles Storage + page_table + total_len for attention layers (returned by bind())
-```
-
-`KVCache.bind(page_table, total_len)` returns a `KvcacheView` used by attention layers via `write()` / `gather()`.
-
-## Continuous Batching
-
-`InferenceScheduler` runs a daemon thread with a 4-phase loop:
-
-```
-1. Cleanup → Remove finished tasks, free KV pages
-2. Refill  → Pop from waiting_queue, task_alloc pages, activate
-3. Prefill → Group by (prompt_len, start_pos), run full forward
-4. Decode  → Pick largest same-position group, single-token forward
-```
-
-## Sampling (Strategy Pattern)
-
-```
-BaseSamplingStrategy (ABC)
-  ├── TemperatureStrategy
-  ├── TopKStrategy
-  ├── TopPStrategy
-  └── SamplingPipeline
-```
-
-`SamplingPipeline` composes them: Temperature → Top-K → Top-P → softmax → multinomial.  
-`sample()` is a convenience shortcut for one-shot usage.
-
-## Protocol Handlers (Strategy Pattern)
-
-```python
-class ProtocolHandler:  # concrete orchestrator
-    def __init__(self, request, engine, builder): ...
-    async def handle(self):
-        prompt, ctx, stops = builder.prepare(request, engine)
-        agen = engine.generate_async(prompt, ...)
-        if stream: self._handle_stream(agen, ctx, stops)
-        else:      return await self._handle_non_stream(agen, ctx, stops)
-```
-
-`ResponseBuilder` (ABC): `prepare()`, `format_stream_start()`, `format_chunk()`, `format_stream_end()`, `format_response()`.
-
-`OpenAIResponseBuilder` → `/v1/chat/completions`, `AnthropicResponseBuilder` → `/v1/messages`.
-
-Adding a protocol = one builder file, no handler subclassing needed.
-
-## Engine & GenerateResult
-
-```
-InferenceEngine
-  ├── generate(prompt, stream, ...) → str | List[str] | Generator
-  ├── generate_with_request(req)    → same
-  ├── generate_async(prompt, ...)   → AsyncGenerator
-  ├── get_stats()                   → Dict
-  └── shutdown()
-```
-
-`GenerateResult` uses `Condition` for non-streaming (`wait_completion()`) and `Event` for streaming (`wait()`). Stream callback is `cb(token)`.
-
-## HTTP API
-
-```
-POST /v1/chat/completions   OpenAI
-POST /v1/messages            Anthropic
-GET  /health                 {"status":"ok","model_loaded":true}
-GET  /stats                  scheduler statistics
-```
-
-### OpenAI
-
-```bash
-curl -X POST http://localhost:8000/v1/chat/completions \
-  -H "Content-Type: application/json" \
-  -d '{"messages":[{"role":"user","content":"Hello"}],"max_tokens":512}'
-```
-
-Response:
-```json
-{
-  "id": "chatcmpl-abc123",
-  "object": "chat.completion",
-  "created": 1717000000,
-  "model": "astrai",
-  "choices": [{"index": 0, "message": {"role": "assistant", "content": "Hello!"}, "finish_reason": "stop"}],
-  "usage": {"prompt_tokens": 5, "completion_tokens": 10, "total_tokens": 15}
-}
-```
-
-Streaming SSE: `object: "chat.completion.chunk"` — starts with role delta, then token chunks, ends with finish chunk + usage stats, then `data: [DONE]`.
-
-### Anthropic
-
-```bash
-curl -X POST http://localhost:8000/v1/messages \
-  -H "Content-Type: application/json" \
-  -d '{"model":"astrai","system":"You are helpful.","messages":[{"role":"user","content":"Hello"}],"max_tokens":512}'
-```
-
-Supports `stop_sequences` and streaming via `event: content_block_delta`.
-
-### GenerationRequest Parameters
-
-| Param | Type | Default | Description |
-|-------|------|---------|-------------|
-| `messages` | List[dict] | required | Chat messages (role, content) |
-| `top_k` | int | 50 | Top-k count |
-| `top_p` | float | 1.0 | Nucleus threshold |
-| `temperature` | float | 1.0 | Sampling temperature (> 0.0) |
-| `max_tokens` | Optional[int] | None | Max generation length |
-| `stream` | bool | False | Stream output |
-
-### SSE Streaming Format
-
-**OpenAI** (`/v1/chat/completions`, `stream=true`):
-
-```
-data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":...,"model":"astrai",
-       "choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}
-
-data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":0,"model":"astrai",
-       "choices":[{"index":0,"delta":{"content":"Hello"},"finish_reason":null}]}
-
-data: {"id":"chatcmpl-...","object":"chat.completion.chunk","created":...,"model":"astrai",
-       "choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
-
-data: {"prompt_tokens":5,"completion_tokens":1,"total_tokens":6}
-
-data: [DONE]
-```
-
-**Anthropic** (`/v1/messages`, `stream=true`):
-
-```
-event: message_start
-data: {"type":"message_start","message":{"id":"msg_...","model":"astrai","role":"assistant",
-       "content":[],"usage":{"input_tokens":0}}}
-
-event: content_block_start
-data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}
-
-event: content_block_delta
-data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Hello"}}
-
-event: content_block_stop
-data: {"type":"content_block_stop","index":0}
-
-event: message_delta
-data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{...}}
-
-event: message_stop
-data: {"type":"message_stop"}
-```
-
-### Error Responses
-
-The server returns standard HTTP status codes. Pydantic validation errors (e.g. missing required fields)
-are handled automatically by FastAPI with 422 status. The only application-level error is engine initialization:
-
-| Status | Meaning |
-|--------|---------|
-| 200 | Success |
-| 422 | Unprocessable entity (Pydantic validation) |
-| 503 | Service unavailable (model not loaded, engine not ready) |
-
-Error response body (503):
-
-```json
-{
-    "detail": "Engine not initialized"
-}
-```
-
-### Stats Endpoint
-
-```
-GET /stats
-```
-
-Response:
-
-```json
-{
-    "total_tasks": 128,
-    "total_tokens": 10240,
-    "active_tasks": 3,
-    "waiting_queue": 2
-}
-```
-
-## Engine API
-
-```python
-# Non-streaming
-engine.generate("Hello", stream=False)          # -> str
-engine.generate(["A", "B"], stream=False)       # -> List[str]
-
-# Streaming
-engine.generate("Hello", stream=True)           # -> Generator[str]
-engine.generate(["A", "B"], stream=True)        # -> Generator[Tuple[int, str]]
-
-# Async
-async for token in engine.generate_async("Hello", ...):    # -> AsyncGenerator[str]
-    print(token)
-```
-
-> Document Update Time: 2026-06-19

astrai/__init__.py

@@ -1,6 +1,9 @@
-__version__ = "1.3.8"
+__version__ = "1.3.12"
 __author__ = "ViperEkura"
 
+import logging
+import os
+
 from astrai.config import (
     AutoRegressiveLMConfig,
     BaseModelConfig,
@@ -12,13 +15,12 @@ from astrai.config import (
 from astrai.dataset import (
     BaseDataset,
     DatasetFactory,
-    ResumableDistributedSampler,
+    RDSampler,
     Store,
     StoreFactory,
 )
 from astrai.factory import BaseFactory
 from astrai.inference import (
-    GenerationRequest,
     InferenceEngine,
     ProtocolHandler,
     SamplingPipeline,
@@ -53,6 +55,30 @@ from astrai.trainer import (
     Trainer,
 )
 
+
+def setup_logging(level: str = "INFO"):
+    """Attach a handler to the ``astrai`` logger (only, not root).
+
+    Call once per process, e.g. at the top of CLI scripts.
+    Set ``ASTR_LOG_LEVEL`` to override the default ``INFO``.
+    """
+    _logger = logging.getLogger("astrai")
+    if _logger.handlers:
+        return
+    _level = getattr(
+        logging, os.environ.get("ASTR_LOG_LEVEL", level).upper(), logging.INFO
+    )
+    _logger.setLevel(_level)
+    _handler = logging.StreamHandler()
+    _handler.setFormatter(
+        logging.Formatter(
+            "%(asctime)s | %(levelname)-7s | %(name)s | %(message)s",
+            datefmt="%Y-%m-%d %H:%M:%S",
+        )
+    )
+    _logger.addHandler(_handler)
+
+
 __all__ = [
     "AutoRegressiveLM",
     "AutoRegressiveLMConfig",
@@ -71,13 +97,12 @@ __all__ = [
     "EmbeddingEncoder",
     "EncoderConfig",
     "ExecutorFactory",
-    "GenerationRequest",
     "InferenceEngine",
     "LoRAConfig",
     "Pipeline",
     "PipelineConfig",
     "ProtocolHandler",
-    "ResumableDistributedSampler",
+    "RDSampler",
     "SamplingPipeline",
     "SchedulerFactory",
     "Store",
@@ -94,5 +119,6 @@ __all__ = [
     "only_on_rank",
     "run_server",
     "sample",
+    "setup_logging",
     "spawn_parallel_fn",
 ]

astrai/config/base.py

@@ -1,92 +1,32 @@
 import json
-from dataclasses import MISSING, dataclass, fields
+from dataclasses import asdict
 from pathlib import Path
-from typing import Any, Dict, Optional, Self, Union, get_type_hints
+from typing import Any, Dict, Self, Union
+
+from pydantic import ConfigDict
+from pydantic.dataclasses import dataclass
 
 
-@dataclass
+@dataclass(config=ConfigDict(use_attribute_docstrings=True))
 class BaseConfig:
     def to_dict(self) -> Dict[str, Any]:
-        d = {}
-        for fld in fields(self):
-            v = getattr(self, fld.name)
-            if isinstance(v, (str, int, float, bool)):
-                d[fld.name] = v
-            elif v is None:
-                d[fld.name] = None
-            elif isinstance(v, (dict, list, tuple)):
-                try:
-                    val = list(v) if isinstance(v, tuple) else v
-                    json.dumps(val)
-                    d[fld.name] = val
-                except (TypeError, ValueError):
-                    pass
-            elif isinstance(v, BaseConfig):
-                d[fld.name] = v.to_dict()
-            elif hasattr(v, "__dataclass_fields__"):
-                sub = {}
-                for f in fields(v):
-                    a = getattr(v, f.name)
-                    sub[f.name] = list(a) if isinstance(a, tuple) else a
-                d[fld.name] = sub
-        return d
+        result = {}
+        for k, v in asdict(self).items():
+            if isinstance(v, tuple):
+                v = list(v)
+            try:
+                json.dumps(v)
+                result[k] = v
+            except (TypeError, ValueError):
+                # Skip non-serializable runtime objects (e.g. model_fn, dataset).
+                # TrainConfig mixes hyperparams with callables/datasets; only the
+                # JSON-serializable subset is written to checkpoint meta.
+                pass
+        return result
 
     @classmethod
     def from_dict(cls, d: Dict[str, Any]) -> Self:
-        hints = get_type_hints(cls)
-        inst = cls.__new__(cls)
-        for fld in fields(cls):
-            if fld.name in d:
-                v = d[fld.name]
-                target = cls._unwrap_optional(hints.get(fld.name))
-                if target is not None:
-                    try:
-                        v = cls._coerce(v, target)
-                    except (TypeError, ValueError):
-                        pass
-                object.__setattr__(inst, fld.name, v)
-            elif fld.default is not MISSING:
-                object.__setattr__(inst, fld.name, fld.default)
-            elif fld.default_factory is not MISSING:
-                object.__setattr__(inst, fld.name, fld.default_factory())
-            else:
-                object.__setattr__(inst, fld.name, None)
-        return inst
-
-    @staticmethod
-    def _unwrap_optional(tp) -> Optional[type]:
-        if tp is None:
-            return None
-        origin = getattr(tp, "__origin__", None)
-        if origin is not None:
-            args = getattr(tp, "__args__", ())
-            non_none = [a for a in args if a is not type(None)]
-            return non_none[0] if non_none else None
-        return tp
-
-    @staticmethod
-    def _coerce(value: Any, target_type: type) -> Any:
-        if target_type is bool and isinstance(value, bool):
-            return value
-        if (
-            target_type is int
-            and isinstance(value, (int, float))
-            and not isinstance(value, bool)
-        ):
-            return int(value)
-        if (
-            target_type is float
-            and isinstance(value, (int, float))
-            and not isinstance(value, bool)
-        ):
-            return float(value)
-        if target_type is str and isinstance(value, str):
-            return value
-        if isinstance(value, target_type):
-            return value
-        if isinstance(value, dict) and issubclass(target_type, BaseConfig):
-            return target_type.from_dict(value)
-        raise TypeError
+        return cls(**d)
 
     @classmethod
     def from_file(cls, path: Union[str, Path]) -> Self:

astrai/config/model_config.py

@@ -1,9 +1,14 @@
-from dataclasses import dataclass
 from typing import Any, Dict, Optional
 
+from pydantic import field_validator
+from pydantic.dataclasses import dataclass
+
 from astrai.config.base import BaseConfig
 from astrai.factory import BaseFactory
 
+_ATTN_TYPES = frozenset({"gqa", "mla"})
+_FFN_TYPES = frozenset({"mlp", "moe"})
+
 
 class ConfigFactory(BaseFactory[BaseConfig]):
     """Factory that dispatches config classes by ``model_type``."""
@@ -17,7 +22,12 @@ class ConfigFactory(BaseFactory[BaseConfig]):
 
 @dataclass
 class BaseModelConfig(BaseConfig):
-    """Base config with ``model_type`` dispatch and file I/O."""
+    """Base config with ``model_type`` dispatch and file I/O.
+
+    Args:
+        model_type (Optional[str]): Model type identifier for AutoModel dispatch. Defaults to None.
+        neftune_alpha (float): NEFTune noise alpha, 0=disabled, typical: 5.0. Defaults to 0.0.
+    """
 
     model_type: Optional[str] = None
     neftune_alpha: float = 0.0
@@ -26,57 +36,143 @@ class BaseModelConfig(BaseConfig):
 @dataclass
 @ConfigFactory.register("autoregressive_lm")
 class AutoRegressiveLMConfig(BaseModelConfig):
-    """Configuration for autoregressive language model."""
+    """Configuration for autoregressive language model.
+
+    Args:
+        model_type (Optional[str]): Model type identifier for AutoModel dispatch. Defaults to None.
+        neftune_alpha (float): NEFTune noise alpha, 0=disabled, typical: 5.0. Defaults to 0.0.
+        vocab_size (Optional[int]): Vocabulary size. Defaults to None.
+        hidden_size (Optional[int]): Hidden dimension size. Defaults to None.
+        num_hidden_layers (Optional[int]): Number of transformer layers. Defaults to None.
+        rms_norm_eps (Optional[float]): Epsilon for RMSNorm. Defaults to None.
+        intermediate_size (Optional[int]): Intermediate size in FFN. Defaults to None.
+        tie_word_embeddings (Optional[bool]): Whether to tie embedding and lm_head weights. Defaults to None.
+        max_position_embeddings (Optional[int]): Maximum sequence length the model was trained with. Defaults to None.
+        rope_theta (Optional[float]): Base frequency for RoPE. Defaults to None.
+        rope_scaling (Optional[dict]): RoPE scaling config, e.g. {"type": "linear", "factor": 4.0}. Defaults to None.
+        attn_type (str): Attention type: 'gqa' or 'mla'. Defaults to "gqa".
+        num_attention_heads (Optional[int]): Number of query attention heads. Defaults to None.
+        num_key_value_heads (Optional[int]): Number of key/value heads for GQA. Defaults to None.
+        use_qk_norm (Optional[bool]): Whether to apply RMSNorm to Q/K. Defaults to None.
+        use_gated_attention (Optional[bool]): Whether to use gated attention. Defaults to None.
+        kv_lora_rank (Optional[int]): KV compression rank, MLA only. Defaults to None.
+        qk_nope_head_dim (Optional[int]): Non-RoPE head dimension, MLA only. Defaults to None.
+        qk_rope_head_dim (Optional[int]): RoPE head dimension, MLA only. Defaults to None.
+        ffn_type (str): FFN type: 'mlp' or 'moe'. Defaults to "mlp".
+        n_routed_experts (Optional[int]): Number of routed experts, MoE only. Defaults to None.
+        n_shared_experts (Optional[int]): Number of shared experts, MoE only. Defaults to None.
+        n_activated_experts (Optional[int]): Number of activated experts per token, MoE only. Defaults to None.
+        topk_method (Optional[str]): Top-k routing method, MoE only. Defaults to None.
+        moe_intermediate_size (Optional[int]): Expert hidden dim, defaults to intermediate_size if None. MoE only.
+        shared_expert_intermediate_size (Optional[int]): Shared expert hidden dim, defaults to intermediate_size if None. MoE only.
+        norm_topk_prob (bool): Normalize top-k routing probabilities. Defaults to True.
+        decoder_sparse_step (int): Frequency of MoE layers, 1=every layer. Defaults to 1.
+        mlp_only_layers (Optional[list[int]]): Layer indices using dense MLP instead of MoE. Defaults to None.
+    """
 
     vocab_size: Optional[int] = None
-    dim: Optional[int] = None
-    n_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
-    tie_weight: Optional[bool] = None
-
-    max_len: Optional[int] = None
+    hidden_size: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
+    rms_norm_eps: Optional[float] = None
+    intermediate_size: Optional[int] = None
+    tie_word_embeddings: Optional[bool] = None
+    max_position_embeddings: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
-
     attn_type: str = "gqa"
-    n_heads: Optional[int] = None
-    n_kv_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
+    num_key_value_heads: Optional[int] = None
     use_qk_norm: Optional[bool] = None
     use_gated_attention: Optional[bool] = None
-
     kv_lora_rank: Optional[int] = None
     qk_nope_head_dim: Optional[int] = None
     qk_rope_head_dim: Optional[int] = None
-
     ffn_type: str = "mlp"
     n_routed_experts: Optional[int] = None
     n_shared_experts: Optional[int] = None
     n_activated_experts: Optional[int] = None
     topk_method: Optional[str] = None
+    moe_intermediate_size: Optional[int] = None
+    shared_expert_intermediate_size: Optional[int] = None
+    norm_topk_prob: bool = True
+    decoder_sparse_step: int = 1
+    mlp_only_layers: Optional[list[int]] = None
+    moe_aux_loss_coef: float = 0.01
+
+    @field_validator("attn_type")
+    def _validate_attn_type(cls, v: str) -> str:
+        if v not in _ATTN_TYPES:
+            raise ValueError(
+                f"attn_type must be one of {sorted(_ATTN_TYPES)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("ffn_type")
+    def _validate_ffn_type(cls, v: str) -> str:
+        if v not in _FFN_TYPES:
+            raise ValueError(f"ffn_type must be one of {sorted(_FFN_TYPES)}, got {v!r}")
+        return v
+
+    @field_validator("decoder_sparse_step")
+    def _validate_decoder_sparse_step(cls, v: int) -> int:
+        if v < 1:
+            raise ValueError(f"decoder_sparse_step must be at least 1, got {v}")
+        return v
 
 
 @dataclass
 @ConfigFactory.register("embedding")
 class EncoderConfig(BaseModelConfig):
-    """Configuration for embedding encoder model."""
+    """Configuration for embedding encoder model.
+
+    Args:
+        model_type (Optional[str]): Model type identifier for AutoModel dispatch. Defaults to None.
+        neftune_alpha (float): NEFTune noise alpha, 0=disabled, typical: 5.0. Defaults to 0.0.
+        vocab_size (Optional[int]): Vocabulary size. Defaults to None.
+        hidden_size (Optional[int]): Hidden dimension size. Defaults to None.
+        num_hidden_layers (Optional[int]): Number of transformer layers. Defaults to None.
+        rms_norm_eps (Optional[float]): Epsilon for RMSNorm. Defaults to None.
+        intermediate_size (Optional[int]): Intermediate size in FFN. Defaults to None.
+        max_position_embeddings (Optional[int]): Maximum sequence length the model was trained with. Defaults to None.
+        rope_theta (Optional[float]): Base frequency for RoPE. Defaults to None.
+        rope_scaling (Optional[dict]): RoPE scaling config, e.g. {"type": "linear", "factor": 4.0}. Defaults to None.
+        attn_type (str): Attention type: 'gqa' or 'mla'. Defaults to "gqa".
+        num_attention_heads (Optional[int]): Number of query attention heads. Defaults to None.
+        num_key_value_heads (Optional[int]): Number of key/value heads for GQA. Defaults to None.
+        use_qk_norm (Optional[bool]): Whether to apply RMSNorm to Q/K. Defaults to None.
+        use_gated_attention (Optional[bool]): Whether to use gated attention. Defaults to None.
+        ffn_type (str): FFN type: 'mlp' or 'moe'. Defaults to "mlp".
+        pooling_type (Optional[str]): Pooling strategy for embedding, e.g. 'mean', 'cls'. Defaults to None.
+        normalize_embeddings (Optional[bool]): Whether to L2-normalize output embeddings. Defaults to None.
+    """
 
     vocab_size: Optional[int] = None
-    dim: Optional[int] = None
-    n_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
-
-    max_len: Optional[int] = None
+    hidden_size: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
+    rms_norm_eps: Optional[float] = None
+    intermediate_size: Optional[int] = None
+    max_position_embeddings: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
-
     attn_type: str = "gqa"
-    n_heads: Optional[int] = None
-    n_kv_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
+    num_key_value_heads: Optional[int] = None
     use_qk_norm: Optional[bool] = None
     use_gated_attention: Optional[bool] = None
-
     ffn_type: str = "mlp"
     pooling_type: Optional[str] = None
     normalize_embeddings: Optional[bool] = None
+
+    @field_validator("attn_type")
+    def _validate_attn_type(cls, v: str) -> str:
+        if v not in _ATTN_TYPES:
+            raise ValueError(
+                f"attn_type must be one of {sorted(_ATTN_TYPES)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("ffn_type")
+    def _validate_ffn_type(cls, v: str) -> str:
+        if v not in _FFN_TYPES:
+            raise ValueError(f"ffn_type must be one of {sorted(_FFN_TYPES)}, got {v!r}")
+        return v

astrai/config/preprocess_config.py

@@ -5,11 +5,19 @@ modes, both driven declaratively through ``input.sections`` or
 ``input.sources``.
 """
 
-from dataclasses import dataclass, field
+from dataclasses import field
 from typing import Dict, List, Optional
 
+from pydantic import field_validator
+from pydantic.dataclasses import dataclass
+
 from astrai.config.base import BaseConfig
 
+_PACKING_STRATEGIES = frozenset({"simple", "bfd", "bfd_split"})
+_TRUNCATION_MODES = frozenset({"keep_start", "keep_end"})
+_STORAGE_FORMATS = frozenset({"bin", "jsonl"})
+_POSITION_IDS_MODES = frozenset({"none", "doc_reset", "continuous"})
+
 
 @dataclass
 class InputConfig(BaseConfig):
@@ -25,6 +33,10 @@ class InputConfig(BaseConfig):
             "chosen": {"sections": [{"field": "chosen", ...}]},
             "rejected": {"sections": [{"field": "rejected", ...}]},
         }}}
+
+    Args:
+        sections (Optional[List[Dict]]): Section list for single-output mode. Defaults to None.
+        sources (Optional[Dict[str, Dict]]): Source map for multi-output mode, DPO/GRPO. Defaults to None.
     """
 
     sections: Optional[List[Dict]] = None
@@ -33,63 +45,67 @@ class InputConfig(BaseConfig):
 
 @dataclass
 class ProcessingConfig(BaseConfig):
-    """Processing configuration.
+    """Processing configuration for tokenization and packing.
 
-    Parameters
-    ----------
-    max_seq_len : int
-        Maximum sequence length (default: 2048).
-    min_chars : int
-        Minimum number of characters to keep (default: 50).
-    max_chars : int
-        Maximum number of characters to keep (default: 2_000_000).
-    max_items : Optional[int]
-        Maximum number of items to process (default: None, unlimited).
-    packing_strategy : str
-        How to pack sequences into a contiguous stream.
-
-        - ``"simple"``: sequential concatenation (default, backward compatible).
-        - ``"bfd"``: best-fit decreasing bin packing, minimises wasted tokens.
-        - ``"bfd_split"``: BFD with over-length sequences split into chunks.
-    max_packed_len : int
-        Maximum length of a packed bin. Sequences longer than this are
-        truncated or split depending on ``packing_strategy`` (default: 8192).
-    truncation_mode : str
-        How to truncate sequences longer than ``max_packed_len``.
-
-        - ``"keep_start"``: keep the first ``max_packed_len`` tokens (default).
-        - ``"keep_end"``: keep the last ``max_packed_len`` tokens.
+    Args:
+        max_seq_len (int): Maximum sequence length. Defaults to 2048.
+        min_chars (int): Minimum number of characters to keep. Defaults to 50.
+        max_chars (int): Maximum number of characters to keep. Defaults to 2_000_000.
+        max_items (Optional[int]): Maximum number of items to process, None=unlimited. Defaults to None.
+        batch_size (int): Number of records tokenized together. Defaults to 256.
+        packing_strategy (str): How to pack sequences: 'simple', 'bfd', or 'bfd_split'. Defaults to "simple".
+        max_packed_len (int): Maximum length of a packed bin. Defaults to 8192.
+        truncation_mode (str): How to truncate over-length sequences: 'keep_start' or 'keep_end'. Defaults to "keep_start".
     """
 
     max_seq_len: int = 2048
     min_chars: int = 50
     max_chars: int = 2_000_000
     max_items: Optional[int] = None
+    batch_size: int = 256
     packing_strategy: str = "simple"
     max_packed_len: int = 8192
     truncation_mode: str = "keep_start"
 
+    @field_validator("packing_strategy")
+    def _validate_packing_strategy(cls, v: str) -> str:
+        if v not in _PACKING_STRATEGIES:
+            raise ValueError(
+                f"packing_strategy must be one of {sorted(_PACKING_STRATEGIES)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("truncation_mode")
+    def _validate_truncation_mode(cls, v: str) -> str:
+        if v not in _TRUNCATION_MODES:
+            raise ValueError(
+                f"truncation_mode must be one of {sorted(_TRUNCATION_MODES)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("max_seq_len", "batch_size", "max_packed_len")
+    def _validate_positive_int(cls, v: int) -> int:
+        if v <= 0:
+            raise ValueError(f"must be positive, got {v}")
+        return v
+
+    @field_validator("min_chars")
+    def _validate_non_negative(cls, v: int) -> int:
+        if v < 0:
+            raise ValueError(f"min_chars must be non-negative, got {v}")
+        return v
+
 
 @dataclass
 class OutputConfig(BaseConfig):
-    """Output configuration.
+    """Output configuration for storage.
 
-    Parameters
-    ----------
-    domain_key : Optional[str]
-        Domain key for the output store (default: None).
-    storage_format : str
-        Storage format, one of ``"bin"``, ``"jsonl"`` (default: ``"bin"``).
-    max_tokens_per_shard : int
-        Maximum tokens per shard before splitting (default: 100_000_000).
-    dtype : Dict[str, str]
-        Per-key dtype overrides, e.g. ``{"input_ids": "int32"}`` (default: {}).
-    position_ids_mode : Optional[str]
-        How to compute position_ids in packed sequences.
-
-         - ``"none"``: do not generate (default).
-        - ``"doc_reset"``: reset to 0 at each document boundary.
-        - ``"continuous"``: sequential 0, 1, 2, ... (pretrain, single doc).
+    Args:
+        domain_key (Optional[str]): Domain key for the output store. Defaults to None.
+        storage_format (str): Storage format: 'bin' or 'jsonl'. Defaults to "bin".
+        max_tokens_per_shard (int): Maximum tokens per shard before splitting. Defaults to 100_000_000.
+        dtype (Dict[str, str]): Per-key dtype overrides, e.g. {"input_ids": "int32"}. Defaults to {}.
+        position_ids_mode (str): Position ids mode: 'none', 'doc_reset', or 'continuous'. Defaults to "doc_reset".
     """
 
     domain_key: Optional[str] = None
@@ -98,9 +114,36 @@ class OutputConfig(BaseConfig):
     dtype: Dict[str, str] = field(default_factory=dict)
     position_ids_mode: str = "doc_reset"
 
+    @field_validator("storage_format")
+    def _validate_storage_format(cls, v: str) -> str:
+        if v not in _STORAGE_FORMATS:
+            raise ValueError(
+                f"storage_format must be one of {sorted(_STORAGE_FORMATS)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("position_ids_mode")
+    def _validate_position_ids_mode(cls, v: str) -> str:
+        if v not in _POSITION_IDS_MODES:
+            raise ValueError(
+                f"position_ids_mode must be one of {sorted(_POSITION_IDS_MODES)}, got {v!r}"
+            )
+        return v
+
 
 @dataclass
 class PipelineConfig(BaseConfig):
+    """Top-level preprocessing pipeline config.
+
+    Args:
+        version (int): Config schema version. Defaults to 1.
+        input (InputConfig): Input mapping config.
+        mask (Dict[str, str]): Per-field mask labels, e.g. {"system": "mask", "assistant": "train"}. Defaults to {}.
+        mask_default (str): Default mask label for unlisted fields. Defaults to "mask".
+        preprocessing (ProcessingConfig): Processing config.
+        output (OutputConfig): Output config.
+    """
+
     version: int = 1
     input: InputConfig = field(default_factory=InputConfig)
     mask: Dict[str, str] = field(default_factory=dict)

astrai/config/train_config.py

@@ -1,7 +1,9 @@
-from dataclasses import dataclass, field, fields
+from dataclasses import field
 from typing import Any, Callable, Dict, List, Optional
 
 import torch.nn as nn
+from pydantic import ConfigDict, field_validator, model_validator
+from pydantic.dataclasses import dataclass
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
 from torch.utils.data import Dataset
@@ -9,142 +11,208 @@ from torch.utils.data import Dataset
 from astrai.config.base import BaseConfig
 from astrai.model.components.lora import LoRAConfig
 
-
-def required(**kw):
-    return {"required": True, **kw}
+_TRAIN_TYPES = frozenset({"seq", "sft", "dpo", "grpo", "online_grpo", "online_dpo"})
+_PARALLEL_MODES = frozenset({"none", "ddp", "fsdp"})
+_BACKENDS = frozenset({"nccl", "gloo"})
+_START_METHODS = frozenset({"spawn", "fork", "forkserver"})
+_COMPILE_MODES = frozenset({"default", "reduce-overhead", "max-autotune"})
 
 
-@dataclass
+@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
 class TrainConfig(BaseConfig):
-    # basic setting
-    model_fn: Callable[[], nn.Module] = field(
-        default=None, metadata=required(help="Model factory for training.")
-    )
-    strategy: str = field(default=None, metadata=required(help="Training strategy."))
-    dataset: Dataset = field(
-        default=None, metadata=required(help="Dataset for training.")
-    )
-    optimizer_fn: Callable[[nn.Module], Optimizer] = field(
-        default=None, metadata=required(help="Optimizer factory for training.")
-    )
-    scheduler_fn: Callable[[Optimizer], LRScheduler] = field(
-        default=None, metadata=required(help="Scheduler factory for training.")
-    )
-    n_epoch: int = field(default=1, metadata={"help": "Number of epochs for training."})
-    batch_per_device: int = field(
-        default=4, metadata={"help": "Batch size per device."}
-    )
-    grad_accum_steps: int = field(
-        default=1, metadata={"help": "Number of iterations between steps."}
-    )
-    max_grad_norm: float = field(
-        default=1.0, metadata={"help": "Maximum gradient norm."}
-    )
-    gradient_checkpointing_modules: List[str] = field(
-        default_factory=list,
-        metadata={"help": "Module types to enable activation checkpointing for."},
-    )
+    """Training configuration.
 
-    # checkpoint setting
-    start_epoch: int = field(default=0, metadata={"help": "Start epoch for training."})
-    start_samples: int = field(
-        default=0,
-        metadata={
-            "help": "Start samples count (per rank). Superseded by checkpoint consumed_samples."
-        },
-    )
-    ckpt_dir: str = field(
-        default="./checkpoint", metadata={"help": "Checkpoint directory."}
-    )
-    ckpt_interval: int = field(
-        default=5000,
-        metadata={"help": "Number of optimizer steps between checkpoints."},
-    )
+    Combines hyperparameters with runtime objects (model_fn, dataset, etc.).
+    Only JSON-serializable fields are written to checkpoint meta via to_dict().
 
-    # lora setting
-    lora: Optional[LoRAConfig] = field(
-        default=None,
-        metadata={"help": "LoRA config. None means full fine-tuning."},
-    )
+    Args:
+        model_fn (Callable[[], nn.Module]): Model factory for training.
+        strategy (str): Training strategy (seq, sft, dpo, grpo, online_*).
+        dataset (Dataset): Dataset for training.
+        optimizer_fn (Callable[[nn.Module], Optimizer]): Optimizer factory for training.
+        optimizer_name (Optional[str]): Serializable built-in optimizer identifier. Defaults to None.
+        optimizer_hyperparameters (Dict[str, Any]): Serializable optimizer settings. Defaults to {}.
+        scheduler_fn (Callable[[Optimizer], LRScheduler]): Scheduler factory for training.
+        n_epoch (int): Number of epochs for training. Defaults to 1.
+        batch_per_device (int): Batch size per device. Defaults to 4.
+        grad_accum_steps (int): Number of iterations between optimizer steps. Defaults to 1.
+        max_grad_norm (Optional[float]): Maximum gradient norm. None disables clipping. Defaults to 1.0.
+        gradient_checkpointing_modules (List[type]): Module types to enable activation checkpointing for. Defaults to [].
+        compile_mode (Optional[str]): torch.compile mode: 'default', 'reduce-overhead', 'max-autotune', or None. Defaults to None.
+        start_epoch (int): Start epoch for training. Defaults to 0.
+        start_samples (int): Start samples count (per rank). Superseded by checkpoint consumed_samples. Defaults to 0.
+        ckpt_dir (str): Checkpoint directory. Defaults to "./checkpoint".
+        ckpt_interval (int): Number of optimizer steps between checkpoints. Defaults to 5000.
+        lora (Optional[LoRAConfig]): LoRA config. None means full fine-tuning. Defaults to None.
+        metrics (List[str]): Metrics to record during training. Defaults to ["loss", "lr", "grad_norm"].
+        random_seed (int): Random seed. Defaults to 3407.
+        num_workers (int): Number of workers for dataloader. Defaults to 0.
+        prefetch_factor (Optional[int]): Prefetch factor for dataloader. Defaults to None.
+        pin_memory (bool): Pin memory for dataloader. Defaults to False.
+        collate_fn (Optional[Callable[[List[Any]], Any]]): Collate function for dataloader (e.g. dpo_collate_fn). Defaults to None.
+        nprocs (int): Number of processes for distributed training. Defaults to 1.
+        backend (str): Distributed training backend. Defaults to "nccl".
+        master_addr (str): Master address for distributed training. Defaults to "localhost".
+        master_port (str): Master port for distributed training. Defaults to "29500".
+        parallel_mode (str): Parallel strategy: none, ddp, fsdp. Defaults to "none".
+        start_method (str): Multiprocessing start method: spawn/fork/forkserver. Defaults to "spawn".
+        device_type (str): Device type for distributed training. Defaults to "cuda".
+        val_dataset (Optional[Dataset]): Dataset for validation. Defaults to None.
+        val_split (Optional[float]): Ratio to split from training dataset for validation, e.g. 0.05. Defaults to None.
+        val_step (int): Number of optimizer steps between validation runs. Defaults to 1000.
+        neftune_alpha (float): NEFTune noise alpha, 0=disabled, typical: 5.0. Defaults to 0.0.
+        moe_aux_loss_coef (float): Weight applied to the MoE load-balancing loss. Defaults to 0.01.
+        rollout_interval (int): Number of optimizer steps between online rollouts. Defaults to 512.
+        rollout_temperature (float): Sampling temperature for online rollout. Defaults to 0.7.
+        rollout_top_k (int): Top-k filtering for online rollout, 0=disable. Defaults to 0.
+        rollout_top_p (float): Top-p (nucleus) filtering for online rollout. Defaults to 0.9.
+        rollout_max_tokens (int): Maximum generated tokens per response in rollout. Defaults to 1024.
+        reward_model_fn (Optional[Callable]): Factory for reward model, required for online RL strategies. Defaults to None.
+        executor_kwargs (Dict[str, Any]): Extra kwargs passed to ExecutorFactory.create(). Defaults to {}.
+        extra_kwargs (Dict[str, Any]): Other arguments. Defaults to {}.
+    """
 
-    # metric setting
-    log_dir: str = field(
-        default="./checkpoint/logs", metadata={"help": "Directory for metric logs."}
-    )
-    metrics: List[str] = field(
-        default_factory=lambda: ["loss", "lr", "grad_norm"],
-        metadata={"help": "Metrics to record during training."},
-    )
+    model_fn: Callable[[], nn.Module]
+    strategy: str
+    dataset: Dataset
+    optimizer_fn: Callable[[nn.Module], Optimizer]
+    scheduler_fn: Callable[[Optimizer], LRScheduler]
+    optimizer_name: Optional[str] = None
+    optimizer_hyperparameters: Dict[str, Any] = field(default_factory=dict)
+    n_epoch: int = 1
+    batch_per_device: int = 4
+    grad_accum_steps: int = 1
+    max_grad_norm: Optional[float] = 1.0
+    gradient_checkpointing_modules: List[type] = field(default_factory=list)
+    compile_mode: Optional[str] = None
 
-    # dataloader setting
-    random_seed: int = field(default=3407, metadata={"help": "Random seed."})
-    num_workers: int = field(
-        default=0, metadata={"help": "Number of workers for dataloader."}
-    )
-    prefetch_factor: Optional[int] = field(
-        default=None, metadata={"help": "Prefetch factor for dataloader."}
-    )
-    pin_memory: bool = field(
-        default=False, metadata={"help": "Pin memory for dataloader."}
-    )
+    start_epoch: int = 0
+    start_samples: int = 0
+    ckpt_dir: str = "./checkpoint"
+    ckpt_interval: int = 5000
 
-    # distributed training
-    nprocs: int = field(
-        default=1, metadata={"help": "Number of processes for distributed training."}
-    )
-    backend: str = field(
-        default="nccl", metadata={"help": "Distributed training backend."}
-    )
-    master_addr: str = field(
-        default="localhost",
-        metadata={"help": "Master address for distributed training."},
-    )
-    master_port: str = field(
-        default="29500", metadata={"help": "Master port for distributed training."}
-    )
-    parallel_mode: str = field(
-        default="none",
-        metadata={"help": "Parallel strategy: none, ddp, fsdp."},
-    )
-    start_method: str = field(
-        default="spawn",
-        metadata={"help": "Multiprocessing start method (spawn/fork/forkserver)."},
-    )
+    lora: Optional[LoRAConfig] = None
 
-    # others
-    device_type: str = field(
-        default="cuda", metadata={"help": "Device type for distributed training."}
-    )
-    val_dataset: Optional[Dataset] = field(
-        default=None, metadata={"help": "Dataset for validation."}
-    )
-    val_split: Optional[float] = field(
-        default=None,
-        metadata={
-            "help": "Ratio to split from training dataset for validation (e.g. 0.05). Ignored if val_dataset is set."
-        },
-    )
-    val_step: int = field(
-        default=1000,
-        metadata={"help": "Number of optimizer steps between validation runs."},
-    )
-    neftune_alpha: float = field(
-        default=0.0,
-        metadata={"help": "NEFTune noise alpha (0=disabled, typical: 5.0)."},
-    )
+    metrics: List[str] = field(default_factory=lambda: ["loss", "lr", "grad_norm"])
 
-    executor_kwargs: Dict[str, Any] = field(
-        default_factory=dict,
-        metadata={"help": "Extra kwargs passed to ExecutorFactory.create()."},
-    )
-    extra_kwargs: Dict[str, Any] = field(
-        default_factory=dict, metadata={"help": "Other arguments."}
-    )
+    random_seed: int = 3407
+    num_workers: int = 0
+    prefetch_factor: Optional[int] = None
+    pin_memory: bool = False
+    collate_fn: Optional[Callable[[List[Any]], Any]] = None
 
-    def __post_init__(self):
-        self.validate()
+    nprocs: int = 1
+    backend: str = "nccl"
+    master_addr: str = "localhost"
+    master_port: str = "29500"
+    parallel_mode: str = "none"
+    start_method: str = "spawn"
 
-    def validate(self):
-        for fld in fields(self):
-            if fld.metadata.get("required") and getattr(self, fld.name) is None:
-                raise ValueError(f"TrainConfig.{fld.name} is required but got None.")
+    device_type: str = "cuda"
+    val_dataset: Optional[Dataset] = None
+    val_split: Optional[float] = None
+    val_step: int = 1000
+    neftune_alpha: float = 0.0
+    moe_aux_loss_coef: float = 0.01
+
+    rollout_interval: int = 512
+    rollout_temperature: float = 0.7
+    rollout_top_k: int = 0
+    rollout_top_p: float = 0.9
+    rollout_max_tokens: int = 1024
+    reward_model_fn: Optional[Callable] = None
+
+    executor_kwargs: Dict[str, Any] = field(default_factory=dict)
+    extra_kwargs: Dict[str, Any] = field(default_factory=dict)
+
+    @field_validator("strategy")
+    def _validate_strategy(cls, v: str) -> str:
+        if v not in _TRAIN_TYPES:
+            raise ValueError(
+                f"strategy must be one of {sorted(_TRAIN_TYPES)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("parallel_mode")
+    def _validate_parallel_mode(cls, v: str) -> str:
+        if v not in _PARALLEL_MODES:
+            raise ValueError(
+                f"parallel_mode must be one of {sorted(_PARALLEL_MODES)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("backend")
+    def _validate_backend(cls, v: str) -> str:
+        if v not in _BACKENDS:
+            raise ValueError(f"backend must be one of {sorted(_BACKENDS)}, got {v!r}")
+        return v
+
+    @field_validator("start_method")
+    def _validate_start_method(cls, v: str) -> str:
+        if v not in _START_METHODS:
+            raise ValueError(
+                f"start_method must be one of {sorted(_START_METHODS)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("compile_mode")
+    def _validate_compile_mode(cls, v: Optional[str]) -> Optional[str]:
+        if v is not None and v not in _COMPILE_MODES:
+            raise ValueError(
+                f"compile_mode must be one of {sorted(_COMPILE_MODES)} or None, got {v!r}"
+            )
+        return v
+
+    @field_validator(
+        "n_epoch",
+        "batch_per_device",
+        "grad_accum_steps",
+        "ckpt_interval",
+        "val_step",
+        "rollout_interval",
+        "rollout_max_tokens",
+    )
+    def _validate_positive_int(cls, v: int) -> int:
+        if v <= 0:
+            raise ValueError(f"must be positive, got {v}")
+        return v
+
+    @field_validator("rollout_temperature")
+    def _validate_positive_float(cls, v: float) -> float:
+        if v <= 0:
+            raise ValueError(f"must be positive, got {v}")
+        return v
+
+    @field_validator("rollout_top_p")
+    def _validate_top_p(cls, v: float) -> float:
+        if not 0 < v <= 1:
+            raise ValueError(f"rollout_top_p must be in (0, 1], got {v}")
+        return v
+
+    @field_validator(
+        "rollout_top_k", "num_workers", "neftune_alpha", "moe_aux_loss_coef"
+    )
+    def _validate_non_negative(cls, v):
+        if v < 0:
+            raise ValueError(f"must be non-negative, got {v}")
+        return v
+
+    @field_validator("max_grad_norm")
+    def _validate_max_grad_norm(cls, v: Optional[float]) -> Optional[float]:
+        if v is not None and v <= 0:
+            raise ValueError(f"max_grad_norm must be positive or None, got {v}")
+        return v
+
+    @field_validator("val_split")
+    def _validate_val_split(cls, v: Optional[float]) -> Optional[float]:
+        if v is not None and not 0 < v < 1:
+            raise ValueError(f"val_split must be in (0, 1) or None, got {v}")
+        return v
+
+    @model_validator(mode="after")
+    def _validate_online_strategy(self) -> "TrainConfig":
+        if self.strategy.startswith("online_") and self.reward_model_fn is None:
+            raise ValueError(
+                f"reward_model_fn is required for online RL strategy {self.strategy!r}"
+            )
+        return self

astrai/dataset/__init__.py

@@ -1,35 +1,37 @@
 from astrai.dataset.dataset import (
     BaseDataset,
     DatasetFactory,
+    dpo_collate_fn,
+    grpo_collate_fn,
 )
-from astrai.dataset.sampler import ResumableDistributedSampler
+from astrai.dataset.sampler import RDSampler
 from astrai.dataset.storage import (
-    H5Store,
     JsonlStore,
     MmapStore,
+    Recordable,
     Store,
     StoreFactory,
+    Streamable,
     detect_format,
 )
 from astrai.serialization import (
     load_bin,
-    load_h5,
     save_bin,
-    save_h5,
 )
 
 __all__ = [
     "BaseDataset",
     "DatasetFactory",
+    "dpo_collate_fn",
+    "grpo_collate_fn",
     "Store",
+    "Streamable",
+    "Recordable",
     "StoreFactory",
-    "H5Store",
     "MmapStore",
     "JsonlStore",
     "detect_format",
-    "save_h5",
-    "load_h5",
     "save_bin",
     "load_bin",
-    "ResumableDistributedSampler",
+    "RDSampler",
 ]

astrai/dataset/dataset.py

@@ -1,7 +1,31 @@
-"""Dataset implementations with factory pattern for training."""
+"""Dataset implementations for training.
+
+Composition over inheritance — every dataset is a thin wrapper that
+binds a :class:`Store` to a particular train-type's key mapping.  All
+sample-id → token/record indexing lives on the Store; datasets never
+know about window/stride math or segment layouts.
+
+Class hierarchy:
+
+    BaseDataset (ABC)         — holds a Store, exposes __len__/keys,
+                                 overrides __getitem__
+    ├── SEQDataset            — next-token prediction (stream)
+    ├── SFTDataset            — loss-mask + position_ids (stream)
+    ├── DPODataset            — chosen/rejected pairs (record)
+    └── GRPODataset           — prompt + response group (record)
+
+``DatasetFactory.load(train_type, load_path, window_size, stride, …)``
+builds the Store (auto-detecting format) before constructing the
+matching dataset.  Passing ``store=`` skips Store construction.
+
+When a record dataset (DPO) reads from raw JSONL, a *processor*
+function (pure ``record -> Dict[str, Tensor]``) is forwarded to
+:class:`JsonlStore` so tokenisation happens on the fly.
+"""
 
 from abc import ABC, abstractmethod
-from typing import Dict, List, Optional
+from functools import partial
+from typing import Callable, Dict, List, Optional
 
 import torch
 from torch import Tensor
@@ -13,202 +37,401 @@ from astrai.dataset.storage import (
     detect_format,
 )
 from astrai.factory import BaseFactory
+from astrai.tokenize import AutoTokenizer
+
+
+def dpo_tokenize(
+    record: dict,
+    tokenizer,
+    max_len: int = 2048,
+) -> Optional[dict]:
+    """Tokenize one DPO record into chosen/rejected + masks.
+
+    Applies the tokenizer's chat template so token sequences match the
+    SFT checkpoint's format.  Prompt is rendered with
+    ``add_generation_prompt=True``; chosen/rejected are appended as a
+    single assistant turn.
+
+    Accepts:
+
+    - Flat:   ``{"prompt": str, "chosen": str, "rejected": str}``
+    - Conv:   ``{"prompt": [{role, content}, ...], "chosen": [...], ...}``
+    - Legacy: ``{"input": str, "chosen": str, "rejected": str}``
+
+    No packing, no ``position_ids`` — DPO sequences are independent.
+    """
+    prompt = record.get("prompt") or record.get("input")
+    chosen = record.get("chosen")
+    rejected = record.get("rejected")
+    if prompt is None or chosen is None or rejected is None:
+        return None
+
+    prompt_messages = _to_messages(prompt)
+    chosen_text = _extract_text(chosen)
+    rejected_text = _extract_text(rejected)
+    if chosen_text is None or rejected_text is None:
+        return None
+    chosen_messages = prompt_messages + [{"role": "assistant", "content": chosen_text}]
+    rejected_messages = prompt_messages + [
+        {"role": "assistant", "content": rejected_text}
+    ]
+
+    prompt_ids = tokenizer.apply_chat_template(
+        prompt_messages, tokenize=True, add_generation_prompt=True
+    )
+    ch_ids = tokenizer.apply_chat_template(
+        chosen_messages, tokenize=True, add_generation_prompt=False
+    )
+    re_ids = tokenizer.apply_chat_template(
+        rejected_messages, tokenize=True, add_generation_prompt=False
+    )
+
+    full_ch = ch_ids[:max_len]
+    full_re = re_ids[:max_len]
+
+    prompt_len = min(len(prompt_ids), max_len)
+    ch_mask = [0] * prompt_len + [1] * max(0, len(full_ch) - prompt_len)
+    ch_mask = ch_mask[:max_len]
+    re_mask = [0] * prompt_len + [1] * max(0, len(full_re) - prompt_len)
+    re_mask = re_mask[:max_len]
+
+    return {
+        "chosen": full_ch,
+        "rejected": full_re,
+        "chosen_mask": ch_mask,
+        "rejected_mask": re_mask,
+    }
+
+
+def _to_messages(value) -> list:
+    """Accept str or conversation list; return message list."""
+    if isinstance(value, str):
+        return [{"role": "user", "content": value}]
+    if isinstance(value, list):
+        return value
+    return [{"role": "user", "content": str(value)}]
+
+
+def _extract_text(value) -> Optional[str]:
+    """Accept str or conversation list; return plain text."""
+    if value is None:
+        return None
+    if isinstance(value, str):
+        return value
+    if isinstance(value, list):
+        return "".join(m.get("content", "") for m in value if isinstance(m, dict))
+    return None
+
+
+def dpo_processor(
+    record: dict,
+    tokenizer,
+    max_len: int = 2048,
+) -> Dict[str, Tensor]:
+    """DPO processor: wraps :func:`dpo_tokenize` and returns tensors."""
+    result = dpo_tokenize(record, tokenizer, max_len=max_len)
+    if result is None:
+        raise ValueError(f"Malformed DPO record: {list(record.keys())}")
+    return {
+        "chosen": torch.tensor(result["chosen"], dtype=torch.int32),
+        "rejected": torch.tensor(result["rejected"], dtype=torch.int32),
+        "chosen_mask": torch.tensor(result["chosen_mask"], dtype=torch.bool),
+        "rejected_mask": torch.tensor(result["rejected_mask"], dtype=torch.bool),
+    }
+
+
+def dpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
+    """Collate variable-length DPO samples into padded 2-D tensors.
+
+    Input: list of dicts, each with:
+      - chosen:        [C_i]
+      - rejected:      [R_i]
+      - chosen_mask:   [C_i]
+      - rejected_mask: [R_i]
+
+    Output (padded to the max length across chosen/rejected within the batch):
+      - chosen:        [B, S_max]
+      - rejected:      [B, S_max]
+      - chosen_mask:   [B, S_max]
+      - rejected_mask: [B, S_max]
+    """
+    B = len(batch)
+    S_max = max(b["chosen"].size(0) for b in batch)
+    S_max = max(S_max, max(b["rejected"].size(0) for b in batch))
+
+    chosen = torch.zeros(B, S_max, dtype=torch.long)
+    rejected = torch.zeros(B, S_max, dtype=torch.long)
+    chosen_mask = torch.zeros(B, S_max, dtype=torch.bool)
+    rejected_mask = torch.zeros(B, S_max, dtype=torch.bool)
+
+    for i, b in enumerate(batch):
+        c_len = b["chosen"].size(0)
+        r_len = b["rejected"].size(0)
+        chosen[i, :c_len] = b["chosen"]
+        rejected[i, :r_len] = b["rejected"]
+        chosen_mask[i, :c_len] = b["chosen_mask"]
+        rejected_mask[i, :r_len] = b["rejected_mask"]
+
+    return {
+        "chosen": chosen,
+        "rejected": rejected,
+        "chosen_mask": chosen_mask,
+        "rejected_mask": rejected_mask,
+    }
+
+
+def grpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
+    """Collate variable-length GRPO samples into padded 3-D tensors.
+
+    Input: list of dicts, each with:
+      - prompts:  [P_i]
+      - responses: list of G tensors, each [R_ij]
+      - masks:     list of G tensors, each [R_ij]
+      - rewards:  [G]
+
+    Output:
+      - prompts:   [B, P_max], left-padded
+      - prompt_mask: [B, P_max]
+      - responses: [B, G, R_max]
+      - masks:     [B, G, R_max]
+      - rewards:   [B, G]
+    """
+    B = len(batch)
+    G = len(batch[0]["responses"])
+    P_max = max(b["prompts"].size(0) for b in batch)
+    R_max = max(r.size(0) for b in batch for r in b["responses"])
+
+    prompts = torch.zeros(B, P_max, dtype=torch.long)
+    prompt_mask = torch.zeros(B, P_max, dtype=torch.bool)
+    responses = torch.zeros(B, G, R_max, dtype=torch.long)
+    masks = torch.zeros(B, G, R_max, dtype=torch.bool)
+    rewards = torch.zeros(B, G, dtype=torch.float32)
+
+    for i, b in enumerate(batch):
+        p_len = b["prompts"].size(0)
+        prompts[i, -p_len:] = b["prompts"]
+        prompt_mask[i, -p_len:] = True
+        rewards[i, : b["rewards"].size(0)] = b["rewards"]
+        for g in range(min(G, len(b["responses"]))):
+            r_len = b["responses"][g].size(0)
+            responses[i, g, :r_len] = b["responses"][g]
+            if g < len(b["masks"]):
+                masks[i, g, :r_len] = b["masks"][g]
+
+    return {
+        "prompts": prompts,
+        "prompt_mask": prompt_mask,
+        "responses": responses,
+        "masks": masks,
+        "rewards": rewards,
+    }
+
+
+def validate_keys(store: Store, required: List[str]) -> None:
+    """Raise ``KeyError`` if *store* is missing any *required* key."""
+    if not required:
+        return
+    actual = set(store.keys)
+    missing = [k for k in required if k not in actual]
+    if missing:
+        raise KeyError(
+            f"Store at {getattr(store, '_load_path', '?')} is missing required "
+            f"keys {missing}; available keys are {sorted(actual)}."
+        )
 
 
 class BaseDataset(Dataset, ABC):
-    """Abstract base class for all dataset types.
+    """Abstract base class for dataset types.
 
-    Implements common functionality for window-based data fetching.
-    Uses a storage abstraction for format-agnostic data loading.
+    Holds a :class:`Store`.  All sample-id indexing is delegated to the
+    store — this class exposes ``__len__`` as ``len(store)`` and the
+    ``keys`` property as ``store.keys``.  Subclasses implement
+    ``__getitem__`` with the train-type-specific key mapping and any
+    training-only index arithmetic (e.g. the next-token ``+1`` shift).
     """
 
-    def __init__(self, window_size: int, stride: int):
+    required_keys: List[str] = []
+
+    def __init__(self, store: Store):
         super().__init__()
-        self.window_size = window_size
-        self.stride = stride
-        self.storage: Optional[Store] = None
+        self.store: Store = store
+        validate_keys(store, self.required_keys)
 
-    @property
-    def required_keys(self) -> List[str]:
-        """Return required storage keys for this dataset type.
-
-        Subclasses should override to specify expected keys.
-        """
-        return []
-
-    def _validate_keys(self):
-        if not self.required_keys:
-            return
-        actual_keys = set(self.storage.keys)
-        missing = [k for k in self.required_keys if k not in actual_keys]
-        if missing:
-            raise KeyError(
-                f"Dataset {type(self).__name__} requires keys {self.required_keys}, "
-                f"but storage at {self._load_path} only has {sorted(actual_keys)}. "
-                f"Missing: {missing}"
-            )
-
-    def load(self, load_path: str, storage_type: Optional[str] = None, **kwargs):
-        """Load dataset from the given path.
-
-        Auto-detects the storage format if not specified.
-
-        Args:
-            load_path: Path to the data directory or file
-            storage_type: Force a specific storage type ("h5", "bin", "jsonl"),
-                          or None for auto-detection
-            **kwargs: Extra arguments forwarded to the store constructor and
-                      to ``store.load()``.
-
-        Raises:
-            KeyError: If the loaded storage is missing required keys.
-        """
-        if storage_type is None:
-            storage_type = detect_format(load_path)
-        self.storage = StoreFactory.create(storage_type, **kwargs)
-        self._load_path = load_path
-        self.storage.load(load_path, **kwargs)
-        self._validate_keys()
-
-    @property
-    def count(self) -> int:
-        """Return the total number of raw elements (tokens) in the dataset."""
-        if self.storage is None:
-            return 0
-        return len(self.storage)
+    def __len__(self) -> int:
+        return len(self.store)
 
     @property
     def keys(self) -> List[str]:
-        """Return the available data keys."""
-        if self.storage is None:
-            return []
-        return self.storage.keys
+        return self.store.keys
 
-    def get_index(self, index: int) -> tuple:
-        """Calculate begin and end indices for a sample.
-
-        Args:
-            index: Sample index
-
-        Returns:
-            Tuple of (begin_idx, end_idx)
-        """
-        if self.storage is None:
-            raise RuntimeError("Dataset not loaded, call load() first")
-        total = len(self.storage)
-        if total <= self.window_size:
-            raise ValueError(
-                f"Data too short: {total} tokens <= window_size {self.window_size}"
-            )
-
-        begin_idx = min(index * self.stride, total - 1 - self.window_size)
-        end_idx = min(begin_idx + self.window_size, total - 1)
-
-        return begin_idx, end_idx
+    @property
+    def token_count(self) -> int:
+        return self.store.token_count
 
     @abstractmethod
     def __getitem__(self, index: int) -> Dict[str, Tensor]:
-        """Get a single sample by index.
-
-        Must be implemented by subclasses.
-        """
         raise NotImplementedError
 
-    def __len__(self) -> int:
-        if self.storage is None:
-            return 0
-        total = len(self.storage)
-        if total <= self.window_size:
-            return 0
-        return (total - 1 - self.window_size) // self.stride + 1
-
 
 class DatasetFactory(BaseFactory["BaseDataset"]):
-    """Factory class for creating dataset instances.
+    """Factory for creating dataset instances by train-type.
 
-    Supports decorator-based registration for extensible dataset types.
-    All default dataset types (seq, sft, dpo, grpo) are registered automatically
-    when their classes are defined with the decorator.
-
-    Example usage:
-        @DatasetFactory.register("custom")
-        class CustomDataset(BaseDataset):
-            ...
-
-        dataset = DatasetFactory.create("custom", window_size, stride)
+    Use :meth:`DatasetFactory.register("custom")` to register new
+    dataset classes; they must inherit from :class:`BaseDataset`.
     """
 
     @classmethod
     def load(
         cls,
         train_type: str,
-        load_path: str,
-        window_size: int,
+        load_path: Optional[str] = None,
+        window_size: int = 0,
         stride: Optional[int] = None,
         storage_type: Optional[str] = None,
+        tokenizer_path: Optional[str] = None,
+        max_len: int = 2048,
+        store: Optional[Store] = None,
         **kwargs,
     ) -> "BaseDataset":
         """Create and load a dataset in one step.
 
+        Two entry points:
+
+        - **store given**: bind it directly — the caller fully controls
+          Store construction and processor setup.  *load_path*,
+          *storage_type*, *tokenizer_path*, *window_size*, *stride* are
+          ignored.
+        - **store is None**: build a Store from *load_path*, auto-detecting
+          format and constructing a processor when *tokenizer_path* is
+          given for a record dataset on JSONL.
+
         Args:
-            train_type: Type of training dataset
-            load_path: Path to the data file
-            window_size: Window size for data sampling
-            stride: Stride between consecutive samples (default: same as window_size)
-            storage_type: Storage type ("h5", "bin", "jsonl") or None for auto-detection
-            **kwargs: Extra arguments forwarded to ``dataset.load()``.
+            train_type: Registered dataset name ("seq", "sft", "dpo",
+                "grpo", …).
+            load_path: Path to the data file or directory (ignored if
+                *store* is given).
+            window_size: Stream window length — only meaningful for
+                stream datasets (SEQ/SFT).  Record datasets ignore it.
+            stride: Stride between consecutive stream samples
+                (default: same as *window_size*).
+            storage_type: Storage backend ("bin", "jsonl") or
+                None for auto-detection.
+            tokenizer_path: Path to tokenizer for lazy JSONL
+                tokenisation (record datasets only).
+            max_len: Max sequence length forwarded to processors.
+            store: Pre-built, already-loaded Store instance.
+            **kwargs: Extra arguments forwarded to ``store.load()``.
 
         Returns:
-            Loaded dataset instance
+            Loaded dataset instance.
         """
+        if store is not None:
+            return cls.create(train_type, store=store)
+
+        if load_path is None:
+            raise ValueError("Either load_path or store must be provided")
+
+        if storage_type is None:
+            storage_type = detect_format(load_path)
+
         if stride is None:
             stride = window_size
 
-        dataset = cls.create(train_type, window_size, stride)
-        dataset.load(load_path, storage_type=storage_type, **kwargs)
+        processor = cls._maybe_build_processor(
+            train_type, storage_type, tokenizer_path, max_len
+        )
 
-        return dataset
+        store_window = cls._store_window_for(train_type, window_size)
+        store = StoreFactory.create(
+            storage_type,
+            window_size=store_window,
+            stride=stride if stride else store_window,
+        )
+        if processor is not None:
+            store.load(load_path, processor=processor, **kwargs)
+        else:
+            load_kwargs = dict(kwargs)
+            if (
+                tokenizer_path is not None
+                and storage_type == "jsonl"
+                and train_type in ("seq", "sft")
+                and "tokenizer_path" not in load_kwargs
+            ):
+                load_kwargs["tokenizer_path"] = tokenizer_path
+            store.load(load_path, **load_kwargs)
+
+        return cls.create(train_type, store=store)
+
+    @staticmethod
+    def _store_window_for(train_type: str, window_size: int) -> int:
+        """Stream datasets consume ``window_size``; record datasets ignore it.
+
+        Record datasets (dpo/grpo) treat each record as an independent
+        training unit and never window, so the store is built with
+        ``window_size=0`` and ``len(store)`` returns the record count.
+        """
+        if train_type in ("seq", "sft"):
+            return window_size
+        return 0
+
+    @staticmethod
+    def _maybe_build_processor(
+        train_type: str,
+        storage_type: str,
+        tokenizer_path: Optional[str],
+        max_len: int,
+    ) -> Optional[Callable[[dict], Dict[str, Tensor]]]:
+        """Build an on-the-fly tokenisation processor if applicable.
+
+        Only raw JSONL + record datasets (DPO/GRPO) need a processor;
+        pre-tokenised backends (bin) and stream datasets (SEQ/SFT)
+        return ``None`` so no tokenizer is loaded.
+        """
+        if tokenizer_path is None or storage_type != "jsonl":
+            return None
+        if train_type == "dpo":
+            tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+            return partial(dpo_processor, tokenizer=tokenizer, max_len=max_len)
+        return None
 
 
 @DatasetFactory.register("seq")
 class SEQDataset(BaseDataset):
-    """Dataset for sequential next-token prediction training."""
+    """Dataset for sequential next-token prediction training.
 
-    @property
-    def required_keys(self) -> List[str]:
-        return ["sequence"]
+    Stream mode: ``store.fetch(begin, end, "sequence")`` returns the
+    input window; the +1 shifted call returns the next-token target.
+    """
 
-    def _fetch_data(self, begin_idx: int, end_idx: int) -> Tensor:
-        return self.storage.fetch(begin_idx, end_idx, "sequence")
+    required_keys = ["sequence"]
 
-    def __getitem__(self, index):
-        begin_idx, end_idx = self.get_index(index)
-
-        x = self._fetch_data(begin_idx, end_idx).to(dtype=torch.long)
-        y = self._fetch_data(begin_idx + 1, end_idx + 1).to(dtype=torch.long)
-
-        return {"input_ids": x, "target_ids": y}
+    def __getitem__(self, index: int):
+        begin, end = self.store.sample_window(index)
+        x = self.store.fetch(begin, end, "sequence")
+        y = self.store.fetch(begin + 1, end + 1, "sequence")
+        return {
+            "input_ids": x.to(dtype=torch.long),
+            "target_ids": y.to(dtype=torch.long),
+        }
 
 
 @DatasetFactory.register("sft")
 class SFTDataset(BaseDataset):
-    """Dataset for supervised fine-tuning with loss masking."""
+    """Dataset for supervised fine-tuning with loss masking.
 
-    @property
-    def required_keys(self) -> List[str]:
-        return ["sequence", "loss_mask", "position_ids"]
+    Stream mode: ``sequence``/``loss_mask``/``position_ids`` are sliced
+    to the window.  ``loss_mask`` and ``target_ids`` use the +1 shifted
+    slice so they align with the predicted positions.
+    """
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
-        return self.storage.fetch(begin_idx, end_idx, key)
-
-    def __getitem__(self, index):
-        begin_idx, end_idx = self.get_index(index)
-
-        x = self._fetch_data(begin_idx, end_idx, "sequence")
-        y = self._fetch_data(begin_idx + 1, end_idx + 1, "sequence")
-        position_ids = self._fetch_data(begin_idx, end_idx, "position_ids")
-        loss_mask = self._fetch_data(begin_idx + 1, end_idx + 1, "loss_mask")
+    required_keys = ["sequence", "loss_mask", "position_ids"]
 
+    def __getitem__(self, index: int):
+        begin, end = self.store.sample_window(index)
+        x = self.store.fetch(begin, end, "sequence")
+        y = self.store.fetch(begin + 1, end + 1, "sequence")
+        position_ids = self.store.fetch(begin, end, "position_ids")
+        loss_mask = self.store.fetch(begin + 1, end + 1, "loss_mask")
         return {
             "input_ids": x.to(dtype=torch.long),
             "target_ids": y.to(dtype=torch.long),
@@ -219,59 +442,65 @@ class SFTDataset(BaseDataset):
 
 @DatasetFactory.register("dpo")
 class DPODataset(BaseDataset):
-    """Dataset for Direct Preference Optimization training."""
+    """Record-structured dataset for Direct Preference Optimization.
 
-    @property
-    def required_keys(self) -> List[str]:
-        return ["chosen", "rejected", "chosen_mask", "rejected_mask"]
+    Each sample is one preference pair (chosen + rejected) and is an
+    independent training unit — no windowing, stride, or cross-record
+    concatenation.  This keeps each sequence self-contained so attention
+    never leaks across preference pairs.
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
-        return self.storage.fetch(begin_idx, end_idx, key)
+    Two loading paths (handled by :class:`DatasetFactory`):
 
-    def __getitem__(self, index: int):
-        begin_idx, end_idx = self.get_index(index)
+    - **Pre-tokenized** (bin): ``store.load(path)`` reads per-record
+      tensors; ``__getitem__`` returns them directly.
+    - **Raw JSONL** (``tokenizer_path=...``): builds a lazy processor
+      via :func:`dpo_processor` that tokenises on the fly — no packing,
+      no ``position_ids``.
+    """
 
-        chosen = self._fetch_data(begin_idx, end_idx, "chosen").to(dtype=torch.long)
-        rejected = self._fetch_data(begin_idx, end_idx, "rejected").to(dtype=torch.long)
-        chosen_mask = self._fetch_data(begin_idx, end_idx, "chosen_mask").to(
-            dtype=torch.bool
-        )
-        rejected_mask = self._fetch_data(begin_idx, end_idx, "rejected_mask").to(
-            dtype=torch.bool
-        )
+    required_keys = ["chosen", "rejected", "chosen_mask", "rejected_mask"]
 
+    def make_processor(self, tokenizer, max_len: int):
+        return partial(dpo_processor, tokenizer=tokenizer, max_len=max_len)
+
+    def __getitem__(self, index: int) -> Dict[str, Tensor]:
         return {
-            "chosen": chosen,
-            "rejected": rejected,
-            "chosen_mask": chosen_mask,
-            "rejected_mask": rejected_mask,
+            "chosen": self.store.fetch_record(index, "chosen").to(dtype=torch.long),
+            "rejected": self.store.fetch_record(index, "rejected").to(dtype=torch.long),
+            "chosen_mask": self.store.fetch_record(index, "chosen_mask").to(
+                dtype=torch.bool
+            ),
+            "rejected_mask": self.store.fetch_record(index, "rejected_mask").to(
+                dtype=torch.bool
+            ),
         }
 
 
 @DatasetFactory.register("grpo")
 class GRPODataset(BaseDataset):
-    """Dataset for Group Relative Policy Optimization training."""
+    """Dataset for offline Group Relative Policy Optimization.
 
-    @property
-    def required_keys(self) -> List[str]:
-        return ["prompts", "responses", "masks", "rewards"]
+    Each sample is one prompt with its group of responses and scalar
+    rewards — an independent training unit with no windowing or stride.
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
-        return self.storage.fetch(begin_idx, end_idx, key)
+    Expected storage layout (produced by JsonlStore or pre-tokenized):
+
+    - ``prompts``:   List[Tensor]  — one 1-D token tensor per record
+    - ``responses``: List[List[Tensor]] — G response tensors per record
+    - ``masks``:     List[List[Tensor]] — G mask tensors per record
+    - ``rewards``:   List[Tensor]  — one 1-D float tensor (len G) per record
+    """
+
+    required_keys = ["prompts", "responses", "masks", "rewards"]
 
     def __getitem__(self, index: int) -> Dict[str, Tensor]:
-        begin_idx, end_idx = self.get_index(index)
-
-        prompts = self._fetch_data(begin_idx, end_idx, "prompts").to(dtype=torch.long)
-        responses = self._fetch_data(begin_idx, end_idx, "responses").to(
-            dtype=torch.long
-        )
-        masks = self._fetch_data(begin_idx, end_idx, "masks").to(dtype=torch.bool)
-        rewards = self._fetch_data(begin_idx, end_idx, "rewards")
-
+        prompts = self.store.fetch_record(index, "prompts")
+        responses = self.store.fetch_record(index, "responses")
+        masks = self.store.fetch_record(index, "masks")
+        rewards = self.store.fetch_record(index, "rewards")
         return {
-            "prompts": prompts,
-            "responses": responses,
-            "masks": masks,
-            "rewards": rewards,
+            "prompts": prompts.to(dtype=torch.long),
+            "responses": [r.to(dtype=torch.long) for r in responses],
+            "masks": [m.to(dtype=torch.bool) for m in masks],
+            "rewards": rewards.to(dtype=torch.float32),
         }

astrai/dataset/sampler.py

@@ -5,7 +5,15 @@ import torch.distributed as dist
 from torch.utils.data import Dataset, Sampler
 
 
-class ResumableDistributedSampler(Sampler[int]):
+class RDSampler(Sampler[int]):
+    """Resumable Distributed Sampler.
+
+    A distributed sampler that supports checkpoint-based resume: iteration
+    state (epoch, position) is tracked so training can continue from the
+    exact sample after a restart.  Shards the dataset across
+    ``dist.world_size`` replicas with optional shuffling.
+    """
+
     def __init__(
         self,
         data_source: Dataset,

astrai/dataset/storage.py

@@ -1,20 +1,47 @@
 """Storage backends for different data formats.
 
-Layers:
-  - I/O layer:       save_* / load_* functions, read/write raw files (HDF5/bin)
-                      return Dict[str, List[Tensor]] — format-specific, no state
-  - Store (ABC):     central abstraction, normalizes multi-segment into
-                      Dict[str, List[Tensor]] per key via _normalize(),
-                      fetch() uses bisect across segments — no forced concat
-  - Dataset layer:   BaseDataset owns a Store, only calls store.fetch(begin, end, key)
+Architecture (composition over inheritance):
 
-Key properties:
-  - Multi-segment:   segments kept as-is, no forced concatenation — safe for
-                      datasets larger than RAM
-  - Explicit length: _length = min(total elements across keys), set at load,
-                      __len__ returns O(1)
-  - Zero-copy mmap:  MmapStore wraps np.memmap(mode="r"), all DataLoader
-                       workers share OS page-cache pages
+    Store (ABC)               — owns _data/_cum/_offsets bookkeeping
+                                  + window_size/stride for sample-id
+                                  indexing.  __getitem__/__len__ produce
+                                  the smallest iterable unit so Dataset
+                                  classes are pure delegators.
+    Streamable (mixin)        — raw token slice  fetch(begin, end, keys)
+    Recordable (mixin)        — raw record slice fetch_record(idx, keys)
+
+    MmapStore(Store, Streamable, Recordable)
+    JsonlStore(Store, Streamable, Recordable)
+
+Each mixin is a stateless trait that relies on ``self._data`` etc.
+provided by :class:`Store`.  Concrete stores mix in whichever access
+primitives they support — ``Store`` is the sole base class, so there is
+no diamond inheritance or MRO ambiguity.
+
+Sample-id indexing lives on :class:`Store`, not on the dataset:
+
+- **Stream mode** (``window_size > 0``):  ``len(store)`` returns the number
+  of ``(window_size, stride)`` windows that fit in the token river;
+  ``store[i]`` returns the *i*-th window as a dict of per-key tensors;
+  ``store.sample_window(i)`` exposes the underlying ``(begin, end)``
+  token slice for callers (e.g. next-token trainers) that need a +1
+  shifted companion window.
+- **Record mode** (``num_records > 0``):  ``len(store)`` returns the
+  record count; ``store[i]`` returns the *i*-th record dict.
+
+Raw token/record access via :meth:`fetch` / :meth:`fetch_record`
+remains available for low-level callers that want explicit index
+control.  ``store.token_count`` is the total stream token count (what
+``len(store)`` used to mean in the legacy stream-only API).
+
+``segments_are_records`` (class attribute on each Store subclass)
+tells ``_normalize`` whether segments are inherently per-record (JSONL)
+or opaque shards (bin).  Record access for bin relies on ``_offsets``
+instead.
+
+:class:`JsonlStore` supports a lazy mode (``processor=fn``) that keeps
+raw records and defers tokenisation to ``fetch_record`` — used by DPO
+to train directly from a ``.jsonl`` file without a pre-tokenised copy.
 """
 
 import bisect
@@ -23,20 +50,18 @@ import json
 import logging
 from abc import ABC, abstractmethod
 from pathlib import Path
-from typing import Dict, List, Union
+from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
 from torch import Tensor
 
 from astrai.config.preprocess_config import PipelineConfig
 from astrai.factory import BaseFactory
-from astrai.preprocessing.builder import MaskBuilderFactory
-from astrai.preprocessing.position_id import PositionIdStrategyFactory
+from astrai.preprocessing.transform import TokenizeTransform
 from astrai.serialization import (
     load_bin,
-    load_h5,
+    load_bin_offsets,
 )
-from astrai.tokenize import AutoTokenizer
 
 logger = logging.getLogger(__name__)
 
@@ -48,7 +73,7 @@ def detect_format(load_path: str) -> str:
         load_path: Directory or file path
 
     Returns:
-        Format string ("h5", "bin", or "jsonl")
+        Format string ("h5", "bin", "jsonl", or "processed")
 
     Raises:
         FileNotFoundError: If no supported data files are found
@@ -56,19 +81,10 @@ def detect_format(load_path: str) -> str:
     root = Path(load_path)
     if root.is_file():
         suffix = root.suffix.lower()
-        if suffix in (".h5", ".hdf5"):
-            return "h5"
         if suffix == ".jsonl":
             return "jsonl"
         raise ValueError(f"Unsupported file format: {suffix}")
 
-    h5_files = [
-        Path(p)
-        for pattern in ("*.h5", "*.hdf5")
-        for p in glob.glob(str(root / "**" / pattern), recursive=True)
-    ]
-    if h5_files:
-        return "h5"
     bin_files = [Path(p) for p in glob.glob(str(root / "**" / "*.bin"), recursive=True)]
     if bin_files:
         has_meta = (root / "meta.json").exists() or len(
@@ -85,228 +101,526 @@ def detect_format(load_path: str) -> str:
 
 
 class Store(ABC):
-    """String keys -> segmented tensors with ``fetch(begin, end, keys)``.
+    """Common base for all storage backends.
 
-    Each key maps to one or more tensor segments (no forced concatenation).
-    ``len(store)`` returns ``self._length`` (explicit, O(1)), the minimum
-    total element count across all keys.
+    A Store owns both its data layout AND its sample-id → token/record
+    index translation.  Datasets are thin wrappers that bind a Store
+    to a particular train-type's key mapping; they never know about
+    window/stride math.
 
-    Subclasses fill ``self._data`` and ``self._cum`` during ``load()``
-    via ``_normalize()``.
+    Two iteration modes:
+
+    - **Stream** (``window_size > 0``): data is treated as one long
+      token river.  ``len(store)`` returns the number of windows;
+      ``store[i]`` slices every stream-compatible key to window ``i``;
+      ``store.sample_window(i)`` returns the ``(begin, end)`` token
+      slice for callers needing a +1 shifted companion window.
+    - **Record** (``num_records > 0``): data is per-record.
+      ``len(store)`` returns ``num_records``; ``store[i]`` returns
+      the *i*-th record as a dict.
+
+    Raw token slicing is still available via :meth:`fetch` (mixed in
+    by :class:`Streamable`) when a store has stream support configured.
+    Raw record slicing via :meth:`fetch_record` (mixed in by
+    :class:`Recordable`) when a store has record support.
+
+    ``token_count`` exposes the raw total stream length — this is what
+    ``len(store)`` returned in the legacy stream-only API and what
+    stream-bound ``fetch`` uses for its bounds check.
     """
 
-    def __init__(self):
+    segments_are_records: bool = False
+
+    def __init__(
+        self,
+        window_size: int = 0,
+        stride: Optional[int] = None,
+    ):
         self._data: Dict[str, List[Tensor]] = {}
         self._cum: Dict[str, List[int]] = {}
+        self._offsets: Dict[str, List[int]] = {}
         self._length: int = 0
+        self._num_records: int = 0
+        self._window_size: int = int(window_size)
+        self._stride: int = int(stride) if stride is not None else int(window_size)
 
     @abstractmethod
-    def load(self, path: str) -> None:
+    def load(self, path: str, **kwargs) -> None:
         raise NotImplementedError
 
     @property
     def keys(self) -> List[str]:
         return list(self._data.keys())
 
-    def __len__(self) -> int:
+    @property
+    def window_size(self) -> int:
+        return self._window_size
+
+    @property
+    def stride(self) -> int:
+        return self._stride
+
+    @property
+    def token_count(self) -> int:
+        """Total tokens across all stream segments.
+
+        Useful for the bounds-checked raw :meth:`fetch` and as the
+        legacy ``len(store)`` value.
+        """
         return self._length
 
+    @property
+    def num_records(self) -> int:
+        """Number of records available via :meth:`fetch_record`.
+
+        Non-zero only when the backing layout provides per-record
+        indexing (JSONL segments or bin ``_offsets``).
+        """
+        return self._num_records
+
+    @property
+    def num_samples(self) -> int:
+        """Number of items produced by ``__getitem__``.
+
+        Stream-mode wins when ``window_size > 0`` and there are tokens
+        to slice; otherwise falls back to ``num_records``.
+        """
+        if self._window_size > 0 and self._length > 0:
+            total = self._length
+            w = self._window_size
+            if total <= w:
+                return 0
+            return (total - 1 - w) // self._stride + 1
+        return self._num_records
+
+    def __len__(self) -> int:
+        return self.num_samples
+
+    def __getitem__(self, index: int) -> Dict[str, Tensor]:
+        if index < 0:
+            index += self.num_samples
+        if not 0 <= index < self.num_samples:
+            raise IndexError(
+                f"Store index out of range: {index}, num_samples={self.num_samples}"
+            )
+        if self._window_size > 0 and self._length > 0:
+            begin, end = self.sample_window(index)
+            keys = self._stream_keys()
+            return {k: self.fetch(begin, end, k) for k in keys}
+        return self.fetch_record(index, self._record_keys())
+
+    def sample_window(self, index: int) -> Tuple[int, int]:
+        """Return ``(begin, end)`` token positions for stream sample *index*.
+
+        The clipped tail keeps the last reachable window inside the
+        token river instead of overshooting.  Caller is responsible
+        for staying within :attr:`num_samples`: an out-of-range index
+        raises ``IndexError``.
+        """
+        if self._window_size <= 0:
+            raise RuntimeError("sample_window() requires window_size > 0 (stream mode)")
+        if self._window_size <= 0 or self._length <= self._window_size:
+            raise IndexError(
+                f"Data too short for window: token_count={self._length}, "
+                f"window_size={self._window_size}"
+            )
+        if not 0 <= index < self.num_samples:
+            raise IndexError(
+                f"Sample index out of range: {index}, num_samples={self.num_samples}"
+            )
+        total = self._length
+        begin = min(index * self._stride, total - 1 - self._window_size)
+        end = min(begin + self._window_size, total - 1)
+        return begin, end
+
+    def _stream_keys(self) -> List[str]:
+        out: List[str] = []
+        for k, tensors in self._data.items():
+            if tensors and isinstance(tensors[0], list):
+                continue
+            out.append(k)
+        return out
+
+    def _record_keys(self) -> List[str]:
+        return list(self._data.keys())
+
+    def _normalize(
+        self,
+        raw: Dict[str, list],
+        offsets: Optional[Dict[str, List[int]]] = None,
+    ):
+        """Register segments and pre-compute indices for both access modes.
+
+        Stream mode: ``_cum[key]`` accumulates per-segment lengths so
+        ``Streamable._fetch_stream_key`` can bisect across segments
+        without concatenation.
+
+        Record mode: if *offsets* is provided (bin layout),
+        ``_offsets[key]`` stores cumulative per-record offsets into the
+        single concatenated segment.  Otherwise, when
+        ``segments_are_records`` is True (JSONL), ``_data[key]`` is
+        a per-record list and ``fetch_record`` indexes it directly.
+
+        Nested keys (GRPO ``responses``/``masks`` as
+        ``List[List[Tensor]]``) are stored as-is and excluded from both
+        cumulative bookkeepings — they are only accessed record-by-record.
+        """
+        flat_lengths = []
+        for key, tensors in raw.items():
+            self._data[key] = tensors
+            if not tensors:
+                self._cum[key] = []
+                flat_lengths.append(0)
+                continue
+            if isinstance(tensors[0], list):
+                self._cum[key] = []
+                continue
+            cum = []
+            total = 0
+            for t in tensors:
+                total += t.shape[0]
+                cum.append(total)
+            self._cum[key] = cum
+            flat_lengths.append(cum[-1] if cum else 0)
+        self._length = min(flat_lengths) if flat_lengths else 0
+
+        valid_offsets: Dict[str, List[int]] = {}
+        if offsets:
+            for key, off in offsets.items():
+                segs = self._data.get(key, [])
+                if len(segs) == 1 and len(off) > 1:
+                    valid_offsets[key] = off
+                elif len(segs) > 1:
+                    logger.warning(
+                        "Key '%s' has %d segments with offsets — record mode "
+                        "disabled for this key (multi-shard bin+offsets not "
+                        "supported). Merge shards or use JSONL.",
+                        key,
+                        len(segs),
+                    )
+        self._offsets = valid_offsets
+        if valid_offsets:
+            record_counts = [len(v) - 1 for v in valid_offsets.values()]
+            self._num_records = min(record_counts) if record_counts else 0
+        elif self.segments_are_records:
+            per_record_counts = []
+            for key, tensors in self._data.items():
+                if tensors and isinstance(tensors[0], list):
+                    continue
+                per_record_counts.append(len(tensors))
+            self._num_records = min(per_record_counts) if per_record_counts else 0
+        else:
+            self._num_records = 0
+
+
+class Streamable:
+    """Mixin granting raw token-stream access via :meth:`fetch`.
+
+    Stateless trait relying on ``self._data``, ``self._cum``,
+    ``self._length`` maintained by :class:`Store`.  Stream mode is
+    active when the owning store has ``window_size > 0``; for stores
+    that can also serve record access (JSONL/bin+offsets), the
+    ``fetch_record`` API from :class:`Recordable` is used instead.
+    """
+
     def fetch(
         self,
         begin: int,
         end: int,
         keys: Union[str, List[str]],
     ):
-        if not self._data:
-            raise RuntimeError("Store not loaded")
-        if not (0 <= begin < self._length and 0 <= end <= self._length):
-            raise ValueError(
-                f"Index out of bounds: begin={begin}, end={end}, length={self._length}"
-            )
-        if isinstance(keys, str):
-            return self._fetch_key(keys, begin, end)
-        return {k: self._fetch_key(k, begin, end) for k in keys}
+        return _stream_fetch(self, begin, end, keys)
 
-    def _fetch_key(self, key: str, begin: int, end: int) -> Tensor:
-        """Fetch slice [begin, end) across potentially multiple segments."""
-        segments = self._data[key]
-        cum = self._cum[key]
-        seg_start = bisect.bisect_right(cum, begin)
-        seg_end = bisect.bisect_left(cum, end)
 
-        results = []
-        for i in range(seg_start, seg_end + 1):
-            prev = cum[i - 1] if i > 0 else 0
-            s = max(begin - prev, 0)
-            e = min(end - prev, segments[i].shape[0])
-            results.append(segments[i][s:e])
-
-        return results[0] if len(results) == 1 else torch.cat(results, dim=0)
-
-    def _normalize(self, raw: Dict[str, List[Tensor]]):
-        """Register segments and pre-compute cumulative lengths.
-
-        Does NOT concatenate — segments are kept as-is to avoid OOM on
-        large datasets.  Sets ``self._length`` to the minimum total
-        element count across all keys.
-        """
-        for key, tensors in raw.items():
-            self._data[key] = tensors
-            cum = []
-            total = 0
-            for t in tensors:
-                total += t.shape[0]
-                cum.append(total)
-            self._cum[key] = cum
-        self._length = (
-            min((cum[-1] if cum else 0) for cum in self._cum.values())
-            if self._cum
-            else 0
+def _stream_fetch(self, begin: int, end: int, keys: Union[str, List[str]]):
+    if not getattr(self, "_data", None):
+        raise RuntimeError("Store not loaded")
+    if not (0 <= begin < self._length and 0 <= end <= self._length):
+        raise ValueError(
+            f"Index out of bounds: begin={begin}, end={end}, length={self._length}"
         )
+    if isinstance(keys, str):
+        return _fetch_stream_key(self, keys, begin, end)
+    return {k: _fetch_stream_key(self, k, begin, end) for k in keys}
+
+
+def _fetch_stream_key(self, key: str, begin: int, end: int) -> Tensor:
+    segments = self._data[key]
+    cum = self._cum[key]
+    seg_start = bisect.bisect_right(cum, begin)
+    seg_end = bisect.bisect_left(cum, end)
+
+    results = []
+    for i in range(seg_start, seg_end + 1):
+        prev = cum[i - 1] if i > 0 else 0
+        s = max(begin - prev, 0)
+        e = min(end - prev, segments[i].shape[0])
+        results.append(segments[i][s:e])
+
+    return results[0] if len(results) == 1 else torch.cat(results, dim=0)
+
+
+class Recordable:
+    """Mixin granting raw record access via :meth:`fetch_record`.
+
+    Stateless trait relying on ``self._data``, ``self._offsets``,
+    ``self._num_records`` maintained by :class:`Store`.
+    """
+
+    def fetch_record(
+        self,
+        index: int,
+        keys: Union[str, List[str]],
+    ):
+        return _record_fetch(self, index, keys)
+
+
+def _record_fetch(self, index: int, keys: Union[str, List[str]]):
+    if not getattr(self, "_data", None) and self._num_records == 0:
+        raise RuntimeError("Store not loaded")
+    if not 0 <= index < self._num_records:
+        raise ValueError(
+            f"Record index out of bounds: {index}, num_records={self._num_records}"
+        )
+    if isinstance(keys, str):
+        return _fetch_record_key(self, keys, index)
+    return {k: _fetch_record_key(self, k, index) for k in keys}
+
+
+def _fetch_record_key(self, key: str, index: int):
+    offsets = self._offsets.get(key)
+    if offsets:
+        start = offsets[index]
+        end = (
+            offsets[index + 1]
+            if index + 1 < len(offsets)
+            else self._data[key][0].shape[0]
+        )
+        return self._data[key][0][start:end]
+    return self._data[key][index]
 
 
 class StoreFactory(BaseFactory["Store"]):
-    """Factory for creating Store instances by type name.
-
-    Example::
-
-        @StoreFactory.register("custom")
-        class CustomStore(Store):
-            ...
-    """
-
-
-@StoreFactory.register("h5")
-class H5Store(Store):
-    """HDF5-based storage backend (pre-tokenized data)."""
-
-    def load(self, path: str):
-        self._normalize(load_h5(path))
+    """Factory for creating Store instances by type name."""
 
 
 @StoreFactory.register("bin")
-class MmapStore(Store):
+class MmapStore(Store, Streamable, Recordable):
     """Memory-mapped binary storage backend.
 
     Each key is a single .bin file backed by ``np.memmap(mode="r")``.
     No per-process memory duplication — all DataLoader workers share the
     same OS page-cache pages.
 
-    Format on disk::
+    Supports both access modes:
 
-        data_root/
-          meta.json          # {key: {shape, dtype}, ...}
-          <key>.bin          # raw numpy array, one per key
+    - **Stream**: always available via :meth:`fetch`.
+    - **Record** (``fetch_record(i, key)``): only when ``meta.json``
+      contains per-record ``offsets`` (written via
+      ``save_bin(..., record_keys=...)``).  Legacy bin files without
+      offsets have ``num_records == 0`` and ``len(store)`` reflects the
+      windowed sample count when ``window_size > 0``.
+
+    ``segments_are_records`` is ``False`` here (bin segments are
+    contiguous streams, not per-record) — record access is driven
+    purely by ``_offsets``.
     """
 
-    def load(self, path: str):
+    segments_are_records = False
+
+    def __init__(
+        self,
+        window_size: int = 0,
+        stride: Optional[int] = None,
+    ):
+        super().__init__(window_size=window_size, stride=stride)
+        self._mmap_refs: List[Tensor] = []
+
+    def load(self, path: str, **kwargs):
         self._mmap_refs = []
         root = Path(path)
         all_raw: Dict[str, List[Tensor]] = {}
+        all_offsets: Dict[str, List[int]] = {}
         meta_paths = [
             Path(p) for p in glob.glob(str(root / "**" / "meta.json"), recursive=True)
         ]
         for meta_path in meta_paths:
             raw = load_bin(str(meta_path.parent))
+            off = load_bin_offsets(str(meta_path.parent))
             for key, tensors in raw.items():
                 if key not in all_raw:
                     all_raw[key] = []
                 all_raw[key].extend(tensors)
+            for key, o in off.items():
+                if key not in all_offsets:
+                    all_offsets[key] = []
+                all_offsets[key].extend(o)
         if not meta_paths:
             raise FileNotFoundError(f"No meta.json found under {path}")
-        self._normalize(all_raw)
+        self._normalize(all_raw, offsets=all_offsets or None)
         for tensors in self._data.values():
             self._mmap_refs.extend(tensors)
 
 
-@StoreFactory.register("jsonl")
-class JsonlStore(Store):
-    """On-the-fly tokenization store for raw JSONL files.
+class JsonlSource:
+    """Read raw JSON records from a ``.jsonl`` file or directory.
 
-    A JSONL dataset directory contains ``*.jsonl`` files plus a
-    ``dataset_config.json`` file that follows the same schema as
-    :class:`PipelineConfig` with an additional ``tokenizer_path`` field.
-    Records are tokenized when the store is loaded and concatenated into
-    segmented tensors matching the key layout expected by the dataset
-    classes (``sequence``, ``loss_mask``, ``position_ids``, ...).
+    A thin reader used by :class:`JsonlStore` in processor mode — holds
+    no tokenizer, performs no tokenisation, just yields dicts.
+    """
+
+    def __init__(self, path: str):
+        self.path = Path(path)
+        self._records: Optional[List[dict]] = None
+
+    def load(self) -> List[dict]:
+        if self._records is None:
+            self._records = self._read(self.path)
+        return self._records
+
+    @staticmethod
+    def _read(root: Path) -> List[dict]:
+        if root.is_file():
+            return JsonlSource._read_file(root)
+        return JsonlSource._read_dir(root)
+
+    @staticmethod
+    def _read_file(path: Path) -> List[dict]:
+        records: List[dict] = []
+        with open(path, "r", encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    records.append(json.loads(line))
+                except json.JSONDecodeError:
+                    logger.warning("Failed to parse JSON line in %s, skipping", path)
+        return records
+
+    @staticmethod
+    def _read_dir(root: Path) -> List[dict]:
+        records: List[dict] = []
+        for jsonl_path in sorted(root.glob("*.jsonl")):
+            records.extend(JsonlSource._read_file(jsonl_path))
+        return records
+
+
+@StoreFactory.register("jsonl")
+class JsonlStore(Store, Streamable, Recordable):
+    """JSONL reader with eager/lazy tokenisation modes.
+
+    A JSONL dataset is a ``.jsonl`` file or a directory of ``*.jsonl``
+    files plus (optionally) a ``dataset_config.json`` describing the
+    tokenization pipeline.
+
+    Three ways to supply an eager transform (first match wins):
+
+    - **Explicit** (``transform=``): caller-built
+      :class:`TokenizeTransform` applied eagerly.
+    - **Config file**: ``dataset_config.json`` alongside the ``*.jsonl``
+      files — loaded via :meth:`TokenizeTransform.from_config_file`.
+    - **Default messages** (``tokenizer_path=`` given, no config file):
+      a built-in chatml config that tokenises the ``messages`` field,
+      masking every role except ``assistant`` (loss on assistant only).
+      Lets SFT/SEQ train straight from a chat-style JSONL directory
+      without a hand-written config.
+
+    Two tokenisation modes, selected at :meth:`load` time:
+
+    - **Eager** (default): applies the transform to every record at load
+      time and registers per-key tensors via ``_normalize``.  Both
+      ``fetch`` (stream) and ``fetch_record`` (record) work.
+    - **Lazy** (``processor=fn`` passed): keeps raw records and defers
+      tokenisation to ``fetch_record``.  Only record access works —
+      ``len(store)`` returns ``num_records``; stream primitives raise.
     """
 
     CONFIG_NAME = "dataset_config.json"
+    segments_are_records = True
 
-    def load(self, path: str):
-        root = Path(path)
-        config_path = root / self.CONFIG_NAME
-        if not config_path.exists():
-            raise FileNotFoundError(
-                f"JSONL dataset config not found: {config_path}. "
-                f"Expected {self.CONFIG_NAME} alongside *.jsonl files."
+    _DEFAULT_MESSAGES_CONFIG = {
+        "version": 1,
+        "input": {
+            "sections": [{"field": "messages", "action": "$role", "template": True}]
+        },
+        "mask": {"system": "mask", "user": "mask", "assistant": "train"},
+        "mask_default": "mask",
+        "output": {"position_ids_mode": "doc_reset"},
+    }
+
+    def __init__(
+        self,
+        window_size: int = 0,
+        stride: Optional[int] = None,
+    ):
+        super().__init__(window_size=window_size, stride=stride)
+        self._source: Optional[JsonlSource] = None
+        self._processor: Optional[Callable[[dict], Dict[str, Tensor]]] = None
+        self._keys_cache: Optional[List[str]] = None
+
+    def load(self, path: str, transform=None, processor=None, **kwargs):
+        self._source = JsonlSource(path)
+        records = self._source.load()
+
+        if processor is not None:
+            self._processor = processor
+            self._num_records = len(records)
+            return
+
+        if transform is None:
+            root = Path(path)
+            config_path = root / self.CONFIG_NAME if root.is_dir() else None
+            if config_path is not None and config_path.exists():
+                transform = TokenizeTransform.from_config_file(str(config_path))
+            else:
+                tokenizer_path = kwargs.get("tokenizer_path")
+                if not tokenizer_path:
+                    raise FileNotFoundError(
+                        f"JSONL dataset config not found. Expected "
+                        f"{self.CONFIG_NAME} alongside *.jsonl files, pass an "
+                        f"explicit transform, pass processor= for lazy "
+                        f"on-the-fly tokenisation, or pass tokenizer_path= to "
+                        f"use the built-in messages config."
+                    )
+                config = PipelineConfig.from_dict(self._DEFAULT_MESSAGES_CONFIG)
+                transform = TokenizeTransform(config, tokenizer_path)
+
+        transformed = transform.apply(records)
+        self._normalize(transformed)
+
+    @property
+    def keys(self) -> List[str]:
+        if self._processor is not None:
+            if self._keys_cache is None and self._num_records > 0:
+                sample = self._processor(self._source.load()[0])
+                self._keys_cache = list(sample.keys())
+            return self._keys_cache or []
+        return list(self._data.keys())
+
+    def fetch_record(self, index: int, keys: Union[str, List[str]]):
+        if self._processor is not None:
+            if not 0 <= index < self._num_records:
+                raise ValueError(
+                    f"Record index out of bounds: {index}, "
+                    f"num_records={self._num_records}"
+                )
+            record = self._source.load()[index]
+            data = self._processor(record)
+            if isinstance(keys, str):
+                return data[keys]
+            return {k: data[k] for k in keys}
+        return _record_fetch(self, index, keys)
+
+    def fetch(self, begin: int, end: int, keys: Union[str, List[str]]):
+        if self._processor is not None:
+            raise RuntimeError(
+                "JsonlStore in lazy (processor) mode does not support "
+                "stream fetch(); use fetch_record() instead."
             )
+        return _stream_fetch(self, begin, end, keys)
 
-        with open(config_path, "r", encoding="utf-8") as f:
-            raw_config = json.load(f)
-
-        tokenizer_path = raw_config.pop("tokenizer_path", None)
-        if tokenizer_path is None:
-            raise ValueError(
-                f"JSONL dataset config must specify 'tokenizer_path': {config_path}"
-            )
-
-        self.config = PipelineConfig.from_dict(raw_config)
-        tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
-        mask_builder = MaskBuilderFactory.create("sectioned")
-        position_strategy = PositionIdStrategyFactory.create(
-            self.config.output.position_ids_mode
-        )
-
-        raw: Dict[str, List[Tensor]] = {}
-        doc_sequences: List[List[int]] = []
-
-        for jsonl_path in sorted(root.glob("*.jsonl")):
-            with open(jsonl_path, "r", encoding="utf-8") as f:
-                for line in f:
-                    line = line.strip()
-                    if not line:
-                        continue
-                    try:
-                        item = json.loads(line)
-                    except json.JSONDecodeError:
-                        logger.warning(
-                            "Failed to parse JSON line in %s, skipping", jsonl_path
-                        )
-                        continue
-
-                    result = mask_builder.build(item, self.config, tokenizer)
-                    if result is None:
-                        continue
-
-                    result.pop("domain", None)
-                    primary_ids = self._primary_ids(result)
-                    if not primary_ids:
-                        continue
-
-                    doc_sequences.append(primary_ids)
-                    for key, ids in result.items():
-                        if key not in raw:
-                            raw[key] = []
-                        raw[key].append(torch.tensor(ids, dtype=self._infer_dtype(ids)))
-
-        pos_ids = position_strategy.generate(doc_sequences)
-        if pos_ids:
-            raw["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
-
-        self._normalize(raw)
-
-    @staticmethod
-    def _primary_ids(result: dict) -> List[int]:
-        """Return the first integer list in *result* as the primary id sequence."""
-        for val in result.values():
-            if isinstance(val, list) and val and isinstance(val[0], int):
-                return val
-        return []
-
-    @staticmethod
-    def _infer_dtype(ids: List) -> torch.dtype:
-        """Infer tensor dtype from the first element of a token/value list."""
-        if ids and isinstance(ids[0], float):
-            return torch.float32
-        return torch.int32
+    def __getitem__(self, index: int) -> Dict[str, Tensor]:
+        if self._processor is not None:
+            return self.fetch_record(index, self._record_keys())
+        return super().__getitem__(index)

astrai/extension/__init__.py

@@ -0,0 +1,55 @@
+"""CUDA attention kernel wrappers with torch fallback.
+
+Public API:
+    - ``attn_decode`` — single-query decode attention
+    - ``attn_prefill`` — multi-query prefill attention
+    - ``attn_paged_decode`` — paged decode attention (direct page-table access)
+    - ``AttentionBackend`` — ABC for attention computation strategies
+    - ``TorchNativeBackend`` — default SDPA backend with KV cache I/O
+    - ``CudaBackend`` — CUDA kernel backend with paged decode + prefill
+
+Layout convention: all q/k/v are ``[batch, seq_len, n_heads, head_dim]``
+(blhd). Scale is always ``1/sqrt(head_dim)``.
+
+Each wrapper calls its compiled CUDA kernel directly. Fallback to torch
+SDPA is handled by the attention backend, not the wrapper functions.
+"""
+
+from astrai.extension.attention_backend import (
+    ATTN_BACKEND,
+    AttentionBackend,
+    AttentionBackendFactory,
+    CudaBackend,
+    FlashAttnBackend,
+    TorchNativeBackend,
+    attention,
+    attn_backend,
+    get_backend,
+)
+from astrai.extension.attention_ops import (
+    TensorLayout,
+    attn_decode,
+    attn_paged_decode,
+    attn_prefill,
+)
+from astrai.extension.loader import KERNEL_NAMES, is_available
+from astrai.extension.rotary_backend import apply_rotary_emb
+
+__all__ = [
+    "ATTN_BACKEND",
+    "AttentionBackend",
+    "AttentionBackendFactory",
+    "CudaBackend",
+    "TorchNativeBackend",
+    "FlashAttnBackend",
+    "TensorLayout",
+    "attention",
+    "attn_backend",
+    "get_backend",
+    "attn_decode",
+    "attn_paged_decode",
+    "attn_prefill",
+    "is_available",
+    "KERNEL_NAMES",
+    "apply_rotary_emb",
+]

astrai/extension/attention_backend.py

@@ -0,0 +1,671 @@
+"""Attention backend abstraction with context-manager switching.
+
+The backend encapsulates KV cache I/O and attention computation. The
+attention module (GQA/MLA) keeps projections, rotary, QK-norm, gating,
+and output projection; the backend handles everything from "write K/V
+to cache" through "SDPA output".
+
+Usage — mirroring ``torch.nn.attention.sdpa_kernel``:
+
+    from astrai.extension import attn_backend, ATTN_BACKEND
+
+    with attn_backend(ATTN_BACKEND.TORCH_NATIVE):
+        engine.generate("hello")
+
+    # or with an instance:
+    with attn_backend(TorchNativeBackend()):
+        ...
+
+    # or the shorthand (instance is itself a context manager):
+    with TorchNativeBackend():
+        ...
+
+Thread-safe via ``contextvars`` — each scheduler thread gets its own
+active backend. ``get_backend()`` returns the active one, falling back
+to a process-wide default (cuda > flash > torch, overridable via
+``ASTR_BACKEND``).
+
+Layout convention: all q/k/v are ``[batch, seq_len, n_heads, head_dim]``
+(blhd). The backend returns ``[batch, seq_len, n_heads * head_dim]``.
+"""
+
+import contextvars
+import enum
+import functools
+import importlib
+import os
+from abc import ABC, abstractmethod
+from contextlib import contextmanager
+from typing import TYPE_CHECKING, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor
+
+from astrai.extension.attention_ops import (
+    attn_paged_decode,
+    attn_paged_prefill,
+)
+from astrai.extension.loader import is_available
+from astrai.factory import BaseFactory
+
+if TYPE_CHECKING:
+    from astrai.inference.core.cache import KVCache
+
+_current_backend: contextvars.ContextVar["AttentionBackend"] = contextvars.ContextVar(
+    "attn_backend"
+)
+
+
+@functools.lru_cache(maxsize=1)
+def flash_attn_available() -> bool:
+    if not torch.cuda.is_available():
+        return False
+    fa = _get_flash_attn()
+    if fa is None:
+        return False
+
+    try:
+        major = int(fa.__version__.split(".")[0])
+        cc = torch.cuda.get_device_capability()
+        cc_num = cc[0] * 10 + cc[1]
+    except Exception:
+        major, cc_num = 0, 0
+    if (major >= 3 and cc_num < 90) or (major < 3 and 0 < cc_num < 70):
+        return False
+
+    try:
+        if not hasattr(fa, "flash_attn_func"):
+            return False
+        x = torch.zeros(1, 1, 1, 64, device="cuda", dtype=torch.bfloat16)
+        out = fa.flash_attn_func(x, x, x, causal=True)
+        return bool(torch.isfinite(out).all().item())
+    except Exception:
+        return False
+
+
+@functools.lru_cache(maxsize=1)
+def _get_flash_attn():
+    try:
+        return importlib.import_module("flash_attn")
+    except Exception:
+        return None
+
+
+class ATTN_BACKEND(enum.Enum):
+    """Backend selector enum, mirroring ``torch.nn.attention.SDPBackend``."""
+
+    TORCH_NATIVE = "torch_native"
+    CUDA = "cuda"
+    FLASH = "flash"
+
+
+_default_backend: Optional["AttentionBackend"] = None
+
+
+def _priority_backends() -> list["AttentionBackend"]:
+    """Available backends in priority order: cuda -> flash -> torch."""
+    backends: list[AttentionBackend] = []
+    if is_available("attn_paged_decode") and is_available("attn_paged_prefill"):
+        backends.append(CudaBackend())
+    if flash_attn_available():
+        backends.append(FlashAttnBackend())
+    backends.append(TorchNativeBackend())
+    return backends
+
+
+def _backend_supports(
+    backend: "AttentionBackend",
+    q: Tensor,
+    kv_cache: Optional["KVCache"],
+    attn_mask: Optional[Tensor],
+    is_causal: bool,
+) -> bool:
+    """Whether ``backend`` can run this attention call.
+
+    The CUDA kernels are bf16-only, support head_dim in 32/64/128/256, and
+    need a KV cache (decode/prefill); everything else falls back to torch.
+    """
+    if isinstance(backend, CudaBackend):
+        return (
+            kv_cache is not None
+            and q.dtype == torch.bfloat16
+            and q.size(-1) in (32, 64, 128, 256)
+        )
+    if isinstance(backend, FlashAttnBackend):
+        if not flash_attn_available():
+            return False
+        if q.size(1) == 1 and kv_cache is not None:
+            return True
+        return not (attn_mask is not None and not is_causal)
+    return True
+
+
+def _resolve_default_backend() -> "AttentionBackend":
+    """Pick the highest-priority available backend (cuda -> flash -> torch).
+
+    Set ``ASTR_BACKEND`` to override: ``ASTR_BACKEND=cuda``, ``torch_native``,
+    or ``flash``.  The value is the registered name (same as the
+    ``ATTN_BACKEND`` enum value).
+
+    Resolved lazily on first ``get_backend()`` and cached.  Per-call
+    capability fallback happens in ``attention()``, so the default is
+    safe for training and fp32 models.
+    """
+    forced = os.environ.get("ASTR_BACKEND", "").strip().lower()
+    if forced:
+        try:
+            return AttentionBackendFactory.create(forced)
+        except (ValueError, RuntimeError):
+            pass
+    return _priority_backends()[0]
+
+
+def get_backend() -> "AttentionBackend":
+    """Return the active backend for the current thread/context.
+
+    Falls back to the highest-priority available backend (cuda -> flash ->
+    torch_native) when no backend has been activated via ``with``.  Set
+    ``ASTR_BACKEND`` to override the default.
+    """
+    try:
+        return _current_backend.get()
+    except LookupError:
+        global _default_backend
+        if _default_backend is None:
+            _default_backend = _resolve_default_backend()
+        return _default_backend
+
+
+@contextmanager
+def attn_backend(backend: Union[str, ATTN_BACKEND, "AttentionBackend", type]):
+    """Context manager to select an attention backend.
+
+    Mirrors ``torch.nn.attention.sdpa_kernel``. Accepts an
+    registered name, ``ATTN_BACKEND`` enum value, backend class, or instance.
+
+    Examples::
+
+        with attn_backend(ATTN_BACKEND.TORCH_NATIVE):
+            ...
+        with attn_backend(TorchNativeBackend):
+            ...
+        with attn_backend(TorchNativeBackend()):
+            ...
+    """
+    if isinstance(backend, ATTN_BACKEND):
+        instance = AttentionBackendFactory.create(backend.value)
+    elif isinstance(backend, str):
+        instance = AttentionBackendFactory.create(backend)
+    elif isinstance(backend, type) and issubclass(backend, AttentionBackend):
+        instance = backend()
+    elif isinstance(backend, AttentionBackend):
+        instance = backend
+    else:
+        raise TypeError(
+            f"expected a registered name, ATTN_BACKEND, AttentionBackend type, "
+            f"or instance, "
+            f"got {type(backend).__name__}"
+        )
+    token = _current_backend.set(instance)
+    try:
+        yield instance
+    finally:
+        _current_backend.reset(token)
+
+
+def repeat_kv(x: Tensor, n_rep: int) -> Tensor:
+    """Expand KV heads to match Q heads for GQA."""
+    bs, slen, n_heads, head_dim = x.shape
+    if n_rep == 1:
+        return x
+    return (
+        x[:, :, :, None, :]
+        .expand(bs, slen, n_heads, n_rep, head_dim)
+        .reshape(bs, slen, n_heads * n_rep, head_dim)
+    )
+
+
+def _write_and_gather_kv(
+    kv_cache: "KVCache",
+    k: Tensor,
+    v: Tensor,
+    layer_id: int,
+    q: Tensor,
+    attn_mask: Optional[Tensor],
+) -> tuple[Tensor, Tensor]:
+    kv_cache.k_buffer[layer_id, kv_cache.out_cache_loc] = k
+    kv_cache.v_buffer[layer_id, kv_cache.out_cache_loc] = v
+    max_len = kv_cache.max_len
+    indices = kv_cache.req_to_token[kv_cache.req_pool_indices, :max_len]
+    if q.size(1) == 1 and attn_mask is not None and attn_mask.dim() == 4:
+        pos_mask = attn_mask[:, 0, 0]
+    else:
+        pos_mask = (
+            torch.arange(max_len, device=q.device)[None, :] < kv_cache.seq_lens[:, None]
+        )
+    indices = torch.where(pos_mask, indices, torch.zeros_like(indices))
+    return kv_cache.k_buffer[layer_id, indices], kv_cache.v_buffer[layer_id, indices]
+
+
+def attention(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    kv_cache: Optional["KVCache"] = None,
+    layer_id: int = 0,
+    attn_mask: Optional[Tensor] = None,
+    is_causal: bool = False,
+) -> Tensor:
+    """Functional attention entry point — mirrors ``F.scaled_dot_product_attention``.
+
+    Delegates to the active backend (set via ``with attn_backend(...)``).
+    Handles KV cache I/O, GQA head expansion, and causal masking so the
+    caller only needs to provide projected q/k/v.
+
+    Args:
+        q: [batch, q_len, n_heads, head_dim] (blhd)
+        k: [batch, q_len, n_kv_heads, head_dim] (blhd)
+        v: [batch, q_len, n_kv_heads, head_dim] (blhd)
+        kv_cache: cache dataclass, or None for training (no cache).
+        layer_id: transformer layer index for buffer access.
+        attn_mask: pre-built attention mask (SDPA-compatible).
+        is_causal: whether to apply causal masking.
+
+    Returns:
+        [batch, q_len, n_heads * head_dim]
+    """
+    backend = get_backend()
+    if not _backend_supports(backend, q, kv_cache, attn_mask, is_causal):
+        # The active backend cannot run this call (e.g. CUDA on a training /
+        # fp32 / unsupported-head_dim input) — fall back to the highest-
+        # priority backend that can, ending at torch SDPA.
+        for candidate in _priority_backends():
+            if isinstance(candidate, type(backend)):
+                continue
+            if _backend_supports(candidate, q, kv_cache, attn_mask, is_causal):
+                backend = candidate
+                break
+    return backend.forward(q, k, v, kv_cache, layer_id, attn_mask, is_causal)
+
+
+class AttentionBackend(ABC):
+    """Abstract base for attention computation strategies.
+
+    Subclasses implement ``fwd_decode`` (q_len == 1, with cache) and
+    ``fwd_prefill`` (q_len > 1, with or without cache). The public
+    ``forward`` method dispatches based on q_len.
+
+    Three equivalent ways to activate a backend::
+
+        with attn_backend(ATTN_BACKEND.TORCH_NATIVE):  # enum
+            ...
+        with attn_backend(TorchNativeBackend):          # class
+            ...
+        with TorchNativeBackend():                       # instance
+            ...
+    """
+
+    def __enter__(self) -> "AttentionBackend":
+        self._token = _current_backend.set(self)
+        return self
+
+    def __exit__(self, *exc) -> None:
+        _current_backend.reset(self._token)
+
+    def forward(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        """Dispatch to decode or extend based on q_len.
+
+        Args:
+            q: [batch, q_len, n_heads, head_dim]
+            k: [batch, q_len, n_kv_heads, head_dim]
+            v: [batch, q_len, n_kv_heads, head_dim]
+            kv_cache: cache dataclass, or None for training (no cache).
+            layer_id: transformer layer index for buffer access.
+            attn_mask: pre-built attention mask compatible with SDPA.
+            is_causal: whether to apply causal masking.
+
+        Returns:
+            [batch, q_len, n_heads * head_dim]
+        """
+        if kv_cache is not None and q.size(1) == 1:
+            return self.fwd_decode(q, k, v, kv_cache, layer_id, attn_mask, is_causal)
+        return self.fwd_prefill(q, k, v, kv_cache, layer_id, attn_mask, is_causal)
+
+    @abstractmethod
+    def fwd_decode(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        """Single-token decode with KV cache."""
+
+    @abstractmethod
+    def fwd_prefill(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        """Multi-token prefill or training forward."""
+
+    @staticmethod
+    def supports_graph() -> bool:
+        """Return True if this backend supports CUDA-graph capture.
+
+        Override in subclasses that can run under ``torch.cuda.graph``.
+
+        Called on the *active* backend instance (or its class) — a cheap
+        boolean check with no side-effects.
+        """
+        return False
+
+
+class AttentionBackendFactory(BaseFactory[AttentionBackend]):
+    """Factory for registered attention backends."""
+
+
+@AttentionBackendFactory.register(ATTN_BACKEND.TORCH_NATIVE.value)
+class TorchNativeBackend(AttentionBackend):
+    """Reference backend using torch SDPA with indirect KV cache indexing.
+
+    Writes new K/V into the cache buffers, gathers the full sequence K/V
+    via ``req_to_token`` indirect indexing, then calls
+    ``F.scaled_dot_product_attention``.
+
+    For training (``kv_cache is None``), skips cache I/O entirely and
+    runs SDPA directly on the projected q/k/v.
+    """
+
+    @staticmethod
+    def supports(**kwargs) -> bool:
+        return True
+
+    def fwd_decode(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        return self._forward(q, k, v, kv_cache, layer_id, attn_mask, is_causal)
+
+    def fwd_prefill(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        return self._forward(q, k, v, kv_cache, layer_id, attn_mask, is_causal)
+
+    def _forward(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        if kv_cache is not None:
+            k, v = _write_and_gather_kv(kv_cache, k, v, layer_id, q, attn_mask)
+
+        n_rep = q.size(2) // k.size(2)
+        if n_rep > 1:
+            k = repeat_kv(k, n_rep)
+            v = repeat_kv(v, n_rep)
+
+        out = F.scaled_dot_product_attention(
+            q.permute(0, 2, 1, 3),
+            k.permute(0, 2, 1, 3),
+            v.permute(0, 2, 1, 3),
+            attn_mask,
+            is_causal=is_causal,
+        )
+        out = out.permute(0, 2, 1, 3).contiguous().flatten(2)
+        return out
+
+
+@AttentionBackendFactory.register(ATTN_BACKEND.CUDA.value)
+class CudaBackend(AttentionBackend):
+    """CUDA kernel backend with direct KV cache access.
+
+    Decode path: writes K/V to the flat pool, then calls
+    ``attn_paged_decode`` with req_to_token + kv_indptr.
+
+    Prefill path: writes K/V to the flat pool, then calls
+    ``attn_paged_prefill`` with ragged-batch support via qo_indptr +
+    kv_indptr.
+
+    ``kv_cache is None`` (training) raises — the per-call fallback to
+    torch SDPA for training / fp32 / unsupported head_dim happens in the
+    ``attention()`` entry point.
+
+    Raises ``RuntimeError`` if the required kernel is not available.
+    """
+
+    @staticmethod
+    def supports(**kwargs) -> bool:
+        head_dim = kwargs.get("head_dim", -1)
+        return (
+            torch.cuda.is_available()
+            and head_dim in (32, 64, 128, 256)
+            and is_available("attn_paged_decode")
+            and is_available("attn_paged_prefill")
+        )
+
+    @staticmethod
+    def supports_graph() -> bool:
+        return True
+
+    def fwd_decode(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        if kv_cache is None:
+            raise RuntimeError("CudaBackend does not support training (kv_cache=None)")
+
+        loc = kv_cache.out_cache_loc[:, 0]
+        kv_cache.k_buffer[layer_id].index_copy_(0, loc, k[:, 0])
+        kv_cache.v_buffer[layer_id].index_copy_(0, loc, v[:, 0])
+
+        q_3d = q.squeeze(1)
+
+        kv_indptr = kv_cache.kv_indptr
+
+        out = attn_paged_decode(
+            q_3d,
+            kv_cache.k_buffer[layer_id],
+            kv_cache.v_buffer[layer_id],
+            kv_cache.req_to_token,
+            kv_cache.req_pool_indices,
+            kv_indptr,
+            kv_cache.max_len,
+            is_causal=True,
+            o_part_buf=kv_cache.decode_o_part,
+            ml_part_buf=kv_cache.decode_ml_part,
+            out_buf=kv_cache.decode_out,
+        )
+        return out.unsqueeze(1).flatten(2)
+
+    def fwd_prefill(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        if kv_cache is None:
+            raise RuntimeError("CudaBackend does not support training (kv_cache=None)")
+
+        loc = kv_cache.out_cache_loc.reshape(-1)
+        kv_cache.k_buffer[layer_id].index_copy_(
+            0, loc, k.reshape(-1, k.size(2), k.size(3))
+        )
+        kv_cache.v_buffer[layer_id].index_copy_(
+            0, loc, v.reshape(-1, v.size(2), v.size(3))
+        )
+
+        b = q.size(0)
+        q_len = q.size(1)
+
+        kv_indptr = kv_cache.kv_indptr
+        qo_indptr = kv_cache.qo_indptr
+
+        q_flat = q.reshape(b * q_len, q.size(2), q.size(3))
+
+        out = attn_paged_prefill(
+            q_flat,
+            kv_cache.k_buffer[layer_id],
+            kv_cache.v_buffer[layer_id],
+            kv_cache.req_to_token,
+            kv_cache.req_pool_indices,
+            kv_indptr,
+            qo_indptr,
+            attn_mask,
+            q_len,
+            is_causal=is_causal,
+        )
+        return out.reshape(b, q_len, q.size(2), q.size(3)).flatten(2)
+
+
+@AttentionBackendFactory.register(ATTN_BACKEND.FLASH.value)
+class FlashAttnBackend(AttentionBackend):
+    """FlashAttention backend via the optional ``flash-attn`` package.
+
+    Decode (q_len=1, contiguous cache): uses ``flash_attn_with_kvcache``,
+    which reads K/V directly from the flat pool via cache_batch_idx +
+    cache_seqlens — no materialized KV gather.
+
+    Prefill / non-contiguous decode: falls back to KV gather +
+    ``flash_attn_func``.
+    """
+
+    @staticmethod
+    def supports(**kwargs) -> bool:
+        return flash_attn_available()
+
+    def fwd_decode(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        return self._forward(q, k, v, kv_cache, layer_id, attn_mask, is_causal)
+
+    def fwd_prefill(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        return self._forward(q, k, v, kv_cache, layer_id, attn_mask, is_causal)
+
+    def _forward(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: Optional["KVCache"],
+        layer_id: int,
+        attn_mask: Optional[Tensor] = None,
+        is_causal: bool = False,
+    ) -> Tensor:
+        if kv_cache is not None:
+            if q.size(1) == 1 and kv_cache.k_buffer.size(
+                1
+            ) == kv_cache.req_to_token.size(0) * kv_cache.req_to_token.size(1):
+                return self._decode_with_kvcache(q, k, v, kv_cache, layer_id)
+            k, v = _write_and_gather_kv(kv_cache, k, v, layer_id, q, attn_mask)
+
+        n_rep = q.size(2) // k.size(2)
+        if n_rep > 1:
+            k = repeat_kv(k, n_rep)
+            v = repeat_kv(v, n_rep)
+
+        if attn_mask is not None and not is_causal:
+            raise ValueError(
+                "FlashAttnBackend does not support a custom attention mask; "
+                "use a causal mask or select TorchNativeBackend."
+            )
+        fa = _get_flash_attn()
+        if fa is None:
+            raise RuntimeError(
+                "FlashAttnBackend requires the optional 'flash-attn' package. "
+                "Install with `pip install flash-attn`."
+            )
+        out = fa.flash_attn_func(
+            q.contiguous(), k.contiguous(), v.contiguous(), causal=is_causal
+        )
+        return out.contiguous().flatten(2)
+
+    def _decode_with_kvcache(
+        self,
+        q: Tensor,
+        k: Tensor,
+        v: Tensor,
+        kv_cache: "KVCache",
+        layer_id: int,
+    ) -> Tensor:
+        max_batch = kv_cache.req_to_token.size(0)
+        max_seq = kv_cache.req_to_token.size(1)
+        n_kv = k.size(2)
+
+        k_cache = kv_cache.k_buffer[layer_id].view(max_batch, max_seq, n_kv, k.size(3))
+        v_cache = kv_cache.v_buffer[layer_id].view(max_batch, max_seq, n_kv, v.size(3))
+
+        fa = _get_flash_attn()
+        out = fa.flash_attn_with_kvcache(
+            q=q,
+            k_cache=k_cache,
+            v_cache=v_cache,
+            k=k,
+            v=v,
+            cache_seqlens=(kv_cache.seq_lens - 1).to(torch.int32),
+            cache_batch_idx=kv_cache.req_pool_indices.to(torch.int32),
+            causal=True,
+        )
+        return out.flatten(2)

astrai/extension/attention_ops.py

@@ -0,0 +1,194 @@
+"""Attention kernel wrapper functions — one entry point per compiled kernel.
+
+Each wrapper calls its CUDA kernel directly. If the kernel is not
+available, raises ``RuntimeError``. Fallback to torch SDPA is the
+responsibility of the attention backend, not this module.
+
+Layout convention: all q/k/v are ``[batch, seq_len, n_heads, head_dim]``
+(blhd). Scale is always ``1/sqrt(head_dim)``.
+
+Interface (all functions):
+    is_causal: True = causal mask; False = non-causal
+    mask:      2D [batch, kv_len] or 3D [batch, q_len, kv_len] (bool, True=keep)
+"""
+
+import enum
+from typing import Optional
+
+import torch
+
+from astrai.extension.loader import _available, _modules
+
+
+class TensorLayout(enum.IntEnum):
+    """Q/K/V tensor layout, mirrors the C++ ``TensorLayout`` enum in ``attn_common.h``.
+
+    Kernels internally operate on BHLD; BLHD inputs are transposed at entry.
+    """
+
+    BHLD = 0  # [batch, n_heads, seq_len, head_dim]
+    BLHD = 1  # [batch, seq_len, n_heads, head_dim]
+
+
+def _check_available(name: str):
+    if not _available.get(name):
+        raise RuntimeError(
+            f"CUDA kernel '{name}' is not available. "
+            f"Build with CSRC_KERNELS=true or use a torch-native backend."
+        )
+
+
+def attn_decode(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    mask: Optional[torch.Tensor] = None,
+    is_causal: bool = False,
+) -> torch.Tensor:
+    """GQA decode attention (q_len == 1).
+
+    Args:
+        q: [batch, 1, n_heads, head_dim] (blhd, bf16)
+        k: [batch, kv_len, n_kv_heads, head_dim] (blhd, bf16)
+        v: [batch, kv_len, n_kv_heads, head_dim] (blhd, bf16)
+        mask: 2D [batch, kv_len] or 3D [batch, 1, kv_len] (bool, True=keep)
+        is_causal: apply causal mask
+
+    Returns:
+        [batch, 1, n_heads, head_dim] (blhd, bf16)
+    """
+    _check_available("attn_decode")
+    causal_offset = (k.size(1) - 1) if is_causal else -1
+    return _modules["attn_decode"].attn_decode(
+        q, k, v, mask=mask, causal_offset=causal_offset, layout=TensorLayout.BLHD
+    )
+
+
+def attn_prefill(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    mask: Optional[torch.Tensor] = None,
+    is_causal: bool = False,
+) -> torch.Tensor:
+    """GQA prefill attention (q_len > 1).
+
+    Args:
+        q: [batch, q_len, n_heads, head_dim] (blhd, bf16)
+        k: [batch, kv_len, n_kv_heads, head_dim] (blhd, bf16)
+        v: [batch, kv_len, n_kv_heads, head_dim] (blhd, bf16)
+        mask: 2D [batch, kv_len] or 3D [batch, q_len, kv_len] (bool, True=keep)
+        is_causal: apply causal mask
+
+    Returns:
+        [batch, q_len, n_heads, head_dim] (blhd, bf16)
+    """
+    _check_available("attn_prefill")
+    causal_offset = (k.size(1) - q.size(1)) if is_causal else -1
+    return _modules["attn_prefill"].attn_prefill(
+        q, k, v, mask=mask, causal_offset=causal_offset, layout=TensorLayout.BLHD
+    )
+
+
+def attn_paged_decode(
+    q: torch.Tensor,
+    k_cache: torch.Tensor,
+    v_cache: torch.Tensor,
+    req_to_token: torch.Tensor,
+    req_pool_indices: torch.Tensor,
+    kv_indptr: torch.Tensor,
+    max_seq_len: int,
+    mask: Optional[torch.Tensor] = None,
+    is_causal: bool = False,
+    o_part_buf: Optional[torch.Tensor] = None,
+    ml_part_buf: Optional[torch.Tensor] = None,
+    out_buf: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """SGLang-style paged decode (q_len == 1, flat KV pool).
+
+    Reads K/V directly from a flat pool [size, kv_head, head_dim] via
+    req_to_token indirect indexing.  Each request has its own seq_len
+    (from kv_indptr), eliminating padding waste.
+
+    Args:
+        q: [batch, n_heads, head_dim] (bf16, 3D — no seq dim)
+        k_cache: [pool_size, n_kv_heads, head_dim] (bf16, flat)
+        v_cache: same as k_cache
+        req_to_token: [num_reqs, max_context_len] (int64) — token -> slot
+        req_pool_indices: [batch] (int64) — rows into req_to_token
+        kv_indptr: [batch+1] (int32) — prefix sum of per-request seq_lens
+        max_seq_len: max per-request seq_len (Python int, for split computation)
+        mask: 2D [batch, max_seq_len] (bool, True=keep) or None
+        is_causal: apply causal mask
+        o_part_buf: pre-allocated split-KV o partial buffer (workflow bypass)
+        ml_part_buf: pre-allocated split-KV m/l buffer (workflow bypass)
+        out_buf: pre-allocated output buffer [batch, n_heads, head_dim] (graph-safe)
+
+    Returns:
+        [batch, n_heads, head_dim] (bf16, 3D)
+    """
+    _check_available("attn_paged_decode")
+    causal_offset = 0 if is_causal else -1
+    return _modules["attn_paged_decode"].attn_paged_decode(
+        q,
+        k_cache,
+        v_cache,
+        req_to_token,
+        req_pool_indices,
+        kv_indptr,
+        max_seq_len,
+        mask=mask,
+        causal_offset=causal_offset,
+        o_part_buf=o_part_buf,
+        ml_part_buf=ml_part_buf,
+        out_buf=out_buf,
+    )
+
+
+def attn_paged_prefill(
+    q: torch.Tensor,
+    k_cache: torch.Tensor,
+    v_cache: torch.Tensor,
+    req_to_token: torch.Tensor,
+    req_pool_indices: torch.Tensor,
+    kv_indptr: torch.Tensor,
+    qo_indptr: torch.Tensor,
+    mask: Optional[torch.Tensor] = None,
+    max_q_len: int = 0,
+    is_causal: bool = False,
+) -> torch.Tensor:
+    """SGLang-style paged prefill (ragged batch, flat KV pool).
+
+    Reads K/V directly from a flat pool [size, kv_head, head_dim] via
+    req_to_token.  Supports ragged batches: each request has its own
+    q_len and kv_len, addressed via qo_indptr and kv_indptr.
+
+    Args:
+        q: [total_q, n_heads, head_dim] (bf16, 3D — flattened across requests)
+        k_cache: [pool_size, n_kv_heads, head_dim] (bf16, flat)
+        v_cache: same as k_cache
+        req_to_token: [num_reqs, max_context_len] (int64)
+        req_pool_indices: [batch] (int64)
+        kv_indptr: [batch+1] (int32) — prefix sum of per-request kv_lens
+        qo_indptr: [batch+1] (int32) — prefix sum of per-request q_lens
+        mask: 4D [batch, 1, q_len, kv_len] (bool, True=keep) or None
+        max_q_len: max per-request q_len (Python int, for grid computation)
+        is_causal: apply causal mask
+
+    Returns:
+        [total_q, n_heads, head_dim] (bf16, 3D)
+    """
+    _check_available("attn_paged_prefill")
+    causal_offset = 0 if is_causal else -1
+    return _modules["attn_paged_prefill"].attn_paged_prefill(
+        q,
+        k_cache,
+        v_cache,
+        req_to_token,
+        req_pool_indices,
+        kv_indptr,
+        qo_indptr,
+        mask,
+        max_q_len,
+        causal_offset=causal_offset,
+    )

astrai/extension/lib/__init__.py

@@ -0,0 +1 @@
+"""Compiled CUDA kernel modules (``*.so``) live here, kept separate from Python source."""

astrai/extension/loader.py

@@ -0,0 +1,42 @@
+"""Dynamic discovery and loading of compiled CUDA kernel modules.
+
+Each kernel is registered in ``csrc/build.py`` and built into a ``.so`` placed
+in this package directory. On import we try to load each one; kernels that
+failed to build (or are running on a CPU-only machine) are marked unavailable
+so the wrapper functions can fall back to ``torch`` SDPA.
+"""
+
+import importlib
+import logging
+
+logger = logging.getLogger(__name__)
+
+KERNEL_NAMES = [
+    "attn_decode",
+    "attn_prefill",
+    "attn_paged_decode",
+    "attn_paged_prefill",
+    "rotary_emb",
+]
+
+_available: dict[str, bool] = {}
+_modules: dict[str, object] = {}
+
+for _name in KERNEL_NAMES:
+    try:
+        _mod = importlib.import_module(f".lib.{_name}", package=__package__)
+        _available[_name] = True
+        _modules[_name] = _mod
+    except ImportError:
+        _available[_name] = False
+        _modules[_name] = None
+
+
+def is_available(name: str) -> bool:
+    """Return ``True`` if the compiled kernel ``name`` was loaded."""
+    return _available.get(name, False)
+
+
+def get_module(name: str) -> object:
+    """Return the loaded kernel module for ``name``, or ``None`` if unavailable."""
+    return _modules.get(name)

astrai/extension/rotary_backend.py

@@ -0,0 +1,54 @@
+"""Rotary embedding with auto-dispatch to CUDA kernel.
+
+Single entry point ``apply_rotary_emb(x, freqs_cis)`` — uses the fused
+CUDA kernel when available, falls back to torch complex multiply otherwise.
+
+Layout: x is [batch, seq_len, n_heads, head_dim] (bf16).
+freqs_cis is [batch, seq_len, dim/2, 2] (f32) — [cos, sin] pairs.
+"""
+
+import torch
+from torch import Tensor
+
+from astrai.extension.loader import is_available
+
+_cache = {"available": None}
+
+
+def _cuda_available() -> bool:
+    if _cache["available"] is None:
+        _cache["available"] = is_available("rotary_emb")
+    return _cache["available"]
+
+
+def _torch_apply(x: Tensor, freqs_cis: Tensor) -> Tensor:
+    cos, sin = freqs_cis[..., 0], freqs_cis[..., 1]
+    dtype = x.dtype
+    x_ = x.float().reshape(*x.shape[:-1], -1, 2)
+    x_complex = torch.view_as_complex(x_)
+    freqs_cis_complex = torch.complex(cos, sin).unsqueeze(2)
+    x_rotated = x_complex * freqs_cis_complex
+    x_out = torch.view_as_real(x_rotated).flatten(-2)
+    return x_out.to(dtype)
+
+
+def apply_rotary_emb(x: Tensor, freqs_cis: Tensor) -> Tensor:
+    """Apply rotary embedding to x.
+
+    Args:
+        x: [batch, seq_len, n_heads, head_dim] (bf16)
+        freqs_cis: [batch, seq_len, dim/2, 2] (f32) — [cos, sin] pairs
+
+    Returns:
+        [batch, seq_len, n_heads, head_dim] (bf16)
+    """
+    if (
+        _cuda_available()
+        and not torch.is_grad_enabled()
+        and x.is_cuda
+        and x.dtype == torch.bfloat16
+    ):
+        from astrai.extension.rotary_ops import rotary_emb as _cuda_rotary
+
+        return _cuda_rotary(x, freqs_cis)
+    return _torch_apply(x, freqs_cis)

astrai/extension/rotary_ops.py

@@ -0,0 +1,39 @@
+"""Rotary embedding CUDA kernel wrapper.
+
+Calls the compiled CUDA kernel directly. If the kernel is not available,
+raises ``RuntimeError``. Fallback to torch complex multiply is the
+responsibility of ``astrai.extension.rotary_backend.apply_rotary_emb``.
+
+Layout: x is [batch, seq_len, n_heads, head_dim] (bf16, contiguous).
+freqs_cis is [batch, seq_len, head_dim/2, 2] (f32, contiguous) — [cos, sin] pairs.
+"""
+
+import torch
+
+from astrai.extension.loader import _available, _modules
+
+
+def _check_available():
+    if not _available.get("rotary_emb"):
+        raise RuntimeError(
+            "CUDA kernel 'rotary_emb' is not available. "
+            "Build with CSRC_KERNELS=true or use the torch fallback."
+        )
+
+
+def rotary_emb(x: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
+    """Fused rotary embedding kernel.
+
+    Args:
+        x: [batch, seq_len, n_heads, head_dim] (bf16, contiguous)
+        freqs_cis: [batch, seq_len, head_dim/2, 2] (f32, contiguous) — [cos, sin] pairs
+
+    Returns:
+        [batch, seq_len, n_heads, head_dim] (bf16)
+    """
+    _check_available()
+    if not x.is_contiguous():
+        x = x.contiguous()
+    if not freqs_cis.is_contiguous():
+        freqs_cis = freqs_cis.contiguous()
+    return _modules["rotary_emb"].rotary_emb(x, freqs_cis)

astrai/factory.py

@@ -13,41 +13,63 @@ from typing import (
     Type,
     TypeVar,
     Union,
+    get_args,
+    get_origin,
 )
-from typing import get_args as _get_args
-from typing import get_origin as _get_origin
 
 T = TypeVar("T")
 
 
-def _resolve_type(
+def _resolve_base_type(
     arg: Union[Type, str, ForwardRef], factory_cls: type
 ) -> Optional[Type]:
-    """Resolve a generic type-arg (str forward-ref, ForwardRef, or class)."""
-    if not isinstance(arg, (str, ForwardRef)):
+    """Resolve the generic type-arg T to a concrete class.
+
+    - Concrete class (``BaseFactory[MyBase]``): returned directly.
+    - Forward reference (``BaseFactory["MyBase"]``): ``Base["X"]``
+      produces a ``ForwardRef("X")`` at class-creation time.  We
+      extract the name and evaluate it in the factory module's
+      global namespace — the same mechanism ``typing.get_type_hints``
+      uses internally.
+    """
+    if isinstance(arg, type):
         return arg
 
-    name = arg if isinstance(arg, str) else arg.__forward_arg__
-    if name == factory_cls.__name__:
-        return factory_cls
+    if isinstance(arg, str):
+        name = arg
+    elif isinstance(arg, ForwardRef):
+        name = arg.__forward_arg__
+    else:
+        return None
 
     mod = sys.modules.get(factory_cls.__module__)
     if mod is None:
         return None
-    ns = vars(mod)
+    try:
+        return eval(name, vars(mod))  # noqa: S307
+    except NameError:
+        return None
 
-    if isinstance(arg, ForwardRef):
-        return arg._evaluate(ns, None, recursive_guard=frozenset())
 
-    return ns.get(name)
+def _validate_component(component_cls: Type, base: Optional[Type]) -> None:
+    """Validate that *component_cls* inherits from *base*.
+
+    No-op when *base* is ``None`` (e.g. forward-ref resolution failed).
+    """
+    if base is not None and not issubclass(component_cls, base):
+        raise TypeError(f"{component_cls.__name__} must inherit from {base.__name__}")
 
 
 class BaseFactory(ABC, Generic[T]):
-    """Generic factory with decorator-based component registration.
+    """Generic factory with decorator-based registration.
+
+    Create a factory by subclassing with the desired base type::
 
         class MyFactory(BaseFactory[MyBase]):
             pass
 
+    Register components with the ``register`` decorator::
+
         @MyFactory.register("custom")
         class CustomComponent(MyBase):
             ...
@@ -64,10 +86,10 @@ class BaseFactory(ABC, Generic[T]):
     def __init_subclass__(cls, **kwargs):
         super().__init_subclass__(**kwargs)
         for orig_base in getattr(cls, "__orig_bases__", ()):
-            if _get_origin(orig_base) is BaseFactory:
-                (arg,) = _get_args(orig_base)
+            if get_origin(orig_base) is BaseFactory:
+                (arg,) = get_args(orig_base)
                 cls._entries = {}
-                cls._component_base = _resolve_type(arg, cls)
+                cls._component_base = _resolve_base_type(arg, cls)
                 return
 
     @classmethod
@@ -79,7 +101,7 @@ class BaseFactory(ABC, Generic[T]):
         """
 
         def decorator(component_cls: Type[T]) -> Type[T]:
-            cls._validate_component(component_cls)
+            _validate_component(component_cls, cls._component_base)
             if name in cls._entries:
                 raise ValueError(f"Component '{name}' is already registered")
             cls._entries[name] = component_cls
@@ -92,12 +114,11 @@ class BaseFactory(ABC, Generic[T]):
         """Create a component instance by name, filtering kwargs to match
         the component's ``__init__`` signature.
         """
-        entry = cls._entries.get(name)
-        if entry is None:
+        component_cls = cls._entries.get(name)
+        if component_cls is None:
             raise ValueError(
                 f"Unknown component: '{name}'. Supported types: {sorted(cls._entries)}"
             )
-        component_cls = entry
         sig = inspect.signature(component_cls.__init__)
         has_var_kwargs = any(
             p.kind == inspect.Parameter.VAR_KEYWORD for p in sig.parameters.values()
@@ -111,18 +132,6 @@ class BaseFactory(ABC, Generic[T]):
             kwargs = {k: v for k, v in kwargs.items() if k in valid}
         return component_cls(*args, **kwargs)
 
-    @classmethod
-    def _validate_component(cls, component_cls: Type[T]):
-        """Validate the decorated class inherits from the factory's base type.
-
-        Override for custom validation beyond ``issubclass``.
-        """
-        base = cls._component_base
-        if base is not None and not issubclass(component_cls, base):
-            raise TypeError(
-                f"{component_cls.__name__} must inherit from {base.__name__}"
-            )
-
     @classmethod
     def get_component_class(cls, name: str) -> Type[T]:
         """Get the registered component class without instantiating it."""

astrai/inference/__init__.py

@@ -5,8 +5,8 @@ Layers:
   - api/:         HTTP orchestration (ProtocolHandler, server)
   - protocols/:   Response builders (OpenAI, Anthropic)
   - transport/:   SSE transport utilities
-  - engine.py:    Facade (InferenceEngine), Value Object (GenerationRequest)
-  - sample.py:    Strategy pattern (TemperatureStrategy, TopKStrategy, TopPStrategy)
+  - engine.py:    Facade (InferenceEngine)
+  - sample.py:    Strategy pattern (TemperatureStrategy, TopKStrategy, TopPStrategy, FrequencyPenaltyStrategy)
 """
 
 from astrai.inference.api import (
@@ -30,26 +30,22 @@ from astrai.inference.api.openai import OpenAIResponseBuilder
 from astrai.inference.core import (
     STOP,
     Allocator,
-    CacheView,
-    ContiguousCache,
-    ContiguousCacheView,
     Executor,
     InferenceScheduler,
     KVCache,
-    PageCache,
-    PageCacheView,
+    KVStorage,
     PagePool,
-    PrefixCache,
-    Storage,
+    RadixCache,
+    ReqToTokenPool,
     Task,
     TaskManager,
     TaskStatus,
-    TaskTable,
     page_hash,
 )
-from astrai.inference.engine import GenerationRequest, InferenceEngine
+from astrai.inference.engine import InferenceEngine
 from astrai.inference.sample import (
     BaseSamplingStrategy,
+    FrequencyPenaltyStrategy,
     SamplingPipeline,
     TemperatureStrategy,
     TopKStrategy,
@@ -59,7 +55,6 @@ from astrai.inference.sample import (
 
 __all__ = [
     "InferenceEngine",
-    "GenerationRequest",
     "InferenceScheduler",
     "Executor",
     "STOP",
@@ -67,22 +62,18 @@ __all__ = [
     "TaskManager",
     "TaskStatus",
     "Allocator",
-    "CacheView",
     "KVCache",
-    "ContiguousCache",
-    "ContiguousCacheView",
-    "PageCache",
-    "PageCacheView",
+    "KVStorage",
     "PagePool",
-    "PrefixCache",
-    "Storage",
-    "TaskTable",
+    "RadixCache",
+    "ReqToTokenPool",
     "page_hash",
     "sample",
     "BaseSamplingStrategy",
     "TemperatureStrategy",
     "TopKStrategy",
     "TopPStrategy",
+    "FrequencyPenaltyStrategy",
     "SamplingPipeline",
     "ProtocolHandler",
     "StopChecker",

astrai/inference/api/openai.py

@@ -21,7 +21,6 @@ logger = logging.getLogger(__name__)
 _UNSUPPORTED_PARAMS = (
     "n",
     "presence_penalty",
-    "frequency_penalty",
     "logit_bias",
     "user",
 )

astrai/inference/api/protocol.py

@@ -125,6 +125,7 @@ class ProtocolHandler:
             temperature=self.request.temperature,
             top_p=self.request.top_p,
             top_k=self.request.top_k,
+            frequency_penalty=getattr(self.request, "frequency_penalty", 0.0),
         )
 
         if self.request.stream:

astrai/inference/api/server.py

@@ -110,6 +110,7 @@ def _create_engine(
     device: str = "cuda",
     dtype: torch.dtype = torch.bfloat16,
     max_batch_size: int = 16,
+    max_seq_len: Optional[int] = None,
 ) -> InferenceEngine:
     if not param_path.exists():
         raise FileNotFoundError(f"Parameter directory not found: {param_path}")
@@ -123,6 +124,7 @@ def _create_engine(
         model=model,
         tokenizer=tokenizer,
         max_batch_size=max_batch_size,
+        max_seq_len=max_seq_len,
     )
     logger.info(f"Inference engine initialized with max_batch_size={max_batch_size}")
     return engine
@@ -186,6 +188,7 @@ def run_server(
     device: str = "cuda",
     dtype: torch.dtype = torch.bfloat16,
     max_batch_size: int = 16,
+    max_seq_len: Optional[int] = None,
 ):
     app = get_app()
     app.state.server_config = {
@@ -193,6 +196,7 @@ def run_server(
         "dtype": dtype,
         "param_path": param_path,
         "max_batch_size": max_batch_size,
+        "max_seq_len": max_seq_len,
     }
     uvicorn.run(
         app,

astrai/inference/api/tool_parser.py

@@ -7,6 +7,7 @@ Subclasses may optionally consume ``token_ids`` for token-level parsing
 (e.g. Harmony / VLM-style parsers).
 """
 
+import json
 import re
 import uuid
 from abc import ABC, abstractmethod
@@ -21,13 +22,10 @@ class BaseToolParser(ABC):
     Maintains streaming state internally so that each call to :meth:`feed`
     can diff against previously emitted content.
 
-    Parameters
-    ----------
-    tools : list of dict, optional
-        Tool definitions from the request.
-    tool_choice : str
-        ``"auto"`` / ``"required"`` / ``"none"`` or a named tool choice
-        dict.
+    Args:
+        tools (list of dict, optional): Tool definitions from the request.
+        tool_choice (str): ``"auto"`` / ``"required"`` / ``"none"`` or a named
+            tool choice dict.
     """
 
     def __init__(self, tools: Optional[List[Dict]] = None, tool_choice: str = "auto"):
@@ -50,14 +48,12 @@ class BaseToolParser(ABC):
 
         Returns an empty list when nothing new should be emitted.
 
-        Parameters
-        ----------
-        body : str
-            The complete accumulated generated text so far.
-        current_token_ids : list of int, optional
-            All token IDs decoded into *body* (cumulative).
-        delta_token_ids : list of int, optional
-            Only the token IDs for this chunk.
+        Args:
+            body (str): The complete accumulated generated text so far.
+            current_token_ids (list of int, optional): All token IDs decoded
+                into *body* (cumulative).
+            delta_token_ids (list of int, optional): Only the token IDs for
+                this chunk.
         """
 
     @abstractmethod
@@ -117,6 +113,29 @@ def _parse_tool_call_json(json_str: str, complete: bool):
 
     Returns ``(name, args, valid)``.
     """
+    if complete:
+        try:
+            obj = json.loads(json_str)
+        except json.JSONDecodeError:
+            return None, "", False
+        name = obj.get("name")
+        if not isinstance(name, str) or not name:
+            return None, "", False
+        args = obj.get("arguments")
+        if isinstance(args, dict):
+            if not args:
+                args = ""
+            else:
+                args = json.dumps(args, ensure_ascii=False)
+                args = args[1:-1].rstrip()
+        elif isinstance(args, list):
+            args = json.dumps(args, ensure_ascii=False) if args else ""
+        elif isinstance(args, str):
+            pass
+        else:
+            args = str(args) if args is not None else ""
+        return name, args, True
+
     name_match = re.search(r'"name"\s*:\s*"([^"]*)"', json_str)
     if not name_match:
         return None, "", False
@@ -127,8 +146,6 @@ def _parse_tool_call_json(json_str: str, complete: bool):
         return name, "", True
 
     raw = args_match.group(1).rstrip()
-    if complete and raw.endswith("}"):
-        raw = raw[:-1].rstrip()
     if raw.startswith("{"):
         inner = raw[1:].rstrip()
         if inner.endswith("}"):
@@ -156,9 +173,6 @@ def _find_tool_calls(text: str, start_pos: int = 0):
             break
 
         json_str = text[brace:end]
-        if not _TOOL_CALL_HEAD_RE.search(json_str):
-            pos = end
-            continue
 
         name, args, valid = _parse_tool_call_json(json_str, complete=True)
         if not valid or name is None:
@@ -186,7 +200,7 @@ def _find_partial_tool_call(text: str, start_pos: int = 0):
         return None
 
     json_str = text[brace:]
-    if not _TOOL_CALL_HEAD_RE.search(json_str):
+    if '"name"' not in json_str:
         return None
 
     name, args, valid = _parse_tool_call_json(json_str, complete=False)

astrai/inference/core/__init__.py

@@ -2,16 +2,11 @@
 
 from astrai.inference.core.cache import (
     Allocator,
-    CacheView,
-    ContiguousCache,
-    ContiguousCacheView,
     KVCache,
-    PageCache,
-    PageCacheView,
+    KVStorage,
     PagePool,
-    PrefixCache,
-    Storage,
-    TaskTable,
+    RadixCache,
+    ReqToTokenPool,
     page_hash,
 )
 from astrai.inference.core.executor import Executor
@@ -20,16 +15,11 @@ from astrai.inference.core.task import STOP, Task, TaskManager, TaskStatus
 
 __all__ = [
     "Allocator",
-    "CacheView",
     "KVCache",
-    "ContiguousCache",
-    "ContiguousCacheView",
-    "PageCache",
-    "PageCacheView",
+    "KVStorage",
     "PagePool",
-    "PrefixCache",
-    "Storage",
-    "TaskTable",
+    "RadixCache",
+    "ReqToTokenPool",
     "page_hash",
     "Executor",
     "InferenceScheduler",

astrai/inference/core/cache.py

@@ -1,16 +1,34 @@
+"""KV cache architecture: three-layer separation (SGLang-inspired).
+
+Layer 1 — KVStorage:      flat token-level K/V buffers [n_layers, size, H, D]
+Layer 2 — ReqToTokenPool:  index table [req_idx, pos] → physical token slot
+Layer 3 — Allocator:       slot/page allocation with ref-counting and LRU
+
+PagePool orchestrates all three plus RadixCache (prefix addressing).
+KVCache is a pure dataclass passed to the model for direct buffer access.
+
+Two modes:
+  - contiguous (default): pre-allocated per-request blocks, no dynamic alloc
+  - paged: shared pool with on-demand allocation, prefix caching support
+"""
+
 import threading
-from abc import ABC, abstractmethod
 from collections import OrderedDict
-from typing import Callable, Dict, List, Optional, Tuple
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional
 
 import torch
 from torch import Tensor
 
+from astrai.inference.core.workspace import InferenceWorkspace
 
-def page_hash(token_ids: List[int], page_idx: int, page_size: int) -> int:
+
+def page_hash(
+    token_ids: List[int], page_idx: int, page_size: int, parent_hash: int = 0
+) -> int:
     start = page_idx * page_size
     end = min(start + page_size, len(token_ids))
-    h = 0
+    h = parent_hash
     for i in range(start, end):
         h = (h * 31 + token_ids[i]) & 0xFFFFFFFFFFFFFFFF
     return h
@@ -67,433 +85,554 @@ class Allocator:
                 self._lru.move_to_end(idx)
 
 
-class PrefixCache:
-    """Hash-based prefix matching: maps page hashes to physical page indices."""
+class RadixNode:
+    """A page-aligned edge in the CPU-side prefix radix."""
+
+    __slots__ = ("parent", "children", "page_idx", "tokens", "lock_ref")
+
+    def __init__(self, parent=None, tokens=(), page_idx=None):
+        self.parent = parent
+        self.children: Dict[tuple, "RadixNode"] = {}
+        self.page_idx = page_idx
+        self.tokens = tuple(tokens)
+        self.lock_ref = 0
+
+
+class RadixCache:
+    """Page-granular radix prefix index with exact token matching."""
 
     def __init__(self, page_size: int):
         self._page_size = page_size
+        self._root = RadixNode()
+        self._page_to_node: Dict[int, RadixNode] = {}
+        # Retained as an introspection-compatible map; matching never relies on
+        # this lossy value.
         self._page_to_hash: Dict[int, int] = {}
-        self._hash_to_page: Dict[int, int] = {}
         self._lock = threading.Lock()
 
     def evict(self, idx: int):
         with self._lock:
-            h = self._page_to_hash.pop(idx, None)
-            if h is not None:
-                self._hash_to_page.pop(h, None)
+            node = self._page_to_node.pop(idx, None)
+            self._page_to_hash.pop(idx, None)
+            if node is None:
+                return
+            node.page_idx = None
+            parent = node.parent
+            if parent is not None:
+                parent.children.pop(node.tokens, None)
 
     def has_page(self, idx: int) -> bool:
         with self._lock:
-            return idx in self._page_to_hash
+            return idx in self._page_to_node
 
     def lookup(self, token_ids: List[int]) -> List[int]:
         with self._lock:
             full_pages = len(token_ids) // self._page_size
             hits: List[int] = []
+            node = self._root
             for i in range(full_pages):
-                h = page_hash(token_ids, i, self._page_size)
-                p = self._hash_to_page.get(h)
-                if p is None:
+                start = i * self._page_size
+                page_tokens = tuple(token_ids[start : start + self._page_size])
+                child = node.children.get(page_tokens)
+                if child is None or child.page_idx is None:
                     break
-                hits.append(p)
+                hits.append(child.page_idx)
+                node = child
             return hits
 
     def record(self, page_idx: int, token_ids: List[int], logical_page_idx: int):
         with self._lock:
-            h = page_hash(token_ids, logical_page_idx, self._page_size)
-            old_h = self._page_to_hash.pop(page_idx, None)
-            if old_h is not None:
-                self._hash_to_page.pop(old_h, None)
-            self._page_to_hash[page_idx] = h
-            self._hash_to_page[h] = page_idx
+            full_pages = len(token_ids) // self._page_size
+            if logical_page_idx >= full_pages:
+                return
+            old = self._page_to_node.pop(page_idx, None)
+            self._page_to_hash.pop(page_idx, None)
+            if old is not None and old.parent is not None:
+                old.parent.children.pop(old.tokens, None)
+
+            node = self._root
+            for i in range(logical_page_idx + 1):
+                start = i * self._page_size
+                page_tokens = tuple(token_ids[start : start + self._page_size])
+                child = node.children.get(page_tokens)
+                if child is None:
+                    child = RadixNode(node, page_tokens)
+                    node.children[page_tokens] = child
+                node = child
+            if node.page_idx is not None and node.page_idx != page_idx:
+                replaced = node.page_idx
+                self._page_to_node.pop(replaced, None)
+                self._page_to_hash.pop(replaced, None)
+            node.page_idx = page_idx
+            self._page_to_node[page_idx] = node
+            self._page_to_hash[page_idx] = page_hash(
+                token_ids, logical_page_idx, self._page_size
+            )
+
+    def release(self, pages: List[int]) -> None:
+        with self._lock:
+            for page_idx in pages:
+                node = self._page_to_node.get(page_idx)
+                if node is not None and node.lock_ref:
+                    node.lock_ref -= 1
+
+
+class ReqToTokenPool:
+    """Maps [req_idx, pos] -> physical token slot in KV storage.
+
+    Each row is one request; each column is a sequence position. The value
+    at [req_idx, pos] is the flat index into the KV storage buffers.
+    """
+
+    def __init__(self, size: int, max_context_len: int, device: torch.device):
+        self.size = size
+        self.max_context_len = max_context_len
+        self.req_to_token = torch.zeros(
+            (size, max_context_len), dtype=torch.long, device=device
+        )
+        self.free_slots = list(range(size))
+        self._lock = threading.Lock()
+
+    def alloc(self, num_reqs: int) -> Optional[List[int]]:
+        with self._lock:
+            if num_reqs > len(self.free_slots):
+                return None
+            slots = self.free_slots[:num_reqs]
+            self.free_slots = self.free_slots[num_reqs:]
+            return slots
+
+    def free(self, req_indices: List[int]):
+        with self._lock:
+            self.free_slots.extend(req_indices)
+
+    def write(self, indices, values):
+        self.req_to_token[indices] = values
+
+
+class KVStorage:
+    """Token-level KV cache storage.
+
+    Buffers: [n_layers, size, n_kv_heads, head_dim]. Each token occupies
+    one slot indexed by ReqToTokenPool.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        n_layers: int,
+        n_kv_heads: int,
+        head_dim: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ):
+        self.size = size
+        self.k_buffer = torch.empty(
+            (n_layers, size, n_kv_heads, head_dim), device=device, dtype=dtype
+        )
+        self.v_buffer = torch.empty(
+            (n_layers, size, n_kv_heads, head_dim), device=device, dtype=dtype
+        )
+
+    def get_key_buffer(self, layer_id: int) -> Tensor:
+        return self.k_buffer[layer_id]
+
+    def get_value_buffer(self, layer_id: int) -> Tensor:
+        return self.v_buffer[layer_id]
+
+    def set_kv_buffer(self, layer_id: int, loc: Tensor, k: Tensor, v: Tensor) -> None:
+        self.k_buffer[layer_id, loc] = k
+        self.v_buffer[layer_id, loc] = v
+
+
+@dataclass
+class KVCache:
+    """Pure data struct passed to model for KV cache I/O.
+
+    The attention layer does raw buffer indexing — no methods, no abstraction.
+
+    Attributes:
+        k_buffer: [n_layers, size, n_kv_heads, head_dim]
+        v_buffer: [n_layers, size, n_kv_heads, head_dim]
+        req_to_token: [num_reqs, max_ctx_len] — index table
+        req_pool_indices: [batch_size] — row indices into req_to_token
+        seq_lens: [batch_size] — per-request total sequence lengths
+        out_cache_loc: [batch, new_seq_len] or [batch, 1] — write indices
+        max_len: max(seq_lens) as Python int — avoids GPU sync in decode
+        kv_indptr: [batch+1] int32 — prefix sum of seq_lens, precomputed once
+            per step so the attention backend avoids rebuilding it per layer.
+        qo_indptr: [batch+1] int32 — prefill qo prefix-sum (None in decode)
+        decode_o_part: split-KV o partial workspace (mirrors FlashInfer)
+        decode_ml_part: split-KV m/l partial workspace (mirrors FlashInfer)
+        decode_out: pre-allocated decode output buffer (graph-safe)
+    """
+
+    k_buffer: Tensor
+    v_buffer: Tensor
+    req_to_token: Tensor
+    req_pool_indices: Tensor
+    seq_lens: Tensor
+    out_cache_loc: Tensor
+    max_len: int = 0
+    kv_indptr: Optional[Tensor] = None
+    qo_indptr: Optional[Tensor] = None
+    decode_o_part: Optional[Tensor] = None
+    decode_ml_part: Optional[Tensor] = None
+    decode_out: Optional[Tensor] = None
 
 
 class PagePool:
-    """Orchestrates allocator (page management) and PrefixCache (content addressing)."""
+    """Top-level KV cache manager.
 
-    def __init__(self, allocator: Allocator, prefix: PrefixCache):
-        self._alloc = allocator
-        self._prefix = prefix
-        self._alloc.on_evict = prefix.evict
+    Combines KVStorage + ReqToTokenPool + Allocator + RadixCache.
 
-    @property
-    def allocator(self) -> Allocator:
-        return self._alloc
-
-    @property
-    def prefix(self) -> PrefixCache:
-        return self._prefix
-
-    def alloc(self) -> int:
-        return self._alloc.alloc()
-
-    def free(self, idx: int):
-        keep = self._prefix.has_page(idx)
-        self._alloc.free(idx, keep_cached=keep)
-        if not keep:
-            self._prefix.evict(idx)
-
-    def inc_ref(self, idx: int):
-        self._alloc.inc_ref(idx)
-
-    def lookup(self, token_ids: List[int]) -> List[int]:
-        hits = self._prefix.lookup(token_ids)
-        for p in hits:
-            self._alloc.touch(p)
-        return hits
-
-    def record(self, page_idx: int, token_ids: List[int], logical_page_idx: int):
-        self._prefix.record(page_idx, token_ids, logical_page_idx)
-
-
-class TaskTable:
-    """Maps task_ids to page tables and cached token counts."""
-
-    def __init__(self, page_size: int):
-        self._page_size = page_size
-        self._pages: Dict[str, List[int]] = {}
-        self._cached: Dict[str, int] = {}
-        self._lock = threading.Lock()
-
-    def set(self, task_id: str, page_table: List[int], cached: int):
-        with self._lock:
-            self._pages[task_id] = page_table
-            self._cached[task_id] = cached
-
-    def get(self, task_id: str) -> List[int]:
-        with self._lock:
-            return self._pages.get(task_id, [])
-
-    def get_cached(self, task_id: str) -> int:
-        with self._lock:
-            return self._cached.get(task_id, 0)
-
-    def pop(self, task_id: str) -> Tuple[List[int], int]:
-        with self._lock:
-            pages = self._pages.pop(task_id, [])
-            cached = self._cached.pop(task_id, 0)
-            return pages, cached
-
-    def get_ref(self, task_id: str) -> List[int]:
-        with self._lock:
-            return self._pages.setdefault(task_id, [])
-
-    def table_tensor(self, task_ids: List[str], device: torch.device) -> Tensor:
-        with self._lock:
-            states = [self._pages.get(tid, []) for tid in task_ids]
-            max_pages = max((len(s) for s in states), default=0)
-            rows = [s + [-1] * (max_pages - len(s)) for s in states]
-            return torch.tensor(rows, dtype=torch.long, device=device)
-
-
-class Storage:
-    """KV-cache tensor storage with paged write/gather."""
+    Args:
+        n_layers: Number of transformer layers.
+        n_kv_heads: Number of KV attention heads.
+        head_dim: Dimension per head.
+        max_batch_size: Maximum concurrent requests.
+        max_seq_len: Maximum sequence length per request.
+        device, dtype: Tensor device and dtype.
+        page_size: Page size for paged mode (1 = token-level).
+        n_tokens: Total token slots for paged mode. None = contiguous mode
+            (pre-allocates max_batch_size * max_seq_len).
+    """
 
     def __init__(
         self,
         n_layers: int,
-        n_pages: int,
-        page_size: int,
         n_kv_heads: int,
         head_dim: int,
-        device: torch.device,
-        dtype: torch.dtype,
-    ):
-        self.page_size = page_size
-        self.k_cache = torch.empty(
-            (n_layers, n_pages, page_size, n_kv_heads, head_dim),
-            device=device,
-            dtype=dtype,
-        )
-        self.v_cache = torch.empty(
-            (n_layers, n_pages, page_size, n_kv_heads, head_dim),
-            device=device,
-            dtype=dtype,
-        )
-
-    def write(
-        self,
-        layer_id: int,
-        page_table: Tensor,
-        start_pos: int,
-        k: Tensor,
-        v: Tensor,
-    ):
-        seq_len = k.size(1)
-        if seq_len == 0:
-            return
-        page_size = self.page_size
-        written = 0
-        first_page = start_pos // page_size
-        last_page = (start_pos + seq_len - 1) // page_size
-        for pi in range(first_page, last_page + 1):
-            phys_pages = page_table[:, pi]
-            page_start = pi * page_size
-            write_start = max(page_start, start_pos)
-            write_end = min(page_start + page_size, start_pos + seq_len)
-            offset = write_start - page_start
-            chunk = write_end - write_start
-            valid = phys_pages >= 0
-            if not valid.all():
-                if valid.any():
-                    valid_pages = phys_pages[valid]
-                    self.k_cache[layer_id, valid_pages, offset : offset + chunk] = k[
-                        valid, written : written + chunk
-                    ]
-                    self.v_cache[layer_id, valid_pages, offset : offset + chunk] = v[
-                        valid, written : written + chunk
-                    ]
-                written += chunk
-                continue
-            self.k_cache[layer_id, phys_pages, offset : offset + chunk] = k[
-                :, written : written + chunk
-            ]
-            self.v_cache[layer_id, phys_pages, offset : offset + chunk] = v[
-                :, written : written + chunk
-            ]
-            written += chunk
-
-    def gather(
-        self, layer_id: int, page_table: Tensor, total_len: int
-    ) -> Tuple[Tensor, Tensor]:
-        safe = page_table.clamp(min=0)
-        k = self.k_cache[layer_id, safe]
-        v = self.v_cache[layer_id, safe]
-        k = k.flatten(1, 2)
-        v = v.flatten(1, 2)
-        if (page_table < 0).any():
-            invalid = (
-                (page_table < 0)
-                .unsqueeze(-1)
-                .expand(-1, -1, self.page_size)
-                .flatten(1, 2)
-            )
-            invalid = invalid[:, :, None, None].expand_as(k)
-            k = k.masked_fill(invalid, 0.0)
-            v = v.masked_fill(invalid, 0.0)
-        k = k[:, :total_len]
-        v = v[:, :total_len]
-        return k, v
-
-
-class CacheView(ABC):
-    """Abstract view passed to attention layers for KV-cache I/O."""
-
-    @abstractmethod
-    def write(self, layer_id: int, k: Tensor, v: Tensor): ...
-
-    @abstractmethod
-    def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]: ...
-
-
-class KVCache(ABC):
-    """Abstract KV-cache facade for scheduler/executor."""
-
-    @abstractmethod
-    def task_alloc(self, task_id: str, prompt_ids: List[int]) -> bool: ...
-
-    @abstractmethod
-    def task_free(self, task_id: str): ...
-
-    @abstractmethod
-    def task_extend(self, task_id: str, pos: int) -> bool: ...
-
-    @abstractmethod
-    def bind_tasks(
-        self, task_ids: List[str], total_len: int, device: torch.device
-    ) -> CacheView: ...
-
-    def task_cached(self, task_id: str) -> int:
-        return 0
-
-    def task_record_hashes(
-        self, task_id: str, prompt_ids: List[int], start_logical_page: int = 0
-    ): ...
-
-
-class PageCacheView(CacheView):
-    """Bundles Storage + page_table + total_len for attention layers."""
-
-    def __init__(self, storage: Storage, page_table: Tensor, total_len: int = 0):
-        self._storage = storage
-        self._page_table = page_table
-        self._total_len = total_len
-
-    def write(self, layer_id: int, k: Tensor, v: Tensor):
-        start_pos = self._total_len - k.size(1)
-        self._storage.write(layer_id, self._page_table, start_pos, k, v)
-
-    def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]:
-        return self._storage.gather(layer_id, self._page_table, self._total_len)
-
-
-class PageCache(KVCache):
-    """Paged KV-cache with prefix sharing."""
-
-    def __init__(
-        self,
-        n_layers: int,
-        n_pages: int,
-        page_size: int,
-        n_kv_heads: int,
-        head_dim: int,
-        device: torch.device,
-        dtype: torch.dtype,
-    ):
-        self.page_size = page_size
-        self._pool = PagePool(Allocator(n_pages), PrefixCache(page_size))
-        self._table = TaskTable(page_size)
-        self._storage = Storage(
-            n_layers, n_pages, page_size, n_kv_heads, head_dim, device, dtype
-        )
-
-    def task_alloc(self, task_id: str, prompt_ids: List[int]) -> bool:
-        hits = self._pool.lookup(prompt_ids)
-        cached = len(hits) * self.page_size
-        for p in hits:
-            self._pool.inc_ref(p)
-
-        remaining = len(prompt_ids) - cached
-        n_new = (
-            (remaining + self.page_size - 1) // self.page_size if remaining > 0 else 0
-        )
-        new_pages: List[int] = []
-        if n_new > 0:
-            for _ in range(n_new):
-                p = self._pool.alloc()
-                if p < 0:
-                    for hp in hits:
-                        self._pool.free(hp)
-                    for np in new_pages:
-                        self._pool.free(np)
-                    return False
-                new_pages.append(p)
-
-        self._table.set(task_id, hits + new_pages, cached)
-        return True
-
-    def task_free(self, task_id: str):
-        page_table, _ = self._table.pop(task_id)
-        for idx in page_table:
-            self._pool.free(idx)
-
-    def task_extend(self, task_id: str, pos: int) -> bool:
-        page_table = self._table.get(task_id)
-        needed = (pos + 1 + self.page_size - 1) // self.page_size
-        while len(page_table) < needed:
-            p = self._pool.alloc()
-            if p < 0:
-                return False
-            page_table.append(p)
-        return True
-
-    def task_cached(self, task_id: str) -> int:
-        return self._table.get_cached(task_id)
-
-    def task_record_hashes(
-        self, task_id: str, prompt_ids: List[int], start_logical_page: int = 0
-    ):
-        page_table = self._table.get(task_id)
-        full_pages = len(prompt_ids) // self.page_size
-        for i in range(start_logical_page, full_pages):
-            self._pool.record(page_table[i], prompt_ids, i)
-
-    def bind_tasks(
-        self, task_ids: List[str], total_len: int, device: torch.device
-    ) -> PageCacheView:
-        page_table = self._table.table_tensor(task_ids, device)
-        return PageCacheView(self._storage, page_table, total_len)
-
-
-class ContiguousCacheView(CacheView):
-    """Contiguous KV-cache view for attention layers."""
-
-    def __init__(
-        self, cache: "ContiguousCache", batch_indices: Tensor, total_len: int = 0
-    ):
-        self._cache = cache
-        self._batch_indices = batch_indices
-        self._total_len = total_len
-
-    def write(self, layer_id: int, k: Tensor, v: Tensor):
-        seq_len = k.size(1)
-        start_pos = self._total_len - seq_len
-        indices = self._batch_indices
-        self._cache.k[layer_id, indices, start_pos : start_pos + seq_len] = k
-        self._cache.v[layer_id, indices, start_pos : start_pos + seq_len] = v
-        new_len = start_pos + seq_len
-        for s in indices.tolist():
-            cur = self._cache._slot_len.get(s, 0)
-            if new_len > cur:
-                self._cache._slot_len[s] = new_len
-
-    def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]:
-        max_len = max(
-            self._cache._slot_len.get(int(s), 0) for s in self._batch_indices.tolist()
-        )
-        indices = self._batch_indices
-        k = self._cache.k[layer_id, indices, :max_len]
-        v = self._cache.v[layer_id, indices, :max_len]
-        return k, v
-
-
-class ContiguousCache(KVCache):
-    """Contiguous per-slot KV cache (default implementation)."""
-
-    def __init__(
-        self,
-        n_layers: int,
         max_batch_size: int,
         max_seq_len: int,
-        n_kv_heads: int,
-        head_dim: int,
         device: torch.device,
         dtype: torch.dtype,
+        page_size: int = 1,
+        n_tokens: Optional[int] = None,
     ):
+        self.page_size = page_size
+        self.max_batch_size = max_batch_size
         self.max_seq_len = max_seq_len
-        self.k = torch.zeros(
-            n_layers,
-            max_batch_size,
-            max_seq_len,
-            n_kv_heads,
-            head_dim,
-            device=device,
-            dtype=dtype,
+        self.device = device
+        self.dtype = dtype
+        self.n_layers = n_layers
+        self.n_kv_heads = n_kv_heads
+        self.head_dim = head_dim
+
+        self.contiguous = n_tokens is None
+        if self.contiguous:
+            self.n_tokens = max_batch_size * max_seq_len
+        else:
+            self.n_tokens = n_tokens
+
+        self._storage = KVStorage(
+            self.n_tokens, n_layers, n_kv_heads, head_dim, device, dtype
         )
-        self.v = torch.zeros(
-            n_layers,
-            max_batch_size,
-            max_seq_len,
-            n_kv_heads,
-            head_dim,
-            device=device,
-            dtype=dtype,
-        )
-        self._slot_len: Dict[int, int] = {}
-        self._task_slot: Dict[str, int] = {}
-        self._free_slots = list(range(max_batch_size))
-        self._device = device
+        self._req_pool = ReqToTokenPool(max_batch_size, max_seq_len, device)
+
+        if self.contiguous:
+            for i in range(max_batch_size):
+                self._req_pool.req_to_token[i] = torch.arange(
+                    i * max_seq_len, (i + 1) * max_seq_len, device=device
+                )
+            self._alloc: Optional[Allocator] = None
+            self._prefix: Optional[RadixCache] = None
+        else:
+            n_pages = self.n_tokens // page_size
+            self._alloc = Allocator(n_pages)
+            self._prefix = RadixCache(page_size) if page_size > 1 else None
+            if self._prefix is not None:
+                self._alloc.on_evict = self._prefix.evict
+
+        self._task_req: Dict[str, int] = {}
+        self._task_len: Dict[int, int] = {}
+        self._task_cached: Dict[str, int] = {}
+        self._task_slots: Dict[str, List[int]] = {}
+        self._task_pages: Dict[str, List[int]] = {}
+        self._lock = threading.Lock()
+
+        # Steady-state decode validation state: the ordered task set and its
+        # Python seq_lens mirror.  When the same set advances every sequence
+        # by exactly one token per step, bind_tasks updates the stable
+        # buffers in-place (+=1 / +=inc) instead of re-cumsumming.  Any
+        # task-set change is a miss and rebuilds.
+        self._bind_sig: Optional[tuple] = None
+        self._bind_seq_lens: Optional[List[int]] = None
+
+    # ---- task lifecycle ----
 
     def task_alloc(self, task_id: str, prompt_ids: List[int]) -> bool:
-        if not self._free_slots:
+        req_slots = self._req_pool.alloc(1)
+        if req_slots is None:
             return False
-        slot = self._free_slots.pop(0)
-        self._task_slot[task_id] = slot
-        self._slot_len[slot] = 0
+        req_idx = req_slots[0]
+        self._task_req[task_id] = req_idx
+
+        if self.contiguous:
+            self._task_len[req_idx] = len(prompt_ids)
+            self._task_cached[task_id] = 0
+            return True
+
+        n_tokens_needed = len(prompt_ids)
+        cached = 0
+
+        if self._prefix is not None:
+            hits = self._prefix.lookup(prompt_ids)
+            cached = len(hits) * self.page_size
+            for p in hits:
+                self._alloc.inc_ref(p)
+            self._task_pages[task_id] = list(hits)
+            self._task_slots[task_id] = []
+        else:
+            self._task_pages[task_id] = []
+            self._task_slots[task_id] = []
+
+        remaining = n_tokens_needed - cached
+        if remaining > 0:
+            if self.page_size == 1:
+                slots = self._alloc_tokens(remaining)
+                if slots is None:
+                    for p in self._task_pages[task_id]:
+                        self._alloc.free(p)
+                    self._req_pool.free([req_idx])
+                    del self._task_req[task_id]
+                    return False
+                self._task_slots[task_id] = slots
+            else:
+                n_new_pages = (remaining + self.page_size - 1) // self.page_size
+                new_pages = []
+                for _ in range(n_new_pages):
+                    p = self._alloc.alloc()
+                    if p < 0:
+                        for hp in self._task_pages[task_id]:
+                            self._alloc.free(hp)
+                        for np_ in new_pages:
+                            self._alloc.free(np_)
+                        self._req_pool.free([req_idx])
+                        del self._task_req[task_id]
+                        return False
+                    new_pages.append(p)
+                self._task_pages[task_id].extend(new_pages)
+
+        self._write_req_to_token(task_id, prompt_ids, cached)
+        self._task_len[req_idx] = len(prompt_ids)
+        self._task_cached[task_id] = cached
         return True
 
     def task_free(self, task_id: str):
-        slot = self._task_slot.pop(task_id, None)
-        if slot is not None:
-            self._slot_len.pop(slot, None)
-            self._free_slots.append(slot)
+        req_idx = self._task_req.pop(task_id, None)
+        if req_idx is None:
+            return
+        self._task_len.pop(req_idx, None)
+        self._task_cached.pop(task_id, None)
+
+        if not self.contiguous:
+            if self._prefix is not None:
+                for p in self._task_pages.get(task_id, []):
+                    keep = self._prefix.has_page(p)
+                    self._alloc.free(p, keep_cached=keep)
+                    if not keep:
+                        self._prefix.evict(p)
+            else:
+                for p in self._task_pages.get(task_id, []):
+                    self._alloc.free(p)
+            self._task_pages.pop(task_id, None)
+            self._task_slots.pop(task_id, None)
+
+        self._req_pool.free([req_idx])
 
     def task_extend(self, task_id: str, pos: int) -> bool:
-        return pos < self.max_seq_len
+        req_idx = self._task_req.get(task_id)
+        if req_idx is None or pos >= self.max_seq_len:
+            return False
+
+        # Paged mode must also claim a physical slot for the new token;
+        # contiguous mode's block is pre-allocated so this is a no-op.
+        if not self.contiguous and not self._extend_slot(task_id, req_idx, pos):
+            return False
+
+        self._task_len[req_idx] = pos + 1
+        return True
+
+    def _extend_slot(self, task_id: str, req_idx: int, pos: int) -> bool:
+        """Allocate the physical slot for one extended token (paged mode)."""
+        if self.page_size == 1:
+            slots = self._alloc_tokens(1)
+            if slots is None:
+                return False
+            self._task_slots.setdefault(task_id, []).extend(slots)
+            self._req_pool.req_to_token[req_idx, pos] = slots[0]
+            return True
+
+        page_idx = pos // self.page_size
+        existing = self._task_pages.get(task_id, [])
+        if page_idx >= len(existing):
+            p = self._alloc.alloc()
+            if p < 0:
+                return False
+            existing.append(p)
+            self._task_pages[task_id] = existing
+        page_offset = pos % self.page_size
+        page = existing[page_idx]
+        token_slot = page * self.page_size + page_offset
+        self._req_pool.req_to_token[req_idx, pos] = token_slot
+        return True
+
+    def task_cached(self, task_id: str) -> int:
+        return self._task_cached.get(task_id, 0)
+
+    def task_record_hashes(
+        self, task_id: str, prompt_ids: List[int], start_logical_page: int = 0
+    ):
+        if self._prefix is None or self.contiguous:
+            return
+        pages = self._task_pages.get(task_id, [])
+        full_pages = len(prompt_ids) // self.page_size
+        for i in range(start_logical_page, min(full_pages, len(pages))):
+            self._prefix.record(pages[i], prompt_ids, i)
+
+    def task_cacheable_ids(
+        self, task_id: str, prompt_ids: List[int], output_ids: List[int]
+    ):
+        """Return the sequence whose KV entries are already materialized.
+
+        The first sampled output is produced by prompt prefill, and the last
+        sampled output has not been decoded into KV yet.  Therefore the cache
+        can safely retain the prompt plus every output except the last one.
+        """
+        return list(prompt_ids) + list(output_ids[:-1])
+
+    # ---- bind for forward ----
 
     def bind_tasks(
-        self, task_ids: List[str], total_len: int, device: torch.device
-    ) -> ContiguousCacheView:
-        slots = [self._task_slot[tid] for tid in task_ids]
-        batch_indices = torch.tensor(slots, dtype=torch.long, device=device)
-        return ContiguousCacheView(self, batch_indices, total_len)
+        self,
+        task_ids: List[str],
+        workspace: InferenceWorkspace,
+        device: Optional[torch.device] = None,
+        start_pos: Optional[int] = None,
+    ) -> KVCache:
+        if device is None:
+            device = workspace.device
+        req_indices = [self._task_req[tid] for tid in task_ids]
+        # Per-request lengths come from the pool's own tracking (task_alloc
+        # sets len(prompt_ids); task_extend sets pos+1), so callers need not
+        # pass them.
+        seq_lens = [self._task_len[req_idx] for req_idx in req_indices]
+        b = len(task_ids)
+        sig = tuple(task_ids)
+
+        # Write into the caller's workspace buffers (fixed addresses, sized
+        # to max_batch/max_seq at init) — the sole owner of the per-step
+        # KV bind tensors.
+        rpi_buf = workspace.req_pool_indices
+        sl_buf = workspace.seq_lens
+        kvp_buf = workspace.kv_indptr
+        inc_buf = workspace.inc
+        ocl_buf = workspace.out_cache_loc
+
+        incremental = (
+            start_pos is None
+            and self._bind_sig is not None
+            and self._bind_sig == sig
+            and self._bind_seq_lens is not None
+            and len(self._bind_seq_lens) == b
+            and all(s == p + 1 for s, p in zip(seq_lens, self._bind_seq_lens))
+        )
+        if incremental:
+            # Steady-state decode: advance the stable buffers in-place.
+            # Normal-mode buffers keep ``+=`` legal regardless of whether
+            # this runs inside ``torch.inference_mode()``.
+            sl_buf[:b] += 1
+            kvp_buf[: b + 1] += inc_buf[: b + 1]
+            req_pool_indices = rpi_buf[:b]
+            seq_lens_t = sl_buf[:b]
+            kv_indptr = kvp_buf[: b + 1]
+        else:
+            # Cold path: fill the stable buffers from fresh host tensors.
+            rpi_buf[:b].copy_(
+                torch.tensor(req_indices, dtype=torch.long, device=device)
+            )
+            sl_buf[:b].copy_(torch.tensor(seq_lens, dtype=torch.long, device=device))
+            kvp_buf[: b + 1].zero_()
+            kvp_buf[1 : b + 1] = sl_buf[:b].cumsum(0).to(torch.int32)
+            req_pool_indices = rpi_buf[:b]
+            seq_lens_t = sl_buf[:b]
+            kv_indptr = kvp_buf[: b + 1]
+        self._bind_sig = sig
+        self._bind_seq_lens = list(seq_lens)
+
+        if start_pos is not None:
+            seq_len = seq_lens[0]
+            out_cache_loc = self._req_pool.req_to_token[
+                req_pool_indices, start_pos:seq_len
+            ]
+            # Ragged query segmentation for the prefill kernel, computed once
+            # (was rebuilt per layer in CudaBackend.fwd_prefill).
+            q_len = seq_len - start_pos
+            workspace.qo_indptr[: b + 1].copy_(
+                torch.arange(b + 1, dtype=torch.int32, device=device) * q_len
+            )
+            qo_indptr = workspace.qo_indptr[: b + 1]
+            decode_o_part, decode_ml_part = None, None
+            decode_out = None
+        else:
+            write_pos = seq_lens_t - 1
+            loc = self._req_pool.req_to_token[req_pool_indices, write_pos].unsqueeze(-1)
+            ocl_buf[:b].copy_(loc)
+            out_cache_loc = ocl_buf[:b]
+            qo_indptr = None
+            decode_o_part = getattr(workspace, "decode_o_part", None)
+            decode_ml_part = getattr(workspace, "decode_ml_part", None)
+            decode_out = getattr(workspace, "decode_out", None)
+
+        return KVCache(
+            k_buffer=self._storage.k_buffer,
+            v_buffer=self._storage.v_buffer,
+            req_to_token=self._req_pool.req_to_token,
+            req_pool_indices=req_pool_indices,
+            seq_lens=seq_lens_t,
+            out_cache_loc=out_cache_loc,
+            max_len=max(seq_lens),
+            kv_indptr=kv_indptr,
+            qo_indptr=qo_indptr,
+            decode_o_part=decode_o_part,
+            decode_ml_part=decode_ml_part,
+            decode_out=decode_out,
+        )
+
+    # ---- internals ----
+
+    def _alloc_tokens(self, n: int) -> Optional[List[int]]:
+        if self.page_size != 1:
+            raise RuntimeError("_alloc_tokens is for page_size=1 only")
+        slots = []
+        for _ in range(n):
+            p = self._alloc.alloc()
+            if p < 0:
+                for s in slots:
+                    self._alloc.free(s)
+                return None
+            slots.append(p)
+        return slots
+
+    def _write_req_to_token(self, task_id: str, prompt_ids: List[int], cached: int):
+        req_idx = self._task_req[task_id]
+        total = len(prompt_ids)
+
+        if self.contiguous:
+            return
+
+        if self.page_size == 1:
+            slots = self._task_slots.get(task_id, [])
+            all_slots = slots[: total - cached]
+            if all_slots:
+                self._req_pool.req_to_token[req_idx, cached:total] = torch.tensor(
+                    all_slots, dtype=torch.long, device=self.device
+                )
+        else:
+            pages = self._task_pages.get(task_id, [])
+            for pos in range(cached, total):
+                page_idx = pos // self.page_size
+                page_offset = pos % self.page_size
+                if page_idx < len(pages):
+                    token_slot = pages[page_idx] * self.page_size + page_offset
+                    self._req_pool.req_to_token[req_idx, pos] = token_slot

astrai/inference/core/executor.py

@@ -1,15 +1,151 @@
 import logging
+import os
+import time
+from contextlib import contextmanager
+from dataclasses import dataclass
 from typing import List, Optional
 
 import torch
+from torch import Tensor
 
-from astrai.inference.core.cache import KVCache
+from astrai.extension.attention_backend import (
+    ATTN_BACKEND,
+    CudaBackend,
+    attn_backend,
+    get_backend,
+)
+from astrai.inference.core.cache import PagePool
+from astrai.inference.core.graph import CudaGraphContext
 from astrai.inference.core.task import Task
+from astrai.inference.core.workspace import InferenceWorkspace
 from astrai.inference.sample import sample
 from astrai.model.automodel import AutoModel
 from astrai.tokenize.tokenizer import AutoTokenizer
 
 logger = logging.getLogger(__name__)
+_TIMED = os.environ.get("ASTRAI_TIMED", "") == "1"
+
+
+@contextmanager
+def timed(label: str, log: Optional[logging.Logger] = None):
+    """Wall-clock debug timer, enabled via ``ASTRAI_TIMED=1``."""
+    if not _TIMED:
+        yield
+        return
+    tic = time.perf_counter()
+    yield
+    elapsed_ms = (time.perf_counter() - tic) * 1000
+    (log or logger).info("%s %.1fms", label, elapsed_ms)
+
+
+@dataclass
+class SamplingBatchInfo:
+    """Per-batch sampling parameters, cached across decode steps.
+
+    Sampling params are constant for a given ordered task set, so they are
+    built once (pinned-memory async H2D) and reused until the task set
+    changes.  ``top_ks`` is int32 to match the native consumers.
+    """
+
+    temperatures: Tensor  # float32 [B]
+    top_ks: Tensor  # int32  [B]
+    top_ps: Tensor  # float32 [B]
+    freq_penalties: Tensor  # float32 [B]
+    has_freq: bool  # any frequency_penalty != 0 (avoids per-step GPU .any())
+
+
+def _build_sampling_batch_info(tasks: List[Task], device) -> SamplingBatchInfo:
+    pin = str(device).startswith("cuda")
+    freq_penalties = torch.tensor(
+        [t.frequency_penalty for t in tasks], dtype=torch.float32, pin_memory=pin
+    ).to(device, non_blocking=True)
+    return SamplingBatchInfo(
+        temperatures=torch.tensor(
+            [t.temperature for t in tasks], dtype=torch.float32, pin_memory=pin
+        ).to(device, non_blocking=True),
+        top_ks=torch.tensor(
+            [t.top_k for t in tasks], dtype=torch.int32, pin_memory=pin
+        ).to(device, non_blocking=True),
+        top_ps=torch.tensor(
+            [t.top_p for t in tasks], dtype=torch.float32, pin_memory=pin
+        ).to(device, non_blocking=True),
+        freq_penalties=freq_penalties,
+        has_freq=bool((freq_penalties != 0).any()),
+    )
+
+
+def _warmup_cuda_graphs(
+    model: AutoModel,
+    pool: PagePool,
+    ws: InferenceWorkspace,
+    gctx: "CudaGraphContext",
+    max_batch_size: int,
+    prompt_len: int = 1,
+    device: Optional[str] = None,
+):
+    batch_sizes = [1]
+    n = 2
+    while n <= max_batch_size:
+        batch_sizes.append(n)
+        n *= 2
+    if max_batch_size not in batch_sizes:
+        batch_sizes.append(max_batch_size)
+
+    dev = device or next(model.parameters()).device
+
+    for b in batch_sizes:
+        task_ids = [f"_gr_{b}_{i}" for i in range(b)]
+        prompt_tokens = [list(range(prompt_len)) for _ in range(b)]
+        alloc_ok = True
+        for tid, pt in zip(task_ids, prompt_tokens):
+            if not pool.task_alloc(tid, pt):
+                alloc_ok = False
+                break
+        if not alloc_ok:
+            for tid in task_ids:
+                pool.task_free(tid)
+            continue
+
+        with (
+            torch.inference_mode(),
+            attn_backend(ATTN_BACKEND.CUDA),
+            timed(f"warmup prefill b={b}", logger),
+        ):
+            kv_cache = pool.bind_tasks(task_ids, ws, start_pos=0)
+            ids_in = torch.tensor(prompt_tokens, dtype=torch.long, device=dev)
+            pos_in = torch.arange(prompt_len, device=dev).unsqueeze(0).expand(b, -1)
+            model(
+                ids_in,
+                input_mask=pos_in.unsqueeze(-1) >= torch.arange(prompt_len, device=dev),
+                kv_cache=kv_cache,
+                position_ids=pos_in,
+            )
+
+        with (
+            torch.inference_mode(),
+            attn_backend(ATTN_BACKEND.CUDA),
+            timed(f"warmup decode b={b}", logger),
+        ):
+            for step in range(2):
+                seq_pos = prompt_len + step
+                ws.position_ids[:b] = seq_pos
+                for tid in task_ids:
+                    pool.task_extend(tid, seq_pos)
+                kv = pool.bind_tasks(task_ids, ws)
+                input_mask = ws.decode_mask(ws.position_ids[:b], ws.max_seq_len)
+                ids_buf = ws.fill_input_ids([step] * b)
+                gctx.forward(
+                    model,
+                    key=(b,),
+                    input_ids=ids_buf.unsqueeze(1),
+                    input_mask=input_mask,
+                    kv_cache=kv,
+                    position_ids=ws.position_ids[:b].unsqueeze(1),
+                )
+
+        for tid in task_ids:
+            pool.task_free(tid)
+        torch.cuda.synchronize()
 
 
 class Executor:
@@ -19,7 +155,7 @@ class Executor:
         self,
         model: AutoModel,
         tokenizer: AutoTokenizer,
-        kv_cache: KVCache,
+        kv_cache: PagePool,
         device: Optional[str] = None,
         dtype: Optional[torch.dtype] = None,
     ):
@@ -29,10 +165,110 @@ class Executor:
         self.device = device or next(model.parameters()).device
         self.dtype = dtype or next(model.parameters()).dtype
 
-    def execute_prefill(self, tasks: List[Task], prompt_len: int, start_pos: int = 0):
-        if start_pos >= prompt_len:
+        # Per-step decode cache for the steady-state case where the same
+        # ordered task set decodes one token per step.  Sampling params are
+        # constant across steps; position_ids grows by exactly 1.  Single-slot:
+        # any task-set change is a cache miss.
+        self._decode_cache: Optional[tuple] = None
+
+        # Pre-allocated fixed-shape buffers for the decode hot path
+        # (input_ids, decode mask, KV bind metadata).  Eagerly sized at init
+        # so the workspace is CUDA-graph-capture friendly — no allocation
+        # during capture.
+        config = model.config
+        max_q_heads = config.num_attention_heads
+        head_dim = config.hidden_size // config.num_attention_heads
+        self._head_dim = head_dim
+        self._graph_supported = CudaBackend.supports(head_dim=head_dim)
+        self._workspace = InferenceWorkspace(
+            max_batch_size=kv_cache.max_batch_size,
+            max_seq_len=kv_cache.max_seq_len,
+            max_q_heads=max_q_heads,
+            head_dim=head_dim,
+            device=self.device,
+            dtype=self.dtype,
+        )
+
+        # CUDA-graph capture: one graph per (batch_size,) key.
+        # Enabled at init-time via _warmup_cuda_graphs for CudaBackend
+        # on supported head_dims; left disabled otherwise.
+        self._graph_ctx = CudaGraphContext()
+        self._try_enable_cuda_graph()
+
+    def _try_enable_cuda_graph(self):
+        if not self._graph_supported:
             return
 
+        self._graph_ctx.set_enabled(True)
+        _warmup_cuda_graphs(
+            self.model,
+            self.kv_cache,
+            self._workspace,
+            self._graph_ctx,
+            max_batch_size=self.kv_cache.max_batch_size,
+            device=self.device,
+        )
+
+    def _sample_logits(
+        self,
+        logits: Tensor,
+        tasks: List[Task],
+        return_logprobs: bool = False,
+        info: Optional[SamplingBatchInfo] = None,
+    ):
+        info = info or _build_sampling_batch_info(tasks, self.device)
+        if info.has_freq:
+            history_lists = [
+                t.prompt_ids[-t.rep_window :] + t.output_ids for t in tasks
+            ]
+            history_lens = [len(ids) for ids in history_lists]
+            max_len = max(history_lens, default=0)
+            padded_ids = torch.zeros(
+                len(tasks), max_len, dtype=torch.long, device=self.device
+            )
+            padded_mask = torch.zeros(
+                len(tasks), max_len, dtype=torch.bool, device=self.device
+            )
+            for i, ids in enumerate(history_lists):
+                length = len(ids)
+                padded_ids[i, :length] = torch.as_tensor(
+                    ids, dtype=torch.long, device=self.device
+                )
+                padded_mask[i, :length] = True
+        else:
+            padded_ids = None
+            padded_mask = None
+
+        result = sample(
+            logits,
+            temperature=info.temperatures,
+            top_k=info.top_ks,
+            top_p=info.top_ps,
+            frequency_penalty=info.freq_penalties,
+            input_ids=padded_ids,
+            input_mask=padded_mask,
+            return_logprobs=return_logprobs,
+        )
+        if not return_logprobs:
+            return result.tolist()
+
+        tokens, logprobs = result
+        tokens_list = tokens.tolist()
+        logprobs_list = logprobs.tolist()
+        for task, logprob in zip(tasks, logprobs_list):
+            task.output_logprobs.append(float(logprob))
+        return list(zip(tokens_list, logprobs_list))
+
+    def execute_prefill(
+        self,
+        tasks: List[Task],
+        prompt_len: int,
+        start_pos: int = 0,
+        return_logprobs: bool = False,
+    ):
+        if start_pos >= prompt_len:
+            return []
+
         tasks = sorted(tasks, key=lambda t: t.task_id)
         batch_sz = len(tasks)
 
@@ -43,50 +279,118 @@ class Executor:
         )
 
         task_ids = [t.task_id for t in tasks]
-
-        with torch.inference_mode():
-            self.model(
-                input_ids,
-                position_ids=torch.arange(
-                    start_pos, prompt_len, dtype=torch.long, device=self.device
-                )
-                .unsqueeze(0)
-                .expand(batch_sz, -1),
-                paged_cache=self.kv_cache.bind_tasks(task_ids, prompt_len, self.device),
-            )
-
-    def execute_decode(self, tasks: List[Task]) -> List[int]:
-        if not tasks:
-            return []
-
-        input_ids = torch.tensor(
-            [t.output_ids[-1] if t.output_ids else t.prompt_ids[-1] for t in tasks],
-            dtype=torch.long,
-            device=self.device,
+        position_ids = (
+            torch.arange(start_pos, prompt_len, dtype=torch.long, device=self.device)
+            .unsqueeze(0)
+            .expand(batch_sz, -1)
+        )
+        input_mask = position_ids.unsqueeze(-1) >= torch.arange(
+            prompt_len, device=self.device
         )
 
-        position_ids = torch.tensor(
-            [t.next_pos for t in tasks], dtype=torch.long, device=self.device
-        )
-        total_len = position_ids.max().item() + 1
-
-        task_ids = [t.task_id for t in tasks]
-
-        temperatures = torch.tensor([t.temperature for t in tasks], device=self.device)
-        top_ks = torch.tensor([t.top_k for t in tasks], device=self.device)
-        top_ps = torch.tensor([t.top_p for t in tasks], device=self.device)
-
-        with torch.inference_mode():
+        with (
+            torch.inference_mode(),
+            timed(f"execute_prefill b={batch_sz} prompt_len={prompt_len}", logger),
+        ):
             outputs = self.model(
-                input_ids.unsqueeze(1),
-                paged_cache=self.kv_cache.bind_tasks(task_ids, total_len, self.device),
-                position_ids=position_ids.unsqueeze(1),
+                input_ids,
+                input_mask=input_mask,
+                position_ids=position_ids,
+                kv_cache=self.kv_cache.bind_tasks(
+                    task_ids,
+                    self._workspace,
+                    start_pos=start_pos,
+                ),
             )
             logits = outputs["logits"][:, -1, :]
 
-        return sample(
-            logits,
-            temperature=temperatures,
-            top_k=top_ks,
-            top_p=top_ps,
-        ).tolist()
+        return tasks, self._sample_logits(logits, tasks, return_logprobs)
+
+    def execute_decode(
+        self, tasks: List[Task], return_logprobs: bool = False
+    ) -> List[int]:
+        """Decode next token for each task.
+
+        Args:
+            return_logprobs: When ``True``, also record (and return)
+                the log-probability of each sampled token under the
+                post-strategy sampling distribution.  The logprob is
+                appended to ``task.output_logprobs`` and the return
+                list becomes ``List[Tuple[int, float]]``.
+
+        Returns:
+            ``List[int]`` of sampled token IDs, or
+            ``List[Tuple[int, float]]`` of ``(token_id, logprob)`` when
+            ``return_logprobs`` is ``True``.
+        """
+        if not tasks:
+            return []
+
+        b = len(tasks)
+        ws = self._workspace
+
+        # ---- pre-replay: update input buffers in-place ----
+
+        input_ids = ws.fill_input_ids(
+            [t.output_ids[-1] if t.output_ids else t.prompt_ids[-1] for t in tasks]
+        )
+
+        task_ids = [t.task_id for t in tasks]
+        cur_positions = [t.next_pos for t in tasks]
+
+        sig = tuple(task_ids)
+        cached = self._decode_cache
+        if (
+            cached is not None
+            and cached[0] == sig
+            and cur_positions == [p + 1 for p in cached[1]]
+        ):
+            info = cached[2]
+            ws.position_ids[:b] += 1
+            self._decode_cache = (sig, cur_positions, info)
+        else:
+            info = _build_sampling_batch_info(tasks, self.device)
+            ws.position_ids[:b].copy_(
+                torch.tensor(cur_positions, dtype=torch.long, device=self.device)
+            )
+            self._decode_cache = (sig, cur_positions, info)
+
+        total_len = max(cur_positions) + 1
+        input_mask = ws.decode_mask(ws.position_ids[:b], total_len)
+
+        kv_cache = self.kv_cache.bind_tasks(task_ids, ws)
+
+        # ---- forward (graph replay or live run + capture) ----
+
+        use_graph = (
+            self._graph_ctx.enabled
+            and self._graph_supported
+            and get_backend().supports_graph()
+        )
+        key = (b,)
+        if use_graph:
+            input_mask = ws.decode_mask(ws.position_ids[:b], ws.max_seq_len)
+
+        with (
+            torch.inference_mode(),
+            timed(f"execute_decode forward b={b}", logger),
+        ):
+            if use_graph:
+                outputs = self._graph_ctx.forward(
+                    self.model,
+                    key=key,
+                    input_ids=input_ids.unsqueeze(1),
+                    input_mask=input_mask,
+                    kv_cache=kv_cache,
+                    position_ids=ws.position_ids[:b].unsqueeze(1),
+                )
+            else:
+                outputs = self.model(
+                    input_ids.unsqueeze(1),
+                    input_mask=input_mask,
+                    kv_cache=kv_cache,
+                    position_ids=ws.position_ids[:b].unsqueeze(1),
+                )
+            logits = outputs["logits"][:, -1, :]
+
+        return self._sample_logits(logits, tasks, return_logprobs, info=info)

astrai/inference/core/graph.py

@@ -0,0 +1,103 @@
+"""CUDA-graph capture for the decode model-forward step.
+
+Mirrors SGLang's cuda-graph manager: one graph per batch size.  The graph
+pair.  The graph captures ``model.forward()`` with workspace-backed inputs
+(all at fixed addresses).  Before each replay the caller updates the input
+buffer content in-place so the graph sees fresh data at the same tensor
+addresses.
+
+Only the model forward is captured — sampling runs outside the graph
+(via ``torch.multinomial`` which consumes a mutable RNG state).
+"""
+
+import torch
+from torch import Tensor
+
+
+class CudaGraphContext:
+    """CUDA-graph capture/replay for decode steps.
+
+    Parameters:
+        enabled: When ``False``, ``forward()`` always runs the live model
+            forward without capture/replay (graphs are cleared).  Toggle at
+            runtime via the ``set_enabled()`` method.
+
+    Usage::
+
+        gctx = CudaGraphContext()
+        with torch.inference_mode():
+            outputs = gctx.forward(
+                model,
+                key=(batch_size,),
+                input_ids=workspace.input_ids[:b].unsqueeze(1),
+                input_mask=input_mask,
+                kv_cache=kv_cache,
+                position_ids=workspace.position_ids[:b].unsqueeze(1),
+            )
+
+    The first call at a given key runs *without* capture (warmup).  The
+    second call captures the graph.  Subsequent calls replay the captured
+    graph.  A ``torch.cuda.synchronize()`` before capture drains in-flight
+    work so the graph trace is clean.
+    """
+
+    def __init__(self, enabled: bool = False):
+        self._enabled = enabled
+        self._graphs: dict[tuple, torch.cuda.CUDAGraph] = {}
+        self._outputs: dict[tuple, dict[str, Tensor]] = {}
+        self._warmed: set[tuple] = set()
+
+    @property
+    def enabled(self) -> bool:
+        return self._enabled
+
+    def set_enabled(self, flag: bool):
+        """Enable or disable CUDA-graph capture at runtime.
+
+        Disabling clears all captured graphs (frees GPU memory) and warmup
+        state.  Re-enabling after disable starts fresh — graphs are
+        re-captured on the next warmup cycle.
+        """
+        if flag == self._enabled:
+            return
+        self._enabled = flag
+        if not flag:
+            self._graphs.clear()
+            self._outputs.clear()
+            self._warmed.clear()
+
+    def forward(self, model, *, key, **kwargs) -> dict[str, Tensor]:
+        """Run ``model(**kwargs)`` via graph replay or live forward.
+
+        Args:
+            model: callable, e.g. ``self.model.forward``.
+            key: ``(batch_size,)`` — the dispatch key (one graph per batch size).
+            **kwargs: arguments forwarded to ``model``.  All tensor arguments
+                must reside at stable addresses (workspace buffers).
+
+        Returns:
+            The dict produced by ``model(**kwargs)``, e.g.
+            ``{"logits": ..., "h0": ...}``.
+        """
+        if not self._enabled:
+            self._outputs[key] = model(**kwargs)
+            return self._outputs[key]
+
+        if key in self._graphs:
+            self._graphs[key].replay()
+        elif key in self._warmed:
+            cap_output = model(**kwargs)
+            torch.cuda.synchronize()
+            graph = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(graph):
+                self._outputs[key] = model(**kwargs)
+            self._graphs[key] = graph
+            self._warmed.discard(key)
+            return cap_output
+        else:
+            self._warmed.add(key)
+            self._outputs[key] = model(**kwargs)
+        return self._outputs[key]
+
+    def has_graph(self, key: tuple) -> bool:
+        return key in self._graphs

astrai/inference/core/scheduler.py

@@ -1,10 +1,11 @@
 import logging
 import threading
+import uuid
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch
 
-from astrai.inference.core.cache import ContiguousCache, KVCache
+from astrai.inference.core.cache import PagePool
 from astrai.inference.core.executor import Executor
 from astrai.inference.core.task import STOP, Task, TaskManager, TaskStatus
 from astrai.model.automodel import AutoModel
@@ -22,45 +23,43 @@ class InferenceScheduler:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 16,
         max_seq_len: Optional[int] = None,
-        max_prompt_len: int = 2048,
         device: Optional[str] = None,
         dtype: Optional[torch.dtype] = None,
-        cache: Optional[KVCache] = None,
+        cache: Optional[PagePool] = None,
     ):
         config = model.config
 
         if max_seq_len is not None:
             self.max_seq_len = max_seq_len
-        elif config.max_len is not None:
-            self.max_seq_len = config.max_len
+        elif config.max_position_embeddings is not None:
+            self.max_seq_len = config.max_position_embeddings
         else:
             raise ValueError(
                 "max_seq_len must be provided either as argument "
-                "or in model config (config.max_len)"
+                "or in model config (config.max_position_embeddings)"
             )
         self.device = device or next(model.parameters()).device
         self.dtype = dtype or next(model.parameters()).dtype
 
-        head_dim = config.dim // config.n_heads
+        head_dim = config.hidden_size // config.num_attention_heads
 
         if cache is not None:
             self._cache = cache
         else:
-            self._cache = ContiguousCache(
-                config.n_layers,
-                max_batch_size,
-                self.max_seq_len,
-                config.n_kv_heads,
-                head_dim,
-                self.device,
-                self.dtype,
+            self._cache = PagePool(
+                n_layers=config.num_hidden_layers,
+                n_kv_heads=config.num_key_value_heads,
+                head_dim=head_dim,
+                max_batch_size=max_batch_size,
+                max_seq_len=self.max_seq_len,
+                device=self.device,
+                dtype=self.dtype,
             )
 
         self._task_mgr = TaskManager(
             tokenizer=tokenizer,
             max_batch_size=max_batch_size,
             max_seq_len=self.max_seq_len,
-            max_prompt_len=max_prompt_len,
         )
 
         self._executor = Executor(
@@ -84,6 +83,80 @@ class InferenceScheduler:
     def get_stats(self) -> Dict[str, Any]:
         return self._task_mgr.get_stats()
 
+    def _step(
+        self, tasks: List[Task], return_logprobs: bool = False
+    ) -> Tuple[List[Task], List[Task]]:
+        """Advance every active task by one token (prefill + decode).
+
+        Single shared primitive for both the continuous-batching loop and
+        the synchronous ``run_batch`` path, so the two cannot drift.
+
+        Tasks must already be allocated in the KV cache. Tasks without output
+        are prefilled first and sample their first token from the final prompt
+        position. Tasks with output extend the cache by one position and decode
+        from their latest generated token.
+
+        Args:
+            tasks: Active tasks to advance by one token.
+            return_logprobs: Forwarded to ``execute_decode``; per-token
+                logprobs are recorded on each task's ``output_logprobs``.
+
+        Returns:
+            ``(decoded, aborted)``: tasks that produced a new token (its ID
+            already appended to ``output_ids``) and tasks that hit the
+            sequence cap and were marked ``ABORTED``.
+        """
+        cache = self._cache
+
+        to_prefill = [t for t in tasks if t.output_tokens == 0 and t.prompt_ids]
+        prefilled_ids = set()
+        produced: List[Task] = []
+        if to_prefill:
+            for t in to_prefill:
+                t.input_tokens = len(t.prompt_ids)
+
+            groups: Dict[Tuple[int, int], List[Task]] = {}
+            for t in to_prefill:
+                start_pos = min(cache.task_cached(t.task_id), len(t.prompt_ids) - 1)
+                groups.setdefault((len(t.prompt_ids), start_pos), []).append(t)
+
+            for (prompt_len, start_pos), group in groups.items():
+                prefilled, step_out = self._executor.execute_prefill(
+                    group, prompt_len, start_pos, return_logprobs=return_logprobs
+                )
+                for t, out in zip(prefilled, step_out):
+                    t.output_ids.append(out[0] if return_logprobs else out)
+                    t.output_tokens += 1
+                    prefilled_ids.add(t.task_id)
+                    produced.append(t)
+                start_logical_page = start_pos // getattr(cache, "page_size", 64)
+                for t in group:
+                    cache.task_record_hashes(
+                        t.task_id, t.prompt_ids, start_logical_page
+                    )
+
+        decoded: List[Task] = []
+        aborted: List[Task] = []
+        for t in tasks:
+            if t.task_id in prefilled_ids:
+                continue
+            if cache.task_extend(t.task_id, t.next_pos):
+                decoded.append(t)
+            else:
+                t.status = TaskStatus.ABORTED
+                aborted.append(t)
+
+        if decoded:
+            step_out = self._executor.execute_decode(
+                decoded, return_logprobs=return_logprobs
+            )
+            for t, out in zip(decoded, step_out):
+                t.output_ids.append(out[0] if return_logprobs else out)
+                t.output_tokens += 1
+                produced.append(t)
+
+        return produced, aborted
+
     def _run_generation_loop(self):
         stop_ids = self._task_mgr.tokenizer.stop_ids
         cache = self._cache
@@ -91,6 +164,13 @@ class InferenceScheduler:
             while not self._stop_event.is_set():
                 finished = self._task_mgr.remove_finished_tasks(stop_ids)
                 for task in finished:
+                    if task.status == TaskStatus.FINISHED:
+                        cache.task_record_hashes(
+                            task.task_id,
+                            cache.task_cacheable_ids(
+                                task.task_id, task.prompt_ids, task.output_ids
+                            ),
+                        )
                     cache.task_free(task.task_id)
 
                 active = self._task_mgr.get_active_tasks()
@@ -110,63 +190,22 @@ class InferenceScheduler:
                     self._task_mgr.wait_for_tasks(timeout=1.0)
                     continue
 
-                to_prefill = [
-                    t
-                    for t in self._task_mgr.get_active_tasks()
-                    if t.output_tokens == 0
-                    and cache.task_cached(t.task_id) < len(t.prompt_ids)
-                ]
-                if to_prefill:
-                    for t in to_prefill:
-                        t.input_tokens = len(t.prompt_ids)
+                active = self._task_mgr.get_active_tasks()
 
-                    groups: Dict[Tuple[int, int], List[Task]] = {}
-                    for t in to_prefill:
-                        key = (
-                            len(t.prompt_ids),
-                            cache.task_cached(t.task_id),
-                        )
-                        groups.setdefault(key, []).append(t)
+                decoded, aborted = self._step(active)
 
-                    for (prompt_len, start_pos), group in groups.items():
-                        self._executor.execute_prefill(group, prompt_len, start_pos)
-                        start_logical_page = start_pos // getattr(
-                            cache, "page_size", 64
-                        )
-                        for t in group:
-                            cache.task_record_hashes(
-                                t.task_id, t.prompt_ids, start_logical_page
-                            )
+                for t in aborted:
+                    self._task_mgr.invoke_callback(t.task_id, STOP)
 
-                pos_groups: Dict[int, List[Task]] = {}
-                for t in self._task_mgr.get_active_tasks():
-                    pos_groups.setdefault(t.next_pos, []).append(t)
-
-                for next_pos in sorted(pos_groups.keys()):
-                    group = sorted(pos_groups[next_pos], key=lambda t: t.task_id)
-
-                    valid: List[Task] = []
-                    for t in group:
-                        if cache.task_extend(t.task_id, t.next_pos):
-                            valid.append(t)
-                        else:
-                            t.status = TaskStatus.ABORTED
-                            self._task_mgr.invoke_callback(t.task_id, STOP)
-
-                    if valid:
-                        next_tokens = self._executor.execute_decode(valid)
-
-                        for t, ntok in zip(valid, next_tokens):
-                            t.output_ids.append(ntok)
-                            t.output_tokens += 1
-                            self._task_mgr.invoke_callback(
-                                t.task_id,
-                                self._task_mgr.tokenizer.decode([ntok]),
-                            )
-
-                        for t in valid:
-                            if t.is_finished(stop_ids):
-                                self._task_mgr.invoke_callback(t.task_id, STOP)
+                for t in decoded:
+                    new_text = t.decode_new_token(self._task_mgr.tokenizer)
+                    if new_text:
+                        self._task_mgr.invoke_callback(t.task_id, new_text)
+                    if t.is_finished(stop_ids):
+                        remaining = t.flush_remaining(self._task_mgr.tokenizer)
+                        if remaining:
+                            self._task_mgr.invoke_callback(t.task_id, remaining)
+                        self._task_mgr.invoke_callback(t.task_id, STOP)
 
         except Exception as e:
             self._stop_event.set()
@@ -197,6 +236,94 @@ class InferenceScheduler:
             self._cache.task_free(task.task_id)
         for task in self._task_mgr.get_waiting_tasks():
             self._task_mgr.invoke_callback(task.task_id, STOP)
+            self._cache.task_free(task.task_id)
         self._task_mgr.clear_queues()
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
+
+    def run_batch(
+        self,
+        prompt_ids_list: List[List[int]],
+        *,
+        max_tokens: Optional[int] = None,
+        temperature: float = 1.0,
+        top_p: float = 1.0,
+        top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
+        return_logprobs: bool = False,
+    ) -> List[List[int]]:
+        """Synchronous batch generation without the scheduler thread.
+
+        Accepts already-tokenized prompts (no string round-trip) and runs
+        prefill + decode to completion on the calling thread.  Designed for
+        RL rollout, where logprobs of the behaviour policy must be collected
+        alongside generated tokens.
+
+        Args:
+            prompt_ids_list: ``B`` prompts, each a list of token IDs.
+            max_tokens: Maximum tokens to generate per prompt.  ``None``
+                uses ``self.max_seq_len - len(prompt_ids)``.
+            temperature/top_p/top_k/frequency_penalty/rep_window: Sampling
+                parameters (uniform across the batch).
+            return_logprobs: If ``True``, return ``(token_ids, logprobs)``
+                tuples per prompt (logprobs aligned 1-to-1 with token_ids).
+
+        Returns:
+            ``List[List[int]]`` of generated token IDs per prompt, or —
+            when ``return_logprobs`` is ``True`` —
+            ``List[Tuple[List[int], List[float]]]``.
+        """
+        stop_ids = self._task_mgr.tokenizer.stop_ids
+        cache = self._cache
+        seq_cap = self.max_seq_len
+
+        tasks: List[Task] = []
+        for ids in prompt_ids_list:
+            if len(ids) >= seq_cap:
+                tasks.append(None)
+                continue
+            t_max = max_tokens
+            if t_max is None:
+                t_max = seq_cap - len(ids)
+            else:
+                t_max = min(t_max, seq_cap - len(ids))
+            if t_max <= 0:
+                tasks.append(None)
+                continue
+            task = Task(
+                task_id=f"batch_{uuid.uuid4().hex[:8]}",
+                prompt_ids=list(ids),
+                max_tokens=t_max,
+                temperature=temperature,
+                top_p=top_p,
+                top_k=top_k,
+                frequency_penalty=frequency_penalty,
+                rep_window=rep_window,
+            )
+            if not cache.task_alloc(task.task_id, task.prompt_ids):
+                tasks.append(None)
+                continue
+            task.input_tokens = len(task.prompt_ids)
+            tasks.append(task)
+
+        try:
+            live = [t for t in tasks if t is not None]
+
+            while live:
+                decoded, _ = self._step(live, return_logprobs=return_logprobs)
+                live = [t for t in decoded if not t.is_finished(stop_ids)]
+        finally:
+            for t in tasks:
+                if t is not None:
+                    cache.task_free(t.task_id)
+
+        results: List[Any] = []
+        for t in tasks:
+            if t is None:
+                results.append(([], []) if return_logprobs else [])
+            elif return_logprobs:
+                results.append((list(t.output_ids), list(t.output_logprobs)))
+            else:
+                results.append(list(t.output_ids))
+        return results

astrai/inference/core/task.py

@@ -6,6 +6,8 @@ from collections import deque
 from enum import Enum
 from typing import Any, Callable, Deque, Dict, List, Optional
 
+from tokenizers.decoders import DecodeStream
+
 from astrai.tokenize.tokenizer import AutoTokenizer
 
 logger = logging.getLogger(__name__)
@@ -13,6 +15,33 @@ logger = logging.getLogger(__name__)
 STOP = object()
 
 
+class StreamDecoder:
+    """Incremental decoder backed by the tokenizers library's DecodeStream.
+
+    Delegates to the Rust-native streaming decoder which maintains an
+    O(1) bounded token buffer internally (via prefix drain), avoiding
+    the O(n²) cost of re-decoding the full history on each step.
+
+    Multi-byte UTF-8 sequences split across token boundaries are
+    buffered until complete; ``push`` returns "" while the trailing
+    sequence is still incomplete.
+    """
+
+    __slots__ = ("_stream", "_tok")
+
+    def __init__(self, tokenizer: AutoTokenizer):
+        self._tok = tokenizer._tokenizer
+        self._stream = DecodeStream(skip_special_tokens=True)
+
+    def push(self, token_id: int) -> str:
+        """Append a token ID and return newly completed text.
+
+        Returns "" while a multi-byte character is still incomplete.
+        """
+        chunk = self._stream.step(self._tok, token_id)
+        return chunk or ""
+
+
 class TaskStatus(Enum):
     """Task lifecycle states."""
 
@@ -33,6 +62,8 @@ class Task:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
     ):
         self.task_id = task_id
         self.prompt_ids = prompt_ids
@@ -40,17 +71,42 @@ class Task:
         self.temperature = temperature
         self.top_p = top_p
         self.top_k = top_k
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
 
         self.status = TaskStatus.PENDING
         self.output_ids: List[int] = []
+        self.output_logprobs: List[float] = []
         self.input_tokens: int = 0
         self.output_tokens: int = 0
         self.arrival_time = time.time()
         self.finish_time: Optional[float] = None
+        self._decoder: Optional[StreamDecoder] = None
+
+    def decode_new_token(self, tokenizer: AutoTokenizer) -> str:
+        """Decode the last appended output token, buffering incomplete
+        multi-byte sequences across calls.
+
+        Lazily creates a :class:`StreamDecoder` on first use.
+        """
+        if self._decoder is None:
+            self._decoder = StreamDecoder(tokenizer)
+        return self._decoder.push(self.output_ids[-1])
+
+    def flush_remaining(self, tokenizer: AutoTokenizer) -> str:
+        """Emit any text still buffered in the decoder.
+
+        With the Rust-native DecodeStream, the stream is always in a
+        correct state — any completed text was already emitted by the
+        last ``push``.  A trailing incomplete multi-byte sequence has no
+        valid text to emit, so this is a no-op.
+        """
+        return ""
 
     @property
     def next_pos(self) -> int:
-        return self.input_tokens + len(self.output_ids)
+        # The first output is sampled from prefill and enters KV on the next step.
+        return self.input_tokens + max(0, len(self.output_ids) - 1)
 
     def is_finished(self, stop_ids: List[int]) -> bool:
         if self.max_tokens is not None and self.output_tokens >= self.max_tokens:
@@ -68,12 +124,10 @@ class TaskManager:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 16,
         max_seq_len: int = 8192,
-        max_prompt_len: int = 512,
     ):
         self.tokenizer = tokenizer
         self.max_batch_size = max_batch_size
         self.max_seq_len = max_seq_len
-        self.max_prompt_len = max_prompt_len
 
         self.waiting_queue: Deque[Task] = deque()
         self.active_tasks: List[Task] = []
@@ -92,14 +146,16 @@ class TaskManager:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
         stream_callback: Optional[Callable[[str], None]] = None,
     ) -> str:
         task_id = f"task_{int(time.time())}_{uuid.uuid4().hex[:8]}"
         prompt_ids = self.tokenizer.encode(prompt)
-        if len(prompt_ids) > self.max_prompt_len:
-            prompt_ids = prompt_ids[-self.max_prompt_len :]
+        if len(prompt_ids) > self.max_seq_len:
+            prompt_ids = prompt_ids[-self.max_seq_len :]
 
-        if len(prompt_ids) >= self.max_seq_len:
+        if len(prompt_ids) > self.max_seq_len:
             if stream_callback:
                 stream_callback(STOP)
             return task_id
@@ -116,6 +172,8 @@ class TaskManager:
             temperature=temperature,
             top_p=top_p,
             top_k=top_k,
+            frequency_penalty=frequency_penalty,
+            rep_window=rep_window,
         )
 
         with self._lock:

astrai/inference/core/workspace.py

@@ -0,0 +1,151 @@
+"""Pre-allocated buffers for the inference decode hot path.
+
+Mirrors FlashInfer / SGLang's global workspace pattern: all per-step tensors
+are allocated eagerly at init (nothing is lazy), so the decode step
+reads/writes fixed-address tensors with zero ``torch.empty`` calls during
+the hot loop — a prerequisite for CUDA-graph capture.
+"""
+
+import torch
+from torch import Tensor
+
+_MAX_SPLITS = 32
+
+
+class InferenceWorkspace:
+    """Reusable fixed-shape per-step buffers for decode.
+
+    Families of buffers, all sized to ``max_batch_size`` / ``max_seq_len``
+    and sliced via views each step:
+
+    - ``decode_mask``: a ``[B, 1, total_len]`` validity mask, the RHS
+      ``arange`` pre-computed so only a single ``torch.ge(out=)`` runs per
+      step.
+    - ``input_ids``: per-step token IDs filled from host (pinned, double-
+      buffered so an in-flight async H2D copy never races the next fill).
+    - KV-cache bind metadata (``req_pool_indices``, ``seq_lens``,
+      ``kv_indptr``, ``inc``, ``out_cache_loc``), written by
+      ``PagePool.bind_tasks`` when the Executor passes this workspace.
+    - ``decode_o_part`` / ``decode_ml_part``: split-KV partial result buffers
+      (mirrors FlashInfer's workspace).  One global alloc, reused by every
+      decode step across all layers.  Sliced views are passed to the CUDA
+      attention kernel so its internal ``torch.empty`` hot-path alloc goes
+      through a stable address (CUDA-graph capturable).
+
+    No re-allocation while the server's bounds are respected.
+    """
+
+    def __init__(
+        self,
+        max_batch_size: int,
+        max_seq_len: int,
+        max_q_heads: int,
+        head_dim: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ):
+        self.max_batch_size = max_batch_size
+        self.max_seq_len = max_seq_len
+        self.max_q_heads = max_q_heads
+        self.head_dim = head_dim
+        self.device = device
+        self.dtype = dtype
+
+        # ``position_ids[:, None, None] >= arange`` RHS, reused every step.
+        self.arange = torch.arange(max_seq_len, device=device)
+        # Decode validity mask: [max_batch, 1, max_seq_len] bool.
+        self.input_mask = torch.empty(
+            (max_batch_size, 1, max_seq_len), dtype=torch.bool, device=device
+        )
+
+        # Per-step token IDs.  Values come from host Python lists every
+        # step, so the device buffer is pre-allocated (stable address for
+        # CUDA-graph capture) and filled via a host staging buffer.  A
+        # double buffer keeps a copy in flight from being overwritten by
+        # the next fill.
+        self.input_ids = torch.empty((max_batch_size,), dtype=torch.long, device=device)
+        self._pin = [
+            torch.empty((max_batch_size,), dtype=torch.long),
+            torch.empty((max_batch_size,), dtype=torch.long),
+        ]
+        self._pin_idx = 0
+
+        # KV-cache bind metadata (fixed shape, written by ``PagePool.bind_tasks``
+        # when the Executor passes this workspace).  Stable addresses make the
+        # decode forward CUDA-graph capturable.
+        self.req_pool_indices = torch.empty(
+            (max_batch_size,), dtype=torch.long, device=device
+        )
+        self.seq_lens = torch.empty((max_batch_size,), dtype=torch.long, device=device)
+        self.kv_indptr = torch.empty(
+            (max_batch_size + 1,), dtype=torch.int32, device=device
+        )
+        self.qo_indptr = torch.empty(
+            (max_batch_size + 1,), dtype=torch.int32, device=device
+        )
+        self.inc = torch.arange(max_batch_size + 1, dtype=torch.int32, device=device)
+        self.out_cache_loc = torch.empty(
+            (max_batch_size, 1), dtype=torch.long, device=device
+        )
+
+        # Per-step position IDs (must be at a fixed address for CUDA-graph capture).
+        self.position_ids = torch.empty(
+            (max_batch_size,), dtype=torch.long, device=device
+        )
+
+        # Split-KV partial-result buffers for decode (persistent, one global
+        # alloc per process — mirrors FlashInfer's workspace pattern).
+        # Shape: [max_batch_size, max_q_heads, _MAX_SPLITS, head_dim] (o_part)
+        #        [max_batch_size, max_q_heads, _MAX_SPLITS, 2]     (ml_part)
+        self.decode_o_part = torch.empty(
+            (max_batch_size, max_q_heads, _MAX_SPLITS, head_dim),
+            dtype=torch.float32,
+            device=device,
+        )
+        self.decode_ml_part = torch.empty(
+            (max_batch_size, max_q_heads, _MAX_SPLITS, 2),
+            dtype=torch.float32,
+            device=device,
+        )
+
+        # Decode output buffer (graph-safe pre-alloc).  Shape matches the
+        # decode kernel's output: [batch, q_head, head_dim].
+        self.decode_out = torch.empty(
+            (max_batch_size, max_q_heads, head_dim),
+            dtype=dtype,
+            device=device,
+        )
+
+    def decode_buffers(self, batch: int, q_heads: int):
+        """Return ``(o_part, ml_part)`` view sliced to live dimensions."""
+        return (
+            self.decode_o_part[:batch, :q_heads],
+            self.decode_ml_part[:batch, :q_heads],
+        )
+
+    def fill_input_ids(self, ids: "list[int]") -> Tensor:
+        """Write ``ids`` into the device buffer and return ``[B]``.
+
+        Host values are staged through the double buffer and copied into the
+        stable device buffer (``copy_`` without pinning is synchronous, so
+        the alternating buffers guard against an in-flight transfer).
+        """
+        b = len(ids)
+        pin = self._pin[self._pin_idx]
+        self._pin_idx ^= 1
+        for i, v in enumerate(ids):
+            pin[i] = v
+        self.input_ids[:b].copy_(pin[:b])
+        return self.input_ids[:b]
+
+    def decode_mask(self, position_ids: Tensor, total_len: int) -> Tensor:
+        """Return the ``[B, 1, total_len]`` validity mask for this step.
+
+        Written into the pre-allocated buffer via ``torch.ge(out=)`` — no
+        new tensor is allocated.  ``position_ids`` is the current step's
+        ``[B]`` positions; ``total_len`` must not exceed ``max_seq_len``.
+        """
+        b = position_ids.size(0)
+        out = self.input_mask[:b, :, :total_len]
+        torch.ge(position_ids[:, None, None], self.arange[:total_len], out=out)
+        return out

astrai/inference/engine.py

@@ -8,7 +8,7 @@ from typing import Any, AsyncGenerator, Dict, Generator, List, Optional, Tuple,
 import torch
 import torch.nn as nn
 
-from astrai.inference.core.cache import KVCache
+from astrai.inference.core.cache import PagePool
 from astrai.inference.core.scheduler import InferenceScheduler
 from astrai.inference.core.task import STOP
 from astrai.tokenize import AutoTokenizer
@@ -64,33 +64,6 @@ class GenerateResult:
             return self.results.copy()
 
 
-class GenerationRequest:
-    """Request parameters for text generation."""
-
-    def __init__(
-        self,
-        messages: List[Dict[str, str]],
-        top_k: int = 50,
-        top_p: float = 1.0,
-        temperature: float = 1.0,
-        max_tokens: Optional[int] = None,
-        stream: bool = False,
-    ):
-        if not (isinstance(top_k, int) and top_k >= 0):
-            raise ValueError("top_k must be a non-negative integer")
-        if not (0.0 <= top_p <= 1.0):
-            raise ValueError("top_p must be a float between 0.0 and 1.0")
-        if not (isinstance(temperature, (int, float)) and temperature > 0):
-            raise ValueError("temperature must be a positive number")
-
-        self.messages = messages
-        self.top_k = top_k
-        self.top_p = top_p
-        self.temperature = temperature
-        self.max_tokens = max_tokens
-        self.stream = stream
-
-
 class InferenceEngine:
     """Unified inference engine backed by continuous-batching scheduler."""
 
@@ -100,9 +73,7 @@ class InferenceEngine:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 1,
         max_seq_len: Optional[int] = None,
-        max_prompt_len: int = 2048,
-        page_size: int = 128,
-        cache: Optional[KVCache] = None,
+        cache: Optional[PagePool] = None,
     ):
         self.model = model
         self.tokenizer = tokenizer
@@ -111,7 +82,6 @@ class InferenceEngine:
             tokenizer=self.tokenizer,
             max_batch_size=max_batch_size,
             max_seq_len=max_seq_len,
-            max_prompt_len=max_prompt_len,
             cache=cache,
         )
 
@@ -132,18 +102,29 @@ class InferenceEngine:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
     ) -> Union[Generator, str, List[str]]:
         is_batch = isinstance(prompt, list)
         prompts = prompt if is_batch else [prompt]
 
-        if stream:
-            return self._generate_streaming(
-                prompts, is_batch, max_tokens, temperature, top_p, top_k
-            )
-        else:
-            return self._generate_non_streaming(
-                prompts, is_batch, max_tokens, temperature, top_p, top_k
-            )
+        if max_tokens is not None and max_tokens <= 0:
+            if stream:
+                return iter(())
+            results = [""] * len(prompts)
+            return results if is_batch else results[0]
+
+        return self._generate(
+            prompts,
+            is_batch,
+            stream,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
+        )
 
     def generate_async(
         self,
@@ -152,85 +133,72 @@ class InferenceEngine:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
     ) -> AsyncGenerator[str, None]:
-        sync_gen = self._generate_streaming(
-            [prompt], False, max_tokens, temperature, top_p, top_k
+        sync_gen = self._generate(
+            [prompt],
+            False,
+            True,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
         )
 
         async def _agen():
             loop = asyncio.get_event_loop()
             while True:
-                token = await loop.run_in_executor(None, self._next_token, sync_gen)
-                if token is None:
+                try:
+                    token = await loop.run_in_executor(None, next, sync_gen)
+                except StopIteration:
                     break
                 yield token
 
         return _agen()
 
-    @staticmethod
-    def _next_token(gen: Generator) -> Optional[str]:
-        try:
-            return next(gen)
-        except StopIteration:
-            return None
-
-    def generate_with_request(
-        self, request: GenerationRequest
-    ) -> Union[Generator[str, None, None], str, List[str]]:
-        prompt = self.tokenizer.apply_chat_template(request.messages, tokenize=False)
-        return self.generate(
-            prompt=prompt,
-            stream=request.stream,
-            max_tokens=request.max_tokens,
-            temperature=request.temperature,
-            top_p=request.top_p,
-            top_k=request.top_k,
-        )
-
-    def _submit_tasks(
+    def _generate(
         self,
         prompts: List[str],
+        is_batch: bool,
+        stream: bool,
         max_tokens: Optional[int],
         temperature: float,
         top_p: float,
         top_k: int,
-    ) -> Tuple[GenerateResult, List[str]]:
+        frequency_penalty: float,
+        rep_window: int,
+    ) -> Union[Generator, str, List[str]]:
         n = len(prompts)
         result = GenerateResult(count=n)
-        task_ids = []
-        for i, p in enumerate(prompts):
-            cb = self._make_callback(result, i)
-            task_id = self.scheduler.add_task(
+        task_ids = [
+            self.scheduler.add_task(
                 prompt=p,
                 max_tokens=max_tokens,
                 temperature=temperature,
                 top_p=top_p,
                 top_k=top_k,
-                stream_callback=cb,
+                frequency_penalty=frequency_penalty,
+                rep_window=rep_window,
+                stream_callback=lambda token, idx=i: result.append(token, idx),
             )
-            task_ids.append(task_id)
-        return result, task_ids
+            for i, p in enumerate(prompts)
+        ]
 
-    @staticmethod
-    def _make_callback(result: GenerateResult, idx: int):
-        def cb(token):
-            result.append(token, idx)
+        if not stream:
+            try:
+                result.wait_completion()
+            except TimeoutError:
+                for tid in task_ids:
+                    self.scheduler.remove_task(tid)
+                raise
+            for tid in task_ids:
+                self.scheduler.remove_task(tid)
+            res = result.get_results()
+            return res if is_batch else res[0]
 
-        return cb
-
-    def _generate_streaming(
-        self,
-        prompts: List[str],
-        is_batch: bool,
-        max_tokens: Optional[int],
-        temperature: float,
-        top_p: float,
-        top_k: int,
-    ) -> Generator:
-        result, task_ids = self._submit_tasks(
-            prompts, max_tokens, temperature, top_p, top_k
-        )
-        n = len(prompts)
         remaining = n
         finished = [False] * n
 
@@ -254,32 +222,6 @@ class InferenceEngine:
 
         return gen()
 
-    def _generate_non_streaming(
-        self,
-        prompts: List[str],
-        is_batch: bool,
-        max_tokens: Optional[int],
-        temperature: float,
-        top_p: float,
-        top_k: int,
-    ) -> Union[str, List[str]]:
-        result, task_ids = self._submit_tasks(
-            prompts, max_tokens, temperature, top_p, top_k
-        )
-
-        try:
-            result.wait_completion()
-        except TimeoutError:
-            for tid in task_ids:
-                self.scheduler.remove_task(tid)
-            raise
-
-        for tid in task_ids:
-            self.scheduler.remove_task(tid)
-
-        res = result.get_results()
-        return res if is_batch else res[0]
-
     def get_stats(self) -> Dict[str, Any]:
         return self.scheduler.get_stats()
 

astrai/inference/sample.py

@@ -1,15 +1,15 @@
 """Composable sampling strategies for logit transformation.
 
 Implements the Strategy pattern: each sampling technique
-(temperature, top-k, top-p) is a pluggable strategy that
-can be composed into a pipeline.
+(temperature, top-k, top-p, frequency penalty) is a pluggable
+strategy that can be composed into a pipeline.
 
 All strategies accept both scalar and per-sample tensor
 parameters, so a single pipeline works for any batch size.
 """
 
 from abc import ABC, abstractmethod
-from typing import List, Union
+from typing import List, Optional, Union
 
 import torch
 from torch import Tensor
@@ -19,12 +19,23 @@ class BaseSamplingStrategy(ABC):
     """Abstract base for a logit transformation strategy."""
 
     @abstractmethod
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         """Applies the strategy to logits.
 
         Args:
             logits: Raw logits tensor (batch, vocab_size).
             filter_value: Value assigned to filtered-out positions.
+            input_ids: Previously generated token IDs ``[batch, seq_len]``,
+                padded with 0. Used by frequency penalty.
+            input_mask: Boolean mask ``[batch, seq_len]``, True for real
+                tokens, False for padding. Used to exclude padding from
+                penalty computation.
 
         Returns:
             Transformed logits tensor.
@@ -42,7 +53,13 @@ class TemperatureStrategy(BaseSamplingStrategy):
     def __init__(self, temperature: Union[float, Tensor] = 1.0):
         self.temperature = temperature
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         t = self.temperature
         if isinstance(t, Tensor):
             t = t.to(logits.device, non_blocking=True).view(-1, 1)
@@ -64,7 +81,13 @@ class TopKStrategy(BaseSamplingStrategy):
     def __init__(self, top_k: Union[int, Tensor] = 0):
         self.top_k = top_k
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         tk = self.top_k
         if isinstance(tk, Tensor):
             tk = tk.to(logits.device, non_blocking=True).long().clamp(min=0)
@@ -114,7 +137,13 @@ class TopPStrategy(BaseSamplingStrategy):
         logits[mask] = filter_value
         return logits
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         tp = self.top_p
         if isinstance(tp, Tensor):
             tp = tp.to(logits.device, non_blocking=True)
@@ -125,6 +154,84 @@ class TopPStrategy(BaseSamplingStrategy):
         return logits
 
 
+class FrequencyPenaltyStrategy(BaseSamplingStrategy):
+    """Penalizes tokens based on how many times they appeared in history.
+
+    Subtracts ``penalty * count(token)`` from each token's logit, where
+    ``count(token)`` is the number of occurrences in the generation history
+    (prompt + output). A penalty of ``0.0`` disables the strategy.
+
+    Unlike repetition penalty (which only checks *presence*), frequency
+    penalty scales linearly with occurrence count: the first use is
+    penalized once, the third use three times. This allows natural
+    repetition of common words while suppressing degenerate loops.
+
+    Reference: OpenAI API ``frequency_penalty`` parameter.
+
+    Args:
+        penalty: Scalar or ``[batch]`` tensor (0.0 disables, range -2.0~2.0).
+    """
+
+    def __init__(self, penalty: Union[float, Tensor] = 0.0):
+        self.penalty = penalty
+
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        if input_ids is None:
+            return logits
+
+        p = self.penalty
+        if isinstance(p, Tensor):
+            p = p.to(logits.device, non_blocking=True).view(-1, 1)
+            if (p == 0.0).all():
+                return logits
+        elif p == 0.0:
+            return logits
+
+        input_ids = input_ids.to(logits.device, non_blocking=True)
+
+        if input_mask is not None:
+            input_mask = input_mask.to(logits.device, non_blocking=True)
+            masked_ids = input_ids.clone()
+            masked_ids[~input_mask] = -1
+        else:
+            masked_ids = input_ids
+
+        batch_sz, seq_len = masked_ids.shape
+        vocab_size = logits.size(-1)
+
+        if isinstance(p, Tensor):
+            penalty_per_row = p.expand(batch_sz, 1)
+        else:
+            penalty_per_row = torch.full(
+                (batch_sz, 1), float(p), device=logits.device, dtype=logits.dtype
+            )
+
+        counts = torch.zeros(
+            batch_sz, vocab_size, device=logits.device, dtype=logits.dtype
+        )
+        valid_mask = masked_ids >= 0
+        if valid_mask.any():
+            valid_ids = masked_ids[valid_mask]
+            row_indices = (
+                torch.arange(batch_sz, device=logits.device)
+                .unsqueeze(1)
+                .expand_as(masked_ids)[valid_mask]
+            )
+            counts.index_put_(
+                (row_indices, valid_ids),
+                torch.ones_like(valid_ids, dtype=logits.dtype),
+                accumulate=True,
+            )
+
+        return logits - penalty_per_row * counts
+
+
 class SamplingPipeline(BaseSamplingStrategy):
     """Composes multiple sampling strategies into a single transformation.
 
@@ -145,25 +252,76 @@ class SamplingPipeline(BaseSamplingStrategy):
     def __init__(self, strategies: List[BaseSamplingStrategy]):
         self.strategies = strategies
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         for strategy in self.strategies:
-            logits = strategy.apply(logits, filter_value)
+            logits = strategy.apply(logits, filter_value, input_ids, input_mask)
         return logits
 
-    @torch.no_grad()
-    def sample(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    @staticmethod
+    def _is_greedy(temperature: Union[float, Tensor]) -> bool:
+        if isinstance(temperature, Tensor):
+            return bool((temperature == 0).all())
+        return temperature == 0
+
+    @torch.inference_mode()
+    def sample(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+        return_logprobs: bool = False,
+    ):
         """Apply strategies then sample (softmax + multinomial).
 
+        Short-circuits to ``argmax`` when temperature is exactly 0
+        (deterministic / greedy decode).
+
         Args:
             logits: Raw logits ``[batch, vocab_size]``.
+            input_ids: Previously generated token IDs ``[batch, seq_len]``.
+            input_mask: Boolean mask for ``input_ids`` padding.
+            return_logprobs: If ``True``, return ``(tokens, logprobs)``
+                where ``logprobs[i]`` is the log-probability of
+                ``tokens[i]`` under the (post-strategy) sampling
+                distribution.
 
         Returns:
-            Sampled token IDs ``[batch]``.
+            Sampled token IDs ``[batch]``, or — when ``return_logprobs``
+            is ``True`` — a ``(token_ids, chosen_logprobs)`` tuple.
         """
-        return torch.multinomial(
-            torch.softmax(self.apply(logits, filter_value), dim=-1),
-            num_samples=1,
+        if self._is_greedy_pipeline():
+            tokens = logits.argmax(dim=-1)
+            if not return_logprobs:
+                return tokens
+            log_probs = torch.log_softmax(logits.float(), dim=-1)
+            chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
+            return tokens, chosen
+
+        transformed = self.apply(logits, filter_value, input_ids, input_mask)
+        tokens = torch.multinomial(
+            torch.softmax(transformed, dim=-1), num_samples=1
         ).squeeze(-1)
+        if not return_logprobs:
+            return tokens
+        log_probs = torch.log_softmax(transformed.float(), dim=-1)
+        chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
+        return tokens, chosen
+
+    def _is_greedy_pipeline(self) -> bool:
+        """True if the first strategy is greedy temperature (temp=0)."""
+        if not self.strategies:
+            return False
+        first = self.strategies[0]
+        return isinstance(first, TemperatureStrategy) and self._is_greedy(
+            first.temperature
+        )
 
 
 @torch.inference_mode()
@@ -172,22 +330,71 @@ def sample(
     temperature: Union[float, Tensor] = 1.0,
     top_k: Union[int, Tensor] = 0,
     top_p: Union[float, Tensor] = 1.0,
+    frequency_penalty: Union[float, Tensor] = 0.0,
+    input_ids: Optional[Tensor] = None,
+    input_mask: Optional[Tensor] = None,
     filter_value: float = -float("inf"),
-) -> Tensor:
+    return_logprobs: bool = False,
+):
     """Apply sampling strategies then sample (softmax + multinomial).
 
-    Shortcut for ``SamplingPipeline(...).sample(logits)``.
+    Shortcut for ``SamplingPipeline(...).sample(logits, return_logprobs=)``.
+
+    When **temperature** is exactly 0 (scalar or single-element tensor)
+    the function short-circuits to ``argmax`` for deterministic decode.
+
+    When **frequency_penalty** is 0 (the common decode case), the entire
+    frequency penalty computation — including the O(batch * vocab) count
+    tensor allocation — is skipped.
 
     Args:
         logits: Raw logits ``[batch, vocab_size]``.
+        frequency_penalty: Penalty per occurrence for repeated tokens
+            (0.0 disables, range -2.0~2.0).
+        input_ids: Previously generated token IDs ``[batch, seq_len]``.
+        input_mask: Boolean mask for ``input_ids`` padding.
+        return_logprobs: If ``True``, also return the log-probability
+            of each sampled token under the (post-strategy) sampling
+            distribution — useful for RL rollout (PPO/GRPO importance
+            ratios).
 
     Returns:
-        Sampled token IDs ``[batch]``.
+        Sampled token IDs ``[batch]``, or — when ``return_logprobs`` is
+        ``True`` — a ``(token_ids, chosen_logprobs)`` tuple where
+        ``chosen_logprobs`` has shape ``[batch]``.
     """
-    return SamplingPipeline(
-        [
-            TemperatureStrategy(temperature),
-            TopKStrategy(top_k),
-            TopPStrategy(top_p),
-        ]
-    ).sample(logits, filter_value)
+    greedy = (
+        bool((temperature == 0).all())
+        if isinstance(temperature, Tensor)
+        else temperature == 0
+    )
+
+    if greedy:
+        tokens = logits.argmax(dim=-1)
+        if not return_logprobs:
+            return tokens
+        log_probs = torch.log_softmax(logits.float(), dim=-1)
+        chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
+        return tokens, chosen
+
+    has_freq = (
+        (isinstance(frequency_penalty, Tensor) and (frequency_penalty != 0).any())
+        if isinstance(frequency_penalty, Tensor)
+        else frequency_penalty != 0
+    )
+
+    strategies: List[BaseSamplingStrategy] = [
+        TemperatureStrategy(temperature),
+        TopKStrategy(top_k),
+        TopPStrategy(top_p),
+    ]
+    if has_freq:
+        strategies.append(FrequencyPenaltyStrategy(frequency_penalty))
+
+    return SamplingPipeline(strategies).sample(
+        logits,
+        filter_value=filter_value,
+        input_ids=input_ids,
+        input_mask=input_mask,
+        return_logprobs=return_logprobs,
+    )

astrai/model/__init__.py

@@ -9,7 +9,7 @@ from astrai.model.components.lora import (
     merge_lora,
     save_lora,
 )
-from astrai.model.components.mlp import MLP
+from astrai.model.components.mlp import MLP, DeepSeekMoE
 from astrai.model.components.norm import RMSNorm
 from astrai.model.encoder import EmbeddingEncoder
 from astrai.model.transformer import AutoRegressiveLM
@@ -19,6 +19,7 @@ __all__ = [
     "Linear",
     "RMSNorm",
     "MLP",
+    "DeepSeekMoE",
     "GQA",
     "DecoderBlock",
     # Models

astrai/model/automodel.py

@@ -40,11 +40,12 @@ def _disable_random_init(enable: bool = True):
             setattr(nn.init, n, fn)
 
 
-class AutoModel(BaseFactory["AutoModel"], nn.Module):
-    """
-    Autoregressive language model base class.
-    Provides model loading/saving, registration, and generation.
-    """
+class ModelFactory(BaseFactory[nn.Module]):
+    """Pure factory for model dispatch, separated from nn.Module state."""
+
+
+class AutoModel(nn.Module):
+    """Model base class with loading/saving and generation."""
 
     def __init__(self, config: BaseModelConfig):
         super().__init__()
@@ -68,7 +69,7 @@ class AutoModel(BaseFactory["AutoModel"], nn.Module):
         config = ConfigFactory.load(raw)
         model_type = config.model_type or "autoregressive_lm"
 
-        actual_cls = AutoModel.get_component_class(model_type)
+        actual_cls = ModelFactory.get_component_class(model_type)
 
         with _disable_random_init(enable=disable_random_init):
             model = actual_cls(config)

astrai/model/components/__init__.py

@@ -1,12 +1,12 @@
-from astrai.model.components.attention import GQA, MLA, repeat_kv
+from astrai.extension.rotary_backend import apply_rotary_emb
+from astrai.model.components.attention import GQA, MLA
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.model.components.embedding import Embedding
 from astrai.model.components.linear import Linear
-from astrai.model.components.mlp import MLP
+from astrai.model.components.mlp import MLP, DeepSeekMoE
 from astrai.model.components.norm import RMSNorm
 from astrai.model.components.rope import (
     RotaryEmbedding,
-    apply_rotary_emb,
     get_rotary_emb,
 )
 
@@ -14,6 +14,7 @@ __all__ = [
     "Linear",
     "RMSNorm",
     "MLP",
+    "DeepSeekMoE",
     "Embedding",
     "GQA",
     "MLA",
@@ -21,5 +22,4 @@ __all__ = [
     "RotaryEmbedding",
     "apply_rotary_emb",
     "get_rotary_emb",
-    "repeat_kv",
 ]

astrai/model/components/attention.py

@@ -5,22 +5,12 @@ import torch.nn as nn
 import torch.nn.functional as F
 from torch import Tensor
 
+from astrai.extension import attention
+from astrai.extension.rotary_backend import apply_rotary_emb
 from astrai.factory import BaseFactory
-from astrai.inference.core.cache import CacheView
+from astrai.inference.core.cache import KVCache
 from astrai.model.components.linear import Linear
 from astrai.model.components.norm import RMSNorm
-from astrai.model.components.rope import apply_rotary_emb
-
-
-def repeat_kv(x: Tensor, n_rep: int) -> Tensor:
-    bs, slen, n_heads, head_dim = x.shape
-    if n_rep == 1:
-        return x
-    return (
-        x[:, :, :, None, :]
-        .expand(bs, slen, n_heads, n_rep, head_dim)
-        .reshape(bs, slen, n_heads * n_rep, head_dim)
-    )
 
 
 class AttnFactory(BaseFactory[nn.Module]):
@@ -75,10 +65,9 @@ class GQA(nn.Module):
         x: Tensor,
         rotary_emb: Tensor,
         attn_mask: Tensor = None,
-        paged_cache: Optional[CacheView] = None,
+        kv_cache: Optional[KVCache] = None,
+        is_causal: bool = False,
     ) -> Tensor:
-        is_causal = attn_mask is None
-
         q = self._split_heads(self.q_proj(x), self.n_heads)
         k = self._split_heads(self.k_proj(x), self.n_kv_heads)
         v = self._split_heads(self.v_proj(x), self.n_kv_heads)
@@ -87,19 +76,7 @@ class GQA(nn.Module):
         if self.use_qk_norm:
             q, k = self.q_norm(q), self.k_norm(k)
 
-        if paged_cache is not None:
-            paged_cache.write(self.layer_id, k, v)
-            k, v = paged_cache.gather(self.layer_id)
-
-        k, v = repeat_kv(k, self.n_rep), repeat_kv(v, self.n_rep)
-
-        q, k, v = q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3)
-        sdqa_out = (
-            F.scaled_dot_product_attention(q, k, v, attn_mask, is_causal=is_causal)
-            .permute(0, 2, 1, 3)
-            .contiguous()
-            .flatten(2)
-        )
+        sdqa_out = attention(q, k, v, kv_cache, self.layer_id, attn_mask, is_causal)
 
         if self.use_gated_attention:
             sdqa_out = sdqa_out * F.sigmoid(self.gate(x))
@@ -162,10 +139,10 @@ class MLA(nn.Module):
         x: Tensor,
         rotary_emb: Tensor,
         attn_mask: Tensor = None,
-        paged_cache: Optional[CacheView] = None,
+        kv_cache: Optional[KVCache] = None,
+        is_causal: bool = False,
     ) -> Tensor:
         bsz, seq_len, _ = x.size()
-        is_causal = attn_mask is None
 
         q = self.q_proj(x)
         q = q.view(bsz, seq_len, self.n_heads, self.head_dim)
@@ -194,18 +171,7 @@ class MLA(nn.Module):
             q = self.q_norm(q)
             k = self.k_norm(k)
 
-        if paged_cache is not None:
-            paged_cache.write(self.layer_id, k, v)
-            k, v = paged_cache.gather(self.layer_id)
-
-        q = q.permute(0, 2, 1, 3)
-        k = k.permute(0, 2, 1, 3)
-        v = v.permute(0, 2, 1, 3)
-
-        attn_out = F.scaled_dot_product_attention(
-            q, k, v, attn_mask, is_causal=is_causal
-        )
-        attn_out = attn_out.permute(0, 2, 1, 3).contiguous().flatten(2)
+        attn_out = attention(q, k, v, kv_cache, self.layer_id, attn_mask, is_causal)
 
         if self.use_gated_attention:
             attn_out = attn_out * F.sigmoid(self.gate(x))

astrai/model/components/decoder_block.py

@@ -1,39 +1,74 @@
 from dataclasses import asdict
-from typing import Optional
+from typing import Optional, TypedDict
 
 import torch.nn as nn
 from torch import Tensor
 
-from astrai.inference.core.cache import CacheView
+from astrai.inference.core.cache import KVCache
 from astrai.model.components.attention import AttnFactory
-from astrai.model.components.mlp import FFNFactory
+from astrai.model.components.mlp import FFNFactory, RouterStats
 from astrai.model.components.norm import RMSNorm
 
 
+class DecoderOutput(TypedDict):
+    hidden_states: Tensor
+    aux_loss: Optional[Tensor]
+    router_stats: Optional[RouterStats]
+
+
 class DecoderBlock(nn.Module):
     def __init__(self, config, layer_id: int):
         super().__init__()
         cfg = asdict(config)
-        cfg["down_init_std"] = 0.02 / (2 * config.n_layers) ** 0.5
+        cfg.update(
+            dim=config.hidden_size,
+            dim_ffn=config.intermediate_size,
+            n_layers=config.num_hidden_layers,
+            n_heads=config.num_attention_heads,
+            n_kv_heads=config.num_key_value_heads,
+            norm_eps=config.rms_norm_eps,
+            down_init_std=0.02 / (2 * config.num_hidden_layers) ** 0.5,
+        )
         self.attention = AttnFactory.create(config.attn_type, **cfg, layer_id=layer_id)
-        self.input_norm = RMSNorm(config.dim, config.norm_eps)
-        self.post_attention_norm = RMSNorm(config.dim, config.norm_eps)
-        self.mlp = FFNFactory.create(config.ffn_type, **cfg)
+        self.input_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        ffn_type = self._resolve_ffn_type(config, layer_id)
+        self.mlp = FFNFactory.create(ffn_type, **cfg)
+
+    @staticmethod
+    def _resolve_ffn_type(config, layer_id: int) -> str:
+        if config.ffn_type != "moe":
+            return config.ffn_type
+        mlp_only = config.mlp_only_layers or []
+        if layer_id in mlp_only:
+            return "mlp"
+        if config.decoder_sparse_step > 1:
+            if (layer_id + 1) % config.decoder_sparse_step != 0:
+                return "mlp"
+        return "moe"
 
     def forward(
         self,
         x: Tensor,
         rotary_emb: Tensor,
         attention_mask: Optional[Tensor] = None,
-        paged_cache: Optional[CacheView] = None,
-    ) -> Tensor:
+        kv_cache: Optional[KVCache] = None,
+        is_causal: bool = False,
+    ) -> DecoderOutput:
         attn_output = self.attention(
             self.input_norm(x),
             rotary_emb,
             attention_mask,
-            paged_cache,
+            kv_cache,
+            is_causal,
         )
         x = attn_output + x
-        x = self.mlp(self.post_attention_norm(x)) + x
+        normalized = self.post_attention_norm(x)
+        mlp_output = self.mlp(normalized)
+        x = mlp_output["hidden_states"] + x
 
-        return x
+        return {
+            "hidden_states": x,
+            "aux_loss": mlp_output["aux_loss"],
+            "router_stats": mlp_output.get("router_stats"),
+        }

astrai/model/components/lora.py

@@ -1,11 +1,12 @@
 import logging
-from dataclasses import asdict, dataclass
+from dataclasses import asdict
 from pathlib import Path
 from typing import Optional, Set
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from pydantic.dataclasses import dataclass
 
 from astrai.model.components.linear import Linear
 from astrai.serialization import (
@@ -39,8 +40,12 @@ class LoRALinear(nn.Module):
 
         self.r = r
         self.scaling = alpha / r
-        self.lora_A = nn.Parameter(torch.randn(r, self.weight.shape[1]) / r)
-        self.lora_B = nn.Parameter(torch.zeros(self.weight.shape[0], r))
+        device = self.weight.device
+        dtype = self.weight.dtype
+        lora_a = torch.randn(r, self.weight.shape[1], device=device, dtype=dtype) / r
+        lora_b = torch.zeros(self.weight.shape[0], r, device=device, dtype=dtype)
+        self.lora_A = nn.Parameter(lora_a)
+        self.lora_B = nn.Parameter(lora_b)
         self._merged = False
 
     def forward(self, x):

astrai/model/components/mlp.py

@@ -1,3 +1,5 @@
+from typing import Optional, TypedDict
+
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -11,6 +13,28 @@ class FFNFactory(BaseFactory[nn.Module]):
     pass
 
 
+class RouterStats(TypedDict):
+    """Per-layer MoE routing statistics for training diagnostics.
+
+    Both tensors are detached monitoring data produced during forward.
+    """
+
+    probs: Tensor
+    topk_indices: Tensor
+
+
+class FFNOutput(TypedDict):
+    hidden_states: Tensor
+    aux_loss: Optional[Tensor]
+    router_stats: Optional[RouterStats]
+
+
+class RoutedOutput(TypedDict):
+    hidden_states: Tensor
+    aux_loss: Optional[Tensor]
+    router_stats: Optional[RouterStats]
+
+
 @FFNFactory.register("mlp")
 class MLP(nn.Module):
     def __init__(self, dim: int, dim_ffn: int, down_init_std: float = 0.02):
@@ -19,10 +43,10 @@ class MLP(nn.Module):
         self.gate = Linear(dim, dim_ffn)
         self.down = Linear(dim_ffn, dim, init_std=down_init_std)
 
-    def forward(self, x: Tensor) -> Tensor:
+    def forward(self, x: Tensor) -> FFNOutput:
         gated = self.up(x) * F.silu(self.gate(x))
         out = self.down(gated)
-        return out
+        return {"hidden_states": out, "aux_loss": None, "router_stats": None}
 
 
 @FFNFactory.register("moe")
@@ -36,6 +60,9 @@ class DeepSeekMoE(nn.Module):
         n_activated_experts: int = 2,
         topk_method: str = "greedy",
         n_layers: int = 1,
+        moe_intermediate_size: Optional[int] = None,
+        shared_expert_intermediate_size: Optional[int] = None,
+        norm_topk_prob: bool = True,
     ):
         super().__init__()
         self.dim = dim
@@ -43,6 +70,16 @@ class DeepSeekMoE(nn.Module):
         self.n_shared_experts = n_shared_experts
         self.n_activated_experts = n_activated_experts
         self.topk_method = topk_method
+        self.norm_topk_prob = norm_topk_prob
+
+        expert_dim_ffn = (
+            moe_intermediate_size if moe_intermediate_size is not None else dim_ffn
+        )
+        shared_dim_ffn = (
+            shared_expert_intermediate_size
+            if shared_expert_intermediate_size is not None
+            else dim_ffn
+        )
 
         self.router = Linear(dim, n_routed_experts, bias=False)
         moe_scale = 1 / max(n_shared_experts, 1) + 1 / n_activated_experts
@@ -50,51 +87,91 @@ class DeepSeekMoE(nn.Module):
 
         self.shared_experts = nn.ModuleList(
             [
-                MLP(dim, dim_ffn, down_init_std=down_init_std)
+                MLP(dim, shared_dim_ffn, down_init_std=down_init_std)
                 for _ in range(n_shared_experts)
             ]
         )
         self.routed_experts = nn.ModuleList(
             [
-                MLP(dim, dim_ffn, down_init_std=down_init_std)
+                MLP(dim, expert_dim_ffn, down_init_std=down_init_std)
                 for _ in range(n_routed_experts)
             ]
         )
 
-    def forward(self, x: Tensor) -> Tensor:
+    def forward(self, x: Tensor) -> FFNOutput:
+        include_aux_loss = self.training and torch.is_grad_enabled()
         bsz, seq_len, dim = x.shape
         x_flat = x.view(-1, dim)
 
         shared_out = self._shared_forward(x_flat)
-        routed_out = self._routed_forward(x_flat)
+        routed_output = self._routed_forward(x_flat, include_aux_loss)
 
-        out = (shared_out + routed_out).view(bsz, seq_len, dim)
-        return out
+        out = (shared_out + routed_output["hidden_states"]).view(bsz, seq_len, dim)
+        return {
+            "hidden_states": out,
+            "aux_loss": routed_output["aux_loss"],
+            "router_stats": routed_output["router_stats"],
+        }
 
     def _shared_forward(self, x: Tensor) -> Tensor:
         if self.n_shared_experts == 0:
             return torch.zeros_like(x)
-        return sum(e(x) for e in self.shared_experts) / self.n_shared_experts
+        return (
+            sum(e(x)["hidden_states"] for e in self.shared_experts)
+            / self.n_shared_experts
+        )
 
-    def _routed_forward(self, x: Tensor) -> Tensor:
+    def _routed_forward(self, x: Tensor, include_aux_loss: bool) -> RoutedOutput:
         N, D = x.shape
         K = self.n_activated_experts
+        E = self.n_routed_experts
 
         router_logits = self.router(x)
         router_probs = torch.softmax(router_logits.float(), dim=-1).to(x.dtype)
 
-        topk_weights, topk_indices = torch.topk(router_probs, K, dim=-1)
-        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+        topk_weights, topk_indices = torch.topk(router_probs, K, dim=-1, sorted=False)
+        if self.norm_topk_prob:
+            topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+
+        aux_loss = None
+        router_stats = None
+        if include_aux_loss:
+            expert_load = F.one_hot(topk_indices, num_classes=E).float()
+            expert_load = expert_load.mean(dim=(0, 1))
+            router_prob = router_probs.float().mean(dim=0)
+            aux_loss = E * (expert_load * router_prob).sum()
+            router_stats = {
+                "probs": router_probs.detach(),
+                "topk_indices": topk_indices,
+            }
+
+        # Grouped dispatch: sort (token, slot) pairs by expert so each expert
+        # consumes one contiguous slice instead of a per-expert mask scan.
+        flat_experts = topk_indices.reshape(-1)
+        sorted_experts, order = torch.sort(flat_experts)
+        flat_tokens = x.repeat_interleave(K, dim=0)[order]
+        flat_weights = topk_weights.reshape(-1, 1)[order]
+        boundaries = torch.cumsum(
+            torch.bincount(sorted_experts, minlength=E), dim=0
+        ).tolist()
 
         output = torch.zeros(N, D, device=x.device, dtype=x.dtype)
-        for expert_idx in range(self.n_routed_experts):
-            expert_mask = topk_indices == expert_idx
-            token_idx, k_idx = expert_mask.nonzero(as_tuple=True)
-            if token_idx.numel() == 0:
+        start = 0
+        for expert_idx, end in enumerate(boundaries):
+            if end == start:
                 continue
-            expert_input = x[token_idx]
-            expert_output = self.routed_experts[expert_idx](expert_input)
-            weights = topk_weights[token_idx, k_idx].unsqueeze(-1)
-            output.index_add_(0, token_idx, expert_output * weights)
+            expert_output = self.routed_experts[expert_idx](flat_tokens[start:end])[
+                "hidden_states"
+            ]
+            output.index_add_(
+                0,
+                order[start:end] // K,
+                expert_output * flat_weights[start:end],
+            )
+            start = end
 
-        return output
+        return {
+            "hidden_states": output,
+            "aux_loss": aux_loss,
+            "router_stats": router_stats,
+        }

astrai/model/components/rope.py

@@ -11,28 +11,23 @@ def get_rotary_emb(
     base: float = 10000,
     device: Optional[torch.device] = None,
 ) -> Tensor:
+    """Precompute cos/sin tables for rotary embedding.
+
+    Returns:
+        [max_len, dim/2, 2] (f32) — [cos, sin] pairs.
+    """
     theta = base ** (-torch.arange(0, dim, 2, dtype=torch.float64, device=device) / dim)
     t = torch.arange(0, max_len, dtype=torch.float64, device=device)
     freqs = torch.outer(t, theta).float()
     cos = torch.cos(freqs)
     sin = torch.sin(freqs)
-    return torch.complex(cos, sin)
+    return torch.stack([cos, sin], dim=-1)
 
 
 def ntk_base(base: float, dim: int, factor: float) -> float:
     return base * (factor ** (dim / (dim - 2)))
 
 
-def apply_rotary_emb(x: torch.Tensor, freqs_cis: Tensor) -> Tensor:
-    dtype = x.dtype
-    x_ = x.float().reshape(*x.shape[:-1], -1, 2)
-    x_complex = torch.view_as_complex(x_)
-    freqs_cis = freqs_cis.unsqueeze(2)
-    x_rotated = x_complex * freqs_cis
-    x_out = torch.view_as_real(x_rotated).flatten(-2)
-    return x_out.to(dtype)
-
-
 class RotaryEmbedding(nn.Module):
     def __init__(
         self,
@@ -56,16 +51,23 @@ class RotaryEmbedding(nn.Module):
         self._set_rotary_buffer(self.max_len)
 
     def _set_rotary_buffer(self, max_len: int):
-        rotary_emb = get_rotary_emb(self.dim, max_len, self.base)
-        freqs_cis = torch.view_as_real(rotary_emb)
+        freqs_cis = get_rotary_emb(self.dim, max_len, self.base)
         self.register_buffer("freqs_cis", freqs_cis, persistent=False)
 
     def forward(self, x: Tensor, position_ids: Optional[Tensor] = None) -> Tensor:
+        """Lookup cos/sin for the given positions.
+
+        Args:
+            x: [batch, seq_len, ...] — only batch and seq_len are used.
+            position_ids: [batch, seq_len] optional position indices.
+
+        Returns:
+            [batch, seq_len, dim/2, 2] (f32) — [cos, sin] pairs.
+        """
         if position_ids is None:
             position_ids = (
                 torch.arange(x.size(1), device=x.device)
                 .unsqueeze(0)
                 .expand(x.size(0), -1)
             )
-        position_freq_cis = self.freqs_cis[position_ids].float()
-        return torch.view_as_complex(position_freq_cis)
+        return self.freqs_cis[position_ids].float()

astrai/model/encoder.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 from torch import Tensor
 
 from astrai.config.model_config import EncoderConfig
-from astrai.model.automodel import AutoModel
+from astrai.model.automodel import AutoModel, ModelFactory
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.model.components.embedding import Embedding
 from astrai.model.components.norm import RMSNorm
@@ -13,25 +13,33 @@ from astrai.model.components.rope import RotaryEmbedding
 from astrai.model.transformer import process_attention_mask
 
 
-@AutoModel.register("embedding")
+@ModelFactory.register("embedding")
 class EmbeddingEncoder(AutoModel):
     def __init__(self, config: EncoderConfig):
         super().__init__(config)
         self.config = config
-        rope_dim = config.dim // config.n_heads
+        rope_dim = config.hidden_size // config.num_attention_heads
         rope_base = config.rope_theta if config.rope_theta is not None else 10000
         self.rotary_embedding = RotaryEmbedding(
-            rope_dim, config.max_len, rope_base, rope_scaling=config.rope_scaling
+            rope_dim,
+            config.max_position_embeddings,
+            rope_base,
+            rope_scaling=config.rope_scaling,
         )
         self.embed_tokens = Embedding(
-            config.vocab_size, config.dim, neftune_alpha=config.neftune_alpha
+            config.vocab_size,
+            config.hidden_size,
+            neftune_alpha=config.neftune_alpha,
         )
 
         self.layers = nn.ModuleList(
-            [DecoderBlock(config, layer_id) for layer_id in range(config.n_layers)]
+            [
+                DecoderBlock(config, layer_id)
+                for layer_id in range(config.num_hidden_layers)
+            ]
         )
 
-        self.norm = RMSNorm(config.dim, config.norm_eps)
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
 
         self.pooling_type = config.pooling_type or "mean"
         self.normalize_embeddings = config.normalize_embeddings or False
@@ -59,10 +67,10 @@ class EmbeddingEncoder(AutoModel):
         x = self.embed_tokens(input_ids)
 
         rotary_emb = self.rotary_embedding(x, position_ids)
-        attn_mask = process_attention_mask(x, position_ids, input_mask, is_causal=False)
+        attn_mask = process_attention_mask(input_mask)
 
         for layer in self.layers:
-            x = layer(x, rotary_emb, attn_mask, paged_cache=None)
+            x = layer(x, rotary_emb, attn_mask)["hidden_states"]
 
         hidden_states = self.norm(x)
 

astrai/model/transformer.py

@@ -5,8 +5,8 @@ import torch.nn as nn
 from torch import Tensor
 
 from astrai.config.model_config import AutoRegressiveLMConfig
-from astrai.inference.core.cache import CacheView
-from astrai.model.automodel import AutoModel
+from astrai.inference.core.cache import KVCache
+from astrai.model.automodel import AutoModel, ModelFactory
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.model.components.embedding import Embedding
 from astrai.model.components.linear import Linear
@@ -15,35 +15,18 @@ from astrai.model.components.rope import RotaryEmbedding
 
 
 def process_attention_mask(
-    input_tensor: Tensor,
-    position_ids: Optional[Tensor],
-    input_mask: Optional[Tensor] = None,
-    is_causal: bool = False,
+    input_mask: Optional[Tensor],
 ) -> Optional[Tensor]:
-    if position_ids is None:
-        return None
-    if input_mask is not None and input_mask.dim() > 2:
-        return input_mask
-
-    device = input_tensor.device
-    B = input_tensor.size(0)
-    T = position_ids.max().item() + 1
-
     if input_mask is None:
-        if position_ids.min().item() == 0 and is_causal:
-            return None
-        attend = torch.ones(B, 1, T, dtype=torch.bool, device=device)
-    else:
-        attend = input_mask[:, :T].to(device=device, dtype=torch.bool).unsqueeze(1)
-
-    if is_causal:
-        causal = position_ids.unsqueeze(-1) >= torch.arange(T, device=device)
-        attend = attend & causal
-
-    return attend.unsqueeze(1)
+        return None
+    if input_mask.dim() == 2:
+        return input_mask[:, None, None, :]
+    if input_mask.dim() == 3:
+        return input_mask[:, None, :, :]
+    return input_mask
 
 
-@AutoModel.register("autoregressive_lm")
+@ModelFactory.register("autoregressive_lm")
 class AutoRegressiveLM(AutoModel):
     """Autoregressive language model with paged KV cache."""
 
@@ -53,24 +36,32 @@ class AutoRegressiveLM(AutoModel):
         rope_dim = (
             config.qk_rope_head_dim
             if config.attn_type == "mla"
-            else config.dim // config.n_heads
+            else config.hidden_size // config.num_attention_heads
         )
         rope_base = config.rope_theta if config.rope_theta is not None else 10000
         self.rotary_embedding = RotaryEmbedding(
-            rope_dim, config.max_len, rope_base, rope_scaling=config.rope_scaling
+            rope_dim,
+            config.max_position_embeddings,
+            rope_base,
+            rope_scaling=config.rope_scaling,
         )
         self.embed_tokens = Embedding(
-            config.vocab_size, config.dim, neftune_alpha=config.neftune_alpha
+            config.vocab_size,
+            config.hidden_size,
+            neftune_alpha=config.neftune_alpha,
         )
 
         self.layers = nn.ModuleList(
-            [DecoderBlock(config, layer_id) for layer_id in range(config.n_layers)]
+            [
+                DecoderBlock(config, layer_id)
+                for layer_id in range(config.num_hidden_layers)
+            ]
         )
 
-        self.norm = RMSNorm(config.dim, config.norm_eps)
-        self.lm_head = Linear(config.dim, config.vocab_size)
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.lm_head = Linear(config.hidden_size, config.vocab_size)
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             self.lm_head.weight = self.embed_tokens.weight
 
         self.apply(self._init_weights)
@@ -85,7 +76,7 @@ class AutoRegressiveLM(AutoModel):
 
         state_dict = dict(state_dict)
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             # same tensor for embed and lm_head
             if embed_key in state_dict:
                 state_dict[lm_head_key] = state_dict[embed_key]
@@ -101,7 +92,7 @@ class AutoRegressiveLM(AutoModel):
             destination=destination, prefix=prefix, keep_vars=keep_vars
         )
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             lm_head_key = prefix + "lm_head.weight"
             if lm_head_key in state_dict:
                 del state_dict[lm_head_key]
@@ -112,19 +103,37 @@ class AutoRegressiveLM(AutoModel):
         self,
         input_ids: Tensor,
         input_mask: Optional[Tensor] = None,
-        paged_cache: Optional[CacheView] = None,
+        kv_cache: Optional[KVCache] = None,
         position_ids: Optional[Tensor] = None,
     ) -> Dict[str, Tensor]:
         assert input_ids.ndim == 2
 
         x = self.embed_tokens(input_ids)
         rotary_emb = self.rotary_embedding(x, position_ids)
-        attn_mask = process_attention_mask(x, position_ids, input_mask, is_causal=True)
+        attn_mask = process_attention_mask(input_mask)
+        use_sdpa_causal_mask = attn_mask is None
 
+        aux_losses = []
+        router_stats_list = []
         for layer in self.layers:
-            x = layer(x, rotary_emb, attn_mask, paged_cache)
+            layer_output = layer(
+                x,
+                rotary_emb,
+                attn_mask,
+                kv_cache,
+                use_sdpa_causal_mask,
+            )
+            x = layer_output["hidden_states"]
+            stats = layer_output.get("router_stats")
+            if stats is not None:
+                aux_losses.append(layer_output["aux_loss"])
+                router_stats_list.append(stats)
 
         hidden_states = self.norm(x)
         logits = self.lm_head(hidden_states)
 
-        return {"logits": logits, "hidden_states": hidden_states}
+        output = {"logits": logits, "hidden_states": hidden_states}
+        if aux_losses:
+            output["aux_loss"] = torch.stack(aux_losses).mean()
+            output["router_stats"] = router_stats_list
+        return output

astrai/optim/__init__.py

@@ -0,0 +1,38 @@
+"""Optimizer implementations and factory registration."""
+
+from astrai.optim.composite import (
+    OptimizerFactory,
+    composite_state_dict,
+    composite_step,
+    composite_zero_grad,
+    refresh_param_groups,
+)
+from astrai.optim.mano_adamw import Mano, ManoAdamW
+from astrai.optim.muon_adamw import MuonAdamW
+from astrai.optim.nora_nadamw import (
+    NAdamW,
+    Nora,
+    NoraNAdamW,
+    OptimizerParameterGroups,
+    nora_direction,
+    nora_lr_scale,
+    partition_optimizer_parameters,
+)
+
+__all__ = [
+    "Mano",
+    "ManoAdamW",
+    "MuonAdamW",
+    "NAdamW",
+    "Nora",
+    "NoraNAdamW",
+    "OptimizerFactory",
+    "OptimizerParameterGroups",
+    "composite_state_dict",
+    "composite_step",
+    "composite_zero_grad",
+    "nora_direction",
+    "nora_lr_scale",
+    "partition_optimizer_parameters",
+    "refresh_param_groups",
+]

astrai/optim/composite.py

@@ -0,0 +1,71 @@
+"""Shared infrastructure for the optim package.
+
+This module hosts two things:
+
+* ``OptimizerFactory`` — the registry for built-in optimizers. Defining it
+  here (rather than in ``__init__.py``) lets each optimizer module import it
+  and register itself with a decorator, avoiding circular imports.
+* Composite-optimizer helpers — ``step``/``zero_grad``/``state_dict``/
+  ``param_groups`` delegation shared by every optimizer that routes different
+  parameter groups through distinct sub-optimizers.
+"""
+
+from typing import Any
+
+import torch
+from torch.optim import Optimizer
+
+from astrai.factory import BaseFactory
+
+
+class OptimizerFactory(BaseFactory[Optimizer]):
+    """Factory for built-in training optimizers."""
+
+
+def composite_step(
+    sub_optimizers: list[Optimizer],
+    closure=None,
+) -> torch.Tensor | None:
+    """Run ``step`` on every sub-optimizer, invoking the closure once.
+
+    The closure (if given) is executed inside ``torch.enable_grad`` exactly
+    once before any sub-optimizer steps, matching the contract of a single
+    ``Optimizer.step``. Sub-optimizers receive ``None`` so they do not
+    re-execute it.
+    """
+    loss = None
+    if closure is not None:
+        with torch.enable_grad():
+            loss = closure()
+    for sub in sub_optimizers:
+        sub.step()
+    return loss
+
+
+def composite_zero_grad(
+    sub_optimizers: list[Optimizer],
+    set_to_none: bool = True,
+) -> None:
+    for sub in sub_optimizers:
+        sub.zero_grad(set_to_none=set_to_none)
+
+
+def composite_state_dict(
+    named_sub_optimizers: dict[str, Optimizer | None],
+) -> dict[str, Any]:
+    """Serialize sub-optimizers, preserving ``None`` slots."""
+    return {
+        name: sub.state_dict() if sub is not None else None
+        for name, sub in named_sub_optimizers.items()
+    }
+
+
+def refresh_param_groups(
+    sub_optimizers: list[Optimizer],
+) -> list[dict]:
+    """Concatenate param_groups from every non-None sub-optimizer."""
+    groups: list[dict] = []
+    for sub in sub_optimizers:
+        if sub is not None:
+            groups.extend(sub.param_groups)
+    return groups

astrai/optim/mano_adamw.py

@@ -0,0 +1,214 @@
+"""Mano manifold optimizer combined with AdamW.
+
+Mano projects the momentum onto the tangent space of the Oblique manifold
+(axis-wise tangent projection) and normalizes it, replacing the expensive
+Newton-Schulz iteration in Muon with a cheaper manifold normalization.
+
+Reference: https://arxiv.org/abs/2601.23000
+"""
+
+import math
+
+import torch
+from torch import nn, optim
+from torch.optim import Optimizer
+
+from astrai.optim.composite import (
+    OptimizerFactory,
+    composite_state_dict,
+    composite_step,
+    composite_zero_grad,
+    refresh_param_groups,
+)
+from astrai.optim.nora_nadamw import partition_optimizer_parameters
+
+
+class Mano(Optimizer):
+    """Manifold Normalized Optimizer for two-dimensional matrices.
+
+    Each step alternates the projection axis (dim 0 / dim 1) to restrike the
+    manifold along both rows and columns. The tangent momentum is computed
+    without normalizing the parameter itself (v2 simplification) and the
+    epsilon is added (not clamped) to the norm denominator.
+    """
+
+    def __init__(
+        self,
+        params,
+        lr: float = 1e-3,
+        weight_decay: float = 0.1,
+        momentum: float = 0.95,
+        nesterov: bool = True,
+        eps: float = 1e-8,
+    ):
+        if lr < 0:
+            raise ValueError(f"Invalid learning rate: {lr}")
+        if weight_decay < 0:
+            raise ValueError(f"Invalid weight decay: {weight_decay}")
+        if not 0 <= momentum <= 1:
+            raise ValueError(f"Invalid momentum: {momentum}")
+        if eps <= 0:
+            raise ValueError(f"Invalid epsilon: {eps}")
+
+        defaults = {
+            "lr": lr,
+            "weight_decay": weight_decay,
+            "momentum": momentum,
+            "nesterov": nesterov,
+            "eps": eps,
+            "steps": 0,
+        }
+        super().__init__(params, defaults)
+        for group in self.param_groups:
+            for param in group["params"]:
+                if param.ndim != 2:
+                    raise ValueError(
+                        f"Mano only supports 2D matrices, got shape {tuple(param.shape)}"
+                    )
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            lr = group["lr"]
+            weight_decay = group["weight_decay"]
+            momentum = group["momentum"]
+            nesterov = group["nesterov"]
+            eps = group["eps"]
+            dim = int(group["steps"] % 2)
+
+            for param in group["params"]:
+                if param.grad is None:
+                    continue
+                if param.grad.is_sparse:
+                    raise RuntimeError("Mano does not support sparse gradients")
+
+                grad = param.grad
+                state = self.state[param]
+                momentum_buffer = state.get("momentum_buffer")
+                if momentum_buffer is None:
+                    momentum_buffer = torch.zeros_like(grad)
+                momentum_buffer.mul_(momentum).add_(grad)
+                update = (
+                    grad.add(momentum_buffer, alpha=momentum)
+                    if nesterov
+                    else momentum_buffer
+                )
+
+                tangent = update - (
+                    torch.sum(update * param.data, dim=dim, keepdim=True) * param.data
+                )
+                direction = tangent / (
+                    torch.norm(tangent, p=2, dim=dim, keepdim=True) + eps
+                )
+
+                if weight_decay != 0:
+                    param.mul_(1 - lr * weight_decay)
+                adjusted_lr = lr * 0.2 * math.sqrt(direction.shape[dim])
+                param.add_(direction, alpha=-adjusted_lr)
+                state["momentum_buffer"] = momentum_buffer
+
+            group["steps"] += 1
+
+        return loss
+
+
+@OptimizerFactory.register("mano_adamw")
+class ManoAdamW(Optimizer):
+    """Mano for internal linear weights and AdamW for remaining parameters."""
+
+    optimizer_name = "mano_adamw"
+
+    def __init__(
+        self,
+        model: nn.Module,
+        lr: float = 3e-4,
+        weight_decay: float = 0.1,
+        momentum: float = 0.95,
+        nesterov: bool = True,
+    ):
+        groups = partition_optimizer_parameters(model)
+        all_params = [
+            *groups.nora,
+            *groups.nadamw_decay,
+            *groups.nadamw_no_decay,
+        ]
+        if not all_params:
+            raise ValueError(
+                "Cannot build an optimizer for a model with no trainable parameters"
+            )
+        super().__init__(all_params, {})
+
+        self.mano = (
+            Mano(
+                groups.nora,
+                lr=lr,
+                weight_decay=weight_decay,
+                momentum=momentum,
+                nesterov=nesterov,
+            )
+            if groups.nora
+            else None
+        )
+
+        adamw_groups = []
+        if groups.nadamw_decay:
+            adamw_groups.append(
+                {"params": groups.nadamw_decay, "weight_decay": weight_decay}
+            )
+        if groups.nadamw_no_decay:
+            adamw_groups.append({"params": groups.nadamw_no_decay, "weight_decay": 0.0})
+        self.adamw = (
+            optim.AdamW(
+                adamw_groups,
+                lr=lr,
+                betas=(0.9, 0.95),
+                fused=True,
+            )
+            if adamw_groups
+            else None
+        )
+        self.param_groups = refresh_param_groups([self.mano, self.adamw])
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        return composite_step(
+            [opt for opt in (self.mano, self.adamw) if opt is not None],
+            closure,
+        )
+
+    def zero_grad(self, set_to_none: bool = True):
+        composite_zero_grad(
+            [opt for opt in (self.mano, self.adamw) if opt is not None],
+            set_to_none,
+        )
+
+    def state_dict(self) -> dict:
+        return composite_state_dict({"mano": self.mano, "adamw": self.adamw})
+
+    def load_state_dict(self, state_dict: dict):
+        if "muon" in state_dict or "nora" in state_dict:
+            raise ValueError(
+                "Checkpoint uses a different optimizer; select the matching "
+                "--optimizer to resume it"
+            )
+        if "mano" not in state_dict or "adamw" not in state_dict:
+            raise ValueError(
+                "Checkpoint optimizer state is not compatible with mano_adamw"
+            )
+
+        saved_mano = state_dict["mano"]
+        saved_adamw = state_dict["adamw"]
+        if (self.mano is None) != (saved_mano is None):
+            raise ValueError("Checkpoint Mano parameter groups do not match the model")
+        if (self.adamw is None) != (saved_adamw is None):
+            raise ValueError("Checkpoint AdamW parameter groups do not match the model")
+        if self.mano is not None:
+            self.mano.load_state_dict(saved_mano)
+        if self.adamw is not None:
+            self.adamw.load_state_dict(saved_adamw)
+        self.param_groups = refresh_param_groups([self.mano, self.adamw])

astrai/optim/muon_adamw.py

@@ -0,0 +1,95 @@
+"""Legacy Muon + AdamW combined optimizer."""
+
+from typing import Any
+
+import torch
+from torch import Tensor, nn, optim
+
+from astrai.optim.composite import (
+    OptimizerFactory,
+    composite_state_dict,
+    composite_step,
+    composite_zero_grad,
+    refresh_param_groups,
+)
+
+
+@OptimizerFactory.register("muon_adamw")
+class MuonAdamW(optim.Optimizer):
+    """Combined Muon (matrix) + AdamW (non-matrix) optimizer."""
+
+    optimizer_name = "muon_adamw"
+
+    def __init__(
+        self,
+        model: nn.Module,
+        lr: float = 3e-4,
+        weight_decay: float = 0.1,
+        momentum: float = 0.95,
+        nesterov: bool = True,
+        ns_steps: int = 5,
+        adjust_lr_fn: str = "match_rms_adamw",
+    ):
+        defaults = {
+            "lr": lr,
+            "weight_decay": weight_decay,
+            "momentum": momentum,
+            "nesterov": nesterov,
+            "ns_steps": ns_steps,
+            "adjust_lr_fn": adjust_lr_fn,
+        }
+        params = [param for param in model.parameters() if param.requires_grad]
+        super().__init__(params, defaults)
+
+        matrix_params: list[Tensor] = []
+        other_params: list[Tensor] = []
+        for name, param in model.named_parameters():
+            if not param.requires_grad:
+                continue
+            if (
+                param.dim() >= 2
+                and "norm" not in name
+                and "bias" not in name
+                and "embed" not in name
+                and "lm_head" not in name
+            ):
+                matrix_params.append(param)
+            else:
+                other_params.append(param)
+
+        self.muon = optim.Muon(
+            matrix_params,
+            lr=lr,
+            weight_decay=weight_decay,
+            momentum=momentum,
+            nesterov=nesterov,
+            ns_steps=ns_steps,
+            adjust_lr_fn=adjust_lr_fn,
+        )
+        self.adamw = optim.AdamW(
+            [{"params": other_params, "weight_decay": 0.0}],
+            lr=lr,
+            betas=(0.9, 0.95),
+            fused=True,
+        )
+
+        self.param_groups = refresh_param_groups([self.muon, self.adamw])
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        return composite_step([self.muon, self.adamw], closure)
+
+    def zero_grad(self, set_to_none: bool = True):
+        composite_zero_grad([self.muon, self.adamw], set_to_none)
+
+    def state_dict(self) -> dict[str, Any]:
+        return composite_state_dict({"muon": self.muon, "adamw": self.adamw})
+
+    def load_state_dict(self, state_dict: dict[str, Any]):
+        if "muon" not in state_dict or "adamw" not in state_dict:
+            raise ValueError(
+                "Checkpoint optimizer state is not compatible with muon_adamw"
+            )
+        self.muon.load_state_dict(state_dict["muon"])
+        self.adamw.load_state_dict(state_dict["adamw"])
+        self.param_groups = refresh_param_groups([self.muon, self.adamw])

astrai/optim/nora_nadamw.py

@@ -0,0 +1,372 @@
+"""Nora matrix optimizer combined with Nesterov AdamW."""
+
+import math
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+from torch import Tensor, nn
+from torch.distributed.tensor import DTensor, Shard
+from torch.optim import Optimizer
+
+from astrai.model.components.embedding import Embedding
+from astrai.model.components.linear import Linear
+from astrai.model.components.lora import LoRALinear
+from astrai.model.components.norm import RMSNorm
+from astrai.optim.composite import (
+    OptimizerFactory,
+    composite_state_dict,
+    composite_step,
+    composite_zero_grad,
+    refresh_param_groups,
+)
+
+NORA_EPS = 1e-10
+
+
+def _row_normalize(tensor: Tensor, eps: float) -> Tensor:
+    return tensor / tensor.norm(dim=-1, keepdim=True).clamp(min=eps)
+
+
+def nora_direction(update: Tensor, param: Tensor, eps: float = NORA_EPS) -> Tensor:
+    """Project an update onto each parameter row's tangent space and normalize."""
+    theta_hat = _row_normalize(param.to(torch.float32), eps)
+    update_fp32 = update.to(torch.float32)
+    radial = (update_fp32 * theta_hat).sum(dim=-1, keepdim=True) * theta_hat
+    direction = _row_normalize(update_fp32 - radial, eps)
+    return direction.to(update.dtype)
+
+
+def nora_lr_scale(lr: float, shape: torch.Size) -> float:
+    """Scale Nora's LR for tall ``[d_out, d_in]`` linear weights."""
+    return lr * math.sqrt(max(1.0, shape[-2] / shape[-1]))
+
+
+def _validate_complete_rows(param: Tensor) -> None:
+    if not isinstance(param, DTensor):
+        return
+    last_dim = param.ndim - 1
+    for placement in param.placements:
+        if isinstance(placement, Shard) and placement.dim % param.ndim == last_dim:
+            raise ValueError(
+                "Nora requires complete parameter rows, but this DTensor is sharded "
+                "along its last dimension"
+            )
+
+
+class Nora(Optimizer):
+    """Normalized Orthogonal Row Alignment for two-dimensional matrices."""
+
+    def __init__(
+        self,
+        params,
+        lr: float = 5e-3,
+        weight_decay: float = 0.0,
+        momentum: float = 0.95,
+        beta: float = 0.95,
+        nesterov: bool = True,
+        eps: float = NORA_EPS,
+    ):
+        if lr < 0:
+            raise ValueError(f"Invalid learning rate: {lr}")
+        if weight_decay < 0:
+            raise ValueError(f"Invalid weight decay: {weight_decay}")
+        if not 0 <= momentum <= 1:
+            raise ValueError(f"Invalid momentum: {momentum}")
+        if not 0 <= beta < 1:
+            raise ValueError(f"Invalid beta: {beta}")
+        if eps <= 0:
+            raise ValueError(f"Invalid epsilon: {eps}")
+
+        defaults = {
+            "lr": lr,
+            "weight_decay": weight_decay,
+            "momentum": momentum,
+            "beta": beta,
+            "nesterov": nesterov,
+            "eps": eps,
+        }
+        super().__init__(params, defaults)
+        for group in self.param_groups:
+            for param in group["params"]:
+                if param.ndim != 2:
+                    raise ValueError(
+                        f"Nora only supports 2D matrices, got shape {tuple(param.shape)}"
+                    )
+                _validate_complete_rows(param)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            lr = group["lr"]
+            weight_decay = group["weight_decay"]
+            momentum = group["momentum"]
+            beta = group["beta"]
+            nesterov = group["nesterov"]
+            eps = group["eps"]
+            for param in group["params"]:
+                if param.grad is None:
+                    continue
+                if param.grad.is_sparse:
+                    raise RuntimeError("Nora does not support sparse gradients")
+
+                grad = param.grad
+                state = self.state[param]
+                momentum_buffer = state.get("momentum_buffer")
+                if momentum_buffer is None:
+                    momentum_buffer = torch.zeros_like(grad)
+                momentum_buffer.lerp_(grad, 1 - beta)
+                update = (
+                    grad.lerp(momentum_buffer, momentum)
+                    if nesterov
+                    else momentum_buffer
+                )
+                direction = nora_direction(update, param, eps)
+
+                if weight_decay != 0:
+                    param.mul_(1 - lr * weight_decay)
+                param.add_(direction, alpha=-nora_lr_scale(lr, param.shape))
+                state["momentum_buffer"] = momentum_buffer
+
+        return loss
+
+
+class NAdamW(Optimizer):
+    """AdamW using the reference Nesterov first-moment update."""
+
+    def __init__(
+        self,
+        params,
+        lr: float = 3e-4,
+        betas: tuple[float, float] = (0.9, 0.999),
+        eps: float = 1e-8,
+        weight_decay: float = 0.1,
+    ):
+        beta1, beta2 = betas
+        if lr < 0:
+            raise ValueError(f"Invalid learning rate: {lr}")
+        if not 0 <= beta1 < 1 or not 0 <= beta2 < 1:
+            raise ValueError(f"Invalid betas: {betas}")
+        if eps <= 0:
+            raise ValueError(f"Invalid epsilon: {eps}")
+        if weight_decay < 0:
+            raise ValueError(f"Invalid weight decay: {weight_decay}")
+        defaults = {
+            "lr": lr,
+            "betas": betas,
+            "eps": eps,
+            "weight_decay": weight_decay,
+        }
+        super().__init__(params, defaults)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            beta1, beta2 = group["betas"]
+            eps = group["eps"]
+            lr = group["lr"]
+            weight_decay = group["weight_decay"]
+            for param in group["params"]:
+                if param.grad is None:
+                    continue
+                if param.grad.is_sparse:
+                    raise RuntimeError("NAdamW does not support sparse gradients")
+
+                grad = param.grad
+                state = self.state[param]
+                if not state:
+                    state["step"] = 0
+                    state["m"] = torch.zeros_like(param)
+                    state["v"] = torch.zeros_like(param)
+
+                state["step"] += 1
+                first_moment = state["m"]
+                second_moment = state["v"]
+                first_moment.mul_(beta1).add_(grad, alpha=1 - beta1)
+                second_moment.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+                bias_correction1 = 1 - beta1 ** state["step"]
+                bias_correction2 = 1 - beta2 ** state["step"]
+                nesterov_moment = (
+                    beta1 * first_moment + (1 - beta1) * grad
+                ) / bias_correction1
+                corrected_second_moment = second_moment / bias_correction2
+
+                if weight_decay != 0:
+                    param.mul_(1 - lr * weight_decay)
+                param.addcdiv_(
+                    nesterov_moment,
+                    corrected_second_moment.sqrt().add_(eps),
+                    value=-lr,
+                )
+
+        return loss
+
+
+@dataclass
+class OptimizerParameterGroups:
+    nora: list[Tensor]
+    nadamw_decay: list[Tensor]
+    nadamw_no_decay: list[Tensor]
+
+
+def partition_optimizer_parameters(model: nn.Module) -> OptimizerParameterGroups:
+    """Partition trainable parameters by module role and parameter identity."""
+    nora_ids: set[int] = set()
+    no_decay_ids: set[int] = set()
+
+    for module_name, module in model.named_modules():
+        if isinstance(module, LoRALinear):
+            for param in module.parameters(recurse=False):
+                if param.requires_grad:
+                    no_decay_ids.add(id(param))
+            continue
+
+        if isinstance(module, (Embedding, RMSNorm)):
+            for param in module.parameters(recurse=False):
+                if param.requires_grad:
+                    no_decay_ids.add(id(param))
+            continue
+
+        if not isinstance(module, Linear):
+            continue
+
+        if module.bias is not None and module.bias.requires_grad:
+            no_decay_ids.add(id(module.bias))
+        if not module.weight.requires_grad:
+            continue
+        if module_name.rsplit(".", 1)[-1] == "lm_head":
+            no_decay_ids.add(id(module.weight))
+        elif module.weight.ndim == 2:
+            nora_ids.add(id(module.weight))
+
+    nora: list[Tensor] = []
+    nadamw_decay: list[Tensor] = []
+    nadamw_no_decay: list[Tensor] = []
+    seen: set[int] = set()
+    for param in model.parameters():
+        param_id = id(param)
+        if not param.requires_grad or param_id in seen:
+            continue
+        seen.add(param_id)
+        if param_id in no_decay_ids or param.ndim <= 1:
+            nadamw_no_decay.append(param)
+        elif param_id in nora_ids:
+            nora.append(param)
+        else:
+            nadamw_decay.append(param)
+
+    trainable_ids = {id(param) for param in model.parameters() if param.requires_grad}
+    grouped_ids = {id(param) for param in [*nora, *nadamw_decay, *nadamw_no_decay]}
+    if grouped_ids != trainable_ids:
+        missing = len(trainable_ids - grouped_ids)
+        extra = len(grouped_ids - trainable_ids)
+        raise RuntimeError(
+            f"Optimizer parameter partition is incomplete: missing={missing}, extra={extra}"
+        )
+
+    return OptimizerParameterGroups(nora, nadamw_decay, nadamw_no_decay)
+
+
+@OptimizerFactory.register("nora_nadamw")
+class NoraNAdamW(Optimizer):
+    """Nora for internal linear weights and NAdamW for remaining parameters."""
+
+    optimizer_name = "nora_nadamw"
+
+    def __init__(
+        self,
+        model: nn.Module,
+        lr: float = 3e-4,
+        weight_decay: float = 0.1,
+        nora_lr: float = 5e-3,
+        nora_weight_decay: float = 0.0,
+        nora_beta: float = 0.95,
+        nora_momentum: float = 0.95,
+    ):
+        groups = partition_optimizer_parameters(model)
+        all_params = [
+            *groups.nora,
+            *groups.nadamw_decay,
+            *groups.nadamw_no_decay,
+        ]
+        if not all_params:
+            raise ValueError(
+                "Cannot build an optimizer for a model with no trainable parameters"
+            )
+        super().__init__(all_params, {})
+
+        self.nora = (
+            Nora(
+                groups.nora,
+                lr=nora_lr,
+                weight_decay=nora_weight_decay,
+                momentum=nora_momentum,
+                beta=nora_beta,
+            )
+            if groups.nora
+            else None
+        )
+
+        nadamw_groups = []
+        if groups.nadamw_decay:
+            nadamw_groups.append(
+                {"params": groups.nadamw_decay, "weight_decay": weight_decay}
+            )
+        if groups.nadamw_no_decay:
+            nadamw_groups.append(
+                {"params": groups.nadamw_no_decay, "weight_decay": 0.0}
+            )
+        self.nadamw = NAdamW(nadamw_groups, lr=lr) if nadamw_groups else None
+        self.param_groups = refresh_param_groups([self.nora, self.nadamw])
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        return composite_step(
+            [opt for opt in (self.nora, self.nadamw) if opt is not None],
+            closure,
+        )
+
+    def zero_grad(self, set_to_none: bool = True):
+        composite_zero_grad(
+            [opt for opt in (self.nora, self.nadamw) if opt is not None],
+            set_to_none,
+        )
+
+    def state_dict(self) -> dict[str, Any]:
+        return composite_state_dict({"nora": self.nora, "nadamw": self.nadamw})
+
+    def load_state_dict(self, state_dict: dict[str, Any]):
+        if "muon" in state_dict or "adamw" in state_dict:
+            raise ValueError(
+                "Checkpoint uses muon_adamw state; select optimizer='muon_adamw' "
+                "to resume it"
+            )
+        if "nora" not in state_dict or "nadamw" not in state_dict:
+            raise ValueError(
+                "Checkpoint optimizer state is not compatible with nora_nadamw"
+            )
+
+        saved_nora = state_dict["nora"]
+        saved_nadamw = state_dict["nadamw"]
+        if (self.nora is None) != (saved_nora is None):
+            raise ValueError("Checkpoint Nora parameter groups do not match the model")
+        if (self.nadamw is None) != (saved_nadamw is None):
+            raise ValueError(
+                "Checkpoint NAdamW parameter groups do not match the model"
+            )
+        if self.nora is not None:
+            self.nora.load_state_dict(saved_nora)
+        if self.nadamw is not None:
+            self.nadamw.load_state_dict(saved_nadamw)
+        self.param_groups = refresh_param_groups([self.nora, self.nadamw])

astrai/parallel/__init__.py

@@ -7,8 +7,9 @@ from astrai.parallel.executor import (
     FSDPExecutor,
     GradientState,
     NoneExecutor,
+    broadcast_state_dict,
+    create_ref_model,
 )
-from astrai.parallel.module import ColumnParallelLinear, RowParallelLinear
 from astrai.parallel.setup import (
     get_current_device,
     get_rank,
@@ -25,8 +26,6 @@ __all__ = [
     "only_on_rank",
     "setup_parallel",
     "spawn_parallel_fn",
-    "RowParallelLinear",
-    "ColumnParallelLinear",
     "ExecutorFactory",
     "BaseExecutor",
     "GradientState",
@@ -35,4 +34,6 @@ __all__ = [
     "NoneExecutor",
     "DDPExecutor",
     "FSDPExecutor",
+    "create_ref_model",
+    "broadcast_state_dict",
 ]

astrai/parallel/executor.py

@@ -4,16 +4,19 @@ import contextlib
 import logging
 import os
 from contextlib import contextmanager
-from typing import Optional, Tuple
+from typing import Any, Callable, Dict, Optional, Tuple
 
 import torch
+import torch.distributed as dist
 import torch.nn as nn
-from torch.distributed.fsdp import FullStateDictConfig, StateDictType
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.distributed.fsdp import (
+    FSDPModule,
+    fully_shard,
+)
+from torch.distributed.tensor import DTensor
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
-from torch.utils.data import DataLoader
 
 from astrai.factory import BaseFactory
 from astrai.parallel.setup import get_rank, get_world_size
@@ -21,6 +24,82 @@ from astrai.parallel.setup import get_rank, get_world_size
 logger = logging.getLogger(__name__)
 
 
+def broadcast_state_dict(
+    state_dict: Optional[Dict[str, torch.Tensor]],
+    src: int = 0,
+) -> Optional[Dict[str, torch.Tensor]]:
+    """Broadcast a state_dict from *src* rank to all ranks.
+
+    Tensors stay on their original device (GPU) for the broadcast.
+    All ranks must call this collectively.
+
+    On non-distributed runs, returns *state_dict* unchanged.
+    """
+    if not dist.is_initialized() or dist.get_world_size() == 1:
+        return state_dict
+
+    rank = dist.get_rank()
+
+    # Broadcast metadata (keys, shapes, dtypes, device) so non-src ranks
+    # can allocate matching empty tensors on the correct device.
+    if rank == src:
+        device = next(iter(state_dict.values())).device
+        metadata = [
+            (k, tuple(v.shape), v.dtype, str(device)) for k, v in state_dict.items()
+        ]
+    else:
+        metadata = None
+    metadata_list = [metadata]
+    dist.broadcast_object_list(metadata_list, src=src)
+    metadata = metadata_list[0]
+
+    # Non-src ranks allocate empty tensors with the broadcasted metadata.
+    if rank != src:
+        state_dict = {
+            k: torch.empty(s, dtype=d, device=torch.device(dev))
+            for k, s, d, dev in metadata
+        }
+
+    # Broadcast each tensor in-place.
+    for tensor in state_dict.values():
+        dist.broadcast(tensor, src=src)
+
+    return state_dict
+
+
+def create_ref_model(
+    model_fn: Callable[[], nn.Module],
+    executor: Optional["BaseExecutor"] = None,
+    model: Optional[nn.Module] = None,
+    state_dict: Optional[Dict[str, torch.Tensor]] = None,
+    device: Optional[str] = None,
+) -> Optional[nn.Module]:
+    """Create a frozen reference model from executor or state dict.
+
+    In distributed mode (FSDP), ``unwrap_model`` returns ``None`` on
+    non-rank-0.  The state_dict is broadcast from rank-0 to all ranks
+    so every rank gets a complete copy.
+    """
+    if state_dict is None and executor is not None and model is not None:
+        state_dict = executor.unwrap_model(model)
+
+    # FSDP's unwrap_model returns None on non-rank-0. Broadcast from
+    # rank-0 so every rank receives a complete state_dict.
+    if executor is not None and executor.use_distributed:
+        state_dict = broadcast_state_dict(state_dict)
+
+    if state_dict is None:
+        return None
+
+    ref_model = model_fn()
+    ref_model.load_state_dict(state_dict)
+    ref_model.requires_grad_(False)
+    ref_model.eval()
+    if device is not None:
+        ref_model = ref_model.to(device=device)
+    return ref_model
+
+
 class GradientState:
     def __init__(self, grad_accum_steps: int = 1):
         self.num_steps = max(grad_accum_steps, 1)
@@ -85,19 +164,28 @@ class BaseExecutor:
 
     def prepare(
         self,
-        model: nn.Module,
-        optimizer: Optional[Optimizer] = None,
-        dataloader: Optional[DataLoader] = None,
-        scheduler: Optional[LRScheduler] = None,
-    ) -> Tuple[
-        nn.Module, Optional[Optimizer], Optional[DataLoader], Optional[LRScheduler]
-    ]:
+        model_fn: Callable[[], nn.Module],
+        optimizer_fn: Optional[Callable[[nn.Module], Optimizer]] = None,
+        scheduler_fn: Optional[Callable[[Optimizer], LRScheduler]] = None,
+        before_wrap: Optional[Callable[[nn.Module], nn.Module]] = None,
+        after_wrap: Optional[Callable[[nn.Module], nn.Module]] = None,
+    ) -> Tuple[nn.Module, Optional[Optimizer], Optional[LRScheduler]]:
+        model = model_fn()
+        if before_wrap is not None:
+            model = before_wrap(model)
         model = self._prepare_model(model)
-        if optimizer is not None:
+        if after_wrap is not None:
+            model = after_wrap(model)
+        optimizer = None
+        scheduler = None
+        if optimizer_fn is not None:
+            optimizer = optimizer_fn(model)
+            if scheduler_fn is not None:
+                scheduler = scheduler_fn(optimizer)
             optimizer = AccumOptimizer(optimizer, self.gradient_state)
-        if scheduler is not None:
-            scheduler = AccumScheduler(scheduler, self.gradient_state)
-        return model, optimizer, dataloader, scheduler
+            if scheduler is not None:
+                scheduler = AccumScheduler(scheduler, self.gradient_state)
+        return model, optimizer, scheduler
 
     def _prepare_model(self, model: nn.Module) -> nn.Module:
         return model
@@ -120,6 +208,21 @@ class BaseExecutor:
     def unwrap_model(self, model: nn.Module):
         return model.state_dict()
 
+    @contextmanager
+    def checkpoint_context(self, model: nn.Module):
+        if self.use_distributed:
+            dist.barrier()
+        state_dict = self._gather_state_dict(model)
+        yield state_dict
+        if self.use_distributed:
+            dist.barrier()
+
+    def _gather_state_dict(self, model: nn.Module):
+        state_dict = self.unwrap_model(model)
+        if self.use_distributed and get_rank() != 0:
+            return None
+        return state_dict
+
     @property
     def use_distributed(self) -> bool:
         return get_world_size() > 1
@@ -211,76 +314,115 @@ class DDPExecutor(BaseExecutor):
 
 @ExecutorFactory.register("fsdp")
 class FSDPExecutor(BaseExecutor):
+    """FSDP executor using `torch.distributed.fsdp.fully_shard` (per-module API).
+
+    Wraps each child module individually via ``fully_shard``.
+    Skips the root model because ``ABC + Generic[T]`` in the MRO makes
+    ``fully_shard``'s dynamic ``__class__`` assignment fail at the CPython level.
+    Original ``Parameter`` objects are preserved (as DTensors) — no
+    ``FlatParameter``, no ``use_orig_params=True`` hack.
+    """
+
     def __init__(
         self,
         grad_accum_steps: int = 1,
-        process_group=None,
-        sharding_strategy=None,
-        cpu_offload=None,
-        auto_wrap_policy=None,
-        backward_prefetch=None,
-        mixed_precision=None,
-        ignored_modules=None,
-        param_init_fn=None,
-        sync_module_states: bool = False,
-        forward_prefetch: bool = False,
-        limit_all_gathers: bool = True,
-        ignored_states=None,
-        device_mesh=None,
+        mesh: Optional[Any] = None,
+        mp_policy: Optional[Any] = None,
+        reshard_after_forward: bool = False,
     ):
         super().__init__(grad_accum_steps=grad_accum_steps)
-        self._fsdp_kwargs = {
-            k: v
-            for k, v in dict(
-                process_group=process_group,
-                sharding_strategy=sharding_strategy,
-                cpu_offload=cpu_offload,
-                auto_wrap_policy=auto_wrap_policy,
-                backward_prefetch=backward_prefetch,
-                mixed_precision=mixed_precision,
-                ignored_modules=ignored_modules,
-                param_init_fn=param_init_fn,
-                sync_module_states=sync_module_states,
-                forward_prefetch=forward_prefetch,
-                limit_all_gathers=limit_all_gathers,
-                use_orig_params=True,
-                ignored_states=ignored_states,
-                device_mesh=device_mesh,
-            ).items()
-            if v is not None
-        }
-        self._original_model: Optional[nn.Module] = None
+        self._mesh = mesh
+        self._mp_policy = mp_policy
+        self._reshard_after_forward = reshard_after_forward
 
     def _prepare_model(self, model: nn.Module) -> nn.Module:
         if not self.use_distributed:
             logger.warning("FSDP backend selected but world_size=1, model not wrapped")
             return model
-        self._original_model = model
-        device_id = torch.device("cuda", get_rank())
-        model = FSDP(model, device_id=device_id, **self._fsdp_kwargs)
-        logger.info("Model wrapped with FSDP (world_size=%d)", get_world_size())
+
+        kwargs = dict(
+            mesh=self._mesh,
+            mp_policy=self._mp_policy,
+            reshard_after_forward=self._reshard_after_forward,
+        )
+        kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+        for child in model.children():
+            if isinstance(child, nn.ModuleList):
+                for sub in child:
+                    fully_shard(sub, **kwargs)
+            else:
+                fully_shard(child, **kwargs)
+
+        logger.info(
+            "FSDP wrapping applied to %d direct children (root skipped for ABC compat)",
+            len(list(model.children())),
+        )
         return model
 
+    @contextmanager
     def _no_sync(self, model: nn.Module):
-        if isinstance(model, FSDP):
-            return model.no_sync()
-        return contextlib.nullcontext()
+        fsdp_modules = [m for m in model.modules() if isinstance(m, FSDPModule)]
+        if fsdp_modules:
+            for m in fsdp_modules:
+                m.set_requires_gradient_sync(False, recurse=True)
+            try:
+                yield
+            finally:
+                for m in fsdp_modules:
+                    m.set_requires_gradient_sync(True, recurse=True)
+        else:
+            yield
 
     def clip_grad_norm(self, model: nn.Module, max_norm: float) -> float:
-        if isinstance(model, FSDP) and self.use_distributed:
-            total_norm = model.clip_grad_norm_(max_norm)
-            if isinstance(total_norm, torch.Tensor):
-                return total_norm.item()
-            return total_norm
-        return super().clip_grad_norm(model, max_norm)
+        if not self.use_distributed:
+            return super().clip_grad_norm(model, max_norm)
+
+        # FSDP params are DTensors (sharded across ranks).
+        # torch.nn.utils.clip_grad_norm_ computes LOCAL norm per rank,
+        # so we must all-reduce to get the global norm before clipping.
+        local_norm = torch.nn.utils.get_total_norm(
+            [p.grad for p in model.parameters() if p.grad is not None],
+        )
+        if isinstance(local_norm, DTensor):
+            local_norm = local_norm.to_local()
+        total_norm_sq = local_norm**2
+        dist.all_reduce(total_norm_sq, op=dist.ReduceOp.SUM)
+        total_norm = total_norm_sq.sqrt()
+
+        clip_coef = max_norm / (total_norm + 1e-6)
+        clip_coef_clamped = torch.clamp(clip_coef, max=1.0)
+        for p in model.parameters():
+            if p.grad is not None:
+                p.grad.mul_(clip_coef_clamped)
+
+        return total_norm.item()
 
     def unwrap_model(self, model: nn.Module):
-        if isinstance(model, FSDP) and self.use_distributed:
-            with FSDP.state_dict_type(
-                model,
-                StateDictType.FULL_STATE_DICT,
-                FullStateDictConfig(offload_to_cpu=True, rank0_only=False),
-            ):
-                return model.state_dict()
+        if not self.use_distributed:
+            return model.state_dict()
 
-        return model.state_dict()
+        # unshard() and full_tensor() are collective ops — all ranks must
+        # participate. Non-rank-0 ranks still call them but discard results.
+        for module in model.modules():
+            if isinstance(module, FSDPModule):
+                module.unshard()
+
+        state_dict = model.state_dict()
+        result = {}
+        for k, v in state_dict.items():
+            if isinstance(v, DTensor):
+                full = v.full_tensor()
+                if get_rank() == 0:
+                    result[k] = full
+            elif get_rank() == 0:
+                result[k] = v
+
+        for module in model.modules():
+            if isinstance(module, FSDPModule):
+                module.reshard()
+
+        if get_rank() != 0:
+            return None
+
+        return result

astrai/parallel/module.py

@@ -1,115 +0,0 @@
-from typing import Dict
-
-import torch
-import torch.distributed as dist
-import torch.nn as nn
-import torch.nn.functional as F
-from torch import Tensor
-
-
-class ParallelModel(nn.Module):
-    def __init__(self, process_group: dist.ProcessGroup):
-        super().__init__()
-        self.process_group = process_group
-        self.rank = dist.get_rank(self.process_group)
-        self.world_size = dist.get_world_size(self.process_group)
-
-
-class RowParallelLinear(ParallelModel):
-    def __init__(
-        self,
-        process_group: dist.ProcessGroup,
-        in_features: int,
-        out_features: int,
-        bias: bool = True,
-        reduce_results: bool = True,
-    ):
-        super().__init__(process_group)
-
-        self.in_features = in_features
-        self.out_features = out_features
-        self.in_features_per_rank = in_features // self.world_size
-        self.reduce_results = reduce_results
-
-        if in_features % self.world_size != 0:
-            raise ValueError(
-                f"in_features must be divisible by world_size. Got {in_features} and {self.world_size}"
-            )
-
-        self.weight = nn.Parameter(torch.empty(out_features, self.in_features_per_rank))
-        self.bias = nn.Parameter(torch.zeros(out_features)) if bias else None
-
-    def forward(self, input: Tensor) -> Tensor:
-        output = F.linear(input, self.weight)
-
-        if self.reduce_results:
-            dist.all_reduce(output, op=dist.ReduceOp.SUM, group=self.process_group)
-
-        if self.bias is not None:
-            output += self.bias
-
-        return output
-
-    def load_state_dict(self, state_dict: Dict[str, Tensor]):
-        full_weight = state_dict.get("weight")
-        full_bias = state_dict.get("bias")
-
-        start_idx = self.rank * self.in_features_per_rank
-        end_idx = start_idx + self.in_features_per_rank
-        weight_slice = full_weight[:, start_idx:end_idx]
-        self.weight.data.copy_(weight_slice)
-
-        if self.bias is not None:
-            self.bias.data.copy_(full_bias)
-
-
-class ColumnParallelLinear(ParallelModel):
-    def __init__(
-        self,
-        process_group: dist.ProcessGroup,
-        in_features: int,
-        out_features: int,
-        bias: bool = True,
-        gather_results: bool = True,
-    ):
-        super().__init__(process_group)
-
-        self.in_features = in_features
-        self.out_features = out_features
-        self.out_features_per_rank = out_features // self.world_size
-        self.gather_results = gather_results
-
-        if out_features % self.world_size != 0:
-            raise ValueError(
-                f"out_features must be divisible by world_size. Got {out_features} and {self.world_size}"
-            )
-
-        self.weight = nn.Parameter(
-            torch.empty(self.out_features_per_rank, self.in_features)
-        )
-        self.bias = (
-            nn.Parameter(torch.zeros(self.out_features_per_rank)) if bias else None
-        )
-
-    def forward(self, input: Tensor) -> Tensor:
-        output = F.linear(input, self.weight, self.bias)
-
-        if self.gather_results:
-            output_list = [torch.empty_like(output) for _ in range(self.world_size)]
-            dist.all_gather(output_list, output, group=self.process_group)
-            output = torch.cat(output_list, dim=-1)
-
-        return output
-
-    def load_state_dict(self, state_dict: Dict[str, Tensor]):
-        full_weight = state_dict.get("weight")
-        full_bias = state_dict.get("bias")
-
-        start_idx = self.rank * self.out_features_per_rank
-        end_idx = start_idx + self.out_features_per_rank
-        weight_slice = full_weight[start_idx:end_idx, :]
-        self.weight.data.copy_(weight_slice)
-
-        if self.bias is not None:
-            bias_slice = full_bias[start_idx:end_idx]
-            self.bias.data.copy_(bias_slice)

astrai/parallel/setup.py

@@ -1,13 +1,27 @@
+import logging
 import os
+import signal
+import socket
+import threading
 from abc import ABC, abstractmethod
 from contextlib import contextmanager
 from functools import wraps
-from typing import Callable
+from typing import Callable, Optional
 
 import torch
 import torch.distributed as dist
 import torch.multiprocessing as mp
 
+from astrai.signal_handler import install_early_signal_handlers
+
+logger = logging.getLogger(__name__)
+
+
+def find_free_port() -> str:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(("", 0))
+        return str(s.getsockname()[1])
+
 
 def get_current_device():
     return os.environ["LOCAL_DEVICE"]
@@ -108,6 +122,7 @@ def _run_single_rank(
     func: Callable,
     kwargs: dict,
 ):
+    install_early_signal_handlers()
     with setup_parallel(
         rank=rank,
         world_size=world_size,
@@ -148,6 +163,7 @@ class TorchrunStrategy(LaunchStrategy):
     """External orchestrator (torchrun, SLURM, K8s) — env vars pre-set."""
 
     def launch(self, func: Callable, **kwargs):
+        install_early_signal_handlers()
         rank = int(os.environ["RANK"])
         world_size = int(os.environ["WORLD_SIZE"])
         local_rank = int(os.environ.get("LOCAL_RANK", rank))
@@ -181,6 +197,7 @@ class LocalStrategy(LaunchStrategy):
             _run_single_rank(0, *args)
             return
 
+        install_early_signal_handlers()
         ctx = mp.start_processes(
             _run_single_rank,
             args=args,
@@ -188,14 +205,46 @@ class LocalStrategy(LaunchStrategy):
             start_method=self.start_method,
             join=False,
         )
+
+        parent_stop = threading.Event()
+        original_handlers = {}
+
+        def _parent_handler(signum, frame):
+            sig = signal.Signals(signum)
+            logger.warning(
+                "Parent (pid=%d) received %s, forwarding to children...",
+                os.getpid(),
+                sig.name,
+            )
+            parent_stop.set()
+            for p in ctx.processes:
+                if p.is_alive():
+                    p.terminate()
+
+        for sig in (signal.SIGTERM, signal.SIGINT):
+            prev = signal.signal(sig, _parent_handler)
+            if prev not in (signal.SIG_DFL, signal.SIG_IGN, None, _parent_handler):
+                original_handlers[sig] = prev
+
         try:
-            while not ctx.join():
+            while not ctx.join() and not parent_stop.is_set():
                 pass
         except BaseException:
+            logger.warning(
+                "Parent received unexpected exception, terminating children..."
+            )
             for p in ctx.processes:
-                p.terminate()
-            ctx.join()
+                if p.is_alive():
+                    p.terminate()
             raise
+        finally:
+            for sig, handler in original_handlers.items():
+                signal.signal(sig, handler)
+
+            for p in ctx.processes:
+                p.join()
+
+            ctx.join()
 
 
 def _detect_launcher() -> str:
@@ -217,11 +266,13 @@ def spawn_parallel_fn(
     world_size: int,
     backend: str = "nccl",
     master_addr: str = "localhost",
-    master_port: str = "29500",
+    master_port: Optional[str] = None,
     device_type: str = "cuda",
     start_method: str = "spawn",
     **kwargs,
 ):
+    if master_port is None:
+        master_port = find_free_port()
     launcher = _detect_launcher()
     if launcher in ("torchelastic", "torchrun", "external"):
         strategy = TorchrunStrategy(

astrai/preprocessing/__init__.py

@@ -1,17 +1,21 @@
 from astrai.preprocessing.builder import (
     BaseMaskBuilder,
     MaskBuilderFactory,
+    MultiOutputMaskBuilder,
     SectionedMaskBuilder,
+    SingleOutputMaskBuilder,
 )
 from astrai.preprocessing.packing import (
     PackingStrategy,
     PackingStrategyFactory,
+    plan_bfd,
 )
 from astrai.preprocessing.pipeline import Pipeline, filter_by_length
 from astrai.preprocessing.position_id import (
     PositionIdStrategy,
     PositionIdStrategyFactory,
 )
+from astrai.preprocessing.transform import TokenizeTransform
 from astrai.preprocessing.writer import (
     StoreWriter,
     StoreWriterFactory,
@@ -20,13 +24,17 @@ from astrai.preprocessing.writer import (
 __all__ = [
     "BaseMaskBuilder",
     "MaskBuilderFactory",
+    "MultiOutputMaskBuilder",
     "PackingStrategy",
     "PackingStrategyFactory",
     "Pipeline",
     "PositionIdStrategy",
     "PositionIdStrategyFactory",
     "SectionedMaskBuilder",
+    "SingleOutputMaskBuilder",
     "StoreWriter",
     "StoreWriterFactory",
+    "TokenizeTransform",
     "filter_by_length",
+    "plan_bfd",
 ]

astrai/preprocessing/builder.py

@@ -1,8 +1,10 @@
 """Mask building for preprocessing pipeline.
 
 :class:`SectionRenderer` converts section specs into token ids and loss
-masks (template / text / value extraction).  :class:`SectionedMaskBuilder`
-orchestrates single-output / multi-output (DPO / GRPO) assembly.
+masks (template / text / value extraction).  :class:`SingleOutputMaskBuilder`
+handles single-output (SFT / pretrain), :class:`MultiOutputMaskBuilder`
+handles multi-output (DPO / GRPO), and :class:`SectionedMaskBuilder`
+orchestrates both modes as a façade.
 """
 
 from abc import ABC, abstractmethod
@@ -92,9 +94,107 @@ class SectionRenderer:
 
         return all_ids, loss_mask
 
+    def process_sections_batch(
+        self,
+        items: list[dict],
+        sections: list,
+        config,
+        tokenizer,
+        *,
+        is_top_level=False,
+        filter_text=True,
+    ):
+        """Render and tokenize a group of records with batched Rust tokenization."""
+        has_template = any(s.get("template") for s in sections)
+        is_text_config = not has_template and all(
+            s["action"] == "train" for s in sections
+        )
+        plans: list[list[tuple[str, str, bool]]] = []
+
+        for item in items:
+            plan: list[tuple[str, str, bool]] = []
+            first_section = True
+            for sec in sections:
+                field = sec["field"]
+                action = sec["action"]
+                use_template = sec.get("template", False)
+                add_special = sec.get(
+                    "add_special_tokens", not use_template and first_section
+                )
+
+                if use_template:
+                    messages = item.get(field)
+                    if not isinstance(messages, list) or not messages:
+                        continue
+                    for msg in messages:
+                        role = msg.get("role", "")
+                        rendered = tokenizer.apply_chat_template(
+                            [msg], tokenize=False, add_generation_prompt=False
+                        )
+                        plan.append(
+                            (rendered, _resolve_action(action, role, config), False)
+                        )
+                else:
+                    text = str(item.get(field, ""))
+                    if not text.strip():
+                        continue
+                    if is_text_config and filter_text:
+                        pp = config.preprocessing
+                        if pp.min_chars > 0 and len(text) < pp.min_chars:
+                            continue
+                        if len(text) > pp.max_chars:
+                            continue
+                    plan.append((text, action, add_special))
+
+                first_section = False
+            plans.append(plan)
+
+        encoded: dict[tuple[int, int], list[int]] = {}
+        for add_special in (False, True):
+            refs = [
+                (item_idx, unit_idx, text)
+                for item_idx, plan in enumerate(plans)
+                for unit_idx, (text, _, add) in enumerate(plan)
+                if add == add_special
+            ]
+            if not refs:
+                continue
+            ids_batch = tokenizer.encode(
+                [text for _, _, text in refs], add_special_tokens=add_special
+            )
+            for (item_idx, unit_idx, _), ids in zip(refs, ids_batch):
+                encoded[(item_idx, unit_idx)] = ids
+
+        outputs = []
+        max_len = config.preprocessing.max_seq_len
+        for item_idx, plan in enumerate(plans):
+            all_ids = []
+            loss_mask = []
+            if is_top_level and has_template and tokenizer.bos_token_id is not None:
+                all_ids.append(tokenizer.bos_token_id)
+                loss_mask.append(0)
+            for unit_idx, (_, action, _) in enumerate(plan):
+                ids = encoded[(item_idx, unit_idx)]
+                all_ids.extend(ids)
+                loss_mask.extend([1 if action == "train" else 0] * len(ids))
+            all_ids = all_ids[:max_len]
+            loss_mask = loss_mask[: len(all_ids)]
+            if not all_ids or (is_top_level and has_template and len(all_ids) <= 1):
+                outputs.append((None, None))
+            else:
+                outputs.append((all_ids, loss_mask))
+        return outputs
+
     def process_list_field(self, item: dict, sections: list, config, tokenizer):
-        all_ids: list[int] = []
-        loss_mask: list[int] = []
+        """Tokenize a list-valued field, preserving per-element boundaries.
+
+        Returns ``(list_of_id_lists, list_of_mask_lists)`` where each
+        inner list corresponds to one element of the source list.  This
+        is critical for GRPO where each response must stay a separate
+        sequence so the strategy can form a ``[G, R]`` tensor.
+        """
+        per_item_ids: list[list[int]] = []
+        per_item_masks: list[list[int]] = []
 
         for sec in sections:
             field = sec["field"]
@@ -106,17 +206,13 @@ class SectionRenderer:
                 continue
 
             for val in values:
+                ids: list[int] = []
+                mask: list[int] = []
                 if use_template:
                     if isinstance(val, list):
                         wrapper = {field: val}
                         self._append_template(
-                            wrapper,
-                            field,
-                            action,
-                            tokenizer,
-                            config,
-                            all_ids,
-                            loss_mask,
+                            wrapper, field, action, tokenizer, config, ids, mask
                         )
                 else:
                     wrapper = {field: str(val)}
@@ -128,17 +224,55 @@ class SectionRenderer:
                         False,
                         False,
                         config,
-                        all_ids,
-                        loss_mask,
+                        ids,
+                        mask,
                     )
+                if ids:
+                    max_len = config.preprocessing.max_seq_len
+                    ids = ids[:max_len]
+                    mask = mask[: len(ids)]
+                    per_item_ids.append(ids)
+                    per_item_masks.append(mask)
 
-        max_len = config.preprocessing.max_seq_len
-        all_ids = all_ids[:max_len]
-        loss_mask = loss_mask[: len(all_ids)]
-
-        if not all_ids:
+        if not per_item_ids:
             return None, None
-        return all_ids, loss_mask
+        return per_item_ids, per_item_masks
+
+    def process_list_field_batch(self, items, sections, config, tokenizer):
+        per_item_ids = [[] for _ in items]
+        per_item_masks = [[] for _ in items]
+
+        for sec in sections:
+            wrappers = []
+            owners = []
+            field = sec["field"]
+            for item_idx, item in enumerate(items):
+                values = item.get(field)
+                if not isinstance(values, list):
+                    continue
+                for val in values:
+                    if sec.get("template", False) and not isinstance(val, list):
+                        continue
+                    wrappers.append({field: val if isinstance(val, list) else str(val)})
+                    owners.append(item_idx)
+
+            rendered = self.process_sections_batch(
+                wrappers,
+                [sec],
+                config,
+                tokenizer,
+                is_top_level=False,
+                filter_text=False,
+            )
+            for owner, (ids, mask) in zip(owners, rendered):
+                if ids:
+                    per_item_ids[owner].append(ids)
+                    per_item_masks[owner].append(mask)
+
+        return [
+            (ids, masks) if ids else (None, None)
+            for ids, masks in zip(per_item_ids, per_item_masks)
+        ]
 
     @staticmethod
     def is_value_section(sections: list) -> bool:
@@ -207,47 +341,25 @@ class BaseMaskBuilder(ABC):
     @abstractmethod
     def build(self, item: dict, config, tokenizer) -> Optional[dict]: ...
 
+    def build_batch(self, items: list[dict], config, tokenizer) -> list[Optional[dict]]:
+        return [self.build(item, config, tokenizer) for item in items]
+
 
 class MaskBuilderFactory(BaseFactory["BaseMaskBuilder"]):
     pass
 
 
-@MaskBuilderFactory.register("sectioned")
-class SectionedMaskBuilder(BaseMaskBuilder):
-    """Config-driven builder supporting single and multi-output modes.
+@MaskBuilderFactory.register("single")
+class SingleOutputMaskBuilder(BaseMaskBuilder):
+    """Build a single output sequence with optional loss mask.
 
-    Single-output::
-
-        {"input": {"sections": [
-            {"field": "messages", "action": "$role", "template": true}
-        ]}}
-        → {"sequence": [...], "loss_mask": [...], "domain": "..."}
-
-    Multi-output (DPO / GRPO)::
-
-        {"input": {"sources": {
-            "chosen": {"sections": [{"field": "chosen", "action": "$role", "template": true}]},
-            "rejected": {"sections": [{"field": "rejected", "action": "$role", "template": true}]},
-        }}}
-        → {"chosen": [...], "chosen_mask": [...], "rejected": [...], "rejected_mask": [...], "domain": "..."}
-
-    Output spec fields::
-
-        sections      – list of section specs (same format as single-output)
-        list_field    – True when JSONL field holds a list (GRPO responses)
-        mask_key      – explicit loss-mask output key (default: ``"{output_key}_mask"``)
+    Expects ``config.input.sections`` (list of section specs).
     """
 
-    def __init__(self):
-        self.renderer = SectionRenderer()
+    def __init__(self, renderer: Optional[SectionRenderer] = None):
+        self.renderer = renderer or SectionRenderer()
 
     def build(self, item: dict, config, tokenizer) -> Optional[dict]:
-        sources_spec = getattr(config.input, "sources", None)
-        if sources_spec:
-            return self._build_multi(item, sources_spec, config, tokenizer)
-        return self._build_single(item, config, tokenizer)
-
-    def _build_single(self, item: dict, config, tokenizer) -> Optional[dict]:
         sections = config.input.sections
         if not sections:
             return None
@@ -266,11 +378,49 @@ class SectionedMaskBuilder(BaseMaskBuilder):
             result["loss_mask"] = mask
         return result
 
-    def _build_multi(
-        self, item: dict, sources_spec: dict, config, tokenizer
-    ) -> Optional[dict]:
+    def build_batch(self, items, config, tokenizer):
+        sections = config.input.sections
+        if not sections:
+            return [None] * len(items)
+        rendered = self.renderer.process_sections_batch(
+            items, sections, config, tokenizer, is_top_level=True
+        )
+        results = []
+        for item, (ids, mask) in zip(items, rendered):
+            if ids is None:
+                results.append(None)
+                continue
+            result = {
+                "sequence": ids,
+                "domain": _extract_domain(item, config.output.domain_key),
+            }
+            if not all(m == 1 for m in mask):
+                result["loss_mask"] = mask
+            results.append(result)
+        return results
+
+
+@MaskBuilderFactory.register("multi")
+class MultiOutputMaskBuilder(BaseMaskBuilder):
+    """Build multiple output sequences (DPO / GRPO).
+
+    Expects ``config.input.sources`` (dict of output_key → spec).
+    """
+
+    def __init__(self, renderer: Optional[SectionRenderer] = None):
+        self.renderer = renderer or SectionRenderer()
+
+    def build(self, item: dict, config, tokenizer) -> Optional[dict]:
+        sources_spec = getattr(config.input, "sources", None)
+        if not sources_spec:
+            return None
+
         result: dict = {}
-        any_output = False
+        required_outputs = {
+            output_key
+            for output_key, spec in sources_spec.items()
+            if spec.get("sections")
+        }
 
         for output_key, spec in sources_spec.items():
             sections = spec.get("sections", [])
@@ -282,7 +432,6 @@ class SectionedMaskBuilder(BaseMaskBuilder):
                 if ids is None:
                     continue
                 result[output_key] = ids
-                any_output = True
                 continue
 
             list_field = spec.get("list_field", False)
@@ -292,10 +441,17 @@ class SectionedMaskBuilder(BaseMaskBuilder):
                 ids, mask = self.renderer.process_list_field(
                     item, sections, config, tokenizer
                 )
-            else:
-                ids, mask = self.renderer.process_sections(
-                    item, sections, config, tokenizer, is_top_level=True
-                )
+                if ids is None:
+                    continue
+                # ids is List[List[int]] — preserve per-response structure
+                result[output_key] = ids
+                if mask is not None:
+                    result[mask_key] = mask
+                continue
+
+            ids, mask = self.renderer.process_sections(
+                item, sections, config, tokenizer, is_top_level=True
+            )
 
             if ids is None:
                 continue
@@ -306,10 +462,81 @@ class SectionedMaskBuilder(BaseMaskBuilder):
             elif "mask_key" in spec:
                 result[mask_key] = mask
 
-            any_output = True
-
-        if not any_output:
+        if not required_outputs or not required_outputs.issubset(result):
             return None
 
         result["domain"] = _extract_domain(item, config.output.domain_key)
         return result
+
+    def build_batch(self, items, config, tokenizer):
+        sources_spec = getattr(config.input, "sources", None)
+        if not sources_spec:
+            return [None] * len(items)
+
+        results = [{} for _ in items]
+        required_outputs = {
+            output_key
+            for output_key, spec in sources_spec.items()
+            if spec.get("sections")
+        }
+        for output_key, spec in sources_spec.items():
+            sections = spec.get("sections", [])
+            if not sections:
+                continue
+            if self.renderer.is_value_section(sections):
+                for item, result in zip(items, results):
+                    value = self.renderer.extract_raw_value(item, sections)
+                    if value is not None:
+                        result[output_key] = value
+                continue
+
+            mask_key = spec.get("mask_key", f"{output_key}_mask")
+            if spec.get("list_field", False):
+                rendered = self.renderer.process_list_field_batch(
+                    items, sections, config, tokenizer
+                )
+            else:
+                rendered = self.renderer.process_sections_batch(
+                    items, sections, config, tokenizer, is_top_level=True
+                )
+
+            for result, (ids, mask) in zip(results, rendered):
+                if ids is None:
+                    continue
+                result[output_key] = ids
+                if spec.get("list_field", False) or not all(m == 1 for m in mask):
+                    result[mask_key] = mask
+                elif "mask_key" in spec:
+                    result[mask_key] = mask
+
+        return [
+            ({**result, "domain": _extract_domain(item, config.output.domain_key)})
+            if required_outputs and required_outputs.issubset(result)
+            else None
+            for item, result in zip(items, results)
+        ]
+
+
+@MaskBuilderFactory.register("sectioned")
+class SectionedMaskBuilder(BaseMaskBuilder):
+    """Façade that dispatches to SingleOutputMaskBuilder or MultiOutputMaskBuilder.
+
+    Preserves backward compatibility for existing configs and code that rely
+    on the ``"sectioned"`` factory name.
+    """
+
+    def __init__(self):
+        self._single = SingleOutputMaskBuilder()
+        self._multi = MultiOutputMaskBuilder()
+
+    def build(self, item: dict, config, tokenizer) -> Optional[dict]:
+        sources_spec = getattr(config.input, "sources", None)
+        if sources_spec:
+            return self._multi.build(item, config, tokenizer)
+        return self._single.build(item, config, tokenizer)
+
+    def build_batch(self, items, config, tokenizer):
+        sources_spec = getattr(config.input, "sources", None)
+        if sources_spec:
+            return self._multi.build_batch(items, config, tokenizer)
+        return self._single.build_batch(items, config, tokenizer)

astrai/preprocessing/core.py

@@ -0,0 +1,124 @@
+"""Shared preprocessing kernel used by both :class:`Pipeline` and
+:class:`TokenizeTransform`.
+
+The two entry points previously duplicated ~60 % of their logic:
+record iteration, mask-builder invocation, primary-id extraction,
+per-key accumulation, dtype inference and position-id generation.
+This module factors out the common core as pure functions so that
+the online (``TokenizeTransform``) and offline (``Pipeline``) paths
+stay in lockstep.
+"""
+
+from itertools import chain
+from typing import Dict, Iterator, List, Optional
+
+import torch
+
+from astrai.config.preprocess_config import PipelineConfig
+from astrai.preprocessing.builder import MaskBuilderFactory
+from astrai.preprocessing.position_id import PositionIdStrategyFactory
+from astrai.tokenize import AutoTokenizer
+
+
+def build_preprocessing_components(config: PipelineConfig, tokenizer_path: str):
+    """Load tokenizer, mask builder and position-id strategy together.
+
+    Both ``Pipeline`` and ``TokenizeTransform`` need the same triple;
+    centralising the construction avoids drift (e.g. one path forgetting
+    to create the position-id strategy).
+    """
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+    mask_builder = MaskBuilderFactory.create("sectioned")
+    position_strategy = PositionIdStrategyFactory.create(
+        config.output.position_ids_mode
+    )
+    return tokenizer, mask_builder, position_strategy
+
+
+def primary_ids(result: dict) -> List[int]:
+    """Return the first flat int-list value in *result*.
+
+    Used for token counting and position-id generation when the
+    primary key name is not known (DPO uses ``chosen``, GRPO uses
+    ``prompts``, SFT uses ``sequence``).
+    """
+    for val in result.values():
+        if isinstance(val, list) and val and isinstance(val[0], int):
+            return val
+    return []
+
+
+def infer_dtype(ids: List) -> torch.dtype:
+    """Float values become float32, everything else int32."""
+    if ids and isinstance(ids[0], float):
+        return torch.float32
+    return torch.int32
+
+
+def iter_raw_records(
+    records: List[dict],
+    mask_builder,
+    config: PipelineConfig,
+    tokenizer,
+) -> Iterator[dict]:
+    """Yield mask-builder output dicts for each record, skipping failures.
+
+    Drops ``domain`` from the result (callers that need it should read
+    it before calling this).  Each yielded dict maps a key
+    (``sequence``, ``chosen``, ``responses``…) to either a flat
+    ``List[int]`` or a nested ``List[List[int]]`` (GRPO responses/masks).
+    """
+    for item in records:
+        result = mask_builder.build(item, config, tokenizer)
+        if result is None:
+            continue
+        result.pop("domain", None)
+        if not primary_ids(result):
+            continue
+        yield result
+
+
+def to_per_record_tensors(
+    raw: Dict[str, list],
+) -> Dict[str, List[torch.Tensor]]:
+    """Convert an accumulated ``{key: [per-record ids]}`` dict to tensors.
+
+    Handles three shapes transparently:
+
+    - ``List[int]`` per record (``sequence``, ``chosen``…) → one tensor per record.
+    - ``List[List[int]]`` per record (GRPO ``responses``/``masks``) → one
+      ``List[Tensor]`` per record (nested), preserving the per-response
+      boundary so downstream code can index responses individually.
+    - ``List[int]`` for the whole shard (pre-packed keys) → single tensor.
+
+    The detection mirrors the previous inline logic in
+    ``Pipeline._flush`` and ``TokenizeTransform.apply``.
+    """
+    tensors: Dict[str, List[torch.Tensor]] = {}
+    for key, ids_list in raw.items():
+        if ids_list and isinstance(ids_list[0], list):
+            tensors[key] = [
+                [torch.tensor(sub, dtype=infer_dtype(sub)) for sub in ids]
+                if ids and isinstance(ids[0], list)
+                else torch.tensor(ids, dtype=infer_dtype(ids))
+                for ids in ids_list
+            ]
+        else:
+            tensors[key] = [
+                torch.tensor(list(chain.from_iterable(ids_list)), dtype=torch.int32)
+            ]
+    return tensors
+
+
+def build_position_ids(
+    sequences: List[List[int]],
+    strategy,
+) -> Optional[List[int]]:
+    """Generate position ids for *sequences* using *strategy*.
+
+    Returns ``None`` when the strategy produces no ids (e.g. ``none``
+    mode), so callers can skip attaching the key instead of storing
+    an empty list.
+    """
+    pos_ids = strategy.generate(sequences)
+    return pos_ids or None

astrai/preprocessing/packing.py

@@ -19,6 +19,43 @@ def _truncate(seq: List[int], max_len: int, mode: str) -> List[int]:
     return seq[:max_len]
 
 
+def plan_bfd(
+    sequences: List[List[int]], max_packed_len: int, truncation_mode: str = "keep_start"
+) -> List[List[int]]:
+    """Best-Fit Decreasing bin packing of *sequences* into bins.
+
+    Returns a list of bins, each bin a list of original indices into
+    *sequences*.  Bin capacities are respected on the *truncated*
+    length of each sequence (so a sequence longer than
+    *max_packed_len* counts at *max_packed_len*).
+
+    Pure index-based so callers can apply the same plan to any
+    aligned key (``loss_mask``, ``position_ids``…).
+    """
+    n = len(sequences)
+    order = sorted(range(n), key=lambda i: len(sequences[i]), reverse=True)
+    bins: List[List[int]] = []
+    bin_lengths: List[int] = []
+
+    for orig_idx in order:
+        seq_len = len(_truncate(sequences[orig_idx], max_packed_len, truncation_mode))
+        best_bin = None
+        best_remain = max_packed_len + 1
+        for i, bl in enumerate(bin_lengths):
+            remain = max_packed_len - bl
+            if seq_len <= remain < best_remain:
+                best_remain = remain
+                best_bin = i
+        if best_bin is not None:
+            bins[best_bin].append(orig_idx)
+            bin_lengths[best_bin] += seq_len
+        else:
+            bins.append([orig_idx])
+            bin_lengths.append(seq_len)
+
+    return bins
+
+
 class PackingStrategy(ABC):
     """Reorder and truncate sequences within a shard."""
 
@@ -70,7 +107,7 @@ class BFDPacking(PackingStrategy):
         sequences = keys.get("sequence", [])
         if not sequences:
             return keys
-        bins = self._plan(sequences, max_packed_len, truncation_mode)
+        bins = plan_bfd(sequences, max_packed_len, truncation_mode)
 
         packed: Dict[str, List[List[int]]] = {}
         for k, vals in keys.items():
@@ -91,31 +128,49 @@ class BFDPacking(PackingStrategy):
             result.extend(vals[i])
         return result
 
+
+@PackingStrategyFactory.register("bfd_split")
+class BFDSplitPacking(BFDPacking):
+    """BFD packing with over-length sequences split into chunks.
+
+    Sequences longer than *max_packed_len* are split into consecutive
+    chunks of at most *max_packed_len* tokens instead of being
+    truncated.  Each chunk becomes an independent sequence that enters
+    BFD planning.  All keys (``loss_mask``, ``position_ids``, …) are
+    split in lockstep so per-token alignment is preserved.
+
+    Note: because each chunk is treated as a separate document, the
+    second chunk of a split sequence loses the preceding context.
+    """
+
+    def apply(
+        self,
+        keys: Dict[str, List[List[int]]],
+        max_packed_len: int,
+        truncation_mode: str,
+    ) -> Dict[str, List[List[int]]]:
+        sequences = keys.get("sequence", [])
+        if not sequences:
+            return keys
+        if max_packed_len <= 0:
+            return super().apply(keys, max_packed_len, truncation_mode)
+
+        split_keys = self._split_all(keys, max_packed_len)
+        return super().apply(split_keys, max_packed_len, truncation_mode)
+
     @staticmethod
-    def _plan(
-        sequences: List[List[int]], max_packed_len: int, truncation_mode: str
-    ) -> List[List[int]]:
-        n = len(sequences)
-        order = sorted(range(n), key=lambda i: len(sequences[i]), reverse=True)
-        bins: List[List[int]] = []
-        bin_lengths: List[int] = []
-
-        for orig_idx in order:
-            seq_len = len(
-                _truncate(sequences[orig_idx], max_packed_len, truncation_mode)
-            )
-            best_bin = None
-            best_remain = max_packed_len + 1
-            for i, bl in enumerate(bin_lengths):
-                remain = max_packed_len - bl
-                if seq_len <= remain < best_remain:
-                    best_remain = remain
-                    best_bin = i
-            if best_bin is not None:
-                bins[best_bin].append(orig_idx)
-                bin_lengths[best_bin] += seq_len
-            else:
-                bins.append([orig_idx])
-                bin_lengths.append(seq_len)
-
-        return bins
+    def _split_all(
+        keys: Dict[str, List[List[int]]], max_packed_len: int
+    ) -> Dict[str, List[List[int]]]:
+        """Split every sequence exceeding *max_packed_len* into chunks,
+        applying the same chunk boundaries to all keys."""
+        sequences = keys["sequence"]
+        chunk_bounds = [list(range(0, len(s), max_packed_len)) for s in sequences]
+        result: Dict[str, List[List[int]]] = {}
+        for key, vals in keys.items():
+            split_vals: List[List[int]] = []
+            for val, starts in zip(vals, chunk_bounds):
+                for start in starts:
+                    split_vals.append(val[start : start + max_packed_len])
+            result[key] = split_vals
+        return result

astrai/preprocessing/pipeline.py

@@ -1,9 +1,13 @@
 """Config-driven JSONL preprocessing pipeline.
 
 Composes a :class:`BaseMaskBuilder` (selected by ``input.type``) with
-sharding and flush to ``.h5`` / ``.bin`` storage.  Packing, position-id
+sharding and flush to ``.bin`` storage.  Packing, position-id
 generation and storage writing are each delegated to pluggable strategies,
 dispatched by configuration keys.
+
+Record iteration, mask building, primary-id extraction and per-key
+accumulation are shared with :class:`TokenizeTransform` via the
+:mod:`astrai.preprocessing.core` helpers.
 """
 
 import json
@@ -17,11 +21,12 @@ import torch
 import tqdm
 
 from astrai.config.preprocess_config import PipelineConfig
-from astrai.preprocessing.builder import MaskBuilderFactory
+from astrai.preprocessing.core import (
+    build_preprocessing_components,
+    primary_ids,
+)
 from astrai.preprocessing.packing import PackingStrategyFactory
-from astrai.preprocessing.position_id import PositionIdStrategyFactory
 from astrai.preprocessing.writer import StoreWriterFactory
-from astrai.tokenize import AutoTokenizer
 
 logger = logging.getLogger(__name__)
 
@@ -64,20 +69,21 @@ class Pipeline:
         self.output_dir = output_dir
         self.tokenizer_path = tokenizer_path
 
-        self.mask_builder = MaskBuilderFactory.create("sectioned")
+        self.tokenizer, self.mask_builder, self._position_id = (
+            build_preprocessing_components(config, tokenizer_path)
+        )
         self._packer = PackingStrategyFactory.create(
             config.preprocessing.packing_strategy
         )
-        self._position_id = PositionIdStrategyFactory.create(
-            config.output.position_ids_mode
-        )
         self._writer = StoreWriterFactory.create(config.output.storage_format)
 
     def transform(self, item: dict) -> Optional[dict]:
-        return self.mask_builder.build(item, self.config, self._tokenizer)
+        return self.mask_builder.build(item, self.config, self.tokenizer)
+
+    def transform_batch(self, items: list[dict]) -> list[Optional[dict]]:
+        return self.mask_builder.build_batch(items, self.config, self.tokenizer)
 
     def run(self):
-        self._tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_path)
         domains: dict = defaultdict(lambda: defaultdict(list))
         total_tokens = 0
         shard_idx: dict[str, int] = defaultdict(int)
@@ -85,59 +91,59 @@ class Pipeline:
 
         pp = self.config.preprocessing
 
-        for item in tqdm.tqdm(
-            self._iter_items(), desc="Tokenizing", unit="docs", mininterval=0.5
-        ):
-            if pp.max_items and count >= pp.max_items:
-                break
-
+        progress = tqdm.tqdm(desc="Tokenizing", unit="docs", mininterval=0.5)
+        stop = False
+        for items in self._iter_batches(pp.batch_size):
+            progress.update(len(items))
             try:
-                result = self.transform(item)
+                results = self.transform_batch(items)
             except Exception:
                 logger.warning(
-                    "Failed to process item #%d, skipping", count + 1, exc_info=True
+                    "Failed to process batch, retrying records individually",
+                    exc_info=True,
                 )
-                continue
-            if result is None:
-                continue
+                results = []
+                for item in items:
+                    try:
+                        results.append(self.transform(item))
+                    except Exception:
+                        logger.warning(
+                            "Failed to process item, skipping", exc_info=True
+                        )
+                        results.append(None)
 
-            domain = result.pop("domain", "__default__")
+            for result in results:
+                if pp.max_items and count >= pp.max_items:
+                    stop = True
+                    break
+                if result is None:
+                    continue
 
-            is_multi = bool(getattr(self.config.input, "sources", None))
-            if is_multi:
-                ids = self._primary_ids(result)
-            else:
-                ids = result.pop("sequence")
-                result["sequence"] = ids
+                domain = result.pop("domain", "__default__")
+                ids = primary_ids(result)
+                if not ids:
+                    continue
 
-            if not ids:
-                continue
+                bucket = domains[domain]
+                self._align_bucket(bucket, result, ids)
+                for key, val in result.items():
+                    bucket[key].append(val)
 
-            bucket = domains[domain]
-            self._align_bucket(bucket, result, ids)
-            for key, val in result.items():
-                bucket[key].append(val)
+                count += 1
+                total_tokens += len(ids)
 
-            count += 1
-            total_tokens += len(ids)
+                if total_tokens >= self.config.output.max_tokens_per_shard:
+                    self._flush(domains, shard_idx)
+                    domains.clear()
+                    total_tokens = 0
+            if stop:
+                break
 
-            if total_tokens >= self.config.output.max_tokens_per_shard:
-                self._flush(domains, shard_idx)
-                domains.clear()
-                total_tokens = 0
+        progress.close()
 
         if total_tokens > 0:
             self._flush(domains, shard_idx)
 
-    @staticmethod
-    def _primary_ids(result: dict) -> list:
-        """Return the first list-valued entry in *result* as the primary id
-        sequence for token counting."""
-        for val in result.values():
-            if isinstance(val, list) and val and isinstance(val[0], int):
-                return val
-        return []
-
     @staticmethod
     def _align_bucket(bucket: dict, result: dict, ids: list):
         """Pad previously-accumulated keys that are missing from *result*."""
@@ -149,11 +155,29 @@ class Pipeline:
     def _iter_items(self):
         for path in self.paths:
             with open(path, "r", encoding="utf-8") as f:
-                for line in f:
-                    line = line.strip()
-                    if not line:
-                        continue
-                    yield json.loads(line)
+                if path.endswith(".json"):
+                    data = json.load(f)
+                    if isinstance(data, dict):
+                        yield data
+                    elif isinstance(data, list):
+                        yield from data
+                else:
+                    for line in f:
+                        line = line.strip()
+                        if not line:
+                            continue
+                        yield json.loads(line)
+
+    def _iter_batches(self, batch_size: int):
+        batch_size = max(1, batch_size)
+        batch = []
+        for item in self._iter_items():
+            batch.append(item)
+            if len(batch) >= batch_size:
+                yield batch
+                batch = []
+        if batch:
+            yield batch
 
     def _flush(self, domains, shard_idx):
         for domain, keys in domains.items():
@@ -163,24 +187,12 @@ class Pipeline:
             original_sequences = keys.get("sequence", [])
             mode = self.config.output.position_ids_mode
 
-            if mode == "doc_reset" and original_sequences:
-                keys["position_ids"] = [list(range(len(s))) for s in original_sequences]
-
+            keys = self._inject_doc_reset_position_ids(keys, mode, original_sequences)
             keys = self._packer.apply(dict(keys), pp.max_packed_len, pp.truncation_mode)
-
-            tensors: Dict[str, List[torch.Tensor]] = {}
-            for key, ids_list in keys.items():
-                dt = _STR_TO_DTYPE.get(
-                    self.config.output.dtype.get(key, "int32"), torch.int32
-                )
-                tensors[key] = [
-                    torch.tensor(list(chain.from_iterable(ids_list)), dtype=dt)
-                ]
-
-            if mode == "continuous" and original_sequences:
-                pos_ids = self._position_id.generate(keys.get("sequence", []))
-                if pos_ids:
-                    tensors["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
+            tensors = self._to_tensors(keys)
+            tensors = self._inject_continuous_position_ids(
+                tensors, mode, keys.get("sequence", [])
+            )
 
             self._writer.save(self.output_dir, domain, idx, tensors)
             shard_idx[domain] = idx + 1
@@ -190,3 +202,76 @@ class Pipeline:
                 f"  saved {domain}/shard_{idx:04d}  "
                 f"({tensors[first_key][0].numel():,} tokens)"
             )
+
+    def _inject_doc_reset_position_ids(
+        self,
+        keys: Dict[str, list],
+        mode: str,
+        original_sequences: List[List[int]],
+    ) -> Dict[str, list]:
+        """Attach per-document position_ids before packing (``doc_reset``).
+
+        ``doc_reset`` position ids must enter the packer so that each
+        packed bin concatenates the per-doc ranges in bin order.  The
+        per-record structure ``[range(len(s)) for s in seqs]`` is required
+        by the packer (it concatenates per-record lists per bin); the
+        ``PositionIdStrategy.generate`` flattens, so it cannot be used
+        directly here — it is only consulted for the ``continuous``
+        post-packing path.
+        """
+        if mode != "doc_reset" or not original_sequences:
+            return keys
+        keys["position_ids"] = [list(range(len(s))) for s in original_sequences]
+        return keys
+
+    def _inject_continuous_position_ids(
+        self,
+        tensors: Dict[str, List[torch.Tensor]],
+        mode: str,
+        packed_sequences: List[List[int]],
+    ) -> Dict[str, List[torch.Tensor]]:
+        """Attach a single continuous position_ids tensor after packing.
+
+        ``continuous`` mode spans the whole shard (post-packing), so it
+        cannot participate in bin packing — it is computed from the
+        packed sequences and appended directly to the tensor dict.
+        """
+        if mode != "continuous" or not packed_sequences:
+            return tensors
+        pos_ids = self._position_id.generate(packed_sequences)
+        if pos_ids:
+            tensors["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
+        return tensors
+
+    def _to_tensors(self, keys: Dict[str, list]) -> Dict[str, List[torch.Tensor]]:
+        """Convert packed per-key id lists to tensors.
+
+        Honours ``config.output.dtype`` overrides per key; falls back to
+        ``int32``.  Handles three shapes (see
+        :func:`astrai.preprocessing.core.to_per_record_tensors` for the
+        equivalent online-path helper):
+        - ``List[int]`` per record → one tensor per record.
+        - ``List[List[int]]`` per record (GRPO responses/masks) → one tensor
+          per record, inner lists flattened.
+        - ``List[int]`` for the whole shard (pre-packed keys) → single tensor.
+        """
+        tensors: Dict[str, List[torch.Tensor]] = {}
+        for key, ids_list in keys.items():
+            dt = _STR_TO_DTYPE.get(
+                self.config.output.dtype.get(key, "int32"), torch.int32
+            )
+            if ids_list and isinstance(ids_list[0], list):
+                tensors[key] = [
+                    torch.tensor(
+                        list(chain.from_iterable(ids))
+                        if ids and isinstance(ids[0], list)
+                        else ids,
+                        dtype=dt,
+                    )
+                    for ids in ids_list
+                ]
+            else:
+                tensors[key] = [
+                    torch.tensor(list(chain.from_iterable(ids_list)), dtype=dt)
+                ]
+        return tensors

astrai/preprocessing/transform.py

@@ -0,0 +1,92 @@
+"""Tokenization transform for JSONL record streams.
+
+Bridges the Reader layer (``JsonlStore`` reads raw JSON records) and the
+Dataset layer (expects per-record tensors).  Holds the tokenizer,
+mask-builder and position-id strategy together so that I/O code stays
+free of model dependencies.
+
+The record-processing core (mask building, primary-id extraction,
+per-key tensorisation, position-id generation) is shared with
+:class:`astrai.preprocessing.pipeline.Pipeline` via the
+:mod:`astrai.preprocessing.core` helpers.
+"""
+
+import json
+from pathlib import Path
+from typing import Dict, List
+
+import torch
+
+from astrai.config.preprocess_config import PipelineConfig
+from astrai.preprocessing.core import (
+    build_position_ids,
+    build_preprocessing_components,
+    iter_raw_records,
+    to_per_record_tensors,
+)
+
+
+class TokenizeTransform:
+    """Tokenize raw JSONL record dicts into per-key tensor lists.
+
+    Owns the three preprocessing concerns that were previously inlined in
+    ``JsonlStore``: tokenization, loss-mask construction and position-id
+    generation.  Constructing it loads the tokenizer, so it is intentionally
+    cheap to pass around once built.
+
+    Args:
+        config: Pipeline config describing sections / masks / position mode.
+        tokenizer_path: Path passed to ``AutoTokenizer.from_pretrained``.
+    """
+
+    def __init__(self, config: PipelineConfig, tokenizer_path: str):
+        self.config = config
+        self.tokenizer, self.mask_builder, self.position_strategy = (
+            build_preprocessing_components(config, tokenizer_path)
+        )
+
+    @classmethod
+    def from_config_file(cls, config_path: str) -> "TokenizeTransform":
+        """Build from a ``dataset_config.json`` file path.
+
+        The config file follows :class:`PipelineConfig` schema with an
+        extra ``tokenizer_path`` field.  When omitted, the config's
+        parent directory is used as the tokenizer path.
+        """
+        root = Path(config_path).parent
+        with open(config_path, "r", encoding="utf-8") as f:
+            raw_config = json.load(f)
+        tokenizer_path = raw_config.pop("tokenizer_path", None) or str(root)
+        config = PipelineConfig.from_dict(raw_config)
+        return cls(config, tokenizer_path)
+
+    def apply(self, records: List[dict]) -> Dict[str, list]:
+        """Tokenize a list of raw record dicts.
+
+        Returns a dict mapping key (``sequence``, ``chosen``, ``responses``,
+        …) to a list of per-record tensors (or nested tensor lists for
+        multi-response keys such as GRPO ``responses``).
+        """
+        raw: Dict[str, list] = {}
+        doc_sequences: List[List[int]] = []
+
+        for result in iter_raw_records(
+            records, self.mask_builder, self.config, self.tokenizer
+        ):
+            primary = None
+            for val in result.values():
+                if isinstance(val, list) and val and isinstance(val[0], int):
+                    primary = val
+                    break
+            if primary is not None:
+                doc_sequences.append(primary)
+            for key, ids in result.items():
+                raw.setdefault(key, []).append(ids)
+
+        tensors = to_per_record_tensors(raw)
+
+        pos_ids = build_position_ids(doc_sequences, self.position_strategy)
+        if pos_ids is not None:
+            tensors["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
+
+        return tensors

astrai/preprocessing/writer.py

@@ -1,7 +1,7 @@
 """Storage writer strategies for pipeline output.
 
 The :class:`StoreWriter` abstraction decouples the pipeline from the
-concrete storage format (bin / h5).  The pipeline builds a ``{key:
+concrete storage format (bin).  The pipeline builds a ``{key:
 List[Tensor]}`` dict and delegates the write to the writer selected
 by ``output.storage_format``.
 """
@@ -15,7 +15,7 @@ from typing import Dict, List
 import torch
 
 from astrai.factory import BaseFactory
-from astrai.serialization import save_bin, save_h5
+from astrai.serialization import save_bin
 
 logger = logging.getLogger(__name__)
 
@@ -54,22 +54,3 @@ class BinWriter(StoreWriter):
                 exc_info=True,
             )
             raise
-
-
-@StoreWriterFactory.register("h5")
-class H5Writer(StoreWriter):
-    def save(self, output_dir, domain, shard_idx, tensors):
-        chunk_dir = os.path.join(output_dir, domain)
-        file_path = os.path.join(chunk_dir, f"data_{shard_idx:04d}.h5")
-        try:
-            save_h5(chunk_dir, f"data_{shard_idx:04d}", tensors)
-        except Exception:
-            if os.path.exists(file_path):
-                os.remove(file_path)
-            logger.error(
-                "Failed to write shard %s/data_%04d.h5, cleaned up partial output",
-                domain,
-                shard_idx,
-                exc_info=True,
-            )
-            raise

astrai/serialization/__init__.py

@@ -19,9 +19,8 @@ from astrai.serialization.checkpoint import (
 )
 from astrai.serialization.dataset import (
     load_bin,
-    load_h5,
+    load_bin_offsets,
     save_bin,
-    save_h5,
 )
 
 __all__ = [
@@ -37,7 +36,6 @@ __all__ = [
     "save_safetensors",
     "save_torch",
     "load_bin",
-    "load_h5",
+    "load_bin_offsets",
     "save_bin",
-    "save_h5",
 ]

astrai/serialization/checkpoint.py

@@ -2,7 +2,6 @@
 
 import io
 import json
-import os
 import time
 from dataclasses import dataclass, field
 from pathlib import Path
@@ -148,9 +147,6 @@ class Checkpoint:
         save_path = Path(save_dir)
         save_path.mkdir(parents=True, exist_ok=True)
 
-        if get_rank() != 0:
-            return
-
         meta = {
             "epoch": self.epoch,
             "consumed_samples": self.consumed_samples,
@@ -181,6 +177,7 @@ class Checkpoint:
             epoch=meta.get("epoch", 0),
             consumed_samples=meta.get("consumed_samples", 0),
             extra=extra,
+            meta=meta,
             config=config,
         )
 

astrai/serialization/dataset.py

@@ -1,58 +1,51 @@
-"""Dataset storage serialization helpers (HDF5 / memory-mapped binary)."""
+"""Dataset storage serialization helpers (memory-mapped binary)."""
 
 import json
 import os
-from pathlib import Path
-from typing import Dict, List
+from typing import Any, Dict, List, Optional
 
-import h5py
 import numpy as np
 import torch
 from torch import Tensor
 
 
-def save_h5(file_path: str, file_name: str, tensor_group: Dict[str, List[Tensor]]):
-    os.makedirs(file_path, exist_ok=True)
-    full_file_path = os.path.join(file_path, f"{file_name}.h5")
-    with h5py.File(full_file_path, "w") as f:
-        for key, tensors in tensor_group.items():
-            grp = f.create_group(key)
-            for idx, tensor in enumerate(tensors):
-                arr = tensor.cpu().numpy()
-                grp.create_dataset(f"data_{idx}", data=arr)
-
-
-def load_h5(file_path: str, share_memory=True) -> Dict[str, List[Tensor]]:
-    tensor_group: Dict[str, List[Tensor]] = {}
-
-    root_path = Path(file_path)
-    h5_files = list(root_path.rglob("*.h5")) + list(root_path.rglob("*.hdf5"))
-
-    for h5_file in h5_files:
-        with h5py.File(h5_file, "r") as f:
-            for key in f.keys():
-                grp = f[key]
-                dsets = []
-                for dset_name in grp.keys():
-                    dset = grp[dset_name]
-                    tensor = torch.from_numpy(dset[:])
-                    if share_memory:
-                        tensor = tensor.share_memory_()
-                    dsets.append(tensor)
-
-                if tensor_group.get(key) is None:
-                    tensor_group[key] = []
-                tensor_group[key].extend(dsets)
-
-    return tensor_group
-
-
-def save_bin(file_path: str, tensor_group: Dict[str, List[Tensor]]):
+def save_bin(
+    file_path: str,
+    tensor_group: Dict[str, List[Tensor]],
+    record_keys: Optional[List[str]] = None,
+):
+    """Save tensors as memory-mapped binary files.
+
+    When *record_keys* is provided, those keys are written with per-record
+    cumulative offsets in ``meta.json`` so that ``MmapStore.fetch_record``
+    can slice individual records from the concatenated binary without
+    cross-record concatenation.  Keys not in *record_keys* (e.g. SEQ
+    ``sequence``) are written as a single contiguous stream without
+    offsets, preserving backward compatibility.
+
+    Nested keys (``List[List[Tensor]]`` such as GRPO ``responses``) are
+    not supported in bin format — use JSONL for those.
+    """
     os.makedirs(file_path, exist_ok=True)
+    record_keys = set(record_keys or [])
     meta = {}
     for key, tensors in tensor_group.items():
+        if tensors and isinstance(tensors[0], list):
+            raise ValueError(
+                f"Nested key '{key}' (List[List[Tensor]]) is not supported "
+                f"in bin format. Use JSONL storage instead."
+            )
         cat = torch.cat(tensors, dim=0)
-        meta[key] = {"shape": list(cat.shape), "dtype": str(cat.dtype).split(".")[-1]}
+        entry: Dict[str, Any] = {
+            "shape": list(cat.shape),
+            "dtype": str(cat.dtype).split(".")[-1],
+        }
+        if key in record_keys:
+            offsets = [0]
+            for t in tensors:
+                offsets.append(offsets[-1] + t.shape[0])
+            entry["offsets"] = offsets
+        meta[key] = entry
         np.asarray(cat.cpu().numpy()).tofile(os.path.join(file_path, f"{key}.bin"))
     with open(os.path.join(file_path, "meta.json"), "w") as f:
         json.dump(meta, f)
@@ -66,8 +59,24 @@ def load_bin(file_path: str) -> Dict[str, List[Tensor]]:
         arr = np.memmap(
             os.path.join(file_path, f"{key}.bin"),
             dtype=info["dtype"],
-            mode="r+",
+            mode="c",
             shape=tuple(info["shape"]),
         )
         segments[key] = [torch.from_numpy(arr)]
     return segments
+
+
+def load_bin_offsets(file_path: str) -> Dict[str, List[int]]:
+    """Read per-record cumulative offsets from ``meta.json``.
+
+    Returns an empty dict when no key has offsets (legacy bin files),
+    in which case record-mode access falls back to per-record segment
+    indexing (JSONL layout).
+    """
+    with open(os.path.join(file_path, "meta.json"), "r") as f:
+        meta = json.load(f)
+    offsets: Dict[str, List[int]] = {}
+    for key, info in meta.items():
+        if "offsets" in info:
+            offsets[key] = info["offsets"]
+    return offsets

astrai/signal_handler.py

@@ -0,0 +1,53 @@
+import logging
+import os
+import signal
+import threading
+
+logger = logging.getLogger(__name__)
+
+_early_stop = threading.Event()
+_active_context = None
+
+
+def _early_handler(signum: int, frame):
+    sig = signal.Signals(signum)
+    logger.warning(
+        "Received %s (pid=%d), requesting graceful training stop...",
+        sig.name,
+        os.getpid(),
+    )
+    _early_stop.set()
+    if _active_context is not None:
+        _active_context.request_stop()
+
+
+def install_early_signal_handlers():
+    for sig in (signal.SIGTERM, signal.SIGINT):
+        signal.signal(sig, _early_handler)
+    _unblock_signals()
+
+
+def _unblock_signals():
+    try:
+        mask = signal.pthread_sigmask(signal.SIG_BLOCK, set())
+        blocked = {signal.SIGTERM, signal.SIGINT} & mask
+        if blocked:
+            signal.pthread_sigmask(signal.SIG_UNBLOCK, blocked)
+    except (AttributeError, OSError):
+        pass
+
+
+def register_signal_handlers(context):
+    global _active_context
+    _active_context = context
+    for sig in (signal.SIGTERM, signal.SIGINT):
+        signal.signal(sig, _early_handler)
+    if _early_stop.is_set():
+        context.request_stop()
+        logger.warning("Signal was received during initialization, stopping...")
+
+
+def unregister_signal_handlers():
+    global _active_context
+    _active_context = None
+    _early_stop.clear()

astrai/tokenize/__init__.py

@@ -1,8 +1,10 @@
 from astrai.tokenize.chat_template import ChatTemplate, MessageType
-from astrai.tokenize.tokenizer import AutoTokenizer
+from astrai.tokenize.tokenizer import AutoTokenizer, Message, Messages
 
 __all__ = [
     "AutoTokenizer",
     "ChatTemplate",
     "MessageType",
+    "Message",
+    "Messages",
 ]

astrai/tokenize/chat_template.py

@@ -1,3 +1,4 @@
+from functools import cached_property
 from typing import Any, Dict, List, Optional
 
 from jinja2 import Template
@@ -29,7 +30,34 @@ class ChatTemplate:
         self.description = description
         self.default_variables = default_variables or {}
         self.special_tokens = special_tokens or {}
-        self._compiled: Template = Template(template_str)
+
+    @cached_property
+    def _compiled(self) -> Template:
+        """Lazy-compiled Jinja2 template, cached on first access.
+
+        The compiled :class:`~jinja2.Template` holds a dynamically-generated
+        ``root`` render function whose ``__module__`` is ``None``; under
+        ``pickle`` it falls back to ``__main__`` and breaks ``spawn``-based
+        multiprocessing.  :meth:`__getstate__` drops the cached template so
+        that pickle serialises only ``template_str``; each worker rebuilds
+        the cache on first render.
+        """
+        return Template(self.template_str)
+
+    def __getstate__(self) -> Dict[str, Any]:
+        """Exclude the cached Jinja2 template from pickling.
+
+        ``Template.root_render_func`` is a dynamically generated closure
+        that cannot be pickled by reference.  Dropping ``_compiled`` here
+        lets :class:`cached_property` rebuild it on first access after
+        unpickle.
+        """
+        state = self.__dict__.copy()
+        state.pop("_compiled", None)
+        return state
+
+    def __setstate__(self, state: Dict[str, Any]) -> None:
+        self.__dict__.update(state)
 
     @classmethod
     def from_string(

astrai/tokenize/tokenizer.py

@@ -10,12 +10,16 @@ from tokenizers import Tokenizer
 
 from astrai.tokenize.chat_template import ChatTemplate
 
+Message = Dict[str, str]
+"""Single chat message with ``role`` and ``content`` keys."""
+
+Messages = List[Message]
+"""Single conversation — a list of messages."""
+
 
 class AutoTokenizer:
     """Base tokenizer class with automatic loading support"""
 
-    TOKENIZER_CLASSES = {}  # Registry for auto-loading
-
     def __init__(
         self,
         path: Optional[Union[str, Path]] = None,
@@ -102,17 +106,6 @@ class AutoTokenizer:
         with open(save_path / "tokenizer_config.json", "w", encoding="utf-8") as f:
             json.dump(config, f, ensure_ascii=False, indent=2)
 
-    @classmethod
-    def register_tokenizer(cls, name: str, tokenizer_class: type):
-        """
-        Register a new tokenizer class.
-
-        Args:
-            name: Name to register the tokenizer class under
-            tokenizer_class: The tokenizer class to register
-        """
-        cls.TOKENIZER_CLASSES[name] = tokenizer_class
-
     def encode(
         self,
         tokens: Union[str, List[str]],
@@ -120,7 +113,16 @@ class AutoTokenizer:
         is_pretokenized: bool = False,
         add_special_tokens: bool = True,
     ) -> List:
-        """Encode text to tokens or token IDs."""
+        """Encode text to token IDs.
+
+        Accepts both single strings and batches:
+
+        - ``encode("hello")`` → ``[123, 456]``
+        - ``encode(["hello", "world"])`` → ``[[123, 456], [789]]``
+
+        Batches are tokenised in parallel via the Rust backend's
+        ``encode_batch`` (uses all available CPU cores).
+        """
         if self._tokenizer is None:
             raise RuntimeError(
                 "Tokenizer not initialized. Load or create a tokenizer first."
@@ -133,15 +135,13 @@ class AutoTokenizer:
                 add_special_tokens=add_special_tokens,
             )
             return encoded.ids if out_ids else encoded.tokens
-        else:
-            encoded_list = self._tokenizer.encode_batch(
-                tokens,
-                is_pretokenized=is_pretokenized,
-                add_special_tokens=add_special_tokens,
-            )
-            return [
-                encoded.ids if out_ids else encoded.tokens for encoded in encoded_list
-            ]
+
+        encoded_list = self._tokenizer.encode_batch(
+            tokens,
+            is_pretokenized=is_pretokenized,
+            add_special_tokens=add_special_tokens,
+        )
+        return [encoded.ids if out_ids else encoded.tokens for encoded in encoded_list]
 
     def decode(self, tokens: List[int], skip_special_tokens: bool = True) -> str:
         """Decode token IDs to text."""
@@ -164,7 +164,14 @@ class AutoTokenizer:
           - tokenizer.bos_token   → returns string
           - tokenizer.bos_token_id → returns corresponding integer ID
           - tokenizer.stop_ids → returns list of corresponding integer IDs for all special tokens
+
+        Internal/private attrs are not intercepted: during unpickle
+        ``__dict__`` is empty, so probing ``self._special_token_map``
+        would recurse infinitely.
         """
+        if key.startswith("_"):
+            raise AttributeError(key)
+
         # Handle stop_ids - return IDs for all special tokens
         if key == "stop_ids":
             stop_ids = []
@@ -220,45 +227,63 @@ class AutoTokenizer:
 
     def apply_chat_template(
         self,
-        messages: List[Dict[str, str]],
+        messages: Union[Messages, List[Messages]],
         system_prompt: Optional[str] = None,
         tokenize: bool = True,
         add_generation_prompt: bool = True,
         **kwargs,
-    ) -> Union[str, List[int]]:
-        """
-        Apply the chat template to messages and optionally tokenize the result.
+    ) -> Union[str, List[int], List[str], List[List[int]]]:
+        """Apply the chat template and optionally tokenize.
+
+        Accepts both single conversations and batches:
+
+        - ``apply_chat_template([msg1, msg2])`` → ``"..."`` or ``[ids]``
+        - ``apply_chat_template([[msg1, msg2], [msg3]])`` → ``["..", ".."]``
+          or ``[[ids], [ids]]``
+
+        Batches render each conversation list and tokenise all at once via
+        :meth:`encode` (``List[str]`` → Rust parallel ``encode_batch``).
 
         Args:
-            messages: List of message dicts with 'role' and 'content'.
-            system_prompt: Optional system prompt string (auto-converted to first message).
+            messages: Single conversation (``Messages``) or batch of
+                conversations (``BatchMessages``).
+            system_prompt: Optional system prompt prepended (single mode only).
             tokenize: Whether to return token IDs (True) or raw string (False).
-            add_generation_prompt: Whether to add the generation prompt (default: True).
-            **kwargs: Additional variables to pass to the template.
+            add_generation_prompt: Whether to add the generation prompt.
+            **kwargs: Additional template variables.
 
         Returns:
-            Either the rendered string or list of token IDs.
-
-        Raises:
-            RuntimeError: If chat template is not set.
+            Single mode: ``str`` or ``List[int]``.
+            Batch mode: ``List[str]`` or ``List[List[int]]``.
         """
         if self._chat_template is None:
             raise RuntimeError(
                 "Chat template not set. Use set_chat_template() to set a template first."
             )
 
-        # Auto-convert system_prompt to first message if provided
+        is_batch = bool(messages) and isinstance(messages[0], list)
+
+        if is_batch:
+            rendered = [
+                self._chat_template.render(
+                    messages=msgs,
+                    add_generation_prompt=add_generation_prompt,
+                    **kwargs,
+                )
+                for msgs in messages
+            ]
+            if tokenize:
+                return self.encode(rendered)  # List[str] → batch encode
+            return rendered
+
+        # Single conversation
         if system_prompt:
             messages = [{"role": "system", "content": system_prompt}] + list(messages)
-
-        # Render the template
         rendered = self._chat_template.render(
             messages=messages,
             add_generation_prompt=add_generation_prompt,
             **kwargs,
         )
-
         if tokenize:
             return self.encode(rendered)
-
         return rendered

astrai/trainer/metric_util.py

@@ -22,6 +22,51 @@ def grad_norm(model: nn.Module, per_param: bool = False) -> float | Dict[str, fl
     return total_sq.sqrt().item()
 
 
+class GradSNRTracker:
+    """Track gradient signal-to-noise ratio via EMA of first/second moments.
+
+    SNR = E[g]^2 / Var(g) = E[g]^2 / (E[g^2] - E[g]^2)
+
+    The tracker accumulates per-parameter EMA moments across optimizer steps.
+    Call ``update`` after backward (before ``optimizer.step``) and read
+    ``snr`` to get the aggregate SNR across all parameters.
+    """
+
+    def __init__(self, beta: float = 0.999, eps: float = 1e-8):
+        self.beta = beta
+        self.eps = eps
+        self._first: Dict[int, torch.Tensor] = {}
+        self._second: Dict[int, torch.Tensor] = {}
+
+    @torch.no_grad()
+    def update(self, model: nn.Module) -> None:
+        beta = self.beta
+        for param in model.parameters():
+            if param.grad is None:
+                continue
+            pid = id(param)
+            g = param.grad.detach()
+            if pid not in self._first:
+                self._first[pid] = g.clone()
+                self._second[pid] = g.pow(2).clone()
+            else:
+                self._first[pid].mul_(beta).add_(g, alpha=1 - beta)
+                self._second[pid].mul_(beta).addcmul_(g, g, value=1 - beta)
+
+    @property
+    def snr(self) -> float:
+        if not self._first:
+            return 0.0
+        total_signal = 0.0
+        total_noise = 0.0
+        for m, v in zip(self._first.values(), self._second.values()):
+            signal = m.pow(2).sum().item()
+            noise = (v - m.pow(2)).clamp(min=0).sum().item()
+            total_signal += signal
+            total_noise += noise
+        return total_signal / (total_noise + self.eps)
+
+
 def ctx_get_loss(ctx):
     return ctx.loss
 
@@ -36,3 +81,30 @@ def ctx_get_val_loss(ctx):
 
 def ctx_get_grad_norm(ctx):
     return ctx.grad_norm
+
+
+def ctx_get_grad_snr(ctx):
+    tracker = getattr(ctx, "grad_snr_tracker", None)
+    if tracker is None:
+        return None
+    return tracker.snr
+
+
+def ctx_get_moe_aux_loss(ctx):
+    return ctx.strategy._moe_metrics.get("aux_loss")
+
+
+def ctx_get_router_entropy(ctx):
+    return ctx.strategy._moe_metrics.get("router_entropy")
+
+
+def ctx_get_dead_expert_fraction(ctx):
+    return ctx.strategy._moe_metrics.get("dead_expert_fraction")
+
+
+def ctx_get_load_imbalance_mean(ctx):
+    return ctx.strategy._moe_metrics.get("load_imbalance_mean")
+
+
+def ctx_get_load_imbalance_max(ctx):
+    return ctx.strategy._moe_metrics.get("load_imbalance_max")

astrai/trainer/rollout.py

@@ -0,0 +1,421 @@
+"""Online rollout runner for RL training.
+
+Provides:
+- :class:`RawRollout` — generation output container (no reward yet)
+- :class:`RolloutResult` — a :class:`RawRollout` with rewards attached
+- :class:`BaseRewardModel` — pluggable reward interface
+- :class:`RolloutGenerator` — KV-cache-backed generation of grouped
+  responses + decoding (no reward); delegates the generation loop to
+  :class:`~astrai.inference.core.scheduler.InferenceScheduler.run_batch`
+  so rollout and the production inference server share one code path
+- :class:`RolloutRunner` — orchestrates generation + scoring with a
+  step-driven cache; its ``__call__`` returns ``(RolloutResult, is_fresh)``
+  so callers do not need to rely on object identity to detect refreshes.
+"""
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Tuple
+
+import torch
+from torch import Tensor
+
+from astrai.inference.core.scheduler import InferenceScheduler
+
+
+@dataclass(kw_only=True)
+class RawRollout:
+    """Generation output before reward scoring.
+
+    Produced by :class:`RolloutGenerator`; consumed by :class:`RolloutRunner`
+    to assemble a :class:`RolloutResult` once rewards are attached.
+
+    Fields are designed to cover all common RL algorithms:
+    GRPO, PPO, Online DPO, Rejection Sampling, etc.
+
+    Fields:
+        prompts: Tokenized prompts, shape ``[B, P_len]``.
+        prompt_mask: Boolean mask for real prompt tokens, shape ``[B, P_len]``.
+        responses: Generated response token IDs, shape ``[B, G, R_max]``.
+        response_mask: Boolean mask for real (non-pad) response tokens,
+            shape ``[B, G, R_max]``.
+        logprobs_old: Per-token log-probs under the behaviour policy,
+            shape ``[B, G, R_max]``.
+        prompt_texts: Decoded prompt strings (for reward models that
+            need text).
+        response_texts: Decoded response strings, shape ``[B, G]``
+            (for reward models).
+    """
+
+    prompts: Tensor
+    prompt_mask: Tensor
+    responses: Tensor
+    response_mask: Tensor
+    logprobs_old: Tensor
+    prompt_texts: List[str] = field(default_factory=list)
+    response_texts: List[List[str]] = field(default_factory=list)
+
+
+@dataclass(kw_only=True)
+class RolloutResult(RawRollout):
+    """A :class:`RawRollout` with reward scoring attached.
+
+    Produced by :class:`RolloutRunner` once the :class:`BaseRewardModel`
+    has scored the decoded responses.
+
+    Fields:
+        rewards: Reward per response, shape ``[B, G]``.
+    """
+
+    rewards: Tensor
+
+
+class BaseRewardModel(ABC):
+    """Pluggable reward model interface.
+
+    Subclasses should implement ``score()`` to return a ``[B, G]`` float
+    tensor of rewards.  Implementations can be:
+    * A loaded reward model (e.g. ArmoRM, Skywork-Reward)
+    * An external API call
+    * A rule-based function (format, length, keyword matching)
+    """
+
+    @abstractmethod
+    def score(self, prompts: List[str], responses: List[List[str]]) -> Tensor:
+        """Score each generated response.
+
+        Args:
+            prompts: Raw prompt strings, length ``B``.
+            responses: Generated response strings, shape ``[B, G]``.
+
+        Returns:
+            Float tensor of shape ``[B, G]``.
+        """
+        ...
+
+
+_PAD = 0
+
+
+class RolloutGenerator:
+    """Pure generation + decoding for a group of responses per prompt.
+
+    Delegates the prefill/decode loop to
+    :meth:`~astrai.inference.core.scheduler.InferenceScheduler.run_batch`,
+    which uses a real KV cache (no O(n²) recompute).  Has no dependency
+    on any reward model; can be reused in isolation for offline
+    generation, qualitative sampling, or eval pipelines.
+    """
+
+    def __init__(
+        self,
+        scheduler: InferenceScheduler,
+        tokenizer,
+        max_tokens: int = 1024,
+        group_size: int = 8,
+        temperature: float = 1.0,
+        top_k: int = 0,
+        top_p: float = 1.0,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
+    ):
+        self.scheduler = scheduler
+        self.tokenizer = tokenizer
+        self.max_tokens = max_tokens
+        self.group_size = group_size
+        self.temperature = temperature
+        self.top_k = top_k
+        self.top_p = top_p
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
+
+    @torch.no_grad()
+    def generate(self, batch: Dict) -> RawRollout:
+        """Expand prompts by ``group_size`` and generate one response each.
+
+        Accepted batch formats (per sample, repeated B times):
+
+        - **messages**: ``{"messages": [{"role": "user", "content": "..."}, ...]}``
+        - **instruction + input + output**: ``{"instruction": "...",
+          "input": "...", "output": "..."}`` — mapped to ``system`` /
+          ``user`` / ``assistant`` messages; ``input`` and ``output``
+          are optional and skipped when empty.
+
+        Both are rendered through the tokenizer's chat template with
+        ``add_generation_prompt=True`` so rollout prompts match the
+        format the policy was SFT-trained on.
+        """
+        model = self.scheduler._executor.model
+        was_training = model.training
+        model.eval()
+        try:
+            return self._generate_eval(batch)
+        finally:
+            model.train(was_training)
+
+    def _generate_eval(self, batch: Dict) -> RawRollout:
+        prompt_texts, flat_prompt_ids = self._prepare_prompts(batch)
+        B = len(prompt_texts)
+        G = self.group_size
+        # Re-expand flat list to G copies per prompt for run_batch.
+        expanded_prompt_ids: List[List[int]] = []
+        for ids in flat_prompt_ids:
+            expanded_prompt_ids.extend([list(ids)] * G)
+
+        results = self.scheduler.run_batch(
+            expanded_prompt_ids,
+            max_tokens=self.max_tokens,
+            temperature=self.temperature,
+            top_k=self.top_k,
+            top_p=self.top_p,
+            frequency_penalty=self.frequency_penalty,
+            rep_window=self.rep_window,
+            return_logprobs=True,
+        )
+        if len(results) != B * G:
+            raise RuntimeError(
+                f"Rollout scheduler returned {len(results)} results, expected {B * G}"
+            )
+        for token_ids, logprobs in results:
+            if len(token_ids) != len(logprobs):
+                raise RuntimeError(
+                    "Rollout scheduler returned misaligned token IDs and logprobs"
+                )
+
+        # Each element is (token_ids, logprobs); pad to max length.
+        max_len = 0
+        for token_ids, _lp in results:
+            max_len = max(max_len, len(token_ids))
+        max_len = max(max_len, 1)
+
+        device = self.scheduler.device
+        P_len = max(len(ids) for ids in flat_prompt_ids)
+        prompts_tensor = torch.zeros(B, P_len, dtype=torch.long, device=device)
+        prompt_mask = torch.zeros(B, P_len, dtype=torch.bool, device=device)
+        for i, ids in enumerate(flat_prompt_ids):
+            prompts_tensor[i, -len(ids) :] = torch.tensor(
+                ids, dtype=torch.long, device=device
+            )
+            prompt_mask[i, -len(ids) :] = True
+
+        responses = torch.full((B, G, max_len), _PAD, dtype=torch.long, device=device)
+        response_mask = torch.zeros((B, G, max_len), dtype=torch.bool, device=device)
+        logprobs_old = torch.zeros((B, G, max_len), dtype=torch.float, device=device)
+
+        flat_idx = 0
+        response_texts: List[List[str]] = [[] for _ in range(B)]
+        for i in range(B):
+            for g in range(G):
+                token_ids, lps = results[flat_idx]
+                flat_idx += 1
+                n = len(token_ids)
+                if n:
+                    responses[i, g, :n] = torch.tensor(
+                        token_ids, dtype=torch.long, device=device
+                    )
+                    response_mask[i, g, :n] = True
+                    logprobs_old[i, g, :n] = torch.tensor(
+                        lps, dtype=torch.float, device=device
+                    )
+                response_texts[i].append(
+                    self.tokenizer.decode(token_ids, skip_special_tokens=True)
+                )
+
+        return RawRollout(
+            prompts=prompts_tensor,
+            prompt_mask=prompt_mask,
+            responses=responses,
+            response_mask=response_mask,
+            logprobs_old=logprobs_old,
+            prompt_texts=prompt_texts,
+            response_texts=response_texts,
+        )
+
+    def _prepare_prompts(self, batch: Dict) -> Tuple[List[str], List[List[int]]]:
+        """Render batch prompts to ``(texts, token_id_lists)``.
+
+        Returns two parallel lists of length B (number of prompts in
+        the batch).  Dispatches by batch keys:
+
+        - ``"messages"``: treated as a pre-built message list per sample.
+        - ``"instruction"`` (optionally ``"input"`` and ``"output"``): mapped
+          to ``system`` / ``user`` / ``assistant`` messages respectively.
+
+        Both paths go through the tokenizer's chat template with
+        ``add_generation_prompt=True``.
+        """
+        if "messages" in batch:
+            messages_list = batch["messages"]
+        elif "instruction" in batch:
+            instructions = batch["instruction"]
+            B = len(instructions)
+            inputs = batch.get("input") or [""] * B
+            outputs = batch.get("output") or [""] * B
+            messages_list = [
+                self._instruction_to_messages(i, u, o)
+                for i, u, o in zip(instructions, inputs, outputs)
+            ]
+        else:
+            raise ValueError(
+                "Rollout batch must contain either 'messages' or "
+                "'instruction' (optionally 'input'/'output'); got keys: "
+                f"{list(batch.keys())}"
+            )
+
+        try:
+            prompt_texts = self.tokenizer.apply_chat_template(
+                messages_list, tokenize=False, add_generation_prompt=True
+            )
+            if (
+                not isinstance(prompt_texts, list)
+                or len(prompt_texts) != len(messages_list)
+                or not all(isinstance(text, str) for text in prompt_texts)
+            ):
+                raise TypeError("Tokenizer does not support batched chat templates")
+            flat_prompt_ids = self.tokenizer.encode(prompt_texts)
+            if len(flat_prompt_ids) != len(messages_list) or not all(
+                isinstance(ids, list) for ids in flat_prompt_ids
+            ):
+                raise TypeError("Tokenizer does not support batched encoding")
+        except (TypeError, IndexError, KeyError):
+            # Keep compatibility with lightweight tokenizer adapters that only
+            # implement the single-conversation template API.
+            prompt_texts = []
+            flat_prompt_ids = []
+            for messages in messages_list:
+                text = self.tokenizer.apply_chat_template(
+                    messages, tokenize=False, add_generation_prompt=True
+                )
+                ids = self.tokenizer.apply_chat_template(
+                    messages, tokenize=True, add_generation_prompt=True
+                )
+                prompt_texts.append(text)
+                flat_prompt_ids.append(list(ids))
+        return prompt_texts, flat_prompt_ids
+
+    @staticmethod
+    def _instruction_to_messages(
+        instruction: str, inp: str = "", output: str = ""
+    ) -> List[Dict[str, str]]:
+        """Map instruction/input/output to chat messages.
+
+        Role mapping follows the convention used throughout the
+        preprocessing pipeline: ``instruction`` → system, ``input`` →
+        user, ``output`` → assistant.  Empty fields are skipped so a
+        bare instruction produces a ``[system]`` list and the chat
+        template's ``add_generation_prompt`` adds the assistant header
+        for sampling.
+        """
+        messages: List[Dict[str, str]] = []
+        if instruction:
+            messages.append({"role": "system", "content": instruction})
+        if inp:
+            messages.append({"role": "user", "content": inp})
+        if output:
+            messages.append({"role": "assistant", "content": output})
+        return messages
+
+
+class RolloutRunner:
+    """Produces :class:`RolloutResult` from a prompt batch.
+
+    Composes a :class:`RolloutGenerator` (generation + decoding) with a
+    :class:`BaseRewardModel` (scoring).  Maintains an internal cache so
+    the same batch prompt can be replayed for multiple gradient steps.
+    A new rollout is triggered every ``rollout_interval`` calls to
+    :meth:`step` (or after :meth:`clear_cache`).
+
+    The ``__call__`` contract returns a ``(RolloutResult, is_fresh)``
+    tuple — callers must use the boolean to detect a refreshed rollout
+    rather than relying on object identity.
+
+    Usage::
+
+        generator = RolloutGenerator(policy, tokenizer, pipeline, ...)
+        runner = RolloutRunner(generator, reward_model, rollout_interval=512)
+        result, is_fresh = runner(prompt_batch)
+        if is_fresh:
+            ...  # e.g. sync behaviour policy
+    """
+
+    def __init__(
+        self,
+        generator: RolloutGenerator,
+        reward_model: BaseRewardModel,
+        rollout_interval: int = 512,
+    ):
+        self.generator = generator
+        self.reward_model = reward_model
+        self.rollout_interval = rollout_interval
+
+        self._cache: Optional[RolloutResult] = None
+        self._cache_key = None
+        self._steps_since_rollout: int = 0
+
+    def step(self):
+        """Advance the internal counter (call once per optimizer step)."""
+        self._steps_since_rollout += 1
+
+    def clear_cache(self):
+        """Force next call to re-run rollout."""
+        self._cache = None
+        self._cache_key = None
+
+    @staticmethod
+    def _batch_key(batch: Dict):
+        """Build a stable key for the prompt fields accepted by the generator."""
+
+        def freeze(value):
+            if isinstance(value, dict):
+                return tuple(sorted((key, freeze(val)) for key, val in value.items()))
+            if isinstance(value, (list, tuple)):
+                return tuple(freeze(item) for item in value)
+            return value
+
+        fields = ("messages", "instruction", "input", "output")
+        return tuple(
+            (field, freeze(batch[field])) for field in fields if field in batch
+        )
+
+    def _score(self, raw: RawRollout) -> RolloutResult:
+        rewards = self.reward_model.score(raw.prompt_texts, raw.response_texts)
+        if not isinstance(rewards, Tensor):
+            rewards = torch.as_tensor(rewards, dtype=torch.float32)
+        expected_shape = raw.responses.shape[:2]
+        if rewards.shape != expected_shape:
+            raise ValueError(
+                f"Reward model returned shape {tuple(rewards.shape)}, "
+                f"expected {tuple(expected_shape)}"
+            )
+        if not torch.isfinite(rewards).all():
+            raise ValueError("Reward model returned non-finite values")
+        device = raw.prompts.device
+        return RolloutResult(
+            prompts=raw.prompts,
+            prompt_mask=raw.prompt_mask,
+            responses=raw.responses,
+            response_mask=raw.response_mask,
+            rewards=rewards.to(device=device),
+            logprobs_old=raw.logprobs_old,
+            prompt_texts=raw.prompt_texts,
+            response_texts=raw.response_texts,
+        )
+
+    def __call__(self, batch: Dict[str, Tensor]) -> Tuple[RolloutResult, bool]:
+        """Return ``(cached or fresh) RolloutResult`` plus an ``is_fresh`` flag.
+
+        Triggers a new rollout when ``_steps_since_rollout >= rollout_interval``
+        or when the cache is empty.
+        """
+        cache_key = self._batch_key(batch)
+        if (
+            self._cache is None
+            or cache_key != self._cache_key
+            or self._steps_since_rollout >= self.rollout_interval
+        ):
+            raw = self.generator.generate(batch)
+            self._cache = self._score(raw)
+            self._cache_key = cache_key
+            self._steps_since_rollout = 0
+            return self._cache, True
+        return self._cache, False

astrai/trainer/strategy.py

@@ -1,7 +1,7 @@
 """Training strategy implementations with factory pattern."""
 
 from abc import ABC, abstractmethod
-from typing import Callable, Dict, Union
+from typing import Callable, Dict, List, Optional, TypedDict, Union
 
 import torch
 import torch.nn as nn
@@ -9,17 +9,20 @@ import torch.nn.functional as F
 from torch import Tensor
 
 from astrai.factory import BaseFactory
+from astrai.model.components.mlp import RouterStats
+from astrai.parallel.executor import broadcast_state_dict
+from astrai.trainer.rollout import RolloutResult
 
 
-def create_ref_model(
-    model_fn: Callable[[], nn.Module], state_dict: Dict[str, Tensor]
-) -> nn.Module:
-    """Create a frozen reference model from model_fn + full state dict."""
-    ref_model = model_fn()
-    ref_model.load_state_dict(state_dict)
-    ref_model.requires_grad_(False)
-    ref_model.eval()
-    return ref_model
+class LossOutput(TypedDict):
+    loss: Tensor
+    metrics: Dict[str, float]
+
+
+class LogprobsOutput(TypedDict):
+    logprobs: Tensor
+    aux_loss: Optional[Tensor]
+    router_stats: Optional[List[RouterStats]]
 
 
 def move_to_device(batch: Dict[str, Tensor], device: str) -> Dict[str, Tensor]:
@@ -28,17 +31,19 @@ def move_to_device(batch: Dict[str, Tensor], device: str) -> Dict[str, Tensor]:
 
 
 def get_logprobs(
-    model: Union[nn.Module, Callable[..., Dict[str, Tensor]]],
+    model: nn.Module,
     input_ids: Tensor,
-    mask: Tensor,
+    attn_mask: Tensor,
+    loss_mask: Tensor,
     reduction: str,
-) -> Tensor:
+) -> LogprobsOutput:
     """Compute token-wise log probabilities from model outputs.
 
     Args:
         model: The language model
         input_ids: Input token IDs of shape [batch_size, seq_len]
-        mask: Attention mask of shape [batch_size, seq_len]
+        attn_mask: Attention mask passed to the model (may include causal).
+        loss_mask: Per-token mask for loss reduction.
         reduction: How to reduce over sequence dimension ("mean", "sum", "none")
 
     Returns:
@@ -51,9 +56,13 @@ def get_logprobs(
         )
 
     shifted_input_ids = input_ids[:, 1:]
-    shifted_mask = mask[:, 1:]
+    shifted_loss_mask = loss_mask[:, 1:]
 
-    logits = model(input_ids[:, :-1], mask[:, :-1])["logits"]
+    outputs = model(
+        input_ids[:, :-1],
+        attn_mask[:, :, :-1, :-1] if attn_mask.dim() == 4 else attn_mask[:, :-1],
+    )
+    logits = outputs["logits"]
     log_probs = torch.log_softmax(logits.float(), dim=-1)
 
     token_logprobs = torch.gather(
@@ -61,13 +70,18 @@ def get_logprobs(
     ).squeeze(-1)
 
     if reduction == "mean":
-        return (token_logprobs * shifted_mask).sum(dim=-1) / shifted_mask.sum(
+        logprobs = (token_logprobs * shifted_loss_mask).sum(
             dim=-1
-        ).clamp(min=1.0)
+        ) / shifted_loss_mask.sum(dim=-1).clamp(min=1.0)
     elif reduction == "sum":
-        return (token_logprobs * shifted_mask).sum(dim=-1)
+        logprobs = (token_logprobs * shifted_loss_mask).sum(dim=-1)
     else:
-        return token_logprobs * shifted_mask
+        logprobs = token_logprobs * shifted_loss_mask
+    return {
+        "logprobs": logprobs,
+        "aux_loss": outputs.get("aux_loss"),
+        "router_stats": outputs.get("router_stats"),
+    }
 
 
 def make_doc_boundary_mask(position_ids: Tensor) -> Tensor:
@@ -86,8 +100,78 @@ def make_doc_boundary_mask(position_ids: Tensor) -> Tensor:
     return (same_doc & causal).unsqueeze(1)
 
 
+def _collect_moe_diagnostics(
+    router_stats_list: List[RouterStats],
+) -> Dict[str, float]:
+    """Collect MoE routing diagnostic metrics from per-layer router stats.
+
+    Args:
+        router_stats_list: One :class:`RouterStats` dict per MoE layer with
+            keys ``probs`` (N, E) and ``topk_indices`` (N, K), both detached.
+
+    Returns:
+        Dict with keys: router_entropy, dead_expert_fraction,
+        load_imbalance_mean, load_imbalance_max.  Values are averaged
+        across layers.
+    """
+    layer_entropies: List[Tensor] = []
+    layer_dead_fractions: List[Tensor] = []
+    layer_imbalance_means: List[Tensor] = []
+    layer_imbalance_maxs: List[Tensor] = []
+
+    for stats in router_stats_list:
+        probs = stats["probs"].float()
+        topk_indices = stats["topk_indices"]
+        num_experts = probs.shape[-1]
+        if num_experts == 0:
+            continue
+        probs = probs.reshape(-1, num_experts)
+        if probs.numel() == 0:
+            continue
+
+        # Router entropy
+        entropy = -(probs * torch.log(probs.clamp_min(1e-8))).sum(dim=-1).mean()
+
+        # Load from the actual dispatch: one-hot sum of top-k assignments.
+        expert_counts = F.one_hot(topk_indices, num_experts).sum(dim=(0, 1)).float()
+        ideal_load = expert_counts.mean()  # N*K / E
+        load_ratios = expert_counts / max(float(ideal_load), 1.0)
+        imbalance_mean = (load_ratios - 1.0).abs().mean()
+        imbalance_max = load_ratios.max()
+        dead_fraction = (expert_counts == 0).float().mean()
+
+        layer_entropies.append(entropy)
+        layer_dead_fractions.append(dead_fraction)
+        layer_imbalance_means.append(imbalance_mean)
+        layer_imbalance_maxs.append(imbalance_max)
+
+    if not layer_entropies:
+        return {}
+
+    return {
+        "router_entropy": float(torch.stack(layer_entropies).mean().cpu().item()),
+        "dead_expert_fraction": float(
+            torch.stack(layer_dead_fractions).mean().cpu().item()
+        ),
+        "load_imbalance_mean": float(
+            torch.stack(layer_imbalance_means).mean().cpu().item()
+        ),
+        "load_imbalance_max": float(
+            torch.stack(layer_imbalance_maxs).mean().cpu().item()
+        ),
+    }
+
+
 class BaseStrategy(ABC):
-    """Abstract base class for training strategies."""
+    """Abstract base class for training strategies.
+
+    When a :class:`~astrai.trainer.rollout.RolloutRunner` is injected via
+    :meth:`set_rollout_runner`, the strategy transparently switches to
+    online mode: each ``__call__`` produces a :class:`RolloutResult`,
+    converts it to a training batch via :meth:`prepare_from_rollout`, and
+    then computes the loss.  Without a runner the strategy runs in
+    offline mode and consumes the batch directly.
+    """
 
     def __init__(
         self,
@@ -98,8 +182,10 @@ class BaseStrategy(ABC):
         self.model = model
         self.device = device
         self.executor = kwargs.pop("executor", None)
-        self.model_fn = kwargs.pop("model_fn", None)
+        self.moe_aux_loss_coef = kwargs.pop("moe_aux_loss_coef", 0.01)
+        self._moe_metrics: Dict[str, float] = {}
         self.extra_kwargs = kwargs
+        self._rollout_runner = None
 
     @abstractmethod
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
@@ -113,9 +199,96 @@ class BaseStrategy(ABC):
         """
         raise NotImplementedError
 
-    def __call__(self, batch: Dict[str, Tensor]) -> Tensor:
-        """Allow calling strategy directly as a callable."""
-        return self.compute_loss(batch)
+    def compute_loss_output(self, batch: Dict[str, Tensor]) -> LossOutput:
+        return self._normalize_output(self.compute_loss(batch))
+
+    def _loss_output(
+        self,
+        task_loss: Tensor,
+        metrics: Dict[str, Tensor],
+        aux_loss: Optional[Tensor] = None,
+        router_stats: Optional[List[RouterStats]] = None,
+    ) -> LossOutput:
+        total_loss = task_loss
+        if aux_loss is not None:
+            weighted_aux_loss = self.moe_aux_loss_coef * aux_loss
+            total_loss = total_loss + weighted_aux_loss
+            metrics["moe_aux_loss"] = aux_loss
+            metrics["moe_aux_loss_weighted"] = weighted_aux_loss
+            self._refresh_moe_diagnostics(aux_loss, router_stats)
+        metrics["loss"] = total_loss
+        return {
+            "loss": total_loss,
+            "metrics": {name: value.detach().item() for name, value in metrics.items()},
+        }
+
+    @staticmethod
+    def _normalize_output(output: Union[LossOutput, Tensor]) -> LossOutput:
+        if isinstance(output, dict):
+            return output
+        return {"loss": output, "metrics": {"loss": output.detach().item()}}
+
+    def supports_online(self) -> bool:
+        """Whether this strategy can operate with a rollout runner.
+
+        Base implementation returns ``False``; strategies that implement
+        :meth:`prepare_from_rollout` should override to return ``True``.
+        """
+        return False
+
+    def set_rollout_runner(self, runner):
+        """Inject a :class:`RolloutRunner` to enable online rollout mode."""
+        self._rollout_runner = runner
+
+    def prepare_from_rollout(self, result: RolloutResult) -> Dict[str, Tensor]:
+        """Map a :class:`RolloutResult` to the batch layout expected by
+        :meth:`compute_loss`.
+
+        Strategies that return ``True`` from :meth:`supports_online` must
+        override this.  Default raises :class:`NotImplementedError`.
+        """
+        raise NotImplementedError(
+            f"{type(self).__name__} does not support online rollout"
+        )
+
+    def _on_rollout_refresh(self):
+        """Hook fired when a fresh rollout result is produced.
+
+        Override to refresh stale state (e.g. syncing the behaviour
+        policy).  Default is a no-op.
+        """
+        pass
+
+    def _refresh_moe_diagnostics(
+        self,
+        aux_loss: Tensor,
+        router_stats: Optional[List[RouterStats]] = None,
+    ) -> None:
+        """Collect MoE routing diagnostics from the latest forward pass.
+
+        Populates ``self._moe_metrics`` with router entropy, dead expert
+        fraction, load imbalance, and aux_loss.  Called from
+        :meth:`_loss_output` when an MoE aux loss is present.
+        """
+        self._moe_metrics = _collect_moe_diagnostics(router_stats or [])
+        self._moe_metrics["aux_loss"] = float(aux_loss.detach().cpu().item())
+
+    def on_optimizer_step(self):
+        """Advance online rollout state after a successful optimizer step."""
+        if self._rollout_runner is not None:
+            self._rollout_runner.step()
+
+    def __call__(self, batch: Dict[str, Tensor]) -> LossOutput:
+        """Run offline or online forward depending on runner injection."""
+        if self._rollout_runner is None:
+            return self.compute_loss_output(batch)
+
+        result, is_fresh = self._rollout_runner(batch)
+        if is_fresh:
+            self._on_rollout_refresh()
+
+        train_batch = self.prepare_from_rollout(result)
+        return self.compute_loss_output(train_batch)
 
 
 class StrategyFactory(BaseFactory["BaseStrategy"]):
@@ -142,6 +315,7 @@ class SEQStrategy(BaseStrategy):
     """Standard next-token prediction training strategy.
 
     Computes cross-entropy loss for next token prediction.
+    Optionally adds MoE load balancing auxiliary loss.
     """
 
     def __init__(
@@ -155,9 +329,13 @@ class SEQStrategy(BaseStrategy):
         self.label_smoothing = label_smoothing
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
+        return self.compute_loss_output(batch)["loss"]
+
+    def compute_loss_output(self, batch: Dict[str, Tensor]) -> LossOutput:
         batch = move_to_device(batch, self.device)
         input_ids, target_ids = batch["input_ids"], batch["target_ids"]
-        logits = self.model(input_ids=input_ids)["logits"]
+        outputs = self.model(input_ids=input_ids)
+        logits = outputs["logits"]
 
         loss = F.cross_entropy(
             input=logits.flatten(0, 1).float(),
@@ -165,7 +343,12 @@ class SEQStrategy(BaseStrategy):
             label_smoothing=self.label_smoothing,
         )
 
-        return loss
+        return self._loss_output(
+            loss,
+            {"task_loss": loss},
+            outputs.get("aux_loss"),
+            outputs.get("router_stats"),
+        )
 
 
 @StrategyFactory.register("sft")
@@ -173,6 +356,7 @@ class SFTStrategy(BaseStrategy):
     """Supervised Fine-tuning strategy with loss masking.
 
     Applies cross-entropy loss only to tokens where loss_mask is True.
+    Optionally adds MoE load balancing auxiliary loss.
     """
 
     def __init__(
@@ -186,6 +370,9 @@ class SFTStrategy(BaseStrategy):
         self.label_smoothing = label_smoothing
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
+        return self.compute_loss_output(batch)["loss"]
+
+    def compute_loss_output(self, batch: Dict[str, Tensor]) -> LossOutput:
         batch = move_to_device(batch, self.device)
         input_ids, target_ids, position_ids, loss_mask = (
             batch["input_ids"],
@@ -197,9 +384,10 @@ class SFTStrategy(BaseStrategy):
         ignore_index = -100
         input_mask = make_doc_boundary_mask(position_ids)
         target_ids = target_ids.masked_fill(~loss_mask, ignore_index)
-        logits = self.model(
+        outputs = self.model(
             input_ids=input_ids, position_ids=position_ids, input_mask=input_mask
-        )["logits"]
+        )
+        logits = outputs["logits"]
 
         loss = F.cross_entropy(
             input=logits.flatten(0, 1).float(),
@@ -208,7 +396,12 @@ class SFTStrategy(BaseStrategy):
             label_smoothing=self.label_smoothing,
         )
 
-        return loss
+        return self._loss_output(
+            loss,
+            {"task_loss": loss},
+            outputs.get("aux_loss"),
+            outputs.get("router_stats"),
+        )
 
 
 @StrategyFactory.register("dpo")
@@ -223,31 +416,54 @@ class DPOStrategy(BaseStrategy):
         self,
         model: nn.Module,
         device: str,
+        ref_model: nn.Module,
         beta: float = 0.1,
-        reduction: str = "mean",
+        reduction: str = "sum",
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(
-            self.model_fn, self.executor.unwrap_model(model)
-        ).to(device=self.device)
+        self.ref_model = ref_model
         self.beta = beta
         self.reduction = reduction
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
+        return self.compute_loss_output(batch)["loss"]
+
+    def compute_loss_output(self, batch: Dict[str, Tensor]) -> LossOutput:
         batch = move_to_device(batch, self.device)
         chosen_ids, rejected_ids = batch["chosen"], batch["rejected"]
         chosen_mask, rejected_mask = batch["chosen_mask"], batch["rejected_mask"]
 
         concat_ids = torch.cat([chosen_ids, rejected_ids], dim=0)
-        concat_mask = torch.cat([chosen_mask, rejected_mask], dim=0)
+        concat_loss_mask = torch.cat([chosen_mask, rejected_mask], dim=0)
 
-        log_pi = get_logprobs(self.model, concat_ids, concat_mask, self.reduction)
+        # Build full attention mask: key-padding + causal
+        key_pad = concat_ids.bool()[:, None, None, :]  # [B*2, 1, 1, S]
+        S = key_pad.shape[-1]
+        causal = torch.tril(
+            torch.ones(S, S, dtype=torch.bool, device=concat_ids.device)
+        )[None, None, :, :]  # [1, 1, S, S]
+        full_mask = key_pad & causal  # [B*2, 1, S, S] — composed
+
+        policy_output = get_logprobs(
+            self.model,
+            concat_ids,
+            full_mask,
+            concat_loss_mask,
+            self.reduction,
+        )
+        log_pi = policy_output["logprobs"]
+        aux_loss = policy_output["aux_loss"]
 
         with torch.no_grad():
-            log_ref = get_logprobs(
-                self.ref_model, concat_ids, concat_mask, self.reduction
+            ref_output = get_logprobs(
+                self.ref_model,
+                concat_ids,
+                full_mask,
+                concat_loss_mask,
+                self.reduction,
             )
+            log_ref = ref_output["logprobs"]
 
         log_pi_chosen = log_pi[: chosen_ids.shape[0]]
         log_pi_rejected = log_pi[chosen_ids.shape[0] :]
@@ -260,49 +476,87 @@ class DPOStrategy(BaseStrategy):
         ratio_diff = pi_log_ratio - ref_log_ratio
         dpo_loss = -F.logsigmoid(self.beta * ratio_diff).mean()
 
-        return dpo_loss
+        return self._loss_output(
+            dpo_loss,
+            {"dpo_loss": dpo_loss},
+            aux_loss,
+            policy_output.get("router_stats"),
+        )
+
+    def supports_online(self) -> bool:
+        return True
+
+    def prepare_from_rollout(self, result: RolloutResult) -> Dict[str, Tensor]:
+        """Pick best/worst response per prompt by reward as chosen/rejected."""
+        rewards = result.rewards
+        responses = result.responses
+        masks = result.response_mask
+        best = rewards.argmax(dim=-1)
+        worst = rewards.argmin(dim=-1)
+        B = responses.shape[0]
+        idx = torch.arange(B, device=responses.device)
+        chosen = responses[idx, best]
+        chosen_mask = masks[idx, best].float()
+        rejected = responses[idx, worst]
+        rejected_mask = masks[idx, worst].float()
+        return {
+            "chosen": chosen,
+            "chosen_mask": chosen_mask,
+            "rejected": rejected,
+            "rejected_mask": rejected_mask,
+        }
 
 
 @StrategyFactory.register("grpo")
 class GRPOStrategy(BaseStrategy):
     """Group Relative Policy Optimization strategy.
 
-    On-policy GRPO following DeepSeek-R1: the policy model is updated while
-    a frozen ref_model stores the old-policy log-probs.  ratio = exp(logπ_θ - logπ_ref),
-    clipped PPO objective.  Call ``sync_ref_model()`` after each data-generation round.
+    Implements GRPO following DeepSeek-R1 with token-level PPO clipping.
+    Advantages are group-normalized from scalar per-response rewards and
+    broadcast across all response tokens.  The loss is computed **only on
+    response tokens** — prompt tokens are masked out.
+
+    Three model roles are distinguished:
+
+    * **Policy** ``self.model`` — the model being trained.
+    * **Old policy** ``self.old_model`` — the behaviour policy that generated
+      the responses.  Used for the importance sampling ratio
+      ``ρ = π_θ / π_old``.  Synced externally after each data-generation round.
+    * **Reference model** ``self.ref_model`` — a frozen copy of the initial
+      policy (typically the SFT checkpoint) used **only** for the KL
+      regularisation term.  It is never updated during training.
     """
 
     def __init__(
         self,
         model: nn.Module,
         device: str,
+        old_model: nn.Module,
+        ref_model: nn.Module,
         clip_eps: float = 0.2,
         kl_coef: float = 0.01,
         group_size: int = 4,
-        reduction: str = "mean",
-        sync_interval: int = 200,
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(
-            self.model_fn, self.executor.unwrap_model(model)
-        ).to(device=self.device)
+        self.old_model = old_model
+        self.ref_model = ref_model
         self.clip_eps = clip_eps
         self.kl_coef = kl_coef
         self.group_size = group_size
-        self.reduction = reduction
-        self.sync_interval = sync_interval
-        self._step = 0
 
-    def sync_ref_model(self):
-        """Copy current model weights to ref model."""
-        self.ref_model.load_state_dict(self.executor.unwrap_model(self.model))
+    def sync_old_model(self):
+        """Copy current policy weights to old model."""
+        state_dict = self.executor.unwrap_model(self.model)
+        if self.executor.use_distributed:
+            state_dict = broadcast_state_dict(state_dict)
+        if state_dict is not None:
+            self.old_model.load_state_dict(state_dict)
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
-        self._step += 1
-        if self._step % self.sync_interval == 0:
-            self.sync_ref_model()
+        return self.compute_loss_output(batch)["loss"]
 
+    def compute_loss_output(self, batch: Dict[str, Tensor]) -> LossOutput:
         batch = move_to_device(batch, self.device)
         prompts = batch["prompts"]
         responses = batch["responses"]
@@ -313,33 +567,112 @@ class GRPOStrategy(BaseStrategy):
         responses_flat = responses.view(-1, response_len)
         masks_flat = masks.view(-1, response_len)
         prompt_expanded = prompts.unsqueeze(1).repeat(1, group_size, 1).flatten(0, 1)
+        prompt_mask = batch.get("prompt_mask")
+        if prompt_mask is None:
+            prompt_mask = prompts.ne(0)
+        prompt_mask_expanded = (
+            prompt_mask.unsqueeze(1).expand(-1, group_size, -1).flatten(0, 1)
+        )
+        prompt_len = prompt_expanded.size(1)
 
         full_sequences = torch.cat([prompt_expanded, responses_flat], dim=-1)
-        full_masks = torch.cat([torch.ones_like(prompt_expanded), masks_flat], dim=-1)
-
-        log_probs_policy = get_logprobs(
-            self.model, full_sequences, full_masks, self.reduction
+        # Prompt tokens are masked out (0) so logprobs are computed only for
+        # response tokens.  get_logprobs shifts the mask by one position, so
+        # the first response token's logprob (predicted from the last prompt
+        # token) is correctly included.
+        full_masks = torch.cat(
+            [torch.zeros_like(prompt_expanded, dtype=torch.bool), masks_flat], dim=-1
         )
-        log_probs_policy = log_probs_policy.view(batch_size, group_size)
 
+        # Build full attention mask: key-padding + causal
+        key_pad = torch.cat([prompt_mask_expanded, masks_flat.bool()], dim=-1)[
+            :, None, None, :
+        ]
+        S = key_pad.shape[-1]
+        causal = torch.tril(
+            torch.ones(S, S, dtype=torch.bool, device=full_sequences.device)
+        )[None, None, :, :]
+        attn_mask = key_pad & causal
+
+        # get_logprobs returns [B*G, S-1] (S = prompt_len + response_len).
+        # Response token logprobs occupy the last ``response_len`` positions
+        # (the first response token is predicted from the last prompt token).
+        policy_output = get_logprobs(
+            self.model, full_sequences, attn_mask, full_masks, "none"
+        )
+        token_log_probs_policy = policy_output["logprobs"]
+        aux_loss = policy_output["aux_loss"]
+        token_log_probs_policy = token_log_probs_policy[:, prompt_len - 1 :]
         with torch.no_grad():
-            log_probs_ref = get_logprobs(
-                self.ref_model, full_sequences, full_masks, self.reduction
+            old_output = get_logprobs(
+                self.old_model, full_sequences, attn_mask, full_masks, "none"
             )
-            log_probs_ref = log_probs_ref.view(batch_size, group_size)
+            token_log_probs_old = old_output["logprobs"]
+            token_log_probs_old = token_log_probs_old[:, prompt_len - 1 :]
+            ref_output = get_logprobs(
+                self.ref_model, full_sequences, attn_mask, full_masks, "none"
+            )
+            token_log_probs_ref = ref_output["logprobs"]
+            token_log_probs_ref = token_log_probs_ref[:, prompt_len - 1 :]
 
-        eps = torch.finfo(log_probs_policy.dtype).eps
+        # Reshape to [B, G, response_len]
+        token_log_probs_policy = token_log_probs_policy.view(batch_size, group_size, -1)
+        token_log_probs_old = token_log_probs_old.view(batch_size, group_size, -1)
+        token_log_probs_ref = token_log_probs_ref.view(batch_size, group_size, -1)
+        token_masks = masks_flat.view(batch_size, group_size, -1).float()
+
+        # Group-normalized advantages from scalar per-response rewards.
+        eps = 1e-8
         mean = rewards.mean(dim=-1, keepdim=True)
-        std = rewards.std(dim=-1, keepdim=True)
+        std = rewards.std(dim=-1, keepdim=True, unbiased=False)
         advantages = (rewards - mean) / (std + eps)
+        # Broadcast scalar advantage to every response token: [B, G, 1]
+        advantages = advantages.unsqueeze(-1)
 
-        ratio = torch.exp(log_probs_policy - log_probs_ref)
+        # Token-level ratio (π_θ / π_old) and PPO clipping.
+        log_ratio = token_log_probs_policy - token_log_probs_old
+        ratio = torch.exp(log_ratio)
 
         surr1 = ratio * advantages
         surr2 = torch.clamp(ratio, 1 - self.clip_eps, 1 + self.clip_eps) * advantages
+        per_token_policy_loss = -torch.min(surr1, surr2)
+        token_count = token_masks.sum().clamp(min=1.0)
+        policy_loss = (per_token_policy_loss * token_masks).sum() / token_count
 
-        policy_loss = -torch.min(surr1, surr2).mean()
-        kl_penalty = self.kl_coef * (log_probs_policy - log_probs_ref).square().mean()
-        total_loss = policy_loss + kl_penalty
+        # KL penalty to frozen reference model with k1 estimator (non-negative):
+        # k1 = π_ref / π_θ - log(π_ref / π_θ) - 1, where π_ref / π_θ = exp(log_ref - log_policy).
+        log_ref_ratio = token_log_probs_ref - token_log_probs_policy
+        r = torch.exp(log_ref_ratio)
+        kl_per_token = r - torch.log(r + eps) - 1.0
+        kl_penalty = self.kl_coef * (kl_per_token * token_masks).sum() / token_count
 
-        return total_loss
+        task_loss = policy_loss + kl_penalty
+        return self._loss_output(
+            task_loss,
+            {"policy_loss": policy_loss, "kl_loss": kl_penalty},
+            aux_loss,
+            policy_output.get("router_stats"),
+        )
+
+    def supports_online(self) -> bool:
+        return True
+
+    def prepare_from_rollout(self, result: RolloutResult) -> Dict[str, Tensor]:
+        return {
+            "prompts": result.prompts,
+            "prompt_mask": result.prompt_mask,
+            "responses": result.responses,
+            "masks": result.response_mask,
+            "rewards": result.rewards,
+        }
+
+    def _on_rollout_refresh(self):
+        """Sync the behaviour policy whenever a fresh rollout arrives."""
+        self.sync_old_model()
+
+
+# Factory aliases: online variants use the same strategy class; the
+# ``RolloutRunner`` is injected by ``TrainContextBuilder`` to enable
+# online mode, so no separate subclass is needed.
+StrategyFactory.register("online_grpo")(GRPOStrategy)
+StrategyFactory.register("online_dpo")(DPOStrategy)

astrai/trainer/train_callback.py

@@ -14,12 +14,18 @@ from tqdm import tqdm
 
 from astrai.factory import BaseFactory
 from astrai.parallel import only_on_rank
-from astrai.parallel.setup import get_current_device, get_rank
+from astrai.parallel.setup import get_current_device
 from astrai.serialization import Checkpoint
 from astrai.trainer.metric_util import (
+    ctx_get_dead_expert_fraction,
     ctx_get_grad_norm,
+    ctx_get_grad_snr,
+    ctx_get_load_imbalance_max,
+    ctx_get_load_imbalance_mean,
     ctx_get_loss,
     ctx_get_lr,
+    ctx_get_moe_aux_loss,
+    ctx_get_router_entropy,
     ctx_get_val_loss,
 )
 from astrai.trainer.train_context import TrainContext
@@ -139,28 +145,31 @@ class CheckpointCallback(TrainCallback):
         self.interval = interval
         self.weight_only = weight_only
         self.save_extra_fn = save_extra_fn or CheckpointCallback.save_extra
-        self.last_ckpt_step = 0
+        self.last_ckpt_step = None
 
-    def _save_checkpoint(self, context: TrainContext):
-        state_dict = context.executor.unwrap_model(context.model)
+    def on_train_begin(self, context: TrainContext):
         self.last_ckpt_step = context.optimizer_step
 
-        if get_rank() == 0:
-            save_path = os.path.join(
-                self.save_dir,
-                f"epoch_{context.epoch}_step_{context.optimizer_step}",
-            )
-            extra = self.save_extra_fn(context)
-            meta = context.config.to_dict()
-            context.checkpoint = Checkpoint(
-                state_dict=state_dict,
-                epoch=context.epoch,
-                consumed_samples=context.consumed_samples,
-                config=context.model_config,
-                extra=extra,
-                meta=meta,
-            )
-            context.checkpoint.save(save_path)
+    def _save_checkpoint(self, context: TrainContext):
+        self.last_ckpt_step = context.optimizer_step
+
+        with context.executor.checkpoint_context(context.model) as state_dict:
+            if state_dict is not None:
+                save_path = os.path.join(
+                    self.save_dir,
+                    f"epoch_{context.epoch}_step_{context.optimizer_step}",
+                )
+                extra = self.save_extra_fn(context)
+                meta = context.config.to_dict()
+                context.checkpoint = Checkpoint(
+                    state_dict=state_dict,
+                    epoch=context.epoch,
+                    consumed_samples=context.consumed_samples,
+                    config=context.model_config,
+                    extra=extra,
+                    meta=meta,
+                )
+                context.checkpoint.save(save_path)
 
     def on_batch_end(self, context: TrainContext):
         if context.optimizer_step - self.last_ckpt_step >= self.interval:
@@ -209,9 +218,8 @@ class ProgressBarCallback(TrainCallback):
 
     @only_on_rank(0)
     def on_optimizer_step(self, context: TrainContext):
-        self.progress_bar.update(1)
         postfix = {
-            "step": context.optimizer_step,
+            "step": f"{context.optimizer_step:d}",
             "loss": f"{context.loss:.4f}",
             "lr": f"{context.optimizer.param_groups[-1]['lr']:.2e}",
         }
@@ -220,6 +228,7 @@ class ProgressBarCallback(TrainCallback):
         if context.val_loss is not None:
             postfix["val_loss"] = f"{context.val_loss:.4f}"
         self.progress_bar.set_postfix(postfix)
+        self.progress_bar.update(1)
 
     @only_on_rank(0)
     def on_epoch_end(self, context: TrainContext):
@@ -232,19 +241,18 @@ class ProgressBarCallback(TrainCallback):
 class MetricCallback(TrainCallback):
     def __init__(
         self,
-        log_dir: str,
+        ckpt_dir: str,
         save_interval: int,
         metrics: List[str] = None,
         val_step: int = 0,
     ):
-        self.last_log_flush_step = 0
+        self.last_log_flush_step = None
         self.save_interval = save_interval
         self.metrics = metrics or ["loss", "lr"]
         self.val_step = val_step
         self._next_val_step = 0
 
-        self.log_dir = Path(log_dir) if log_dir else Path.cwd() / "logs"
-        self.log_dir.mkdir(parents=True, exist_ok=True)
+        self.ckpt_dir = Path(ckpt_dir) if ckpt_dir else Path.cwd() / "checkpoint"
 
         self.log_cache = []
 
@@ -253,14 +261,37 @@ class MetricCallback(TrainCallback):
             "lr": ctx_get_lr,
             "val_loss": ctx_get_val_loss,
             "grad_norm": ctx_get_grad_norm,
+            "grad_snr": ctx_get_grad_snr,
+            "moe_aux_loss": ctx_get_moe_aux_loss,
+            "router_entropy": ctx_get_router_entropy,
+            "dead_expert_fraction": ctx_get_dead_expert_fraction,
+            "load_imbalance_mean": ctx_get_load_imbalance_mean,
+            "load_imbalance_max": ctx_get_load_imbalance_max,
         }
 
     def _metrics(self, context: TrainContext, names):
-        return {
-            m: self._metric_funcs[m](context)
-            for m in names
-            if self._metric_funcs[m](context) is not None
-        }
+        metrics = dict(context.metrics)
+        for name in names:
+            metric_fn = self._metric_funcs.get(name)
+            if metric_fn is None:
+                continue
+            value = metric_fn(context)
+            if value is not None:
+                metrics[name] = value
+        selected = set(context.metrics) | set(names)
+        selected.discard("*")
+        result = {name: metrics[name] for name in selected if name in metrics}
+        if context.world_size > 1 and dist.is_initialized() and result:
+            metric_names = sorted(result)
+            values = torch.tensor(
+                [result[name] for name in metric_names],
+                dtype=torch.float32,
+                device=get_current_device(),
+            )
+            dist.all_reduce(values, op=dist.ReduceOp.SUM)
+            values /= context.world_size
+            result.update(zip(metric_names, values.tolist()))
+        return result
 
     @only_on_rank(0)
     def _append(self, event_type: str, context: TrainContext, **extra):
@@ -282,8 +313,8 @@ class MetricCallback(TrainCallback):
 
         with torch.no_grad():
             for batch in context.val_dataloader:
-                loss = context.strategy(batch)
-                total_loss += loss.item()
+                loss_output = context.strategy(batch)
+                total_loss += loss_output["loss"].item()
                 num_batches += 1
 
         if context.world_size > 1 and dist.is_initialized():
@@ -298,15 +329,20 @@ class MetricCallback(TrainCallback):
         context.model.train()
         return avg_loss
 
+    def on_train_begin(self, context: TrainContext):
+        self.last_log_flush_step = context.optimizer_step
+
     @only_on_rank(0)
     def _flush(self, epoch, step):
-        log_file = self.log_dir / f"epoch_{epoch}_step_{step}_metric.jsonl"
+        log_file = self.ckpt_dir / f"epoch_{epoch}_step_{step}" / "metric.jsonl"
         log_file.parent.mkdir(parents=True, exist_ok=True)
         with open(log_file, "w") as f:
             for log in self.log_cache:
                 f.write(json.dumps(log) + "\n")
 
     def on_optimizer_step(self, context):
+        context.grad_snr_tracker.update(context.model)
+
         if (
             context.val_dataloader is not None
             and self.val_step > 0
@@ -327,8 +363,12 @@ class MetricCallback(TrainCallback):
         self._append("epoch", context)
 
     def on_train_end(self, context):
-        if context.optimizer_step != self.last_log_flush_step:
+        if (
+            self.last_log_flush_step is None
+            or context.optimizer_step != self.last_log_flush_step
+        ):
             self._flush(context.epoch, context.optimizer_step)
+            self.last_log_flush_step = context.optimizer_step
 
     def on_error(self, context):
         self._flush(context.epoch, context.optimizer_step)

astrai/trainer/train_context.py

@@ -1,3 +1,5 @@
+import logging
+import threading
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Dict, Optional, Self
@@ -7,14 +9,20 @@ import torch.nn as nn
 from torch.utils.data import DataLoader, random_split
 
 from astrai.config.train_config import TrainConfig
-from astrai.dataset import ResumableDistributedSampler
+from astrai.dataset import RDSampler
+from astrai.inference.core.scheduler import InferenceScheduler
 from astrai.model.components.lora import inject_lora
-from astrai.parallel.executor import BaseExecutor, ExecutorFactory
+from astrai.parallel.executor import BaseExecutor, ExecutorFactory, create_ref_model
 from astrai.parallel.setup import get_current_device, get_rank, get_world_size
 from astrai.protocols import OptimizerProtocol, SchedulerProtocol
 from astrai.serialization import Checkpoint, load_json
+from astrai.tokenize import AutoTokenizer
+from astrai.trainer.metric_util import GradSNRTracker
+from astrai.trainer.rollout import RolloutGenerator, RolloutRunner
 from astrai.trainer.strategy import BaseStrategy, StrategyFactory
 
+logger = logging.getLogger(__name__)
+
 
 @dataclass
 class TrainContext:
@@ -27,11 +35,12 @@ class TrainContext:
     config: TrainConfig = field(default=None)
     model_config: dict = field(default_factory=dict)
     executor: BaseExecutor = field(default=None)
-
     epoch: int = field(default=0)
     consumed_samples: int = field(default=0)
     loss: float = field(default=0.0)
+    metrics: Dict[str, float] = field(default_factory=dict)
     grad_norm: Optional[float] = field(default=None)
+    grad_snr_tracker: GradSNRTracker = field(default_factory=GradSNRTracker)
     val_dataloader: Optional[DataLoader] = field(default=None)
     val_loss: Optional[float] = field(default=None)
 
@@ -39,6 +48,15 @@ class TrainContext:
     rank: int = field(default=0)
     kwargs: Dict[str, Any] = field(default_factory=dict)
 
+    _stop_event: threading.Event = field(default_factory=threading.Event)
+
+    @property
+    def stop_requested(self) -> bool:
+        return self._stop_event.is_set()
+
+    def request_stop(self) -> None:
+        self._stop_event.set()
+
     @property
     def optimizer_step(self) -> int:
         return self.consumed_samples // (
@@ -54,10 +72,12 @@ class TrainContextBuilder:
         config: TrainConfig,
     ):
         self.config = config
-        self._resume_dir: Optional[str] = None
+        self._param_path: Optional[str] = None
+        self._resume: bool = False
 
-    def with_resume_dir(self, resume_dir: Optional[str]) -> Self:
-        self._resume_dir = resume_dir
+    def with_param_path(self, param_path: Optional[str], resume: bool = False) -> Self:
+        self._param_path = param_path
+        self._resume = resume
         return self
 
     def build(self) -> TrainContext:
@@ -70,50 +90,72 @@ class TrainContextBuilder:
             **cfg.executor_kwargs,
         )
 
-        model = cfg.model_fn()
-        model = model.to(device=device)
-
         model_config = {}
-        if self._resume_dir:
-            config_path = Path(self._resume_dir) / "config.json"
+        if self._param_path:
+            config_path = Path(self._param_path) / "config.json"
             if config_path.exists():
                 model_config = load_json(config_path)
 
-        if not model_config and hasattr(model, "config"):
-            model_config = model.config.to_dict()
+        preloaded_state_dict = None
+        preloaded_epoch = cfg.start_epoch
+        preloaded_consumed = cfg.start_samples * get_world_size()
+        preloaded_checkpoint = None
+        if self._param_path:
+            checkpoint = Checkpoint.load_any(self._param_path)
+            if checkpoint is not None:
+                preloaded_state_dict = checkpoint.state_dict
+                if checkpoint.config:
+                    model_config = checkpoint.config
+                if self._resume:
+                    preloaded_epoch = checkpoint.epoch
+                    per_step = (
+                        cfg.batch_per_device * get_world_size() * cfg.grad_accum_steps
+                    )
+                    preloaded_consumed = (
+                        checkpoint.consumed_samples // per_step
+                    ) * per_step
+                    preloaded_checkpoint = checkpoint
+
+        if not model_config and hasattr(cfg.model_fn(), "config"):
+            model_config = cfg.model_fn().config.to_dict()
+
+        def _before_wrap(m):
+            m = m.to(device=device)
+            if cfg.lora is not None:
+                inject_lora(
+                    m,
+                    r=cfg.lora.r,
+                    alpha=cfg.lora.alpha,
+                    target_modules=set(cfg.lora.target_modules),
+                )
+            if preloaded_state_dict is not None:
+                m.load_state_dict(preloaded_state_dict, strict=False)
+            return m
+
+        def _after_wrap(m):
+            if cfg.compile_mode is not None:
+                logger.info("torch.compile enabled (mode=%s)", cfg.compile_mode)
+                m = torch.compile(m, mode=cfg.compile_mode)
+            return m
 
         context = TrainContext(
-            model=model,
             world_size=get_world_size(),
             rank=get_rank(),
             config=cfg,
             model_config=model_config,
             executor=executor,
+            epoch=preloaded_epoch,
+            consumed_samples=preloaded_consumed,
+            checkpoint=preloaded_checkpoint,
         )
 
-        if self._resume_dir:
-            checkpoint = Checkpoint.load_any(self._resume_dir)
-            if checkpoint is not None:
-                model.load_state_dict(checkpoint.state_dict, strict=False)
-                if checkpoint.config:
-                    context.model_config = checkpoint.config
-                context.epoch = checkpoint.epoch or cfg.start_epoch
-                if checkpoint.consumed_samples > 0:
-                    context.consumed_samples = checkpoint.consumed_samples
-                else:
-                    context.consumed_samples = cfg.start_samples * context.world_size
-                context.checkpoint = checkpoint
-
-        if cfg.lora is not None:
-            inject_lora(
-                model,
-                r=cfg.lora.r,
-                alpha=cfg.lora.alpha,
-                target_modules=set(cfg.lora.target_modules),
-            )
-
-        context.optimizer = cfg.optimizer_fn(model)
-        context.scheduler = cfg.scheduler_fn(context.optimizer)
+        context.model, context.optimizer, context.scheduler = executor.prepare(
+            cfg.model_fn,
+            cfg.optimizer_fn,
+            cfg.scheduler_fn,
+            before_wrap=_before_wrap,
+            after_wrap=_after_wrap,
+        )
 
         train_dataset = cfg.dataset
         val_dataset = cfg.val_dataset
@@ -128,7 +170,16 @@ class TrainContextBuilder:
             )
 
         sampler_offset = context.consumed_samples // context.world_size
-        sampler = ResumableDistributedSampler(
+
+        if self._resume and sampler_offset > 0:
+            offset = context.world_size - 1
+            num_samples_per_replica = (
+                len(train_dataset) + offset
+            ) // context.world_size
+            if num_samples_per_replica > 0:
+                context.epoch = sampler_offset // num_samples_per_replica
+
+        sampler = RDSampler(
             data_source=train_dataset,
             start_epoch=context.epoch,
             start_iter=sampler_offset,
@@ -141,10 +192,11 @@ class TrainContextBuilder:
             num_workers=cfg.num_workers,
             pin_memory=cfg.pin_memory,
             prefetch_factor=cfg.prefetch_factor,
+            collate_fn=cfg.collate_fn,
         )
 
         if val_dataset is not None:
-            val_sampler = ResumableDistributedSampler(
+            val_sampler = RDSampler(
                 data_source=val_dataset,
                 start_epoch=0,
                 start_iter=0,
@@ -158,17 +210,9 @@ class TrainContextBuilder:
                 num_workers=cfg.num_workers,
                 pin_memory=cfg.pin_memory,
                 prefetch_factor=cfg.prefetch_factor,
+                collate_fn=cfg.collate_fn,
             )
 
-        context.model, context.optimizer, context.dataloader, context.scheduler = (
-            executor.prepare(
-                model,
-                context.optimizer,
-                context.dataloader,
-                context.scheduler,
-            )
-        )
-
         if context.checkpoint and context.checkpoint.extra:
             extra = context.checkpoint.extra
             for name in ("optimizer", "scheduler"):
@@ -177,13 +221,73 @@ class TrainContextBuilder:
                     if obj is not None:
                         obj.load_state_dict(extra[name])
 
+        strategy_kwargs = dict(cfg.extra_kwargs)
+        strategy_kwargs.setdefault("moe_aux_loss_coef", cfg.moe_aux_loss_coef)
+
+        needs_ref = cfg.strategy in (
+            "dpo",
+            "grpo",
+            "online_grpo",
+            "online_dpo",
+        )
+        needs_old = cfg.strategy in ("grpo", "online_grpo")
+
+        if needs_ref:
+            strategy_kwargs["ref_model"] = create_ref_model(
+                cfg.model_fn, executor=executor, model=context.model, device=device
+            )
+
+        if needs_old:
+            strategy_kwargs["old_model"] = create_ref_model(
+                cfg.model_fn, executor=executor, model=context.model, device=device
+            )
+
         context.strategy = StrategyFactory.create(
             cfg.strategy,
             model=context.model,
             device=device,
             executor=executor,
-            model_fn=cfg.model_fn,
-            **cfg.extra_kwargs,
+            **strategy_kwargs,
         )
 
+        # Enable online rollout when the train_type is an ``online_*`` variant.
+        is_online = cfg.strategy.startswith("online_")
+        if is_online:
+            if not context.strategy.supports_online():
+                raise ValueError(
+                    f"Strategy '{cfg.strategy}' does not support online rollout"
+                )
+            if cfg.reward_model_fn is None:
+                raise ValueError("reward_model_fn is required for online RL strategies")
+
+            tokenizer = AutoTokenizer.from_pretrained(self._param_path)
+            reward_model = cfg.reward_model_fn()
+
+            group_size = strategy_kwargs.get("group_size", 1)
+            rollout_batch_size = group_size * max(1, cfg.batch_per_device)
+            max_seq_len = getattr(context.model.config, "max_position_embeddings", None)
+
+            scheduler = InferenceScheduler(
+                model=context.model,
+                tokenizer=tokenizer,
+                max_batch_size=rollout_batch_size,
+                max_seq_len=max_seq_len,
+            )
+
+            generator = RolloutGenerator(
+                scheduler=scheduler,
+                tokenizer=tokenizer,
+                max_tokens=cfg.rollout_max_tokens,
+                group_size=group_size,
+                temperature=cfg.rollout_temperature,
+                top_k=cfg.rollout_top_k,
+                top_p=cfg.rollout_top_p,
+            )
+            runner = RolloutRunner(
+                generator=generator,
+                reward_model=reward_model,
+                rollout_interval=cfg.rollout_interval,
+            )
+            context.strategy.set_rollout_runner(runner)
+
         return context

astrai/trainer/trainer.py

@@ -1,8 +1,14 @@
 import logging
 from typing import List, Optional
 
+import torch.distributed as dist
+
 from astrai.config import TrainConfig
 from astrai.parallel.setup import spawn_parallel_fn
+from astrai.signal_handler import (
+    register_signal_handlers,
+    unregister_signal_handlers,
+)
 from astrai.trainer.train_callback import (
     CallbackFactory,
     TrainCallback,
@@ -36,7 +42,7 @@ class Trainer:
             ),
             CallbackFactory.create(
                 "metric",
-                log_dir=cfg.log_dir,
+                ckpt_dir=cfg.ckpt_dir,
                 save_interval=cfg.ckpt_interval,
                 metrics=cfg.metrics,
                 val_step=cfg.val_step,
@@ -52,10 +58,13 @@ class Trainer:
             if method:
                 method(context)
 
-    def _trainer_loop(self, resume_dir: Optional[str] = None):
+    def _trainer_loop(self, param_path: Optional[str] = None, resume: bool = False):
         context = (
-            TrainContextBuilder(self.train_config).with_resume_dir(resume_dir).build()
+            TrainContextBuilder(self.train_config)
+            .with_param_path(param_path, resume=resume)
+            .build()
         )
+        register_signal_handlers(context)
         executor = context.executor
         self._call_callbacks("on_train_begin", context)
 
@@ -63,15 +72,20 @@ class Trainer:
             context.model.train()
 
             for epoch in range(context.epoch, context.config.n_epoch):
+                if context.stop_requested:
+                    break
                 context.epoch = epoch
                 self._call_callbacks("on_epoch_begin", context)
 
                 for batch in context.dataloader:
+                    if context.stop_requested:
+                        break
                     with executor.accumulate(context.model):
                         self._call_callbacks("on_batch_begin", context)
-                        loss = context.strategy(batch)
-                        context.loss = loss.item()
-                        stand_loss = loss / executor.grad_accum_steps
+                        loss_output = context.strategy(batch)
+                        context.loss = loss_output["loss"].item()
+                        context.metrics = loss_output["metrics"]
+                        stand_loss = loss_output["loss"] / executor.grad_accum_steps
                         executor.backward(stand_loss)
                         context.consumed_samples += (
                             context.config.batch_per_device * context.world_size
@@ -81,6 +95,7 @@ class Trainer:
                         if executor.sync_gradients:
                             self._call_callbacks("on_optimizer_step", context)
                             context.optimizer.step()
+                            context.strategy.on_optimizer_step()
                             context.optimizer.zero_grad()
 
                             if context.scheduler:
@@ -88,14 +103,23 @@ class Trainer:
 
                 self._call_callbacks("on_epoch_end", context)
 
+            if context.stop_requested:
+                logger.warning(
+                    "Training interrupted by signal, saving emergency checkpoint..."
+                )
+                self._call_callbacks("on_error", context)
+
         except Exception as e:
             logger.error("Training failed: %s", str(e), exc_info=True)
             self._call_callbacks("on_error", context)
             raise
         finally:
             self._call_callbacks("on_train_end", context)
+            if executor.use_distributed and dist.is_initialized():
+                dist.barrier()
+            unregister_signal_handlers()
 
-    def train(self, resume_dir: Optional[str] = None):
+    def train(self, param_path: Optional[str] = None, resume: bool = False):
         cfg = self.train_config
         spawn_parallel_fn(
             self._trainer_loop,
@@ -105,5 +129,6 @@ class Trainer:
             master_port=cfg.master_port,
             device_type=cfg.device_type,
             start_method=cfg.start_method,
-            resume_dir=resume_dir,
+            param_path=param_path,
+            resume=resume,
         )

csrc/CMakeLists.txt

@@ -0,0 +1,74 @@
+cmake_minimum_required(VERSION 3.18)
+project(astrai_kernels LANGUAGES CUDA CXX)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CUDA_STANDARD 17)
+
+find_package(CUDAToolkit REQUIRED)
+
+if(NOT DEFINED TORCH_HOME)
+    set(TORCH_HOME "$ENV{TORCH_HOME}")
+endif()
+if(NOT TORCH_HOME)
+    message(FATAL_ERROR "TORCH_HOME must point at the torch install dir (site-packages/torch)")
+endif()
+
+if(NOT DEFINED PYTHON_INCLUDE_DIR)
+    set(PYTHON_INCLUDE_DIR "/usr/include/python${PYTHON_VERSION_MAJOR}.${PYTHON_VERSION_MINOR}")
+endif()
+
+if(NOT DEFINED ASTRAI_CUDA_ARCH)
+    if(DEFINED ENV{ASTRAI_CUDA_ARCH})
+        set(ASTRAI_CUDA_ARCH "$ENV{ASTRAI_CUDA_ARCH}")
+    else()
+        set(ASTRAI_CUDA_ARCH 80)
+    endif()
+endif()
+
+set(TORCH_LIB_DIR "${TORCH_HOME}/lib")
+set(CUDA_LIB_DIR "/usr/local/cuda/lib64")
+
+set(CXX_FLAGS -O3 -funroll-loops)
+set(NVCC_FLAGS -O3
+    --expt-relaxed-constexpr
+    --use_fast_math
+    "--ptxas-options=-O3,-v"
+    --extra-device-vectorization
+    --threads=16)
+
+set(TORCH_LIBS
+    "${TORCH_LIB_DIR}/libtorch_python.so"
+    "${TORCH_LIB_DIR}/libtorch_cuda.so"
+    "${TORCH_LIB_DIR}/libc10_cuda.so"
+    "${TORCH_LIB_DIR}/libtorch_cpu.so"
+    "${TORCH_LIB_DIR}/libtorch.so"
+    "${TORCH_LIB_DIR}/libc10.so"
+    CUDA::cudart)
+
+set(CMAKE_CUDA_ARCHITECTURES "${ASTRAI_CUDA_ARCH}")
+
+set(KERNELS attn_decode attn_prefill attn_paged_decode attn_paged_prefill rotary_emb)
+
+foreach(name ${KERNELS})
+    add_library(${name} MODULE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${name}.cu")
+
+    target_compile_definitions(${name} PRIVATE TORCH_EXTENSION_NAME=${name})
+
+    target_include_directories(${name} PRIVATE
+        "${TORCH_HOME}/include"
+        "${TORCH_HOME}/include/torch/csrc/api/include"
+        "${PYTHON_INCLUDE_DIR}")
+
+    target_link_libraries(${name} PRIVATE ${TORCH_LIBS})
+    target_link_options(${name} PRIVATE "-Wl,-rpath,${TORCH_LIB_DIR}")
+
+    target_compile_options(${name} PRIVATE
+        $<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS}>
+        $<$<COMPILE_LANGUAGE:CUDA>:${NVCC_FLAGS}>)
+
+    set_target_properties(${name} PROPERTIES
+        PREFIX ""
+        SUFFIX ".${PY_SOABI}.so"
+        LIBRARY_OUTPUT_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/../astrai/extension/lib")
+endforeach()

csrc/__init__.py

@@ -0,0 +1,2 @@
+# Source directory for CUDA kernels — build-time only.
+# Compiled .so files live in astrAI/_ext/.

csrc/kernels/attn_common.h

@@ -0,0 +1,66 @@
+#pragma once
+
+// Tensor layout for Q/K/V tensors passed to attention kernels.
+// Internally, kernels always operate on BHLD [batch, n_heads, seq_len, head_dim].
+// When the caller passes BLHD, dims 1 and 2 are transposed at entry.
+enum TensorLayout : int {
+    BHLD = 0,  // [batch, n_heads, seq_len, head_dim]
+    BLHD = 1,  // [batch, seq_len, n_heads, head_dim]
+};
+
+
+// Unified attention params covering BOTH addressing modes:
+//   - Contiguous K/V: dense [batch, kv_head, kv_len, head_dim] tensors (k/v).
+//   - Paged (SGLang-style): flat pool [size, kv_head, head_dim] + req_to_token.
+// Each kernel selects the addressing via a KVSource policy (see
+// attn_kv_source.cuh); a given call only touches the fields of one mode, so
+// this is a POD shared by both paths rather than two parallel structs that
+// drift out of sync.
+template<typename T, typename AT = float>
+struct AttentionParams {
+    // ---- shared across all paths ----
+    int batch;
+    int q_head;
+    int kv_head;
+    int head_dim;
+    int use_mask;
+    int causal_offset;   // -1 = non-causal; >=0 = absolute position of first Q token
+    int num_splits;
+    float scale;
+
+    // Q strides (element offsets for each dim — layout-agnostic)
+    int q_stride_b, q_stride_h, q_stride_l, q_stride_d;
+
+    // Mask: 2D [batch, kv_len], 3D [batch, q_len, kv_len],
+    // or 4D [batch, n_heads, q_len, kv_len] (head dim broadcasts when stride=0)
+    int mask_b_stride;   // batch stride
+    int mask_h_stride;   // head stride (0 = broadcast across heads)
+    int mask_q_stride;   // q stride (0 = all q rows share)
+    const bool* __restrict__ mask;
+
+    const T* __restrict__ q;
+    T* __restrict__ o;
+    AT* __restrict__ o_part;
+    AT* __restrict__ ml_part;
+
+    // ---- contiguous K/V mode ----
+    int q_len;
+    int kv_len;
+    int kv_stride_b, kv_stride_h, kv_stride_l, kv_stride_d;
+    const T* __restrict__ k;
+    const T* __restrict__ v;
+
+    // ---- paged (SGLang flat pool) mode ----
+    const T* __restrict__ k_cache;
+    const T* __restrict__ v_cache;
+
+    // Indexing
+    const int64_t* __restrict__ req_to_token;      // [num_reqs, max_context_len]
+    const int64_t* __restrict__ req_pool_indices;   // [batch]
+    const int* __restrict__ kv_indptr;               // [batch+1]
+    const int* __restrict__ qo_indptr;               // [batch+1] or nullptr (decode)
+    int max_context_len;  // req_to_token stride (dim 1)
+    int max_seq_len;      // max per-request seq_len (host-side, for split computation)
+    int total_q;          // total Q tokens across all requests (host-side, for grid)
+    int max_q_len;        // max per-request q_len (host-side, for prefill grid)
+};

csrc/kernels/attn_decode.cu

@@ -0,0 +1,56 @@
+#include "attn_dispatchers.cuh"
+#include "attn_entry_utils.cuh"
+
+torch::Tensor attn_decode(
+    torch::Tensor q,
+    torch::Tensor k,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout,
+    c10::optional<torch::Tensor> o_part_buf,
+    c10::optional<torch::Tensor> ml_part_buf
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+    auto stream = at::cuda::getCurrentCUDAStream();
+
+    AttentionParams<bf16> p;
+    attn_pack_params(q, k, v, mask, causal_offset, scale, layout, p);
+    TORCH_CHECK(p.q_len == 1, "Q seq_len must be 1");
+    TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
+
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
+    auto O_view = (layout == BLHD) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    if (o_part_buf.has_value() && ml_part_buf.has_value()
+        && o_part_buf->defined() && ml_part_buf->defined()) {
+        TORCH_CHECK(o_part_buf->scalar_type() == torch::kFloat32, "o_part_buf must be f32");
+        TORCH_CHECK(ml_part_buf->scalar_type() == torch::kFloat32, "ml_part_buf must be f32");
+        int64_t o_needed = (int64_t)p.batch * p.q_head * MAX_SPLITS * p.head_dim;
+        TORCH_CHECK(o_part_buf->numel() >= o_needed,
+                     "o_part_buf too small: need ", o_needed, " got ", o_part_buf->numel());
+        p.o_part = (float*)o_part_buf->data_ptr();
+        p.ml_part = (float*)ml_part_buf->data_ptr();
+    } else {
+        alloc_split_partials(p);
+    }
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_decode, p, stream);
+    C10_CUDA_CHECK(cudaGetLastError());
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_decode", &attn_decode,
+        py::arg("q"),
+        py::arg("k"),
+        py::arg("v"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = (int64_t)BHLD,
+        py::arg("o_part_buf") = py::none(),
+        py::arg("ml_part_buf") = py::none(),
+        "GQA decode (tensor-core head-packing on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_decode_split_kv.cuh

@@ -0,0 +1,143 @@
+#pragma once
+#include <cuda_bf16.h>
+#include <float.h>
+#include "attn_common.h"
+#include "attn_kv_source.cuh"
+#include "attn_warp_utils.cuh"
+constexpr int DC_CHUNK = 64;
+
+// Scalar split-KV decode (fallback for sm < 80, no tensor cores), unified
+// across contiguous and paged (SGLang flat-pool) K/V via the KV template
+// parameter.  For decode the query is the last token, so its valid range
+// [0, seq_len) IS the causal range; KV::decode_attend_len expresses that
+// bound per addressing mode (contig clips to causal_offset, paged = seq_len).
+template <int HEAD_DIM, typename KV, bool IsCausal, bool HasMask>
+__global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
+    int batch = blockIdx.x / p.kv_head;
+    int kv_head = blockIdx.x % p.kv_head;
+    int split = blockIdx.z;
+    int group_size = blockDim.y;
+    int q_head = kv_head * group_size + threadIdx.y;
+    int lane = threadIdx.x;
+    int hd_per_thread = p.head_dim / 32;
+
+    const int seq_len = KV::kv_len(p, batch);
+    const KVContext kctx = KV::template make_ctx<HEAD_DIM>(p, batch, kv_head);
+
+    // Q: [batch, q_head, q_len=1, head_dim] — stride-based
+    float q_reg[8];
+    int q_off = KV::q_decode_base(p, batch, q_head)
+              + lane * hd_per_thread * p.q_stride_d;
+    for (int i = 0; i < hd_per_thread; i++)
+        q_reg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
+
+    int mask_base = batch * p.mask_b_stride + q_head * p.mask_h_stride;
+
+    float m = -FLT_MAX, d = 0.0f, acc_reg[8] = {0.0f};
+
+    extern __shared__ __align__(16) bf16 smem[];
+    bf16* k_smem = smem;
+    bf16* v_smem = smem + DC_CHUNK * p.head_dim;
+
+    // Split-KV: each split processes a contiguous subset of chunks
+    int chunks_total = (seq_len + DC_CHUNK - 1) / DC_CHUNK;
+    int chunks_per_split = (chunks_total + p.num_splits - 1) / p.num_splits;
+    int ch_begin = split * chunks_per_split;
+    int ch_end = min(chunks_total, ch_begin + chunks_per_split);
+
+    for (int ci = ch_begin; ci < ch_end; ci++) {
+        int chunk_start = ci * DC_CHUNK;
+        int this_chunk = min(DC_CHUNK, seq_len - chunk_start);
+
+        // Load K and V into shared memory (addressing via KV policy;
+        // paged guards empty slots with zero-fill).
+        int total = this_chunk * p.head_dim;
+        for (int i = threadIdx.y * 32 + lane; i < total;
+             i += blockDim.x * blockDim.y) {
+            int s = i / p.head_dim;
+            int d_dim = i % p.head_dim;
+            int kc = chunk_start + s;
+            KVAddr a = KV::kv_addr(p, kctx, kc, d_dim, true);
+            k_smem[i] = a.valid ? *reinterpret_cast<const bf16*>(a.k) : (bf16)0.f;
+            v_smem[i] = a.valid ? *reinterpret_cast<const bf16*>(a.v) : (bf16)0.f;
+        }
+        __syncthreads();
+
+        for (int s = 0; s < this_chunk; s++) {
+            float partial = 0.0f;
+            for (int i = 0; i < hd_per_thread; i++)
+                partial += q_reg[i] * __bfloat162float(
+                    k_smem[s * p.head_dim + lane * hd_per_thread + i]);
+            partial = warp_reduce_sum(partial) * p.scale;
+
+            int kv_idx = chunk_start + s;
+            if constexpr (HasMask) {
+                if (!p.mask[mask_base + kv_idx])
+                    partial = -FLT_MAX;
+            }
+            if constexpr (IsCausal) {
+                if (kv_idx >= KV::decode_attend_len(p, batch))
+                    partial = -FLT_MAX;
+            }
+
+            float new_m = fmaxf(m, partial);
+            float alpha = __expf(m - new_m);
+            float beta  = __expf(partial - new_m);
+            d = d * alpha + beta;
+
+            for (int i = 0; i < hd_per_thread; i++) {
+                float vv = __bfloat162float(v_smem[s * p.head_dim + lane * hd_per_thread + i]);
+                acc_reg[i] = fmaf(acc_reg[i], alpha, vv * beta);
+            }
+            m = new_m;
+        }
+        __syncthreads();
+    }
+
+    // ---- write UN-normalised partials for this split ----
+    size_t bh = (size_t)batch * p.q_head + q_head;
+    size_t slot = bh * MAX_SPLITS + split;
+    int d0 = lane * hd_per_thread;
+    for (int i = 0; i < hd_per_thread; i++) {
+        int dd = d0 + i;
+        p.o_part[slot * p.head_dim + dd] = acc_reg[i];
+    }
+    if (lane == 0) {
+        p.ml_part[slot * 2] = m;
+        p.ml_part[slot * 2 + 1] = d;
+    }
+}
+
+// Split-combine: merges the per-split partials (o_part/ml_part) into the
+// final normalised O.  KV selects the O addressing (contig batch stride vs
+// paged row stride).
+template <typename KV>
+__global__ void attn_decode_combine_kernel(AttentionParams<bf16> p) {
+    int bh = blockIdx.x;
+    int d = threadIdx.x;
+    if (d >= p.head_dim) return;
+
+    int batch = bh / p.q_head;
+    int q_head = bh % p.q_head;
+
+    size_t split_base = (size_t)bh * MAX_SPLITS;
+    const float* mlp = p.ml_part + split_base * 2;
+    const float* op = p.o_part + split_base * p.head_dim;
+
+    float m = -FLT_MAX, l = 0.0f, acc = 0.0f;
+    for (int s = 0; s < p.num_splits; s++) {
+        float mi = mlp[s * 2];
+        if (mi <= -FLT_MAX) continue;
+        float li = mlp[s * 2 + 1];
+        float nm = fmaxf(m, mi);
+        float corr = __expf(m - nm);
+        float e = __expf(mi - nm);
+        acc = fmaf(acc, corr, op[s * p.head_dim + d] * e);
+        l = fmaf(l, corr, li * e);
+        m = nm;
+    }
+
+    float inv = (l > 1e-20f) ? (1.0f / l) : 0.0f;
+    int o_off = KV::q_decode_base(p, batch, q_head) + d * p.q_stride_d;
+    p.o[o_off] = __float2bfloat16(acc * inv);
+}

csrc/kernels/attn_decode_split_kv_mma.cuh

@@ -0,0 +1,176 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_kv_source.cuh"
+#include "attn_mma_utils.cuh"
+#include "attn_warp_utils.cuh"
+
+// Split-K (FlashDecoding) tensor-core decode via GQA head-packing, unified
+// across contiguous and paged (SGLang flat-pool) K/V via the KV template
+// parameter.  Decode has q_len == 1, so we pack G = q_head/kv_head query
+// heads into the M=16 rows of mma.sync.m16n8k16, turning G independent GEMVs
+// into a single GEMM that reuses each loaded K/V tile across all G heads.
+//
+// KV = ContigKV (dense tensors) or PagedKV (flat pool + req_to_token).
+// IsCausal and HasMask are compile-time bools — no runtime branch in the
+// inner compute loop.
+//
+// Traits = KernelTraits<HEAD_DIM, BC=16, WARPS=1, STAGES=2>.
+template <typename Traits, typename KV, bool IsCausal, bool HasMask>
+__global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
+    const int lane = threadIdx.x;
+    const int gid = lane >> 2;
+    const int tid4 = lane & 3;
+
+    const int pass = blockIdx.x / p.kv_head;
+    const int kv_head = blockIdx.x % p.kv_head;
+    const int batch = blockIdx.y;
+    const int split = blockIdx.z;
+
+    constexpr int MAX_G = 16;
+    const int G_total = p.q_head / p.kv_head;
+    const int g_begin = pass * MAX_G;
+    const int G = min(MAX_G, G_total - g_begin);
+    const int q_head0 = kv_head * G_total + g_begin;
+
+    // Per-request seq_len (paged reads kv_indptr; contig uses p.kv_len).
+    const int seq_len = KV::kv_len(p, batch);
+    const KVContext kctx = KV::template make_ctx<Traits::HEAD_DIM>(p, batch, kv_head);
+
+    // Double-buffered shared memory for K/V (no sQ needed)
+    __shared__ __align__(16) bf16 sK[Traits::STAGES * Traits::BC * Traits::LD];
+    __shared__ __align__(16) bf16 sV[Traits::STAGES * Traits::BC * Traits::LD];
+
+    // Load Q directly from global into mma A-operand registers.
+    const int q_base = KV::q_decode_base(p, batch, q_head0);
+    const int qra = gid;
+    const int qrb = gid + 8;
+    const bool va = qra < G, vb = qrb < G;
+    unsigned Qa[Traits::KD][4];
+    load_q_mma_frags<Traits::KD>(p.q + q_base, p.q_stride_h, p.q_stride_d,
+                                  qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[Traits::DN8][4];
+    #pragma unroll
+    for (int j = 0; j < Traits::DN8; j++)
+        Oacc[j][0] = Oacc[j][1] = Oacc[j][2] = Oacc[j][3] = 0.0f;
+    float m0 = -FLT_MAX, m1 = -FLT_MAX, l0 = 0.0f, l1 = 0.0f;
+
+    const int tiles_total = (seq_len + Traits::BC - 1) / Traits::BC;
+    const int tiles_per_split = (tiles_total + p.num_splits - 1) / p.num_splits;
+    const int ti_begin = split * tiles_per_split;
+    const int ti_end = min(tiles_total, ti_begin + tiles_per_split);
+
+    // ---- Load tile lambda: predicated cp.async (addressing via KV policy) ----
+    auto load_tile = [&](int ti, int buf) {
+        int kv0 = ti * Traits::BC;
+        bf16* dK = sK + buf * Traits::BC * Traits::LD;
+        bf16* dV = sV + buf * Traits::BC * Traits::LD;
+        #pragma unroll
+        for (int i = lane * Traits::VEC; i < Traits::TOTAL;
+             i += Traits::NUM_THREADS * Traits::VEC) {
+            int r = i / Traits::HEAD_DIM, d = i % Traits::HEAD_DIM;
+            int kc = kv0 + r;
+            bool valid = kc < seq_len;
+            KVAddr a = KV::kv_addr(p, kctx, kc, d, valid);
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            cp_async_16_pred(&dK[off], a.k, a.valid);
+            cp_async_16_pred(&dV[off], a.v, a.valid);
+        }
+        cp_async_commit();
+    };
+
+    // ---- Multi-stage cp.async pipeline ----
+    // Prologue loads STAGES tiles; each loop iteration waits only for the
+    // oldest outstanding group (wait_group<STAGES-1>) so the STAGES-1 newer
+    // tile loads stay in flight and overlap with the current tile's compute.
+    constexpr int STAGES = Traits::STAGES;
+    const int ntiles = ti_end - ti_begin;
+
+    auto process_tile = [&](int it, int buf) {
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = (ti_begin + it) * Traits::BC;
+
+        float Sacc[Traits::NC8][4];
+        mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
+
+        #pragma unroll
+        for (int n8 = 0; n8 < Traits::NC8; n8++)
+            Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
+            Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
+
+        // Decode: q_len=1, so qrow0=qrow1=0.  Paged treats [0, seq_len) as
+        // the causal range (query is the last token); contig clips to the
+        // causal_offset bound.  Dead code eliminated when IsCausal == false.
+        int maxc = IsCausal ? KV::decode_attend_len(p, batch) : seq_len;
+        mma_softmax_tile<Traits, HasMask>(kv0, maxc, maxc,
+                                           0, 0,
+                                           p.mask_b_stride, 0, 0,
+                                           batch, 0,
+                                           p.mask,
+                                           Sacc, Oacc, m0, m1, l0, l1, lane);
+
+        mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+    };
+
+    if (ntiles >= STAGES) {
+        #pragma unroll
+        for (int i = 0; i < STAGES; i++)
+            load_tile(ti_begin + i, i);
+
+        for (int it = 0; it < ntiles; it++) {
+            cp_async_wait_group<STAGES - 1>();
+            __syncwarp();
+            process_tile(it, it & (STAGES - 1));
+            __syncwarp();
+            if (it + STAGES < ntiles)
+                load_tile(ti_begin + it + STAGES, (it + STAGES) & (STAGES - 1));
+        }
+    } else {
+        // Fewer tiles than stages: load all, wait for all, process.
+        for (int i = 0; i < ntiles; i++)
+            load_tile(ti_begin + i, i);
+        cp_async_wait_group<0>();
+        __syncwarp();
+        for (int it = 0; it < ntiles; it++)
+            process_tile(it, it);
+    }
+
+    // ---- write UN-normalised partials for this split ----
+    auto split_slot = [&](int h) -> size_t {
+        size_t bh = (size_t)batch * p.q_head + h;
+        return bh * MAX_SPLITS + split;
+    };
+    #pragma unroll
+    for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+        int d = dn8 * 8 + 2 * tid4;
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
+            op[d] = Oacc[dn8][0];
+            op[d + 1] = Oacc[dn8][1];
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
+            op[d] = Oacc[dn8][2];
+            op[d + 1] = Oacc[dn8][3];
+        }
+    }
+    if (tid4 == 0) {
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m0; mp[1] = l0;
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m1; mp[1] = l1;
+        }
+    }
+}

csrc/kernels/attn_dispatchers.cuh

@@ -0,0 +1,205 @@
+#pragma once
+// Shared attention dispatchers — used by both production .cu and test .cu.
+// No torch dependency; pure CUDA.
+//
+// The paged and contiguous kernels are unified by the KVSource policy
+// (ContigKV / PagedKV from attn_kv_source.cuh), so each launcher struct
+// below is templated on KV and the paged dispatch is just the same launcher
+// instantiated with PagedKV.  Only the grid/split math differs, and that is
+// covered by KV::host_q_len / KV::host_kv_len.
+
+#include <cuda_runtime.h>
+#include <algorithm>
+#include "attn_warp_utils.cuh"
+#include "attn_kv_source.cuh"
+#include "attn_prefill_split_q.cuh"
+#include "attn_decode_split_kv.cuh"
+#ifndef ASTRAI_NO_MMA
+#include "attn_prefill_split_q_mma.cuh"
+#include "attn_decode_split_kv_mma.cuh"
+#endif
+
+// Split-KV: compute number of splits to fill all SMs for small-batch decode.
+// Caps splits so each split processes at least `min_tiles_per_split` tiles,
+// avoiding excessive loop/prologue overhead when tiles are small.
+//
+// Target total grid blocks (`TARGET_BLOCKS`) rather than scaling splits by SM
+// count.  Decode blocks are single-warp (32 threads) and a SM hosts ~11 of
+// them, so the old `2*sm/base` cap badly undersplit at large batch (B=16 got
+// 3 splits, optimal ~8).  Measured (L20, grid search): bandwidth saturates
+// near 256-512 total blocks; 512 minimizes worst-case latency across the
+// B x kv grid; more is pure oversplit overhead.
+constexpr int DECODE_TARGET_BLOCKS = 512;
+inline int compute_num_splits(int base_blocks, int tiles_total,
+                               int min_tiles_per_split = 1) {
+    int n = (DECODE_TARGET_BLOCKS + base_blocks - 1) / base_blocks;
+    int max_by_work = tiles_total / min_tiles_per_split;
+    return std::max(1, std::min(n, std::min(max_by_work, MAX_SPLITS)));
+}
+
+// Dispatch IsCausal × HasMask — eliminates the duplicated 4-way if/else
+// ladder that appeared in each dispatch_* function.  FN must be a function
+// template <int HEAD_DIM, bool IsCausal, bool HasMask>; HEAD_DIM is forwarded
+// as the first template argument so callers only spell it once.
+//
+// Usage:  DISPATCH_CAUSAL_MASK(is_causal, has_mask, launcher<KV>::template launch, HEAD_DIM, p, stream);
+#define DISPATCH_CAUSAL_MASK(is_causal, has_mask, FN, HEAD_DIM, ...) \
+    do { \
+        if (is_causal) { \
+            if (has_mask) FN<HEAD_DIM, true,  true>(__VA_ARGS__); \
+            else          FN<HEAD_DIM, true,  false>(__VA_ARGS__); \
+        } else { \
+            if (has_mask) FN<HEAD_DIM, false, true>(__VA_ARGS__); \
+            else          FN<HEAD_DIM, false, false>(__VA_ARGS__); \
+        } \
+    } while (0)
+
+// ======================================================================
+// Prefill launchers (KV selects ContigKV or PagedKV addressing)
+// ======================================================================
+
+#ifndef ASTRAI_NO_MMA
+template <typename KV>
+struct PrefillLauncherMMA {
+    template <int HEAD_DIM, bool IsCausal, bool HasMask>
+    static void launch(AttentionParams<bf16>& p, cudaStream_t stream) {
+        constexpr int WARPS = 4;
+        constexpr int BC = (HEAD_DIM <= 128) ? 32 : 16;
+        using Traits = KernelTraits<HEAD_DIM, BC, WARPS, 2>;
+        int q_len = KV::host_q_len(p);
+        dim3 grid((q_len + Traits::BR * WARPS - 1) / (Traits::BR * WARPS),
+                  p.q_head, p.batch);
+        dim3 block(Traits::NUM_THREADS);
+        attn_prefill_split_q_mma_kernel<Traits, KV, IsCausal, HasMask>
+            <<<grid, block, 0, stream>>>(p);
+    }
+};
+#endif
+
+template <typename KV>
+struct PrefillLauncherScalar {
+    template <int HEAD_DIM, bool IsCausal, bool HasMask>
+    static void launch(AttentionParams<bf16>& p, cudaStream_t stream) {
+        constexpr int G = 8, ROWS = 32, P_BC = 32;
+        int q_len = KV::host_q_len(p);
+        dim3 grid((q_len + ROWS - 1) / ROWS, p.q_head, p.batch);
+        dim3 block(G, ROWS);
+        attn_prefill_split_q_kernel_t<HEAD_DIM, KV, G, ROWS, P_BC, IsCausal, HasMask>
+            <<<grid, block, 0, stream>>>(p);
+    }
+};
+
+template <int HEAD_DIM>
+static inline void dispatch_prefill(AttentionParams<bf16>& p, cudaStream_t stream) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+
+#ifndef ASTRAI_NO_MMA
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         PrefillLauncherMMA<ContigKV>::template launch,
+                         HEAD_DIM, p, stream);
+#else
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         PrefillLauncherScalar<ContigKV>::template launch,
+                         HEAD_DIM, p, stream);
+#endif
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_paged_prefill(AttentionParams<bf16>& p, cudaStream_t stream) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+
+#ifndef ASTRAI_NO_MMA
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         PrefillLauncherMMA<PagedKV>::template launch,
+                         HEAD_DIM, p, stream);
+#else
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         PrefillLauncherScalar<PagedKV>::template launch,
+                         HEAD_DIM, p, stream);
+#endif
+}
+
+// ======================================================================
+// Decode launchers (KV selects ContigKV or PagedKV addressing)
+// ======================================================================
+
+#ifndef ASTRAI_NO_MMA
+// BC=16: halves smem (16KB vs 32KB) → doubles occupancy (6 vs 3 blocks/SM).
+// For D=256, BC=16 also reduces register pressure (fewer Sacc/PV frags),
+// enabling STAGES=2 (double-buffer) within the 32KB smem budget — eliminates
+// the 176-byte spill that STAGES=1+BC=32 suffered.
+template <typename KV>
+struct DecodeLauncherMMA {
+    template <int HEAD_DIM, bool IsCausal, bool HasMask>
+    static void launch(AttentionParams<bf16>& p, int group_size, cudaStream_t stream) {
+        int G = p.q_head / p.kv_head;
+        constexpr int MAX_G = 16;
+        int num_passes = (G + MAX_G - 1) / MAX_G;
+        constexpr int BC = 16;
+        int kv_len = KV::host_kv_len(p);
+        int tiles_total = (kv_len + BC - 1) / BC;
+        p.num_splits = compute_num_splits(p.batch * p.kv_head * num_passes, tiles_total, 2);
+        constexpr int STAGES = 2;
+        using Traits = KernelTraits<HEAD_DIM, BC, 1, STAGES>;
+        dim3 grid(p.kv_head * num_passes, p.batch, p.num_splits);
+        attn_decode_split_kv_mma_kernel<Traits, KV, IsCausal, HasMask>
+            <<<grid, 32, 0, stream>>>(p);
+    }
+};
+#endif
+
+template <typename KV>
+struct DecodeLauncherScalar {
+    template <int HEAD_DIM, bool IsCausal, bool HasMask>
+    static void launch(AttentionParams<bf16>& p, int group_size, cudaStream_t stream) {
+        int kv_len = KV::host_kv_len(p);
+        int chunks_total = (kv_len + DC_CHUNK - 1) / DC_CHUNK;
+        p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
+        size_t smem = 2 * DC_CHUNK * p.head_dim * sizeof(bf16);
+        int g = min(group_size, 32);  // cap at 32 to respect 1024-thread limit
+        dim3 grid(p.batch * p.kv_head, 1, p.num_splits);
+        dim3 block(32, g);
+        attn_decode_split_kv_kernel<HEAD_DIM, KV, IsCausal, HasMask>
+            <<<grid, block, smem, stream>>>(p);
+    }
+};
+
+template <int HEAD_DIM>
+static inline void dispatch_decode(AttentionParams<bf16>& p, cudaStream_t stream) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+    int group_size = p.q_head / p.kv_head;
+
+#ifndef ASTRAI_NO_MMA
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         DecodeLauncherMMA<ContigKV>::template launch,
+                         HEAD_DIM, p, group_size, stream);
+#else
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         DecodeLauncherScalar<ContigKV>::template launch,
+                         HEAD_DIM, p, group_size, stream);
+#endif
+
+    attn_decode_combine_kernel<ContigKV><<<p.batch * p.q_head, p.head_dim, 0, stream>>>(p);
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_paged_decode(AttentionParams<bf16>& p, cudaStream_t stream) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+    int group_size = p.q_head / p.kv_head;
+
+#ifndef ASTRAI_NO_MMA
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         DecodeLauncherMMA<PagedKV>::template launch,
+                         HEAD_DIM, p, group_size, stream);
+#else
+    DISPATCH_CAUSAL_MASK(is_causal, has_mask,
+                         DecodeLauncherScalar<PagedKV>::template launch,
+                         HEAD_DIM, p, group_size, stream);
+#endif
+
+    attn_decode_combine_kernel<PagedKV><<<p.batch * p.q_head, p.head_dim, 0, stream>>>(p);
+}

csrc/kernels/attn_entry_utils.cuh

@@ -0,0 +1,311 @@
+#pragma once
+#include <float.h>
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "attn_common.h"
+#include "attn_warp_utils.cuh"
+
+using bf16 = __nv_bfloat16;
+
+// Dispatch head_dim: shared macro — avoids C++20 lambda template syntax.
+// Usage: DISPATCH_HEAD_DIM(hd, fn, args...)
+//   Expands to: fn<32>(args...); fn<64>(args...); etc.
+#define DISPATCH_HEAD_DIM(hd, fn, ...) \
+    switch (hd) { \
+        case 32:  fn<32>(__VA_ARGS__); break; \
+        case 64:  fn<64>(__VA_ARGS__); break; \
+        case 128: fn<128>(__VA_ARGS__); break; \
+        case 256: fn<256>(__VA_ARGS__); break; \
+        default: \
+            TORCH_CHECK(false, "unsupported head_dim ", hd, \
+                         " (supported: 32, 64, 128, 256)"); \
+    }
+
+// The split kernel unconditionally writes every (batch, q_head, split) slot it
+// owns — including empty split ranges, which store m = -FLT_MAX so the combine
+// skips them. Allocators are therefore left uninitialized (torch::empty); the
+// per-call memset (torch::zeros / torch::full) was pure overhead.
+template<typename P>
+inline void alloc_split_partials(P& p) {
+    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
+    auto o_part = torch::empty(at::IntArrayRef{p.batch, p.q_head, MAX_SPLITS, p.head_dim}, fopt);
+    auto ml_part = torch::empty(at::IntArrayRef{p.batch, p.q_head, MAX_SPLITS, 2}, fopt);
+    p.o_part = (float*)o_part.data_ptr();
+    p.ml_part = (float*)ml_part.data_ptr();
+}
+
+// ---- Shared Q-dims + strides extraction ----
+template <typename P>
+inline void extract_q_dims_and_strides(torch::Tensor& q, int64_t layout, P& p) {
+    if (layout == BLHD) q = q.transpose(1, 2);
+    p.batch = (int)q.size(0);
+    p.q_head = (int)q.size(1);
+    p.q_len = (int)q.size(2);
+    p.head_dim = (int)q.size(3);
+    p.q_stride_b = (int)q.stride(0);
+    p.q_stride_h = (int)q.stride(1);
+    p.q_stride_l = (int)q.stride(2);
+    p.q_stride_d = (int)q.stride(3);
+}
+
+// ---- Shared mask packing ----
+// Accepts 2D [batch, kv_len], 3D [batch, q_len, kv_len],
+// or 4D [batch, n_heads, q_len, kv_len].
+// Head/q dimensions with size 1 broadcast (stride set to 0).
+template <typename P>
+inline void pack_mask(const c10::optional<torch::Tensor>& mask, P& p) {
+    if (p.use_mask) {
+        auto m = mask.value();
+        TORCH_CHECK(m.is_cuda(), "mask must be on CUDA");
+        TORCH_CHECK(m.dtype() == torch::kBool, "mask must be bool");
+        TORCH_CHECK(m.size(0) == p.batch, "mask batch mismatch");
+        TORCH_CHECK(m.size(m.dim() - 1) == p.kv_len, "mask kv_len mismatch");
+        if (m.dim() == 2) {
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_h_stride = 0;
+            p.mask_q_stride = 0;
+        } else if (m.dim() == 3) {
+            TORCH_CHECK(m.size(1) == 1 || m.size(1) == p.q_len, "mask q_len mismatch");
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_h_stride = 0;
+            p.mask_q_stride = (m.size(1) == 1) ? 0 : (int)m.stride(1);
+        } else if (m.dim() == 4) {
+            TORCH_CHECK(m.size(2) == 1 || m.size(2) == p.q_len, "mask q_len mismatch");
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_h_stride = (m.size(1) == 1) ? 0 : (int)m.stride(1);
+            p.mask_q_stride = (m.size(2) == 1) ? 0 : (int)m.stride(2);
+        } else {
+            TORCH_CHECK(false, "mask must be 2D, 3D, or 4D");
+        }
+        p.mask = m.data_ptr<bool>();
+    } else {
+        p.mask = nullptr;
+        p.mask_b_stride = 0;
+        p.mask_h_stride = 0;
+        p.mask_q_stride = 0;
+    }
+}
+
+// ---- attn_pack_params (contiguous KV) ----
+template<typename T>
+inline void attn_pack_params(
+    torch::Tensor q,
+    torch::Tensor k,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout,
+    AttentionParams<T>& p
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+
+    TORCH_CHECK(q.is_cuda() && k.is_cuda() && v.is_cuda());
+    TORCH_CHECK(q.dtype() == torch::kBFloat16);
+    TORCH_CHECK(k.dtype() == torch::kBFloat16);
+    TORCH_CHECK(v.dtype() == torch::kBFloat16);
+    TORCH_CHECK(k.sizes() == v.sizes(), "K and V must have identical shapes");
+    TORCH_CHECK(q.dim() == 4 && k.dim() == 4, "Q/K/V must be 4D");
+
+    extract_q_dims_and_strides(q, layout, p);
+
+    if (layout == BLHD) k = k.transpose(1, 2), v = v.transpose(1, 2);
+
+    p.kv_head = (int)k.size(1);
+    p.kv_len = (int)k.size(2);
+    TORCH_CHECK(k.size(3) == p.head_dim, "K/V head_dim must match Q");
+
+    p.kv_stride_b = (int)k.stride(0);
+    p.kv_stride_h = (int)k.stride(1);
+    p.kv_stride_l = (int)k.stride(2);
+    p.kv_stride_d = (int)k.stride(3);
+
+    p.causal_offset = (int)causal_offset;
+    p.use_mask = mask.has_value() ? 1 : 0;
+    p.scale = (scale > 0.0) ? (float)scale : 1.0f / sqrtf((float)p.head_dim);
+
+    p.q = (const T*)q.data_ptr();
+    p.k = (const T*)k.data_ptr();
+    p.v = (const T*)v.data_ptr();
+    p.o = nullptr;
+    p.o_part = nullptr;
+    p.ml_part = nullptr;
+
+    pack_mask(mask, p);
+}
+
+// ---- attn_pack_paged_decode_params ----
+// SGLang-style: flat KV pool + req_to_token indexing + variable
+// seq_lens via kv_indptr.  Q is [batch, q_head, head_dim] (q_len=1 per req).
+template<typename T>
+inline void attn_pack_paged_decode_params(
+    torch::Tensor q,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    torch::Tensor req_to_token,
+    torch::Tensor req_pool_indices,
+    torch::Tensor kv_indptr,
+    int64_t max_seq_len,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    AttentionParams<T>& p
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+
+    TORCH_CHECK(q.is_cuda() && k_cache.is_cuda() && v_cache.is_cuda());
+    TORCH_CHECK(req_to_token.is_cuda() && req_pool_indices.is_cuda() && kv_indptr.is_cuda());
+    TORCH_CHECK(q.dtype() == torch::kBFloat16, "q must be bf16");
+    TORCH_CHECK(k_cache.dtype() == torch::kBFloat16, "k_cache must be bf16");
+    TORCH_CHECK(v_cache.dtype() == torch::kBFloat16, "v_cache must be bf16");
+    TORCH_CHECK(req_to_token.dtype() == torch::kLong, "req_to_token must be int64");
+    TORCH_CHECK(req_pool_indices.dtype() == torch::kLong, "req_pool_indices must be int64");
+    TORCH_CHECK(kv_indptr.dtype() == torch::kInt32, "kv_indptr must be int32");
+    TORCH_CHECK(k_cache.sizes() == v_cache.sizes(), "k_cache and v_cache must match");
+    TORCH_CHECK(k_cache.dim() == 3, "k_cache must be 3D [size, kv_head, head_dim]");
+    TORCH_CHECK(q.dim() == 3, "q must be 3D [batch, q_head, head_dim]");
+
+    p.batch = (int)q.size(0);
+    p.q_head = (int)q.size(1);
+    p.head_dim = (int)q.size(2);
+    p.kv_head = (int)k_cache.size(1);
+    TORCH_CHECK(k_cache.size(2) == p.head_dim, "k_cache head_dim mismatch");
+    TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
+    TORCH_CHECK(p.q_head % p.kv_head == 0, "q_head must be divisible by kv_head");
+
+    p.q_stride_l = (int)q.stride(0);
+    p.q_stride_h = (int)q.stride(1);
+    p.q_stride_d = (int)q.stride(2);
+
+    p.k_cache = (const T*)k_cache.data_ptr();
+    p.v_cache = (const T*)v_cache.data_ptr();
+    p.q = (const T*)q.data_ptr();
+    p.req_to_token = req_to_token.data_ptr<int64_t>();
+    p.req_pool_indices = req_pool_indices.data_ptr<int64_t>();
+    p.kv_indptr = kv_indptr.data_ptr<int>();
+    p.qo_indptr = nullptr;
+    p.max_context_len = (int)req_to_token.size(1);
+    p.max_seq_len = (int)max_seq_len;
+    p.total_q = p.batch;  // decode: 1 Q token per request
+    p.max_q_len = 1;
+
+    p.causal_offset = (int)causal_offset;
+    p.use_mask = (mask.has_value() && mask.value().defined()) ? 1 : 0;
+    p.scale = (scale > 0.0) ? (float)scale : 1.0f / sqrtf((float)p.head_dim);
+
+    if (p.use_mask) {
+        auto m = mask.value();
+        TORCH_CHECK(m.is_cuda() && m.dtype() == torch::kBool, "mask must be bool CUDA");
+        TORCH_CHECK(m.size(0) == p.batch, "mask batch mismatch");
+        p.mask_b_stride = (int)m.stride(0);
+        p.mask_h_stride = 0;
+        p.mask_q_stride = 0;
+        p.mask = m.data_ptr<bool>();
+    } else {
+        p.mask = nullptr;
+        p.mask_b_stride = 0;
+        p.mask_h_stride = 0;
+        p.mask_q_stride = 0;
+    }
+
+    p.o = nullptr;
+    p.o_part = nullptr;
+    p.ml_part = nullptr;
+}
+
+// ---- attn_pack_paged_prefill_params ----
+// SGLang-style: flat KV pool + req_to_token + ragged batch via qo_indptr.
+// Q is [total_q, q_head, head_dim] (flattened across all requests).
+template<typename T>
+inline void attn_pack_paged_prefill_params(
+    torch::Tensor q,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    torch::Tensor req_to_token,
+    torch::Tensor req_pool_indices,
+    torch::Tensor kv_indptr,
+    torch::Tensor qo_indptr,
+    c10::optional<torch::Tensor> mask,
+    int64_t max_q_len,
+    int64_t causal_offset,
+    double scale,
+    AttentionParams<T>& p
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+
+    TORCH_CHECK(q.is_cuda() && k_cache.is_cuda() && v_cache.is_cuda());
+    TORCH_CHECK(req_to_token.is_cuda() && req_pool_indices.is_cuda());
+    TORCH_CHECK(kv_indptr.is_cuda() && qo_indptr.is_cuda());
+    TORCH_CHECK(q.dtype() == torch::kBFloat16, "q must be bf16");
+    TORCH_CHECK(k_cache.dtype() == torch::kBFloat16, "k_cache must be bf16");
+    TORCH_CHECK(v_cache.dtype() == torch::kBFloat16, "v_cache must be bf16");
+    TORCH_CHECK(req_to_token.dtype() == torch::kLong, "req_to_token must be int64");
+    TORCH_CHECK(req_pool_indices.dtype() == torch::kLong, "req_pool_indices must be int64");
+    TORCH_CHECK(kv_indptr.dtype() == torch::kInt32, "kv_indptr must be int32");
+    TORCH_CHECK(qo_indptr.dtype() == torch::kInt32, "qo_indptr must be int32");
+    TORCH_CHECK(k_cache.sizes() == v_cache.sizes(), "k_cache and v_cache must match");
+    TORCH_CHECK(k_cache.dim() == 3, "k_cache must be 3D [size, kv_head, head_dim]");
+    TORCH_CHECK(q.dim() == 3, "q must be 3D [total_q, q_head, head_dim]");
+
+    p.q_head = (int)q.size(1);
+    p.head_dim = (int)q.size(2);
+    p.kv_head = (int)k_cache.size(1);
+    p.batch = (int)req_pool_indices.size(0);
+    TORCH_CHECK(k_cache.size(2) == p.head_dim, "k_cache head_dim mismatch");
+    TORCH_CHECK(p.head_dim % 16 == 0, "head_dim must be multiple of 16");
+    TORCH_CHECK(p.q_head % p.kv_head == 0, "q_head must be divisible by kv_head");
+    TORCH_CHECK(kv_indptr.size(0) == p.batch + 1, "kv_indptr must be [batch+1]");
+    TORCH_CHECK(qo_indptr.size(0) == p.batch + 1, "qo_indptr must be [batch+1]");
+
+    p.q_stride_l = (int)q.stride(0);
+    p.q_stride_h = (int)q.stride(1);
+    p.q_stride_d = (int)q.stride(2);
+
+    p.k_cache = (const T*)k_cache.data_ptr();
+    p.v_cache = (const T*)v_cache.data_ptr();
+    p.q = (const T*)q.data_ptr();
+    p.req_to_token = req_to_token.data_ptr<int64_t>();
+    p.req_pool_indices = req_pool_indices.data_ptr<int64_t>();
+    p.kv_indptr = kv_indptr.data_ptr<int>();
+    p.qo_indptr = qo_indptr.data_ptr<int>();
+    p.max_context_len = (int)req_to_token.size(1);
+    p.total_q = (int)q.size(0);  // prefill: flattened Q across all requests
+    p.max_q_len = (int)max_q_len;
+    // max_seq_len is unused by the prefill path (decode uses it for split
+    // computation); fill with max_q_len only to keep the POD struct defined.
+    p.max_seq_len = p.max_q_len;
+
+    p.causal_offset = (int)causal_offset;
+    p.use_mask = (mask.has_value() && mask.value().defined()) ? 1 : 0;
+    if (p.use_mask) {
+        auto m = mask.value();
+        TORCH_CHECK(m.is_cuda() && m.dtype() == torch::kBool, "mask must be bool CUDA");
+        TORCH_CHECK(m.size(0) == p.batch, "mask batch mismatch");
+        if (m.dim() == 2) {
+            TORCH_CHECK(m.size(1) <= p.max_context_len, "mask kv_len mismatch");
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_h_stride = 0;
+            p.mask_q_stride = 0;
+        } else if (m.dim() == 4) {
+            TORCH_CHECK(m.size(1) == 1 || m.size(1) == p.q_head, "mask head mismatch");
+            TORCH_CHECK(m.size(2) == 1 || m.size(2) == p.max_q_len, "mask q_len mismatch");
+            TORCH_CHECK(m.size(3) <= p.max_context_len, "mask kv_len mismatch");
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_h_stride = (m.size(1) == 1) ? 0 : (int)m.stride(1);
+            p.mask_q_stride = (m.size(2) == 1) ? 0 : (int)m.stride(2);
+        } else {
+            TORCH_CHECK(false, "mask must be 2D or 4D");
+        }
+        p.mask = m.data_ptr<bool>();
+    } else {
+        p.mask = nullptr;
+        p.mask_b_stride = 0;
+        p.mask_h_stride = 0;
+        p.mask_q_stride = 0;
+    }
+    p.scale = (scale > 0.0) ? (float)scale : 1.0f / sqrtf((float)p.head_dim);
+
+    p.o = nullptr;
+    p.o_part = nullptr;
+    p.ml_part = nullptr;
+}

csrc/kernels/attn_kv_source.cuh

@@ -0,0 +1,159 @@
+#pragma once
+#include <cuda_bf16.h>
+#include "attn_common.h"
+
+// ============================================================================
+// KVSource policies — the single dimension along which the paged and
+// non-paged attention kernels differ.  Each kernel is templated on one of
+// these (ContigKV / PagedKV) and stays fully generic: the policy owns every
+// place where "where does K/V live" and "what is this request's seq_len"
+// are answered.  All methods are __host__ __device__ so the same policy
+// serves both the device kernels (addressing, seq_len) and the host-side
+// launchers (grid / split computation).
+//
+//   ContigKV:  K/V are dense [batch, kv_head, kv_len, head_dim] tensors.
+//              Params fields used: k, v, kv_stride_*, kv_len, q_len,
+//              q_stride_b, causal_offset.
+//   PagedKV:   K/V live in a flat pool [size, kv_head, head_dim] indexed via
+//              req_to_token.  Params fields used: k_cache, v_cache,
+//              req_to_token, req_pool_indices, kv_indptr, qo_indptr,
+//              max_context_len, q_stride_l.
+//
+// Addressing state that is constant across a whole kernel invocation for one
+// (batch, kv_head) pair is captured once by make_ctx<HEAD_DIM>() and passed
+// to kv_addr, so the load loops never redo the hoistable base computation
+// (e.g. the req_pool_indices global read) element-by-element.
+// ============================================================================
+
+// Every policy method is static + callable from both host and device code.
+#define HOST_DEV_FORCEINLINE static __host__ __device__ __forceinline__
+
+using bf16 = __nv_bfloat16;
+
+// Hoisted per-(batch, kv_head) addressing context.
+struct KVContext {
+    int kv_base;          // contig: batch*kv_stride_b + kv_head*kv_stride_h
+    int64_t req_idx;      // paged: req_pool_indices[batch]
+    int64_t rtt_stride;   // paged: max_context_len
+    int64_t pool_stride;  // paged: kv_head * HEAD_DIM
+    int64_t head_off;     // paged: kv_head * HEAD_DIM
+};
+
+// Per-element K/V global addresses for one (kc, d) position of a K/V tile.
+// The pointers are ALWAYS the computed addresses (never nullptr) — callers
+// gate on `valid` (cp.async src_size=0, or a guarded scalar deref).  `valid`
+// starts as "within the request's seq_len"; the paged policy further degrades
+// it when req_to_token maps the position to a negative slot (empty padding).
+// This matches the original hand-rolled load loops, where the address was
+// always formed and the predicate decided whether anything was read.
+struct KVAddr {
+    const void* k;
+    const void* v;
+    bool valid;
+};
+
+// ---- Contiguous K/V ----
+struct ContigKV {
+    static constexpr bool kPaged = false;
+
+    // host-side length hooks (grid + split computation in the launchers)
+    HOST_DEV_FORCEINLINE int host_q_len(const AttentionParams<bf16>& p) {
+        return p.q_len;
+    }
+    HOST_DEV_FORCEINLINE int host_kv_len(const AttentionParams<bf16>& p) {
+        return p.kv_len;
+    }
+
+    // prefill: element offset of the request's Q rows (kernel adds qrow*q_stride_l)
+    HOST_DEV_FORCEINLINE int q_base(
+        const AttentionParams<bf16>& p, int batch, int q_head) {
+        return batch * p.q_stride_b + q_head * p.q_stride_h;
+    }
+    // decode: same offset (q_len == 1, so there is no row stride component)
+    HOST_DEV_FORCEINLINE int q_decode_base(
+        const AttentionParams<bf16>& p, int batch, int q_head) {
+        return batch * p.q_stride_b + q_head * p.q_stride_h;
+    }
+
+    HOST_DEV_FORCEINLINE int kv_len(const AttentionParams<bf16>& p, int batch) {
+        return p.kv_len;
+    }
+    HOST_DEV_FORCEINLINE int q_len(const AttentionParams<bf16>& p, int batch) {
+        return p.q_len;
+    }
+    HOST_DEV_FORCEINLINE int causal_offset(const AttentionParams<bf16>& p, int batch) {
+        return p.causal_offset;
+    }
+    // decode: exclusive bound of the single query's attend range
+    HOST_DEV_FORCEINLINE int decode_attend_len(const AttentionParams<bf16>& p, int batch) {
+        return (p.kv_len < p.causal_offset + 1) ? p.kv_len : (p.causal_offset + 1);
+    }
+
+    template <int HEAD_DIM>
+    HOST_DEV_FORCEINLINE KVContext make_ctx(
+        const AttentionParams<bf16>& p, int batch, int kv_head) {
+        KVContext c = {};
+        c.kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+        return c;
+    }
+    HOST_DEV_FORCEINLINE KVAddr kv_addr(
+        const AttentionParams<bf16>& p, const KVContext& c, int kc, int d, bool valid) {
+        const int g_off = c.kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
+        return {&p.k[g_off], &p.v[g_off], valid};
+    }
+};
+
+// ---- Paged (SGLang-style flat pool) K/V ----
+struct PagedKV {
+    static constexpr bool kPaged = true;
+
+    HOST_DEV_FORCEINLINE int host_q_len(const AttentionParams<bf16>& p) {
+        return p.max_q_len;
+    }
+    HOST_DEV_FORCEINLINE int host_kv_len(const AttentionParams<bf16>& p) {
+        return p.max_context_len;
+    }
+
+    // prefill: Q rows start at qo_indptr[batch] (ragged batch base)
+    HOST_DEV_FORCEINLINE int q_base(
+        const AttentionParams<bf16>& p, int batch, int q_head) {
+        return p.qo_indptr[batch] * p.q_stride_l + q_head * p.q_stride_h;
+    }
+    // decode: Q is [batch, q_head, head_dim], so batch is the outer row
+    HOST_DEV_FORCEINLINE int q_decode_base(
+        const AttentionParams<bf16>& p, int batch, int q_head) {
+        return batch * p.q_stride_l + q_head * p.q_stride_h;
+    }
+
+    HOST_DEV_FORCEINLINE int kv_len(const AttentionParams<bf16>& p, int batch) {
+        return p.kv_indptr[batch + 1] - p.kv_indptr[batch];
+    }
+    HOST_DEV_FORCEINLINE int q_len(const AttentionParams<bf16>& p, int batch) {
+        return p.qo_indptr[batch + 1] - p.qo_indptr[batch];
+    }
+    HOST_DEV_FORCEINLINE int causal_offset(const AttentionParams<bf16>& p, int batch) {
+        return kv_len(p, batch) - q_len(p, batch);
+    }
+    // decode: the query is the last token, so [0, seq_len) IS its causal range
+    HOST_DEV_FORCEINLINE int decode_attend_len(const AttentionParams<bf16>& p, int batch) {
+        return kv_len(p, batch);
+    }
+
+    template <int HEAD_DIM>
+    HOST_DEV_FORCEINLINE KVContext make_ctx(
+        const AttentionParams<bf16>& p, int batch, int kv_head) {
+        KVContext c = {};
+        c.req_idx = p.req_pool_indices[batch];
+        c.rtt_stride = (int64_t)p.max_context_len;
+        c.pool_stride = (int64_t)p.kv_head * HEAD_DIM;
+        c.head_off = (int64_t)kv_head * HEAD_DIM;
+        return c;
+    }
+    HOST_DEV_FORCEINLINE KVAddr kv_addr(
+        const AttentionParams<bf16>& p, const KVContext& c, int kc, int d, bool valid) {
+        const int64_t slot = valid ? p.req_to_token[c.req_idx * c.rtt_stride + kc] : 0;
+        const bool ok = valid && (slot >= 0);
+        const int64_t gmem_off = slot * c.pool_stride + c.head_off + d;
+        return {&p.k_cache[gmem_off], &p.v_cache[gmem_off], ok};
+    }
+};

csrc/kernels/attn_mma_utils.cuh

@@ -0,0 +1,297 @@
+#pragma once
+#include <cfloat>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+
+// Predicated cp.async (4-operand form) requires CUDA 11.2+.
+// bf16 mma.sync requires sm_80+ (guarded at build time by ASTRAI_NO_MMA).
+#if CUDART_VERSION < 11020
+#error "AstrAI CUDA kernels require CUDA 11.2 or later (CUDART_VERSION >= 11020)."
+#endif
+
+// ============================================================================
+// KernelTraits — FlashAttention-v2 style compile-time configuration bundle.
+//
+// Bundles all dimension-dependent constants so device functions only need a
+// single Traits template parameter rather than scattered <KD, NC8, KT2, ...>.
+// ============================================================================
+template <int HEAD_DIM_, int BC_, int WARPS_, int STAGES_>
+struct KernelTraits {
+    static constexpr int HEAD_DIM = HEAD_DIM_;
+    static constexpr int BC      = BC_;       // K/V tile size along seq dim
+    static constexpr int WARPS   = WARPS_;    // warps per block
+    static constexpr int STAGES  = STAGES_;   // double-buffer stages (1 or 2)
+
+    static constexpr int BR = 16;             // Q rows per warp (mma M=16)
+
+    // Derived: mma.sync.m16n8k16 tile counts
+    static constexpr int KD  = HEAD_DIM / 16;  // Q/K k-slides
+    static constexpr int NC8 = BC / 8;          // S n-tiles (N=8)
+    static constexpr int KT2 = BC / 16;         // P k-tiles (K=16)
+    static constexpr int DN8 = HEAD_DIM / 8;    // O n-tiles (N=8)
+
+    static constexpr int LD = HEAD_DIM;         // smem leading dim
+
+    // XOR swizzle chunk bits for ldmatrix bank-conflict avoidance.
+    // mask = log2(LD/8) bits, clamped to stay within LD.
+    static constexpr int SWIZ_MASK = (HEAD_DIM >= 64) ? 7 : (HEAD_DIM / 8 - 1);
+
+    static constexpr int NUM_THREADS = WARPS * 32;
+    static constexpr int VEC = 8;               // bf16 per cp.async unit (16 bytes)
+    static constexpr int TOTAL = BC * HEAD_DIM; // total elements per tile
+};
+
+// ---- PTX wrappers ----
+using bf16 = __nv_bfloat16;
+
+__device__ __forceinline__ void mma16816(float* d, const unsigned* a,
+                                          const unsigned* b, const float* c) {
+    asm volatile(
+        "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+        "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};"
+        : "=f"(d[0]), "=f"(d[1]), "=f"(d[2]), "=f"(d[3])
+        : "r"(a[0]), "r"(a[1]), "r"(a[2]), "r"(a[3]), "r"(b[0]), "r"(b[1]),
+          "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]));
+}
+
+// read two adjacent bf16 from smem as one packed .b32 (elem0 low, elem1 high)
+__device__ __forceinline__ unsigned ld2(const bf16* p) {
+    return *reinterpret_cast<const unsigned*>(p);
+}
+
+// pack two floats into one bf16x2 as .b32
+__device__ __forceinline__ unsigned pk2(float a, float b) {
+    __nv_bfloat162 v = __floats2bfloat162_rn(a, b);
+    return *reinterpret_cast<unsigned*>(&v);
+}
+
+// pack two (non-contiguous) bf16 into one .b32
+__device__ __forceinline__ unsigned pkb(bf16 a, bf16 b) {
+    __nv_bfloat162 v;
+    v.x = a;
+    v.y = b;
+    return *reinterpret_cast<unsigned*>(&v);
+}
+
+// ldmatrix: cooperatively load mma fragments from smem (one instruction per
+// 16x16 / 16x8 tile) with the exact register layout mma expects.
+__device__ __forceinline__ void ldmatrix_x4(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x4.shared.b16 {%0,%1,%2,%3}, [%4];"
+                 : "=r"(r[0]), "=r"(r[1]), "=r"(r[2]), "=r"(r[3])
+                 : "r"(a));
+}
+__device__ __forceinline__ void ldmatrix_x2(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x2.shared.b16 {%0,%1}, [%2];"
+                 : "=r"(r[0]), "=r"(r[1])
+                 : "r"(a));
+}
+__device__ __forceinline__ void ldmatrix_x2_trans(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x2.trans.shared.b16 {%0,%1}, [%2];"
+                 : "=r"(r[0]), "=r"(r[1])
+                 : "r"(a));
+}
+
+// XOR swizzle for shared-memory column at 8-bf16 chunk granularity.
+__device__ __forceinline__ int swiz_col(int d, int r, int mask = 7) {
+    return ((d >> 3) ^ (r & mask)) << 3 | (d & 7);
+}
+
+// cp.async: copy 16 bytes (8 bf16) from global to shared memory directly.
+__device__ __forceinline__ void cp_async_16(bf16* smem_ptr, const void* gmem_ptr) {
+    unsigned smem_addr = __cvta_generic_to_shared(smem_ptr);
+    asm volatile("cp.async.ca.shared.global [%0], [%1], 16;"
+                 :: "r"(smem_addr), "l"(gmem_ptr));
+}
+
+// Predicated cp.async: copy 16 bytes when `pred`, otherwise zero-fill.
+// src_size=0 → no bytes read from src, so out-of-bounds src address is safe.
+__device__ __forceinline__ void cp_async_16_pred(bf16* smem_ptr,
+                                                  const void* gmem_ptr,
+                                                  bool pred) {
+    unsigned smem_addr = __cvta_generic_to_shared(smem_ptr);
+    int src_size = pred ? 16 : 0;
+    asm volatile("cp.async.ca.shared.global [%0], [%1], 16, %2;"
+                 :: "r"(smem_addr), "l"(gmem_ptr), "r"(src_size));
+}
+
+__device__ __forceinline__ void cp_async_commit() {
+    asm volatile("cp.async.commit_group;");
+}
+
+__device__ __forceinline__ void cp_async_wait_all() {
+    asm volatile("cp.async.wait_all;");
+}
+
+template <int N>
+__device__ __forceinline__ void cp_async_wait_group() {
+    asm volatile("cp.async.wait_group %0;" :: "n"(N));
+}
+
+// ---------------------------------------------------------------------------
+// Q-load: load query rows directly from global memory into mma A-operand
+// register layout. One call replaces ~15 duplicated lines in each MMA kernel.
+// stride_row is p.q_stride_h for decode (q_len=1, G heads) or
+//              p.q_stride_l for prefill (multi-q rows).
+// ---------------------------------------------------------------------------
+template <int KD>
+__device__ inline void load_q_mma_frags(
+    const bf16* __restrict__ q,
+    int stride_row,
+    int stride_d,
+    int qra, int qrb,
+    bool va, bool vb,
+    int tid4,
+    unsigned Qa[KD][4])
+{
+    #pragma unroll
+    for (int kt = 0; kt < KD; kt++) {
+        int c = kt * 16 + tid4 * 2;
+        const unsigned* pau = reinterpret_cast<const unsigned*>(
+            &q[qra * stride_row + c * stride_d]);
+        const unsigned* pbu = reinterpret_cast<const unsigned*>(
+            &q[qrb * stride_row + c * stride_d]);
+        Qa[kt][0] = va ? pau[0] : 0u;
+        Qa[kt][1] = vb ? pbu[0] : 0u;
+        Qa[kt][2] = va ? pau[4] : 0u;
+        Qa[kt][3] = vb ? pbu[4] : 0u;
+    }
+}
+
+// ---------------------------------------------------------------------------
+// S = Q @ K^T  (Qa pre-loaded by the caller; scale applied post-mma in the
+// caller to avoid bf16 precision loss).
+// Traits provides KD, NC8, LD, and SWIZ_MASK.
+// ---------------------------------------------------------------------------
+template <typename Traits>
+__device__ inline void mma_compute_scores(
+    const unsigned Qa[Traits::KD][4],
+    const bf16* __restrict__ sK,
+    int lane,
+    float Sacc[Traits::NC8][4])
+{
+    #pragma unroll
+    for (int n8 = 0; n8 < Traits::NC8; n8++) {
+        Sacc[n8][0] = Sacc[n8][1] = Sacc[n8][2] = Sacc[n8][3] = 0.0f;
+        int krow_l = n8 * 8 + (lane & 7);
+        int kcol_h = (lane & 8) ? 8 : 0;
+        #pragma unroll
+        for (int kt = 0; kt < Traits::KD; kt++) {
+            unsigned b[2];
+            ldmatrix_x2(b, &sK[krow_l * Traits::LD
+                + swiz_col(kt * 16 + kcol_h, krow_l, Traits::SWIZ_MASK)]);
+            mma16816(Sacc[n8], Qa[kt], b, Sacc[n8]);
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Online softmax + Oacc rescale for one K/V tile.
+//
+// HasMask is a compile-time template bool: when false, the mask branch is
+// entirely dead-code-eliminated from the inner unrolled loop.
+// ---------------------------------------------------------------------------
+template <typename Traits, bool HasMask>
+__device__ inline void mma_softmax_tile(
+    int kv0,
+    int maxc0, int maxc1,
+    int qrow0, int qrow1,
+    int mask_b_stride, int mask_h_stride, int mask_q_stride,
+    int mask_batch, int mask_head,
+    const bool* __restrict__ mask,
+    float Sacc[Traits::NC8][4],
+    float Oacc[Traits::DN8][4],
+    float& m0, float& m1,
+    float& l0, float& l1,
+    int lane)
+{
+    int tid4 = lane & 3;
+
+    float rmax0 = -FLT_MAX, rmax1 = -FLT_MAX;
+    int mask_base0 = mask_batch * mask_b_stride + mask_head * mask_h_stride + qrow0 * mask_q_stride;
+    int mask_base1 = mask_batch * mask_b_stride + mask_head * mask_h_stride + qrow1 * mask_q_stride;
+    #pragma unroll
+    for (int n8 = 0; n8 < Traits::NC8; n8++) {
+        int cc = kv0 + n8 * 8 + 2 * tid4;
+        int c1 = cc + 1;
+        bool b0 = (cc >= maxc0) || (HasMask && !mask[mask_base0 + cc]);
+        bool b1 = (c1 >= maxc0) || (HasMask && !mask[mask_base0 + c1]);
+        bool b2 = (cc >= maxc1) || (HasMask && !mask[mask_base1 + cc]);
+        bool b3 = (c1 >= maxc1) || (HasMask && !mask[mask_base1 + c1]);
+        float s0 = b0 ? -FLT_MAX : Sacc[n8][0];
+        float s1 = b1 ? -FLT_MAX : Sacc[n8][1];
+        float s2 = b2 ? -FLT_MAX : Sacc[n8][2];
+        float s3 = b3 ? -FLT_MAX : Sacc[n8][3];
+        Sacc[n8][0] = s0; Sacc[n8][1] = s1;
+        Sacc[n8][2] = s2; Sacc[n8][3] = s3;
+        rmax0 = fmaxf(rmax0, fmaxf(s0, s1));
+        rmax1 = fmaxf(rmax1, fmaxf(s2, s3));
+    }
+    rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 1));
+    rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 2));
+    rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 1));
+    rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 2));
+
+    float nm0 = fmaxf(m0, rmax0), nm1 = fmaxf(m1, rmax1);
+    float corr0 = __expf(m0 - nm0);
+    float corr1 = __expf(m1 - nm1);
+    float pn0 = (nm0 == -FLT_MAX) ? 0.0f : 1.0f;
+    float pn1 = (nm1 == -FLT_MAX) ? 0.0f : 1.0f;
+
+    float rsum0 = 0.0f, rsum1 = 0.0f;
+    #pragma unroll
+    for (int n8 = 0; n8 < Traits::NC8; n8++) {
+        float p0 = pn0 * __expf(Sacc[n8][0] - nm0);
+        float p1 = pn0 * __expf(Sacc[n8][1] - nm0);
+        float p2 = pn1 * __expf(Sacc[n8][2] - nm1);
+        float p3 = pn1 * __expf(Sacc[n8][3] - nm1);
+        Sacc[n8][0] = p0; Sacc[n8][1] = p1;
+        Sacc[n8][2] = p2; Sacc[n8][3] = p3;
+        rsum0 += p0 + p1;
+        rsum1 += p2 + p3;
+    }
+    rsum0 += __shfl_xor_sync(0xFFFFFFFF, rsum0, 1);
+    rsum0 += __shfl_xor_sync(0xFFFFFFFF, rsum0, 2);
+    rsum1 += __shfl_xor_sync(0xFFFFFFFF, rsum1, 1);
+    rsum1 += __shfl_xor_sync(0xFFFFFFFF, rsum1, 2);
+    l0 = l0 * corr0 + rsum0;
+    l1 = l1 * corr1 + rsum1;
+    m0 = nm0; m1 = nm1;
+
+    #pragma unroll
+    for (int j = 0; j < Traits::DN8; j++) {
+        Oacc[j][0] *= corr0; Oacc[j][1] *= corr0;
+        Oacc[j][2] *= corr1; Oacc[j][3] *= corr1;
+    }
+}
+
+// ---------------------------------------------------------------------------
+// O += P @ V  (Sacc must contain P = attention weights after softmax).
+// Traits provides DN8, KT2, LD, and SWIZ_MASK.
+// ---------------------------------------------------------------------------
+template <typename Traits>
+__device__ inline void mma_pv_accumulate(
+    float Sacc[][4],
+    const bf16* __restrict__ sV,
+    int lane,
+    float Oacc[Traits::DN8][4])
+{
+    #pragma unroll
+    for (int kt2 = 0; kt2 < Traits::KT2; kt2++) {
+        unsigned Pa[4];
+        Pa[0] = pk2(Sacc[kt2 * 2][0], Sacc[kt2 * 2][1]);
+        Pa[1] = pk2(Sacc[kt2 * 2][2], Sacc[kt2 * 2][3]);
+        Pa[2] = pk2(Sacc[kt2 * 2 + 1][0], Sacc[kt2 * 2 + 1][1]);
+        Pa[3] = pk2(Sacc[kt2 * 2 + 1][2], Sacc[kt2 * 2 + 1][3]);
+        int vrow_l = kt2 * 16 + (lane & 15);
+        #pragma unroll
+        for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+            unsigned b[2];
+            ldmatrix_x2_trans(b, &sV[vrow_l * Traits::LD
+                + swiz_col(dn8 * 8, vrow_l, Traits::SWIZ_MASK)]);
+            mma16816(Oacc[dn8], Pa, b, Oacc[dn8]);
+        }
+    }
+}

csrc/kernels/attn_paged_decode.cu

@@ -0,0 +1,74 @@
+#include "attn_dispatchers.cuh"
+#include "attn_entry_utils.cuh"
+
+torch::Tensor attn_paged_decode(
+    torch::Tensor q,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    torch::Tensor req_to_token,
+    torch::Tensor req_pool_indices,
+    torch::Tensor kv_indptr,
+    int64_t max_seq_len,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    c10::optional<torch::Tensor> o_part_buf,
+    c10::optional<torch::Tensor> ml_part_buf,
+    c10::optional<torch::Tensor> out_buf
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+    auto stream = at::cuda::getCurrentCUDAStream();
+
+    AttentionParams<bf16> p;
+    attn_pack_paged_decode_params(q, k_cache, v_cache,
+                                   req_to_token, req_pool_indices, kv_indptr,
+                                   max_seq_len, mask, causal_offset, scale, p);
+
+    torch::Tensor O;
+    if (out_buf.has_value() && out_buf->defined()) {
+        TORCH_CHECK(out_buf->dtype() == q.dtype(), "out_buf dtype must match q");
+        TORCH_CHECK(out_buf->size(0) >= q.size(0), "out_buf batch too small");
+        TORCH_CHECK(out_buf->size(1) >= q.size(1), "out_buf heads too small");
+        TORCH_CHECK(out_buf->size(2) >= q.size(2), "out_buf head_dim too small");
+        O = out_buf.value().slice(0, 0, q.size(0))
+                            .slice(1, 0, q.size(1))
+                            .slice(2, 0, q.size(2));
+    } else {
+        O = torch::empty({q.size(0), q.size(1), q.size(2)}, q.options());
+    }
+    p.o = (bf16*)O.data_ptr();
+
+    if (o_part_buf.has_value() && ml_part_buf.has_value()
+        && o_part_buf->defined() && ml_part_buf->defined()) {
+        TORCH_CHECK(o_part_buf->scalar_type() == torch::kFloat32, "o_part_buf must be f32");
+        TORCH_CHECK(ml_part_buf->scalar_type() == torch::kFloat32, "ml_part_buf must be f32");
+        int64_t o_needed = (int64_t)p.batch * p.q_head * MAX_SPLITS * p.head_dim;
+        TORCH_CHECK(o_part_buf->numel() >= o_needed,
+                     "o_part_buf too small: need ", o_needed, " got ", o_part_buf->numel());
+        p.o_part = (float*)o_part_buf->data_ptr();
+        p.ml_part = (float*)ml_part_buf->data_ptr();
+    } else {
+        alloc_split_partials(p);
+    }
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_decode, p, stream);
+    C10_CUDA_CHECK(cudaGetLastError());
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_paged_decode", &attn_paged_decode,
+        py::arg("q"),
+        py::arg("k_cache"),
+        py::arg("v_cache"),
+        py::arg("req_to_token"),
+        py::arg("req_pool_indices"),
+        py::arg("kv_indptr"),
+        py::arg("max_seq_len"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("o_part_buf") = py::none(),
+        py::arg("ml_part_buf") = py::none(),
+        py::arg("out_buf") = py::none(),
+        "SGLang-style paged decode: flat KV pool + req_to_token + kv_indptr.");
+}

csrc/kernels/attn_paged_prefill.cu

@@ -0,0 +1,48 @@
+#include "attn_dispatchers.cuh"
+#include "attn_entry_utils.cuh"
+
+torch::Tensor attn_paged_prefill(
+    torch::Tensor q,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    torch::Tensor req_to_token,
+    torch::Tensor req_pool_indices,
+    torch::Tensor kv_indptr,
+    torch::Tensor qo_indptr,
+    c10::optional<torch::Tensor> mask,
+    int64_t max_q_len,
+    int64_t causal_offset,
+    double scale
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+    auto stream = at::cuda::getCurrentCUDAStream();
+
+    AttentionParams<bf16> p;
+    attn_pack_paged_prefill_params(q, k_cache, v_cache,
+                                    req_to_token, req_pool_indices,
+                                    kv_indptr, qo_indptr, mask,
+                                    max_q_len, causal_offset, scale, p);
+
+    auto O = torch::empty({q.size(0), q.size(1), q.size(2)}, q.options());
+    p.o = (bf16*)O.data_ptr();
+
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_prefill, p, stream);
+    C10_CUDA_CHECK(cudaGetLastError());
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_paged_prefill", &attn_paged_prefill,
+        py::arg("q"),
+        py::arg("k_cache"),
+        py::arg("v_cache"),
+        py::arg("req_to_token"),
+        py::arg("req_pool_indices"),
+        py::arg("kv_indptr"),
+        py::arg("qo_indptr"),
+        py::arg("mask") = py::none(),
+        py::arg("max_q_len"),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        "SGLang-style paged prefill: flat KV pool + ragged batch.");
+}

csrc/kernels/attn_prefill.cu

@@ -0,0 +1,39 @@
+#include "attn_dispatchers.cuh"
+#include "attn_entry_utils.cuh"
+
+torch::Tensor attn_prefill(
+    torch::Tensor q,
+    torch::Tensor k,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+    auto stream = at::cuda::getCurrentCUDAStream();
+
+    AttentionParams<bf16> p;
+    attn_pack_params(q, k, v, mask, causal_offset, scale, layout, p);
+    TORCH_CHECK(p.head_dim % 16 == 0, "head_dim must be multiple of 16");
+
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
+    auto O_view = (layout == BLHD) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_prefill, p, stream);
+    C10_CUDA_CHECK(cudaGetLastError());
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_prefill", &attn_prefill,
+        py::arg("q"),
+        py::arg("k"),
+        py::arg("v"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = (int64_t)BHLD,
+        "GQA prefill (tensor-core mma on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_prefill_split_q.cuh

@@ -0,0 +1,154 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_kv_source.cuh"
+
+using bf16 = __nv_bfloat16;
+
+// v9: group-split register blocking. G threads cooperate on one query row,
+// each owning HEAD_DIM/G dims of qreg[]/acc[]. IsCausal and HasMask are
+// compile-time bools — the compiler eliminates dead branches.
+// Unified across contiguous and paged (SGLang flat-pool) K/V via KV.
+// Templated on <HEAD_DIM, KV, G, ROWS, P_BC, IsCausal, HasMask>.
+
+template <int G>
+__device__ __forceinline__ float group_reduce_sum(float v, unsigned mask) {
+#pragma unroll
+    for (int o = G / 2; o > 0; o >>= 1)
+        v += __shfl_xor_sync(mask, v, o);
+    return v;
+}
+
+// load 8 contiguous bf16 from (16-byte aligned) smem as one float4
+__device__ __forceinline__ void ld8(const bf16* p, float* o) {
+    float4 raw = *reinterpret_cast<const float4*>(p);
+    const __nv_bfloat162* h = reinterpret_cast<const __nv_bfloat162*>(&raw);
+#pragma unroll
+    for (int j = 0; j < 4; j++) {
+        float2 f = __bfloat1622float2(h[j]);
+        o[2 * j] = f.x;
+        o[2 * j + 1] = f.y;
+    }
+}
+
+template <int HEAD_DIM, typename KV, int G, int ROWS, int P_BC, bool IsCausal, bool HasMask>
+__global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
+    constexpr int DPT = HEAD_DIM / G;
+
+    int q_tile = blockIdx.x;
+    int q_head = blockIdx.y;
+    int batch  = blockIdx.z;
+    int gpos   = threadIdx.x;  // 0..G-1  (which d-chunk)
+    int row    = threadIdx.y;  // 0..ROWS-1
+    int q_row  = q_tile * ROWS + row;
+
+    // Per-request dims (from KV policy — paged reads kv_indptr/qo_indptr).
+    const int seq_len    = KV::kv_len(p, batch);
+    const int q_len      = KV::q_len(p, batch);
+    const int causal_off = KV::causal_offset(p, batch);
+    const int kv_head = q_head / (p.q_head / p.kv_head);
+    const KVContext kctx = KV::template make_ctx<HEAD_DIM>(p, batch, kv_head);
+
+    __shared__ __align__(16) bf16 sK[P_BC * HEAD_DIM];
+    __shared__ __align__(16) bf16 sV[P_BC * HEAD_DIM];
+
+    // Q: stride-based load [batch, q_head, q_len, head_dim]
+    const int q_base = KV::q_base(p, batch, q_head);
+    float qreg[DPT];
+    if (q_row < q_len) {
+        int q_off = q_base + q_row * p.q_stride_l + gpos * DPT * p.q_stride_d;
+#pragma unroll
+        for (int i = 0; i < DPT; i++)
+            qreg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
+    }
+
+    float m = -FLT_MAX, l = 0.0f;
+    float acc[DPT];
+#pragma unroll
+    for (int i = 0; i < DPT; i++)
+        acc[i] = 0.0f;
+
+    int mask_batch_base = batch * p.mask_b_stride + q_head * p.mask_h_stride;
+    int tiles   = (seq_len + P_BC - 1) / P_BC;
+    int tt      = G * ROWS;
+    int lid     = row * G + gpos;
+
+    int lane_in_warp = lid & 31;
+    unsigned gmask = (G == 32) ? 0xFFFFFFFFu
+                               : (((1u << G) - 1u) << (lane_in_warp & ~(G - 1)));
+
+    for (int ti = 0; ti < tiles; ti++) {
+        int kv0  = ti * P_BC;
+        int tlen = min(P_BC, seq_len - kv0);
+
+        // Load K/V into shared memory (addressing via KV policy; paged
+        // guards empty slots with zero-fill).
+        for (int i = lid; i < tlen * HEAD_DIM; i += tt) {
+            int s = i / HEAD_DIM;
+            int d_dim = i % HEAD_DIM;
+            int kc = kv0 + s;
+            KVAddr a = KV::kv_addr(p, kctx, kc, d_dim, true);
+            sK[i] = a.valid ? *reinterpret_cast<const bf16*>(a.k) : (bf16)0.f;
+            sV[i] = a.valid ? *reinterpret_cast<const bf16*>(a.v) : (bf16)0.f;
+        }
+        __syncthreads();
+
+        int lim = tlen;
+        if constexpr (IsCausal) {
+            if (q_row < q_len) {
+                int ep = causal_off + q_row + 1;
+                if (kv0 >= ep)
+                    lim = 0;
+                else if (kv0 + tlen > ep)
+                    lim = ep - kv0;
+            }
+        }
+
+        int mask_row_base = mask_batch_base + q_row * p.mask_q_stride;
+        for (int s = 0; s < lim; s++) {
+            const bf16* kr = sK + s * HEAD_DIM + gpos * DPT;
+            float part = 0.0f;
+#pragma unroll
+            for (int i = 0; i < DPT; i += 8) {
+                float k8[8];
+                ld8(kr + i, k8);
+#pragma unroll
+                for (int j = 0; j < 8; j++)
+                    part = fmaf(qreg[i + j], k8[j], part);
+            }
+            float dot = group_reduce_sum<G>(part, gmask) * p.scale;
+
+            int kv_idx = kv0 + s;
+            if constexpr (HasMask) {
+                if (!p.mask[mask_row_base + kv_idx])
+                    dot = -FLT_MAX;
+            }
+
+            float nm = fmaxf(m, dot);
+            float al = __expf(m - nm);
+            float be = __expf(dot - nm);
+            l = l * al + be;
+
+            const bf16* vr = sV + s * HEAD_DIM + gpos * DPT;
+#pragma unroll
+            for (int i = 0; i < DPT; i += 8) {
+                float v8[8];
+                ld8(vr + i, v8);
+#pragma unroll
+                for (int j = 0; j < 8; j++)
+                    acc[i + j] = fmaf(v8[j], be, acc[i + j] * al);
+            }
+            m = nm;
+        }
+        __syncthreads();
+    }
+
+    if (q_row < q_len) {
+        int o_off = q_base + q_row * p.q_stride_l + gpos * DPT * p.q_stride_d;
+        float rl = (l > 1e-20f) ? (1.0f / l) : 0.0f;
+#pragma unroll
+        for (int i = 0; i < DPT; i++)
+            p.o[o_off + i * p.q_stride_d] = __float2bfloat16(acc[i] * rl);
+    }
+}

csrc/kernels/attn_prefill_split_q_mma.cuh

@@ -0,0 +1,154 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_kv_source.cuh"
+#include "attn_mma_utils.cuh"
+
+// Tensor-core prefill flash attention (raw mma.sync PTX), unified across
+// contiguous and paged (SGLang flat-pool) K/V via the KV template parameter.
+// One warp owns BR=16 query rows. S = Q@K^T and O = P@V run on bf16 tensor
+// cores via mma.sync.m16n8k16 (f32 accumulate).
+//
+// KV = ContigKV (dense [batch, kv_head, kv_len, head_dim]) or PagedKV
+//      (flat pool + req_to_token, ragged batches via qo_indptr/kv_indptr).
+// IsCausal and HasMask are compile-time bools — the compiler eliminates all
+// dead branches in the inner compute loop (FA2-style).
+//
+// Traits = KernelTraits<HEAD_DIM, BC, WARPS=4, STAGES=2>.
+template <typename Traits, typename KV, bool IsCausal, bool HasMask>
+__global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
+    const int warp = threadIdx.x / 32;
+    const int lane = threadIdx.x % 32;
+    const int gid = lane >> 2;   // 0..7
+    const int tid4 = lane & 3;   // 0..3
+
+    const int q_head = blockIdx.y;
+    const int batch = blockIdx.z;
+    const int kv_head = q_head / (p.q_head / p.kv_head);
+    const int qrow0 = (blockIdx.x * Traits::WARPS + warp) * Traits::BR;
+
+    // Per-request dims (from KV policy — paged reads kv_indptr/qo_indptr).
+    const int seq_len    = KV::kv_len(p, batch);
+    const int q_len      = KV::q_len(p, batch);
+    const int causal_off = KV::causal_offset(p, batch);
+    const KVContext kctx = KV::template make_ctx<Traits::HEAD_DIM>(p, batch, kv_head);
+
+    // Static shared memory: double-buffered K/V (no sQ — Q goes direct
+    // to registers in mma A-operand layout).
+    __shared__ __align__(16) bf16 sK[Traits::STAGES * Traits::BC * Traits::LD];
+    __shared__ __align__(16) bf16 sV[Traits::STAGES * Traits::BC * Traits::LD];
+
+    // Load Q fragments straight from global into mma A-operand layout.
+    const int q_base = KV::q_base(p, batch, q_head);
+    const int qra = qrow0 + gid;
+    const int qrb = qrow0 + gid + 8;
+    const bool va = qra < q_len, vb = qrb < q_len;
+    unsigned Qa[Traits::KD][4];
+    load_q_mma_frags<Traits::KD>(p.q + q_base, p.q_stride_l, p.q_stride_d,
+                                  qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[Traits::DN8][4];
+    #pragma unroll
+    for (int j = 0; j < Traits::DN8; j++)
+        Oacc[j][0] = Oacc[j][1] = Oacc[j][2] = Oacc[j][3] = 0.0f;
+    float m0 = -FLT_MAX, m1 = -FLT_MAX, l0 = 0.0f, l1 = 0.0f;
+
+    const int tiles = (seq_len + Traits::BC - 1) / Traits::BC;
+    const int qr0 = qrow0 + gid;
+    const int qr1 = qrow0 + gid + 8;
+
+    // Causal tile-skip bounds (dead code when IsCausal == false)
+    const int max_kv = qrow0 + Traits::BR - 1 + causal_off;
+    const int block_max_kv =
+        blockIdx.x * Traits::WARPS * Traits::BR + Traits::WARPS * Traits::BR - 1
+        + causal_off;
+
+    int t_end = tiles - 1;
+    if constexpr (IsCausal) {
+        int bt = block_max_kv / Traits::BC;
+        if (bt < t_end) t_end = bt;
+    }
+
+    // ---- Load tile lambda: predicated cp.async (addressing via KV policy) ----
+    auto load_tile = [&](int ti, int buf) {
+        int kv0 = ti * Traits::BC;
+        bf16* dK = sK + buf * Traits::BC * Traits::LD;
+        bf16* dV = sV + buf * Traits::BC * Traits::LD;
+        #pragma unroll
+        for (int i = threadIdx.x * Traits::VEC; i < Traits::TOTAL;
+             i += Traits::NUM_THREADS * Traits::VEC) {
+            int r = i / Traits::HEAD_DIM, d = i % Traits::HEAD_DIM;
+            int kc = kv0 + r;
+            bool valid = kc < seq_len;
+            KVAddr a = KV::kv_addr(p, kctx, kc, d, valid);
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            cp_async_16_pred(&dK[off], a.k, a.valid);
+            cp_async_16_pred(&dV[off], a.v, a.valid);
+        }
+        cp_async_commit();
+    };
+
+    // ---- Prologue: issue first tile load ----
+    load_tile(0, 0);
+
+    for (int ti = 0; ti <= t_end; ti++) {
+        int buf = ti & 1;
+
+        // Wait for current tile, then publish cross-warp + guard buffer reuse.
+        cp_async_wait_group<0>();
+        __syncthreads();
+        if (ti < t_end) load_tile(ti + 1, (ti + 1) & 1);
+
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = ti * Traits::BC;
+
+        // Warp-level causal skip (dead branch eliminated when IsCausal == false)
+        if (!IsCausal || kv0 <= max_kv) {
+
+            float Sacc[Traits::NC8][4];
+            mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
+
+            // Post-multiply scale in float (no bf16 precision loss)
+            #pragma unroll
+            for (int n8 = 0; n8 < Traits::NC8; n8++)
+                Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
+                Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
+
+            int maxc0 = IsCausal ? min(seq_len, causal_off + qr0 + 1)
+                                 : seq_len;
+            int maxc1 = IsCausal ? min(seq_len, causal_off + qr1 + 1)
+                                 : seq_len;
+            mma_softmax_tile<Traits, HasMask>(kv0, maxc0, maxc1,
+                                               qr0, qr1,
+                                               p.mask_b_stride, p.mask_h_stride, p.mask_q_stride,
+                                               batch, q_head,
+                                               p.mask,
+                                               Sacc, Oacc, m0, m1, l0, l1, lane);
+
+            mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+        }
+    }
+
+    // ---- write output: packed bf16x2 stores ----
+    float rl0 = (l0 > 1e-20f) ? (1.0f / l0) : 0.0f;
+    float rl1 = (l1 > 1e-20f) ? (1.0f / l1) : 0.0f;
+    const int o_base = KV::q_base(p, batch, q_head);
+    #pragma unroll
+    for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+        int d = dn8 * 8 + 2 * tid4;
+        if (qr0 < q_len) {
+            __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][0] * rl0,
+                                                      Oacc[dn8][1] * rl0);
+            *reinterpret_cast<__nv_bfloat162*>(
+                &p.o[o_base + qr0 * p.q_stride_l + d * p.q_stride_d]) = v;
+        }
+        if (qr1 < q_len) {
+            __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][2] * rl1,
+                                                      Oacc[dn8][3] * rl1);
+            *reinterpret_cast<__nv_bfloat162*>(
+                &p.o[o_base + qr1 * p.q_stride_l + d * p.q_stride_d]) = v;
+        }
+    }
+}

csrc/kernels/attn_warp_utils.cuh

@@ -0,0 +1,13 @@
+#pragma once
+#include <cuda_bf16.h>
+
+using bf16 = __nv_bfloat16;
+
+static constexpr int MAX_SPLITS = 32;
+
+__device__ inline float warp_reduce_sum(float val) {
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        val += __shfl_xor_sync(0xFFFFFFFF, val, offset);
+    return val;
+}

csrc/kernels/rotary_emb.cu

@@ -0,0 +1,98 @@
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <c10/cuda/CUDAException.h>
+#include <cuda_bf16.h>
+
+__global__ void rotary_emb_kernel(
+    const __nv_bfloat16* __restrict__ x,
+    const float* __restrict__ freqs_cis,
+    __nv_bfloat16* __restrict__ out,
+    int batch,
+    int seq_len,
+    int n_heads,
+    int head_dim
+) {
+    const int half_dim = head_dim >> 1;
+    const int total = batch * seq_len * n_heads * half_dim;
+
+    for (int idx = blockIdx.x * blockDim.x + threadIdx.x;
+         idx < total;
+         idx += gridDim.x * blockDim.x) {
+
+        int pair = idx % half_dim;
+        int tmp = idx / half_dim;
+        int head = tmp % n_heads;
+        tmp /= n_heads;
+        int seq = tmp % seq_len;
+        int b = tmp / seq_len;
+
+        int x_offset = ((b * seq_len + seq) * n_heads + head) * head_dim + (pair << 1);
+        int cs_offset = ((b * seq_len + seq) * half_dim + pair) * 2;
+
+        __nv_bfloat162 x_pair = *reinterpret_cast<const __nv_bfloat162*>(x + x_offset);
+        float x_even = __bfloat162float(__low2bfloat16(x_pair));
+        float x_odd  = __bfloat162float(__high2bfloat16(x_pair));
+
+        float c = freqs_cis[cs_offset];
+        float s = freqs_cis[cs_offset + 1];
+
+        float out_even = x_even * c - x_odd * s;
+        float out_odd  = x_even * s + x_odd * c;
+
+        __nv_bfloat162 out_pair = __floats2bfloat162_rn(out_even, out_odd);
+        *reinterpret_cast<__nv_bfloat162*>(out + x_offset) = out_pair;
+    }
+}
+
+torch::Tensor rotary_emb(
+    torch::Tensor x,
+    torch::Tensor freqs_cis
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
+    auto stream = at::cuda::getCurrentCUDAStream();
+
+    TORCH_CHECK(x.is_cuda(), "x must be on CUDA");
+    TORCH_CHECK(freqs_cis.is_cuda(), "freqs_cis must be on CUDA");
+    TORCH_CHECK(x.scalar_type() == torch::kBFloat16, "x must be bf16");
+    TORCH_CHECK(x.dim() == 4, "x must be 4D [batch, seq_len, n_heads, head_dim]");
+    TORCH_CHECK(x.is_contiguous(), "x must be contiguous");
+    TORCH_CHECK(freqs_cis.dim() == 4, "freqs_cis must be 4D [batch, seq_len, dim/2, 2]");
+    TORCH_CHECK(freqs_cis.is_contiguous(), "freqs_cis must be contiguous");
+    TORCH_CHECK(freqs_cis.scalar_type() == torch::kFloat32, "freqs_cis must be f32");
+
+    int batch = x.size(0);
+    int seq_len = x.size(1);
+    int n_heads = x.size(2);
+    int head_dim = x.size(3);
+
+    TORCH_CHECK(head_dim % 2 == 0, "head_dim must be even");
+    TORCH_CHECK(freqs_cis.size(0) == batch, "freqs_cis batch mismatch");
+    TORCH_CHECK(freqs_cis.size(1) == seq_len, "freqs_cis seq_len mismatch");
+    TORCH_CHECK(freqs_cis.size(2) == head_dim / 2, "freqs_cis dim/2 mismatch");
+    TORCH_CHECK(freqs_cis.size(3) == 2, "freqs_cis last dim must be 2 [cos, sin]");
+
+    auto out = torch::empty_like(x);
+
+    int half_dim = head_dim / 2;
+    int total = batch * seq_len * n_heads * half_dim;
+    int block = 256;
+    int grid = std::min((total + block - 1) / block, 1024);
+
+    rotary_emb_kernel<<<grid, block, 0, stream>>>(
+        reinterpret_cast<const __nv_bfloat16*>(x.data_ptr()),
+        freqs_cis.data_ptr<float>(),
+        reinterpret_cast<__nv_bfloat16*>(out.data_ptr()),
+        batch, seq_len, n_heads, head_dim
+    );
+    C10_CUDA_CHECK(cudaGetLastError());
+
+    return out;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("rotary_emb", &rotary_emb,
+        py::arg("x"),
+        py::arg("freqs_cis"),
+        "Fused rotary embedding (bf16 x, f32 freqs_cis [b,s,d/2,2], bf16 out)"
+    );
+}

csrc/tests/attn_paged_test.cu

@@ -0,0 +1,945 @@
+// Compile:
+//   nvcc -I csrc -arch=sm_89 -O3 --use_fast_math --ptxas-options=-O3 \
+//        --extra-device-vectorization -Xcompiler -fopenmp \
+//        csrc/tests/attn_paged_test.cu \
+//        -o /tmp/test_paged && /tmp/test_paged
+
+#include <cstring>
+#include <vector>
+#include "test_utils.cuh"
+#include "../kernels/attn_dispatchers.cuh"
+
+// ---- CPU reference: paged decode with variable seq_lens ----
+// Q: [B, Hq, D], K/V pool: [pool_size, Hkv, D]
+// req_to_token: [num_reqs, max_ctx_len], req_pool_indices: [B]
+// kv_indptr: [B+1].  mask: [B, max_seq_len] bool (True=keep) or NULL.
+static void cpu_paged_decode_ref(
+    const float* Q, const float* K_pool, const float* V_pool,
+    const int64_t* req_to_token, const int64_t* req_pool_indices,
+    const int* kv_indptr, const bool* mask, int mask_b_stride,
+    int B, int Hq, int Hkv, int D, int max_ctx_len,
+    float* O)
+{
+    float scale = 1.0f / sqrtf((float)D);
+    int n_rep = Hq / Hkv;
+    for (int b = 0; b < B; b++) {
+        int seq_len = kv_indptr[b + 1] - kv_indptr[b];
+        int64_t req_idx = req_pool_indices[b];
+        #pragma omp parallel for schedule(dynamic)
+        for (int h = 0; h < Hq; h++) {
+            int kv_h = h / n_rep;
+            float mv = -INFINITY, sv = 0.0f;
+            float accum[256] = {0.0f};
+            for (int kj = 0; kj < seq_len; kj++) {
+                if (mask && !mask[b * mask_b_stride + kj]) continue;
+                int64_t slot = req_to_token[req_idx * max_ctx_len + kj];
+                float dot = 0.0f;
+                for (int d = 0; d < D; d++)
+                    dot += Q[(b * Hq + h) * D + d] *
+                           K_pool[slot * Hkv * D + kv_h * D + d];
+                dot *= scale;
+                float nm = fmaxf(mv, dot);
+                float a = expf(mv - nm);
+                float be = expf(dot - nm);
+                sv = sv * a + be;
+                for (int d = 0; d < D; d++)
+                    accum[d] = accum[d] * a +
+                               V_pool[slot * Hkv * D + kv_h * D + d] * be;
+                mv = nm;
+            }
+            float inv = 1.0f / sv;
+            for (int d = 0; d < D; d++)
+                O[(b * Hq + h) * D + d] = accum[d] * inv;
+        }
+    }
+}
+
+// ---- CPU reference: paged prefill with ragged batch ----
+// Q: [total_q, Hq, D], K/V pool: [pool_size, Hkv, D]
+// req_to_token: [num_reqs, max_ctx_len], req_pool_indices: [B]
+// kv_indptr: [B+1], qo_indptr: [B+1].
+// mask: [B, max_q_len, max_seq_len] bool (True=keep, q-local + kv-local
+//   positions) or NULL.  Used only when causal==0 to apply an arbitrary
+//   attention mask on top of the (unused) causal logic.
+static void cpu_paged_prefill_ref(
+    const float* Q, const float* K_pool, const float* V_pool,
+    const int64_t* req_to_token, const int64_t* req_pool_indices,
+    const int* kv_indptr, const int* qo_indptr,
+    const bool* mask, int mask_q_stride, int mask_kv_stride,
+    int B, int Hq, int Hkv, int D, int max_ctx_len, int causal,
+    float* O)
+{
+    float scale = 1.0f / sqrtf((float)D);
+    int n_rep = Hq / Hkv;
+    for (int b = 0; b < B; b++) {
+        int seq_len = kv_indptr[b + 1] - kv_indptr[b];
+        int q_len = qo_indptr[b + 1] - qo_indptr[b];
+        int causal_off = seq_len - q_len;
+        int64_t req_idx = req_pool_indices[b];
+        #pragma omp parallel for collapse(2) schedule(dynamic)
+        for (int h = 0; h < Hq; h++) {
+            for (int qi = 0; qi < q_len; qi++) {
+                int kv_h = h / n_rep;
+                float mv = -INFINITY, sv = 0.0f;
+                float accum[256] = {0.0f};
+                int lim = causal ? min(seq_len, causal_off + qi + 1) : seq_len;
+                for (int kj = 0; kj < lim; kj++) {
+                    if (mask && !mask[b * mask_q_stride * mask_kv_stride
+                                     + qi * mask_kv_stride + kj]) continue;
+                    int64_t slot = req_to_token[req_idx * max_ctx_len + kj];
+                    float dot = 0.0f;
+                    for (int d = 0; d < D; d++)
+                        dot += Q[(qo_indptr[b] + qi) * Hq * D + h * D + d] *
+                               K_pool[slot * Hkv * D + kv_h * D + d];
+                    dot *= scale;
+                    float nm = fmaxf(mv, dot);
+                    float a = expf(mv - nm);
+                    float be = expf(dot - nm);
+                    sv = sv * a + be;
+                    for (int d = 0; d < D; d++)
+                        accum[d] = accum[d] * a +
+                                   V_pool[slot * Hkv * D + kv_h * D + d] * be;
+                    mv = nm;
+                }
+                float inv = 1.0f / sv;
+                for (int d = 0; d < D; d++)
+                    O[(qo_indptr[b] + qi) * Hq * D + h * D + d] = accum[d] * inv;
+            }
+        }
+    }
+}
+
+// ---- paged validation table (kernel vs CPU ref, abs error only) ----
+inline void print_paged_header() {
+    printf("%-58s | %11s | %6s\n",
+           "config", "max_err", "result");
+    printf("----------------------------------------------------------------"
+           "--------------------------------\n");
+}
+
+inline void print_paged_row(const char* cfg, float max_err, bool pass) {
+    printf("%-58s | %11.3e | %s\n",
+           cfg, max_err, pass ? "PASS" : "FAIL");
+}
+
+// ======================================================================
+// DECODE TEST
+// ======================================================================
+template <int HEAD_DIM>
+static int run_decode_test(int B, int Hq, int Hkv, int max_seq,
+                            int causal, int seed) {
+    // Variable seq_lens per request
+    srand(seed);
+    std::vector<int> seq_lens(B);
+    for (int b = 0; b < B; b++)
+        seq_lens[b] = 8 + rand() % (max_seq - 8);
+    int max_sl = *std::max_element(seq_lens.begin(), seq_lens.end());
+    int max_ctx = max_sl + 16;
+
+    int pool_size = B * max_ctx;
+    int num_reqs = B + 4;
+
+    char cfg[80];
+    snprintf(cfg, sizeof(cfg), "DECODE B=%d Hq=%d Hkv=%d D=%d max_sl=%d causal=%d",
+             B, Hq, Hkv, HEAD_DIM, max_sl, causal);
+
+    size_t sz_q  = (size_t)B * Hq * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)pool_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_rtt = (size_t)num_reqs * max_ctx * sizeof(int64_t);
+    size_t sz_rpi = (size_t)B * sizeof(int64_t);
+    size_t sz_kvi = (size_t)(B + 1) * sizeof(int);
+    size_t sz_op = (size_t)B * Hq * MAX_SPLITS * HEAD_DIM * sizeof(float);
+    size_t sz_ml = (size_t)B * Hq * MAX_SPLITS * 2 * sizeof(float);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t *d_rtt, *d_rpi;
+    int *d_kvi;
+    float *d_op, *d_ml;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_rtt, sz_rtt); cudaMalloc(&d_rpi, sz_rpi);
+    cudaMalloc(&d_kvi, sz_kvi);
+    cudaMalloc(&d_op, sz_op); cudaMalloc(&d_ml, sz_ml);
+
+    auto rnd = [&]() { return (rand() / (float)RAND_MAX) * 2.0f - 1.0f; };
+
+    bf16* h_q = (bf16*)malloc(sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) h_q[i] = f2bf(rnd());
+    cudaMemcpy(d_q, h_q, sz_q, cudaMemcpyHostToDevice);
+
+    bf16* h_k_pool = (bf16*)malloc(sz_kv);
+    bf16* h_v_pool = (bf16*)malloc(sz_kv);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) {
+        h_k_pool[i] = f2bf(rnd());
+        h_v_pool[i] = f2bf(rnd());
+    }
+    cudaMemcpy(d_k_pool, h_k_pool, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, h_v_pool, sz_kv, cudaMemcpyHostToDevice);
+
+    // req_to_token: assign unique slots per request (scattered, not contiguous)
+    int64_t* h_rtt = (int64_t*)malloc(sz_rtt);
+    int next_slot = 0;
+    for (int r = 0; r < num_reqs; r++)
+        for (int p = 0; p < max_ctx; p++) {
+            h_rtt[r * max_ctx + p] = next_slot % pool_size;
+            next_slot++;
+        }
+    cudaMemcpy(d_rtt, h_rtt, sz_rtt, cudaMemcpyHostToDevice);
+
+    // req_pool_indices: pick B random request rows
+    int64_t* h_rpi = (int64_t*)malloc(sz_rpi);
+    for (int b = 0; b < B; b++) h_rpi[b] = b;
+    cudaMemcpy(d_rpi, h_rpi, sz_rpi, cudaMemcpyHostToDevice);
+
+    // kv_indptr: prefix sum of seq_lens
+    int* h_kvi = (int*)malloc(sz_kvi);
+    h_kvi[0] = 0;
+    for (int b = 0; b < B; b++) h_kvi[b + 1] = h_kvi[b] + seq_lens[b];
+    cudaMemcpy(d_kvi, h_kvi, sz_kvi, cudaMemcpyHostToDevice);
+
+    // CPU reference
+    float* h_q_f = (float*)malloc(B * Hq * HEAD_DIM * sizeof(float));
+    float* h_k_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    float* h_v_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) h_q_f[i] = bf2f(h_q[i]);
+    for (int i = 0; i < pool_size * Hkv * HEAD_DIM; i++) {
+        h_k_f[i] = bf2f(h_k_pool[i]);
+        h_v_f[i] = bf2f(h_v_pool[i]);
+    }
+    float* h_o_ref = (float*)calloc(B * Hq * HEAD_DIM, sizeof(float));
+    cpu_paged_decode_ref(h_q_f, h_k_f, h_v_f, h_rtt, h_rpi, h_kvi,
+                            nullptr, 0,
+                            B, Hq, Hkv, HEAD_DIM, max_ctx, h_o_ref);
+
+    // Kernel launch
+    AttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv;
+    p.head_dim = HEAD_DIM; p.total_q = B;
+    p.q_stride_l = Hq * HEAD_DIM; p.q_stride_h = HEAD_DIM; p.q_stride_d = 1;
+    p.max_context_len = max_ctx; p.max_seq_len = max_sl;
+    p.causal_offset = causal ? 0 : -1; p.use_mask = 0;
+    p.mask = nullptr; p.mask_b_stride = 0;
+    p.mask_h_stride = 0; p.mask_q_stride = 0;
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.q = d_q; p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.req_to_token = d_rtt; p.req_pool_indices = d_rpi;
+    p.kv_indptr = d_kvi; p.qo_indptr = nullptr;
+    p.o = d_o; p.o_part = d_op; p.ml_part = d_ml;
+
+    dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(p, 0); });
+    cudaDeviceSynchronize();
+
+    bf16* h_o_bf = (bf16*)malloc(sz_q);
+    cudaMemcpy(h_o_bf, d_o, sz_q, cudaMemcpyDeviceToHost);
+    float* h_o_got = (float*)malloc(B * Hq * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) h_o_got[i] = bf2f(h_o_bf[i]);
+
+    const float atol = 0.02f, rtol = 0.02f;
+    bool pass = true;
+    float max_err = 0.0f;
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_got[i] - h_o_ref[i]);
+        if (e > max_err) max_err = e;
+        if (e > atol + rtol * fabsf(h_o_ref[i])) { pass = false; break; }
+    }
+
+    print_paged_row(cfg, max_err, pass);
+
+    free(h_q); free(h_k_pool); free(h_v_pool); free(h_rtt); free(h_rpi);
+    free(h_kvi); free(h_q_f); free(h_k_f); free(h_v_f);
+    free(h_o_ref); free(h_o_bf); free(h_o_got);
+    cudaFree(d_q); cudaFree(d_o); cudaFree(d_k_pool); cudaFree(d_v_pool);
+    cudaFree(d_rtt); cudaFree(d_rpi); cudaFree(d_kvi); cudaFree(d_op); cudaFree(d_ml);
+    return pass ? 0 : 1;
+}
+
+// ======================================================================
+// DECODE WITH MASK TEST (regression: 2D mask on mixed seq_lens)
+// ======================================================================
+template <int HEAD_DIM>
+static int run_decode_mask_test(int B, int Hq, int Hkv, int max_seq,
+                                int seed) {
+    srand(seed);
+    std::vector<int> seq_lens(B);
+    for (int b = 0; b < B; b++)
+        seq_lens[b] = 8 + rand() % (max_seq - 8);
+    int max_sl = *std::max_element(seq_lens.begin(), seq_lens.end());
+    int max_ctx = max_sl + 16;
+    int pool_size = B * max_ctx;
+    int num_reqs = B + 4;
+
+    char cfg[80];
+    snprintf(cfg, sizeof(cfg), "DECODE-MASK B=%d Hq=%d Hkv=%d D=%d max_sl=%d",
+             B, Hq, Hkv, HEAD_DIM, max_sl);
+
+    size_t sz_q  = (size_t)B * Hq * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)pool_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_rtt = (size_t)num_reqs * max_ctx * sizeof(int64_t);
+    size_t sz_rpi = (size_t)B * sizeof(int64_t);
+    size_t sz_kvi = (size_t)(B + 1) * sizeof(int);
+    size_t sz_mask = (size_t)B * max_sl * sizeof(bool);
+    size_t sz_op = (size_t)B * Hq * MAX_SPLITS * HEAD_DIM * sizeof(float);
+    size_t sz_ml = (size_t)B * Hq * MAX_SPLITS * 2 * sizeof(float);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t *d_rtt, *d_rpi;
+    int *d_kvi;
+    bool *d_mask;
+    float *d_op, *d_ml;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_rtt, sz_rtt); cudaMalloc(&d_rpi, sz_rpi);
+    cudaMalloc(&d_kvi, sz_kvi);
+    cudaMalloc(&d_mask, sz_mask);
+    cudaMalloc(&d_op, sz_op); cudaMalloc(&d_ml, sz_ml);
+
+    auto rnd = [&]() { return (rand() / (float)RAND_MAX) * 2.0f - 1.0f; };
+
+    bf16* h_q = (bf16*)malloc(sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) h_q[i] = f2bf(rnd());
+    cudaMemcpy(d_q, h_q, sz_q, cudaMemcpyHostToDevice);
+
+    bf16* h_k_pool = (bf16*)malloc(sz_kv);
+    bf16* h_v_pool = (bf16*)malloc(sz_kv);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) {
+        h_k_pool[i] = f2bf(rnd());
+        h_v_pool[i] = f2bf(rnd());
+    }
+    cudaMemcpy(d_k_pool, h_k_pool, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, h_v_pool, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_rtt = (int64_t*)malloc(sz_rtt);
+    int next_slot = 0;
+    for (int r = 0; r < num_reqs; r++)
+        for (int p = 0; p < max_ctx; p++) {
+            h_rtt[r * max_ctx + p] = next_slot % pool_size;
+            next_slot++;
+        }
+    cudaMemcpy(d_rtt, h_rtt, sz_rtt, cudaMemcpyHostToDevice);
+
+    int64_t* h_rpi = (int64_t*)malloc(sz_rpi);
+    for (int b = 0; b < B; b++) h_rpi[b] = b;
+    cudaMemcpy(d_rpi, h_rpi, sz_rpi, cudaMemcpyHostToDevice);
+
+    int* h_kvi = (int*)malloc(sz_kvi);
+    h_kvi[0] = 0;
+    for (int b = 0; b < B; b++) h_kvi[b + 1] = h_kvi[b] + seq_lens[b];
+    cudaMemcpy(d_kvi, h_kvi, sz_kvi, cudaMemcpyHostToDevice);
+
+    // Mask: keep first half of each request's kv range, drop the rest —
+    // exercises the HasMask path with per-request seq_len.
+    bool* h_mask = (bool*)malloc(sz_mask);
+    for (int b = 0; b < B; b++)
+        for (int k = 0; k < max_sl; k++)
+            h_mask[b * max_sl + k] = (k < seq_lens[b]) && (k % 2 == 0);
+    cudaMemcpy(d_mask, h_mask, sz_mask, cudaMemcpyHostToDevice);
+
+    float* h_q_f = (float*)malloc(B * Hq * HEAD_DIM * sizeof(float));
+    float* h_k_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    float* h_v_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) h_q_f[i] = bf2f(h_q[i]);
+    for (int i = 0; i < pool_size * Hkv * HEAD_DIM; i++) {
+        h_k_f[i] = bf2f(h_k_pool[i]);
+        h_v_f[i] = bf2f(h_v_pool[i]);
+    }
+    float* h_o_ref = (float*)calloc(B * Hq * HEAD_DIM, sizeof(float));
+    cpu_paged_decode_ref(h_q_f, h_k_f, h_v_f, h_rtt, h_rpi, h_kvi,
+                            h_mask, max_sl,
+                            B, Hq, Hkv, HEAD_DIM, max_ctx, h_o_ref);
+
+    AttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv;
+    p.head_dim = HEAD_DIM; p.total_q = B;
+    p.q_stride_l = Hq * HEAD_DIM; p.q_stride_h = HEAD_DIM; p.q_stride_d = 1;
+    p.max_context_len = max_ctx; p.max_seq_len = max_sl;
+    p.causal_offset = -1; p.use_mask = 1;
+    p.mask = d_mask; p.mask_b_stride = max_sl;
+    p.mask_h_stride = 0; p.mask_q_stride = 0;
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.q = d_q; p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.req_to_token = d_rtt; p.req_pool_indices = d_rpi;
+    p.kv_indptr = d_kvi; p.qo_indptr = nullptr;
+    p.o = d_o; p.o_part = d_op; p.ml_part = d_ml;
+
+    dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(p, 0); });
+    cudaDeviceSynchronize();
+
+    bf16* h_o_bf = (bf16*)malloc(sz_q);
+    cudaMemcpy(h_o_bf, d_o, sz_q, cudaMemcpyDeviceToHost);
+    float* h_o_got = (float*)malloc(B * Hq * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) h_o_got[i] = bf2f(h_o_bf[i]);
+
+    const float atol = 0.02f, rtol = 0.02f;
+    bool pass = true;
+    float max_err = 0.0f;
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_got[i] - h_o_ref[i]);
+        if (e > max_err) max_err = e;
+        if (e > atol + rtol * fabsf(h_o_ref[i])) { pass = false; break; }
+    }
+
+    print_paged_row(cfg, max_err, pass);
+
+    free(h_q); free(h_k_pool); free(h_v_pool); free(h_rtt); free(h_rpi);
+    free(h_kvi); free(h_mask); free(h_q_f); free(h_k_f); free(h_v_f);
+    free(h_o_ref); free(h_o_bf); free(h_o_got);
+    cudaFree(d_q); cudaFree(d_o); cudaFree(d_k_pool); cudaFree(d_v_pool);
+    cudaFree(d_rtt); cudaFree(d_rpi); cudaFree(d_kvi); cudaFree(d_mask);
+    cudaFree(d_op); cudaFree(d_ml);
+    return pass ? 0 : 1;
+}
+
+// ======================================================================
+// PREFILL TEST
+// ======================================================================
+template <int HEAD_DIM>
+static int run_prefill_test(int B, int Hq, int Hkv,
+                             std::vector<int>& q_lens,
+                             std::vector<int>& kv_lens,
+                             int causal, int seed) {
+    int total_q = 0;
+    int max_sl = 0;
+    for (int b = 0; b < B; b++) {
+        total_q += q_lens[b];
+        max_sl = max(max_sl, kv_lens[b]);
+    }
+    int max_ctx = max_sl + 16;
+    int pool_size = B * max_ctx;
+    int num_reqs = B + 4;
+
+    char cfg[80];
+    snprintf(cfg, sizeof(cfg), "PREFILL B=%d Hq=%d Hkv=%d D=%d max_sl=%d causal=%d",
+             B, Hq, Hkv, HEAD_DIM, max_sl, causal);
+
+    size_t sz_q  = (size_t)total_q * Hq * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)pool_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_rtt = (size_t)num_reqs * max_ctx * sizeof(int64_t);
+    size_t sz_rpi = (size_t)B * sizeof(int64_t);
+    size_t sz_kvi = (size_t)(B + 1) * sizeof(int);
+    size_t sz_qoi = (size_t)(B + 1) * sizeof(int);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t *d_rtt, *d_rpi;
+    int *d_kvi, *d_qoi;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_rtt, sz_rtt); cudaMalloc(&d_rpi, sz_rpi);
+    cudaMalloc(&d_kvi, sz_kvi); cudaMalloc(&d_qoi, sz_qoi);
+
+    srand(seed);
+    auto rnd = [&]() { return (rand() / (float)RAND_MAX) * 2.0f - 1.0f; };
+
+    bf16* h_q = (bf16*)malloc(sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) h_q[i] = f2bf(rnd());
+    cudaMemcpy(d_q, h_q, sz_q, cudaMemcpyHostToDevice);
+
+    bf16* h_k_pool = (bf16*)malloc(sz_kv);
+    bf16* h_v_pool = (bf16*)malloc(sz_kv);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) {
+        h_k_pool[i] = f2bf(rnd());
+        h_v_pool[i] = f2bf(rnd());
+    }
+    cudaMemcpy(d_k_pool, h_k_pool, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, h_v_pool, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_rtt = (int64_t*)malloc(sz_rtt);
+    int next_slot = 0;
+    for (int r = 0; r < num_reqs; r++)
+        for (int p = 0; p < max_ctx; p++) {
+            h_rtt[r * max_ctx + p] = next_slot % pool_size;
+            next_slot++;
+        }
+    cudaMemcpy(d_rtt, h_rtt, sz_rtt, cudaMemcpyHostToDevice);
+
+    int64_t* h_rpi = (int64_t*)malloc(sz_rpi);
+    for (int b = 0; b < B; b++) h_rpi[b] = b;
+    cudaMemcpy(d_rpi, h_rpi, sz_rpi, cudaMemcpyHostToDevice);
+
+    int* h_kvi = (int*)malloc(sz_kvi);
+    h_kvi[0] = 0;
+    for (int b = 0; b < B; b++) h_kvi[b + 1] = h_kvi[b] + kv_lens[b];
+    cudaMemcpy(d_kvi, h_kvi, sz_kvi, cudaMemcpyHostToDevice);
+
+    int* h_qoi = (int*)malloc(sz_qoi);
+    h_qoi[0] = 0;
+    for (int b = 0; b < B; b++) h_qoi[b + 1] = h_qoi[b] + q_lens[b];
+    cudaMemcpy(d_qoi, h_qoi, sz_qoi, cudaMemcpyHostToDevice);
+
+    // CPU reference
+    float* h_q_f = (float*)malloc(total_q * Hq * HEAD_DIM * sizeof(float));
+    float* h_k_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    float* h_v_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < total_q * Hq * HEAD_DIM; i++) h_q_f[i] = bf2f(h_q[i]);
+    for (int i = 0; i < pool_size * Hkv * HEAD_DIM; i++) {
+        h_k_f[i] = bf2f(h_k_pool[i]);
+        h_v_f[i] = bf2f(h_v_pool[i]);
+    }
+    float* h_o_ref = (float*)calloc(total_q * Hq * HEAD_DIM, sizeof(float));
+    cpu_paged_prefill_ref(h_q_f, h_k_f, h_v_f, h_rtt, h_rpi, h_kvi, h_qoi,
+                             nullptr, 0, 0,
+                             B, Hq, Hkv, HEAD_DIM, max_ctx, causal, h_o_ref);
+
+    // Kernel launch
+    AttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv;
+    p.head_dim = HEAD_DIM; p.total_q = total_q;
+    p.q_stride_l = Hq * HEAD_DIM; p.q_stride_h = HEAD_DIM; p.q_stride_d = 1;
+    p.max_context_len = max_ctx; p.max_seq_len = max_sl;
+    int max_ql = 0;
+    for (int b = 0; b < B; b++) max_ql = max(max_ql, q_lens[b]);
+    p.max_q_len = max_ql;
+    p.causal_offset = causal ? 0 : -1; p.use_mask = 0;
+    p.mask = nullptr; p.mask_b_stride = 0;
+    p.mask_h_stride = 0; p.mask_q_stride = 0;
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.q = d_q; p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.req_to_token = d_rtt; p.req_pool_indices = d_rpi;
+    p.kv_indptr = d_kvi; p.qo_indptr = d_qoi;
+    p.o = d_o; p.o_part = nullptr; p.ml_part = nullptr;
+
+    dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_prefill<H>(p, 0); });
+    cudaDeviceSynchronize();
+
+    bf16* h_o_bf = (bf16*)malloc(sz_q);
+    cudaMemcpy(h_o_bf, d_o, sz_q, cudaMemcpyDeviceToHost);
+    float* h_o_got = (float*)malloc(total_q * Hq * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < total_q * Hq * HEAD_DIM; i++) h_o_got[i] = bf2f(h_o_bf[i]);
+
+    const float atol = 0.02f, rtol = 0.02f;
+    bool pass = true;
+    float max_err = 0.0f;
+    for (int i = 0; i < total_q * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_got[i] - h_o_ref[i]);
+        if (e > max_err) max_err = e;
+        if (e > atol + rtol * fabsf(h_o_ref[i])) { pass = false; break; }
+    }
+
+    print_paged_row(cfg, max_err, pass);
+
+    free(h_q); free(h_k_pool); free(h_v_pool); free(h_rtt); free(h_rpi);
+    free(h_kvi); free(h_qoi); free(h_q_f); free(h_k_f); free(h_v_f);
+    free(h_o_ref); free(h_o_bf); free(h_o_got);
+    cudaFree(d_q); cudaFree(d_o); cudaFree(d_k_pool); cudaFree(d_v_pool);
+    cudaFree(d_rtt); cudaFree(d_rpi); cudaFree(d_kvi); cudaFree(d_qoi);
+    return pass ? 0 : 1;
+}
+
+// ======================================================================
+// PREFILL WITH MASK TEST (regression: 4D causal mask on single request)
+// ======================================================================
+template <int HEAD_DIM>
+static int run_prefill_mask_test(int Hq, int Hkv, int q_len, int seed) {
+    srand(seed);
+    int B = 1;
+    int total_q = q_len;
+    int seq_len = q_len;  // pure prefill: kv_len == q_len
+    int max_ctx = seq_len + 16;
+    int pool_size = B * max_ctx;
+    int num_reqs = B + 4;
+
+    char cfg[80];
+    snprintf(cfg, sizeof(cfg), "PREFILL-MASK Hq=%d Hkv=%d D=%d q_len=%d",
+             Hq, Hkv, HEAD_DIM, q_len);
+    fflush(stdout);
+
+    size_t sz_q  = (size_t)total_q * Hq * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)pool_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_rtt = (size_t)num_reqs * max_ctx * sizeof(int64_t);
+    size_t sz_rpi = (size_t)B * sizeof(int64_t);
+    size_t sz_kvi = (size_t)(B + 1) * sizeof(int);
+    size_t sz_qoi = (size_t)(B + 1) * sizeof(int);
+    size_t sz_mask = (size_t)B * q_len * q_len * sizeof(bool);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t *d_rtt, *d_rpi;
+    int *d_kvi, *d_qoi;
+    bool *d_mask;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_rtt, sz_rtt); cudaMalloc(&d_rpi, sz_rpi);
+    cudaMalloc(&d_kvi, sz_kvi); cudaMalloc(&d_qoi, sz_qoi);
+    cudaMalloc(&d_mask, sz_mask);
+
+    auto rnd = [&]() { return (rand() / (float)RAND_MAX) * 2.0f - 1.0f; };
+
+    bf16* h_q = (bf16*)malloc(sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) h_q[i] = f2bf(rnd());
+    cudaMemcpy(d_q, h_q, sz_q, cudaMemcpyHostToDevice);
+
+    bf16* h_k_pool = (bf16*)malloc(sz_kv);
+    bf16* h_v_pool = (bf16*)malloc(sz_kv);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) {
+        h_k_pool[i] = f2bf(rnd());
+        h_v_pool[i] = f2bf(rnd());
+    }
+    cudaMemcpy(d_k_pool, h_k_pool, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, h_v_pool, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_rtt = (int64_t*)malloc(sz_rtt);
+    int next_slot = 0;
+    for (int r = 0; r < num_reqs; r++)
+        for (int p = 0; p < max_ctx; p++) {
+            h_rtt[r * max_ctx + p] = next_slot % pool_size;
+            next_slot++;
+        }
+    cudaMemcpy(d_rtt, h_rtt, sz_rtt, cudaMemcpyHostToDevice);
+
+    int64_t* h_rpi = (int64_t*)malloc(sz_rpi);
+    h_rpi[0] = 0;
+    cudaMemcpy(d_rpi, h_rpi, sz_rpi, cudaMemcpyHostToDevice);
+
+    int* h_kvi = (int*)malloc(sz_kvi);
+    h_kvi[0] = 0; h_kvi[1] = seq_len;
+    cudaMemcpy(d_kvi, h_kvi, sz_kvi, cudaMemcpyHostToDevice);
+
+    int* h_qoi = (int*)malloc(sz_qoi);
+    h_qoi[0] = 0; h_qoi[1] = q_len;
+    cudaMemcpy(d_qoi, h_qoi, sz_qoi, cudaMemcpyHostToDevice);
+
+    // 4D causal mask [B, 1, q_len, q_len], True=keep.
+    bool* h_mask = (bool*)malloc(sz_mask);
+    for (int qi = 0; qi < q_len; qi++)
+        for (int kj = 0; kj < q_len; kj++)
+            h_mask[qi * q_len + kj] = (kj <= qi);
+    cudaMemcpy(d_mask, h_mask, sz_mask, cudaMemcpyHostToDevice);
+
+    float* h_q_f = (float*)malloc(total_q * Hq * HEAD_DIM * sizeof(float));
+    float* h_k_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    float* h_v_f = (float*)malloc(pool_size * Hkv * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < total_q * Hq * HEAD_DIM; i++) h_q_f[i] = bf2f(h_q[i]);
+    for (int i = 0; i < pool_size * Hkv * HEAD_DIM; i++) {
+        h_k_f[i] = bf2f(h_k_pool[i]);
+        h_v_f[i] = bf2f(h_v_pool[i]);
+    }
+    float* h_o_ref = (float*)calloc(total_q * Hq * HEAD_DIM, sizeof(float));
+    // CPU ref with causal=0 so it consults the mask (not the causal flag).
+    cpu_paged_prefill_ref(h_q_f, h_k_f, h_v_f, h_rtt, h_rpi, h_kvi, h_qoi,
+                             h_mask, q_len, q_len,
+                             B, Hq, Hkv, HEAD_DIM, max_ctx, 0, h_o_ref);
+
+    AttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv;
+    p.head_dim = HEAD_DIM; p.total_q = total_q;
+    p.q_stride_l = Hq * HEAD_DIM; p.q_stride_h = HEAD_DIM; p.q_stride_d = 1;
+    p.max_context_len = max_ctx; p.max_seq_len = q_len;
+    p.max_q_len = q_len;
+    p.causal_offset = -1; p.use_mask = 1;
+    p.mask = d_mask; p.mask_b_stride = q_len * q_len;
+    p.mask_h_stride = 0; p.mask_q_stride = q_len;
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.q = d_q; p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.req_to_token = d_rtt; p.req_pool_indices = d_rpi;
+    p.kv_indptr = d_kvi; p.qo_indptr = d_qoi;
+    p.o = d_o; p.o_part = nullptr; p.ml_part = nullptr;
+
+    dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_prefill<H>(p, 0); });
+    cudaDeviceSynchronize();
+
+    bf16* h_o_bf = (bf16*)malloc(sz_q);
+    cudaMemcpy(h_o_bf, d_o, sz_q, cudaMemcpyDeviceToHost);
+    float* h_o_got = (float*)malloc(total_q * Hq * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < total_q * Hq * HEAD_DIM; i++) h_o_got[i] = bf2f(h_o_bf[i]);
+
+    const float atol = 0.02f, rtol = 0.02f;
+    bool pass = true;
+    float max_err = 0.0f;
+    for (int i = 0; i < total_q * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_got[i] - h_o_ref[i]);
+        if (e > max_err) max_err = e;
+        if (e > atol + rtol * fabsf(h_o_ref[i])) { pass = false; break; }
+    }
+
+    print_paged_row(cfg, max_err, pass);
+
+    free(h_q); free(h_k_pool); free(h_v_pool); free(h_rtt); free(h_rpi);
+    free(h_kvi); free(h_qoi); free(h_mask); free(h_q_f); free(h_k_f); free(h_v_f);
+    free(h_o_ref); free(h_o_bf); free(h_o_got);
+    cudaFree(d_q); cudaFree(d_o); cudaFree(d_k_pool); cudaFree(d_v_pool);
+    cudaFree(d_rtt); cudaFree(d_rpi); cudaFree(d_kvi); cudaFree(d_qoi);
+    cudaFree(d_mask);
+    return pass ? 0 : 1;
+}
+
+// ======================================================================
+// BENCH
+// ======================================================================
+template <int HEAD_DIM>
+static void bench_decode(int B, int Hq, int Hkv, int seq_len) {
+    int max_ctx = seq_len + 16;
+    int pool_size = B * max_ctx;
+    int num_reqs = B;
+
+    size_t sz_q  = (size_t)B * Hq * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)pool_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_rtt = (size_t)num_reqs * max_ctx * sizeof(int64_t);
+    size_t sz_rpi = (size_t)B * sizeof(int64_t);
+    size_t sz_kvi = (size_t)(B + 1) * sizeof(int);
+    size_t sz_op = (size_t)B * Hq * MAX_SPLITS * HEAD_DIM * sizeof(float);
+    size_t sz_ml = (size_t)B * Hq * MAX_SPLITS * 2 * sizeof(float);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t *d_rtt, *d_rpi;
+    int *d_kvi;
+    float *d_op, *d_ml;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_rtt, sz_rtt); cudaMalloc(&d_rpi, sz_rpi);
+    cudaMalloc(&d_kvi, sz_kvi);
+    cudaMalloc(&d_op, sz_op); cudaMalloc(&d_ml, sz_ml);
+
+    bf16* tmp = (bf16*)malloc(sz_kv > sz_q ? sz_kv : sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_q, tmp, sz_q, cudaMemcpyHostToDevice);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_k_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_rtt = (int64_t*)malloc(sz_rtt);
+    for (int r = 0; r < num_reqs; r++)
+        for (int p = 0; p < max_ctx; p++)
+            h_rtt[r * max_ctx + p] = (r * max_ctx + p) % pool_size;
+    cudaMemcpy(d_rtt, h_rtt, sz_rtt, cudaMemcpyHostToDevice);
+    int64_t* h_rpi = (int64_t*)malloc(sz_rpi);
+    for (int b = 0; b < B; b++) h_rpi[b] = b;
+    cudaMemcpy(d_rpi, h_rpi, sz_rpi, cudaMemcpyHostToDevice);
+    int* h_kvi = (int*)malloc(sz_kvi);
+    h_kvi[0] = 0;
+    for (int b = 0; b < B; b++) h_kvi[b + 1] = h_kvi[b] + seq_len;
+    cudaMemcpy(d_kvi, h_kvi, sz_kvi, cudaMemcpyHostToDevice);
+
+    AttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv;
+    p.head_dim = HEAD_DIM; p.total_q = B;
+    p.q_stride_l = Hq * HEAD_DIM; p.q_stride_h = HEAD_DIM; p.q_stride_d = 1;
+    p.max_context_len = max_ctx; p.max_seq_len = seq_len;
+    p.causal_offset = 0; p.use_mask = 0;
+    p.mask = nullptr; p.mask_b_stride = 0;
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.q = d_q; p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.req_to_token = d_rtt; p.req_pool_indices = d_rpi;
+    p.kv_indptr = d_kvi; p.qo_indptr = nullptr;
+    p.o = d_o; p.o_part = d_op; p.ml_part = d_ml;
+
+    auto launch = [&]() {
+        dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(p, 0); });
+    };
+    // Decode: q_len=1, query is the last token → attends to all [0, seq_len).
+    // FLOPs = 2 * (QK^T + PV) = 4 * B * Hq * seq_len * D.
+    double flops = 4.0 * B * Hq * (double)seq_len * HEAD_DIM;
+    BenchResult r = bench_kernel(launch, 3, 10, flops);
+
+    char cfg[64];
+    snprintf(cfg, sizeof(cfg), "DEC B=%2d Hq=%2d Hk=%d kv=%4d D=%3d",
+             B, Hq, Hkv, seq_len, HEAD_DIM);
+    print_bench_row(cfg, r);
+
+    free(tmp); free(h_rtt); free(h_rpi); free(h_kvi);
+    cudaFree(d_q); cudaFree(d_o); cudaFree(d_k_pool); cudaFree(d_v_pool);
+    cudaFree(d_rtt); cudaFree(d_rpi); cudaFree(d_kvi); cudaFree(d_op); cudaFree(d_ml);
+}
+
+template <int HEAD_DIM>
+static void bench_prefill(int B, int Hq, int Hkv, int q_len, int kv_len, int causal) {
+    int total_q = B * q_len;
+    int max_ctx = kv_len + 16;
+    int pool_size = B * max_ctx;
+    int num_reqs = B;
+
+    size_t sz_q  = (size_t)total_q * Hq * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)pool_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_rtt = (size_t)num_reqs * max_ctx * sizeof(int64_t);
+    size_t sz_rpi = (size_t)B * sizeof(int64_t);
+    size_t sz_kvi = (size_t)(B + 1) * sizeof(int);
+    size_t sz_qoi = (size_t)(B + 1) * sizeof(int);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t *d_rtt, *d_rpi;
+    int *d_kvi, *d_qoi;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_rtt, sz_rtt); cudaMalloc(&d_rpi, sz_rpi);
+    cudaMalloc(&d_kvi, sz_kvi); cudaMalloc(&d_qoi, sz_qoi);
+
+    bf16* tmp = (bf16*)malloc(sz_kv > sz_q ? sz_kv : sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_q, tmp, sz_q, cudaMemcpyHostToDevice);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_k_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_rtt = (int64_t*)malloc(sz_rtt);
+    for (int r = 0; r < num_reqs; r++)
+        for (int p = 0; p < max_ctx; p++)
+            h_rtt[r * max_ctx + p] = (r * max_ctx + p) % pool_size;
+    cudaMemcpy(d_rtt, h_rtt, sz_rtt, cudaMemcpyHostToDevice);
+    int64_t* h_rpi = (int64_t*)malloc(sz_rpi);
+    for (int b = 0; b < B; b++) h_rpi[b] = b;
+    cudaMemcpy(d_rpi, h_rpi, sz_rpi, cudaMemcpyHostToDevice);
+    int* h_kvi = (int*)malloc(sz_kvi);
+    h_kvi[0] = 0;
+    for (int b = 0; b < B; b++) h_kvi[b + 1] = h_kvi[b] + kv_len;
+    cudaMemcpy(d_kvi, h_kvi, sz_kvi, cudaMemcpyHostToDevice);
+    int* h_qoi = (int*)malloc(sz_qoi);
+    h_qoi[0] = 0;
+    for (int b = 0; b < B; b++) h_qoi[b + 1] = h_qoi[b] + q_len;
+    cudaMemcpy(d_qoi, h_qoi, sz_qoi, cudaMemcpyHostToDevice);
+
+    AttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv;
+    p.head_dim = HEAD_DIM; p.total_q = total_q;
+    p.q_stride_l = Hq * HEAD_DIM; p.q_stride_h = HEAD_DIM; p.q_stride_d = 1;
+    p.max_context_len = max_ctx; p.max_seq_len = kv_len;
+    p.total_q = total_q; p.max_q_len = q_len;
+    p.causal_offset = causal ? 0 : -1; p.use_mask = 0;
+    p.mask = nullptr; p.mask_b_stride = 0;
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.q = d_q; p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.req_to_token = d_rtt; p.req_pool_indices = d_rpi;
+    p.kv_indptr = d_kvi; p.qo_indptr = d_qoi;
+    p.o = d_o; p.o_part = nullptr; p.ml_part = nullptr;
+
+    auto launch = [&]() {
+        dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_prefill<H>(p, 0); });
+    };
+    // FLOPs = 2 * (QK^T + PV) = 4 * effective_qk_pairs * Hq * D.
+    // Non-causal: effective = q_len * kv_len.
+    // Causal: Q row qi attends to [0, causal_off + qi + 1) where
+    //   causal_off = kv_len - q_len.  Total KV accesses per request:
+    //   sum_{qi=0}^{q_len-1} (kv_len - q_len + qi + 1)
+    //   = q_len * (kv_len - q_len) + q_len * (q_len + 1) / 2.
+    double eff_kv;
+    if (causal) {
+        eff_kv = (double)q_len * (kv_len - q_len)
+               + (double)q_len * (q_len + 1) / 2.0;
+    } else {
+        eff_kv = (double)q_len * kv_len;
+    }
+    double flops = 4.0 * B * Hq * eff_kv * HEAD_DIM;
+    BenchResult r = bench_kernel(launch, 3, 10, flops);
+
+    char cfg[80];
+    snprintf(cfg, sizeof(cfg), "PRE B=%d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d c=%d",
+             B, Hq, Hkv, q_len, kv_len, HEAD_DIM, causal);
+    print_bench_row(cfg, r);
+
+    free(tmp); free(h_rtt); free(h_rpi); free(h_kvi); free(h_qoi);
+    cudaFree(d_q); cudaFree(d_o); cudaFree(d_k_pool); cudaFree(d_v_pool);
+    cudaFree(d_rtt); cudaFree(d_rpi); cudaFree(d_kvi); cudaFree(d_qoi);
+}
+
+int main() {
+    int fail = 0;
+
+    // ===== DECODE TESTS =====
+    printf("=== Paged Decode Tests ===\n");
+    print_paged_header();
+    fail += run_decode_test<128>(1, 32, 4, 512, 0, 1);
+    fail += run_decode_test<128>(1, 32, 4, 1024, 0, 2);
+    fail += run_decode_test<128>(4, 32, 4, 512, 0, 3);
+    fail += run_decode_test<128>(8, 32, 4, 1024, 0, 4);
+    fail += run_decode_test<128>(4, 32, 8, 2048, 0, 5);
+    fail += run_decode_test<128>(1, 16, 1, 256, 0, 6);
+    fail += run_decode_test<128>(2, 8, 2, 512, 1, 7);
+    fail += run_decode_test<64>(1, 4, 2, 256, 0, 8);
+    fail += run_decode_test<256>(1, 2, 1, 256, 0, 9);
+    fail += run_decode_test<128>(16, 32, 4, 2048, 0, 10);
+    fail += run_decode_test<128>(32, 32, 4, 1024, 0, 11);
+
+    // Decode with 2D mask (regression: mixed seq_lens + HasMask)
+    fail += run_decode_mask_test<128>(2, 8, 2, 256, 30);
+    fail += run_decode_mask_test<128>(4, 32, 4, 512, 31);
+    fail += run_decode_mask_test<64>(2, 4, 2, 128, 32);
+
+    if (fail) { printf("\nFAILED decode tests\n"); return fail; }
+
+    // ===== PREFILL TESTS =====
+    printf("\n=== Paged Prefill Tests ===\n");
+    print_paged_header();
+    // Single request, pure prefill (q_len == kv_len)
+    {
+        std::vector<int> ql = {512};
+        std::vector<int> kl = {512};
+        fail += run_prefill_test<128>(1, 32, 4, ql, kl, 1, 20);
+    }
+    {
+        std::vector<int> ql = {1024};
+        std::vector<int> kl = {1024};
+        fail += run_prefill_test<128>(1, 32, 4, ql, kl, 1, 21);
+    }
+    {
+        std::vector<int> ql = {2048};
+        std::vector<int> kl = {2048};
+        fail += run_prefill_test<128>(1, 32, 4, ql, kl, 1, 22);
+    }
+    // Ragged batch: different q_lens and kv_lens
+    {
+        std::vector<int> ql = {128, 256, 64};
+        std::vector<int> kl = {128, 256, 64};
+        fail += run_prefill_test<128>(3, 32, 4, ql, kl, 1, 23);
+    }
+    {
+        std::vector<int> ql = {64, 128, 256, 32};
+        std::vector<int> kl = {64, 128, 256, 32};
+        fail += run_prefill_test<128>(4, 32, 4, ql, kl, 1, 24);
+    }
+    // Extend: kv_len > q_len (append to existing cache)
+    {
+        std::vector<int> ql = {64, 128};
+        std::vector<int> kl = {256, 512};
+        fail += run_prefill_test<128>(2, 32, 4, ql, kl, 1, 25);
+    }
+    // Non-causal
+    {
+        std::vector<int> ql = {256, 128};
+        std::vector<int> kl = {256, 128};
+        fail += run_prefill_test<128>(2, 32, 4, ql, kl, 0, 26);
+    }
+    // Single token (q_len=1 per request, like decode but via prefill path)
+    {
+        std::vector<int> ql = {1, 1, 1, 1};
+        std::vector<int> kl = {128, 256, 64, 512};
+        fail += run_prefill_test<128>(4, 32, 4, ql, kl, 1, 27);
+    }
+    // D=64
+    {
+        std::vector<int> ql = {128, 64};
+        std::vector<int> kl = {128, 64};
+        fail += run_prefill_test<64>(2, 4, 2, ql, kl, 1, 28);
+    }
+    // D=256
+    {
+        std::vector<int> ql = {128, 64};
+        std::vector<int> kl = {128, 64};
+        fail += run_prefill_test<256>(2, 2, 1, ql, kl, 1, 29);
+    }
+
+    // Prefill with 4D causal mask (regression: single-request mask path)
+    fail += run_prefill_mask_test<128>(32, 4, 512, 40);
+    fail += run_prefill_mask_test<128>(32, 4, 1024, 41);
+    fail += run_prefill_mask_test<64>(4, 2, 256, 42);
+
+    if (fail) { printf("\nFAILED prefill tests\n"); return fail; }
+    printf("\nAll tests passed!\n");
+
+    // ===== BENCH =====
+    printf("\n===== PAGED DECODE BENCH =====\n");
+    print_bench_header();
+    bench_decode<128>(1, 32, 4, 512);
+    bench_decode<128>(1, 32, 4, 1024);
+    bench_decode<128>(1, 32, 4, 2048);
+    bench_decode<128>(1, 32, 4, 4096);
+    bench_decode<128>(4, 32, 4, 2048);
+    bench_decode<128>(16, 32, 4, 2048);
+    bench_decode<128>(32, 32, 4, 1024);
+
+    printf("\n===== PAGED PREFILL BENCH =====\n");
+    print_bench_header();
+    bench_prefill<128>(1, 32, 4, 512, 512, 0);
+    bench_prefill<128>(1, 32, 4, 1024, 1024, 0);
+    bench_prefill<128>(1, 32, 4, 2048, 2048, 0);
+    bench_prefill<128>(1, 32, 4, 2048, 2048, 1);
+    bench_prefill<128>(4, 32, 4, 2048, 2048, 1);
+    bench_prefill<128>(1, 32, 4, 4096, 4096, 1);
+
+    return 0;
+}

csrc/tests/attn_test.cu

@@ -0,0 +1,346 @@
+/*
+Pure-C test — uses shared dispatcher.  Combines the decode (split-KV) and
+prefill (split-Q) correctness checks + benchmarks into one binary.
+nvcc -I csrc -arch=sm_89 -O3 \
+    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
+    -Xcompiler -fopenmp csrc/tests/attn_test.cu -o test && ./test
+*/
+
+#include "test_utils.cuh"
+#include "../kernels/attn_dispatchers.cuh"
+
+// Split-K scratch (torch-free)
+struct DecodeScratch {
+    float* o_part = nullptr;
+    float* ml_part = nullptr;
+};
+
+static void setup_scratch(AttentionParams<bf16>& p, DecodeScratch& sc) {
+    int max_splits = 32;
+    cudaMalloc(&sc.o_part, (size_t)p.batch * p.q_head * max_splits * p.head_dim * sizeof(float));
+    cudaMalloc(&sc.ml_part, (size_t)p.batch * p.q_head * max_splits * 2 * sizeof(float));
+}
+
+static void free_scratch(DecodeScratch& sc) {
+    cudaFree(sc.o_part); cudaFree(sc.ml_part);
+}
+
+// ======================================================================
+// DECODE
+// ======================================================================
+
+static int run_decode_test(int B, int Hq, int Hk, int sl, int D, int causal) {
+    int gs = Hq / Hk;
+
+    size_t nQ = B*Hq*1*D, nKV = B*Hk*sl*D;
+    float *hQ=new float[nQ], *hK=new float[nKV], *hV=new float[nKV];
+    for (size_t i=0;i<nQ;i++) hQ[i]=randf();
+    for (size_t i=0;i<nKV;i++){hK[i]=randf();hV[i]=randf();}
+
+    bool* hMask=new bool[B*sl];
+    for (int i=0;i<B*sl;i++) hMask[i]=true;
+
+    bf16 *dQ,*dK,*dV,*dO,*tmp;
+    bool* dMask;
+    cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+    cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+    cudaMalloc(&dMask,B*sl);
+
+    tmp=new bf16[max(nQ,nKV)];
+    for (size_t i=0;i<nQ;i++) tmp[i]=f2bf(hQ[i]);
+    cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hK[i]);
+    cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hV[i]);
+    cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+    cudaMemcpy(dMask,hMask,B*sl,cudaMemcpyHostToDevice);
+
+    AttentionParams<bf16> p;
+    p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=1; p.kv_len=sl; p.head_dim=D;
+    p.use_mask=0; p.causal_offset=causal?0:-1;
+    p.scale=1.0f/sqrtf((float)D);
+    set_default_strides(p);
+    p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
+
+    DecodeScratch sc;
+    setup_scratch(p, sc);
+    p.o_part = sc.o_part; p.ml_part = sc.ml_part;
+
+    double t0=now_ms();
+    dispatch_by_head_dim(D, [&]<int H>() { dispatch_decode<H>(p, 0); });
+    cudaDeviceSynchronize();
+    (void)t0;
+    cudaError_t err=cudaGetLastError();
+    if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return 1;}
+
+    bf16* hOut=new bf16[nQ];
+    cudaMemcpy(hOut,dO,nQ*2,cudaMemcpyDeviceToHost);
+
+    float* ref=new float[nQ];
+    cpu_attention_ref(hQ, hK, hV, hMask, ref, B, Hq, Hk, 1, sl, D, causal ? 0 : -1);
+
+    float max_abs_err=0, max_rel_err=0;
+    for (size_t i=0;i<nQ;i++){
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if(err>max_abs_err) max_abs_err=err;
+        float rel=err/fmaxf(fabsf(ref[i]), 1e-4f);
+        if(rel>max_rel_err) max_rel_err=rel;
+    }
+    const float atol=0.01f, rtol=0.01f;
+    bool pass=true;
+    for (size_t i=0;i<nQ;i++){
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if (err > atol + rtol * fabsf(ref[i])) { pass=false; break; }
+    }
+    char cfg[64];
+    snprintf(cfg, sizeof(cfg), "B=%2d Hq=%2d Hk=%d seq=%4d D=%3d causal=%d",
+             B, Hq, Hk, sl, D, causal);
+    print_test_row(cfg, max_abs_err, max_rel_err, pass);
+
+    cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);cudaFree(dMask);
+    free_scratch(sc);
+    delete[]hQ;delete[]hK;delete[]hV;delete[]hMask;delete[]hOut;delete[]ref;delete[]tmp;
+
+    return pass ? 0 : 1;
+}
+
+static void bench_decode() {
+    const int cfgs[][5] = {
+        {1, 32, 4, 512, 128},
+        {1, 32, 4, 1024, 128},
+        {1, 32, 4, 2048, 128},
+        {1, 32, 4, 4096, 128},
+        {16, 32, 4, 2048, 128},
+        {32, 32, 4, 1024, 128},
+    };
+    const int WARMUP = 3, ITERS = 10;
+    printf("\n===== DECODE BENCH (warmup=%d iters=%d) =====\n", WARMUP, ITERS);
+    print_bench_header();
+
+    for (int ci = 0; ci < 6; ci++) {
+        int B = cfgs[ci][0], Hq = cfgs[ci][1], Hk = cfgs[ci][2];
+        int sl = cfgs[ci][3], D = cfgs[ci][4];
+        size_t nQ = (size_t)B * Hq * D;
+        size_t nKV = (size_t)B * Hk * sl * D;
+
+        bf16 *dQ, *dK, *dV, *dO;
+        cudaMalloc(&dQ, nQ*2); cudaMalloc(&dK, nKV*2);
+        cudaMalloc(&dV, nKV*2); cudaMalloc(&dO, nQ*2);
+        size_t big = nQ > nKV ? nQ : nKV; bf16* tmp = new bf16[big];
+        for (size_t i = 0; i < nQ; i++)  tmp[i] = f2bf(randf());
+        cudaMemcpy(dQ, tmp, nQ*2, cudaMemcpyHostToDevice);
+        for (size_t i = 0; i < nKV; i++) tmp[i] = f2bf(randf());
+        cudaMemcpy(dK, tmp, nKV*2, cudaMemcpyHostToDevice);
+        for (size_t i = 0; i < nKV; i++) tmp[i] = f2bf(randf());
+        cudaMemcpy(dV, tmp, nKV*2, cudaMemcpyHostToDevice);
+        delete[] tmp;
+
+        AttentionParams<bf16> p;
+        p.batch = B; p.q_head = Hq; p.kv_head = Hk; p.q_len = 1; p.kv_len = sl;
+        p.head_dim = D; p.use_mask = 0; p.causal_offset = -1;
+        p.scale = 1.0f / sqrtf((float)D);
+        set_default_strides(p);
+        p.q = dQ; p.k = dK; p.v = dV; p.mask = nullptr; p.o = dO;
+
+        DecodeScratch sc;
+        setup_scratch(p, sc);
+        p.o_part = sc.o_part; p.ml_part = sc.ml_part;
+
+        auto launch = [&]() { dispatch_by_head_dim(D, [&]<int H>() { dispatch_decode<H>(p, 0); }); };
+        double flops = 4.0 * B * Hq * (double)sl * D;
+        BenchResult r = bench_kernel(launch, WARMUP, ITERS, flops);
+
+        char cfg[64];
+        snprintf(cfg, sizeof(cfg),
+                 "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d causal=%d",
+                 B, Hq, Hk, 1, sl, D, 0);
+        print_bench_row(cfg, r);
+
+        cudaFree(dQ); cudaFree(dK); cudaFree(dV); cudaFree(dO);
+        free_scratch(sc);
+    }
+}
+
+// ======================================================================
+// PREFILL
+// ======================================================================
+
+static int run_prefill_test(int B, int Hq, int Hk, int ql, int kl, int D, int causal) {
+    size_t nQ = B*Hq*ql*D, nKV = B*Hk*kl*D;
+    float *hQ=new float[nQ], *hK=new float[nKV], *hV=new float[nKV];
+    for (size_t i=0;i<nQ;i++) hQ[i]=randf();
+    for (size_t i=0;i<nKV;i++){hK[i]=randf();hV[i]=randf();}
+
+    bf16 *dQ,*dK,*dV,*dO,*tmp;
+    cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+    cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+    tmp=new bf16[max(nQ,nKV)];
+    for (size_t i=0;i<nQ;i++) tmp[i]=f2bf(hQ[i]);
+    cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hK[i]);
+    cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hV[i]);
+    cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+
+    AttentionParams<bf16> p;
+    p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=ql; p.kv_len=kl; p.head_dim=D;
+    p.use_mask=0; p.causal_offset=causal?0:-1;
+    set_default_strides(p);
+    p.scale=1.0f/sqrtf((float)D);
+    p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
+
+    double t0=now_ms();
+    dispatch_by_head_dim(D, [&]<int H>() { dispatch_prefill<H>(p, 0); });
+    cudaDeviceSynchronize();
+    (void)t0;
+    cudaError_t err=cudaGetLastError();
+    if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return 1;}
+
+    bf16* hOut=new bf16[nQ];
+    cudaMemcpy(hOut,dO,nQ*2,cudaMemcpyDeviceToHost);
+
+    float* ref=new float[nQ];
+    cpu_attention_ref(hQ, hK, hV, nullptr, ref, B, Hq, Hk, ql, kl, D, causal ? 0 : -1);
+
+    float max_abs_err=0, max_rel_err=0;
+    for (size_t i=0;i<nQ;i++) {
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if(err>max_abs_err) max_abs_err=err;
+        float rel=err/fmaxf(fabsf(ref[i]), 1e-4f);
+        if(rel>max_rel_err) max_rel_err=rel;
+    }
+    const float atol=0.01f, rtol=0.01f;
+    bool pass=true;
+    for (size_t i=0;i<nQ;i++) {
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if (err > atol + rtol * fabsf(ref[i])) { pass=false; break; }
+    }
+    char cfg[64];
+    snprintf(cfg, sizeof(cfg), "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d causal=%d",
+             B, Hq, Hk, ql, kl, D, causal);
+    print_test_row(cfg, max_abs_err, max_rel_err, pass);
+
+    cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);
+    delete[]hQ;delete[]hK;delete[]hV;delete[]hOut;delete[]ref;delete[]tmp;
+
+    return pass ? 0 : 1;
+}
+
+static void bench_prefill() {
+    const int cfgs[][7] = {
+        {1,32,4,512,512,128,0},
+        {1,32,4,1024,1024,128,0},
+        {1,32,4,2048,2048,128,0},
+        {1,32,4,2048,2048,128,1},
+        {4,32,4,2048,2048,128,1},
+        {1,32,4,4096,4096,128,1},
+    };
+    int n = sizeof(cfgs)/sizeof(cfgs[0]);
+    const int WARMUP = 3, ITERS = 10;
+    printf("\n===== PREFILL BENCH (warmup=%d iters=%d) =====\n", WARMUP, ITERS);
+    print_bench_header();
+
+    for (int ci = 0; ci < n; ci++) {
+        int B=cfgs[ci][0], Hq=cfgs[ci][1], Hk=cfgs[ci][2];
+        int ql=cfgs[ci][3], kl=cfgs[ci][4], D=cfgs[ci][5], causal=cfgs[ci][6];
+        size_t nQ=(size_t)B*Hq*ql*D, nKV=(size_t)B*Hk*kl*D;
+
+        bf16 *dQ,*dK,*dV,*dO,*tmp;
+        cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+        cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+        size_t big = nQ>nKV?nQ:nKV; tmp=new bf16[big];
+        for (size_t i=0;i<nQ;i++)  tmp[i]=f2bf(randf());
+        cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(randf());
+        cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(randf());
+        cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+
+        AttentionParams<bf16> p;
+        p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=ql; p.kv_len=kl; p.head_dim=D;
+        p.use_mask=0; p.causal_offset=causal?0:-1;
+        set_default_strides(p);
+        p.scale=1.0f/sqrtf((float)D);
+        p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
+
+        auto launch = [&]() { dispatch_by_head_dim(D, [&]<int H>() { dispatch_prefill<H>(p, 0); }); };
+        for (int i=0;i<WARMUP;i++) launch();
+        cudaDeviceSynchronize();
+        cudaError_t err=cudaGetLastError();
+        if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return;}
+
+        cudaEvent_t s,e; cudaEventCreate(&s); cudaEventCreate(&e);
+        cudaEventRecord(s);
+        for (int i=0;i<ITERS;i++) launch();
+        cudaEventRecord(e); cudaEventSynchronize(e);
+        float ms=0; cudaEventElapsedTime(&ms,s,e); ms/=ITERS;
+
+        double flops = 4.0*B*Hq*(double)ql*kl*D;
+        if (causal) flops *= 0.5;
+        double tflops = flops/(ms*1e-3)/1e12;
+        BenchResult r{ms, tflops};
+
+        char cfg[64];
+        snprintf(cfg, sizeof(cfg),
+                 "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d causal=%d",
+                 B,Hq,Hk,ql,kl,D,causal);
+        print_bench_row(cfg, r);
+
+        cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);
+        delete[]tmp; cudaEventDestroy(s); cudaEventDestroy(e);
+    }
+}
+
+// ======================================================================
+// MAIN
+// ======================================================================
+
+int main() {
+    int fail = 0;
+
+    // ---- DECODE ----
+    {
+        const int configs[][6] = {
+            {1, 2, 1, 64, 32, 0},
+            {1, 32, 4, 512, 128, 0},
+            {1, 32, 4, 1024, 128, 0},
+            {1, 32, 4, 512, 128, 1},
+        };
+        int n_cfgs = sizeof(configs) / sizeof(configs[0]);
+        printf("=== DECODE TESTS ===\n");
+        print_test_header();
+        for (int ci = 0; ci < n_cfgs; ci++) {
+            int B = configs[ci][0], Hq = configs[ci][1], Hk = configs[ci][2];
+            int sl = configs[ci][3], D = configs[ci][4], causal = configs[ci][5];
+            fail += run_decode_test(B, Hq, Hk, sl, D, causal);
+            if (fail) break;
+        }
+        if (fail) { printf("FAILED decode tests\n"); return fail; }
+        bench_decode();
+    }
+
+    // ---- PREFILL ----
+    {
+        const int configs[][7] = {
+            {1,2,1,64,128,64,0},     // tiny: B,Hq,Hk,q,kv,D,causal
+            {1,32,4,512,512,128,0},  // standard
+            {1,32,4,128,256,128,0},  // medium
+            {1,4,2,256,256,128,1},   // causal
+        };
+        int n_configs = sizeof(configs) / sizeof(configs[0]);
+        printf("\n=== PREFILL TESTS ===\n");
+        print_test_header();
+        for (int ci = 0; ci < n_configs; ci++) {
+            int B=configs[ci][0], Hq=configs[ci][1], Hk=configs[ci][2];
+            int ql=configs[ci][3], kl=configs[ci][4], D=configs[ci][5];
+            int causal=configs[ci][6];
+            fail += run_prefill_test(B, Hq, Hk, ql, kl, D, causal);
+            if (fail) break;
+        }
+        if (fail) { printf("FAILED prefill tests\n"); return fail; }
+        bench_prefill();
+    }
+
+    printf("\nAll tests passed!\n");
+    return 0;
+}

csrc/tests/test_utils.cuh

@@ -0,0 +1,187 @@
+#pragma once
+
+#include <cstdio>
+#include <cstdlib>
+#include <cmath>
+#include <chrono>
+#include <cuda_bf16.h>
+
+using bf16 = __nv_bfloat16;
+
+inline bf16 f2bf(float x) { return __float2bfloat16(x); }
+inline float bf2f(bf16 x) { return __bfloat162float(x); }
+
+inline float randf() { return (float)rand() / (float)RAND_MAX - 0.5f; }
+
+inline double now_ms() {
+    using namespace std::chrono;
+    return duration_cast<milliseconds>(steady_clock::now().time_since_epoch()).count();
+}
+
+#define CUDA_CHECK(call) \
+    do { \
+        cudaError_t _e = (call); \
+        if (_e != cudaSuccess) { \
+            printf("CUDA error %s at %s:%d\n", cudaGetErrorString(_e), __FILE__, __LINE__); \
+            exit(1); \
+        } \
+    } while (0)
+
+struct BenchResult {
+    float ms;
+    double tflops;
+};
+
+template <typename Fn>
+BenchResult bench_kernel(Fn launch, int warmup, int iters,
+                         double flops) {
+    for (int i = 0; i < warmup; i++) launch();
+    cudaDeviceSynchronize();
+    cudaError_t err = cudaGetLastError();
+    if (err != cudaSuccess) {
+        printf("CUDA error before bench: %s\n", cudaGetErrorString(err));
+        return {0, 0};
+    }
+
+    cudaEvent_t s, e;
+    cudaEventCreate(&s); cudaEventCreate(&e);
+    cudaEventRecord(s);
+    for (int i = 0; i < iters; i++) launch();
+    cudaEventRecord(e); cudaEventSynchronize(e);
+    float ms = 0; cudaEventElapsedTime(&ms, s, e); ms /= iters;
+    cudaEventDestroy(s); cudaEventDestroy(e);
+
+    return {ms, flops / (ms * 1e-3) / 1e12};
+}
+
+inline void print_bench_header() {
+    printf("%-46s | %10s | %10s\n",
+           "config", "latency", "TFLOP/s");
+    printf("---------------------------------------------------------------"
+           "----------------------------\n");
+}
+
+inline void print_bench_row(const char* cfg, const BenchResult& r) {
+    printf("%-46s | %7.4f ms | %6.2f\n",
+           cfg, r.ms, r.tflops);
+}
+
+// ---- validation table (kernel vs CPU reference) ----
+inline void print_test_header() {
+    printf("%-46s | %11s | %11s | %6s\n",
+           "config", "max_abs_err", "max_rel_err", "result");
+    printf("----------------------------------------------------------------"
+           "----------------------------\n");
+}
+
+inline void print_test_row(const char* cfg, float max_abs_err,
+                           float max_rel_err, bool pass) {
+    printf("%-46s | %11.3e | %11.3e | %s\n",
+           cfg, max_abs_err, max_rel_err, pass ? "PASS" : "FAIL");
+}
+
+template <int... Ds>
+struct _HeadSwitch;
+
+template <int D>
+struct _HeadSwitch<D> {
+    template <typename Fn>
+    static void call(int hd, Fn&& fn) { if (hd == D) fn.template operator()<D>(); }
+};
+
+template <int D, int... Rest>
+struct _HeadSwitch<D, Rest...> {
+    template <typename Fn>
+    static void call(int hd, Fn&& fn) {
+        if (hd == D) fn.template operator()<D>();
+        else _HeadSwitch<Rest...>::call(hd, fn);
+    }
+};
+
+// Default set: 32, 64, 128, 256
+template <typename Fn>
+void dispatch_by_head_dim(int head_dim, Fn&& fn) {
+    _HeadSwitch<32, 64, 128, 256>::call(head_dim, fn);
+}
+
+// Set default strides for contiguous b h l d layout on AttentionParams.
+template<typename P>
+inline void set_default_strides(P& p) {
+    p.q_stride_b  = p.q_head * p.q_len * p.head_dim;
+    p.q_stride_h  = p.q_len * p.head_dim;
+    p.q_stride_l  = p.head_dim;
+    p.q_stride_d  = 1;
+    p.kv_stride_b = p.kv_head * p.kv_len * p.head_dim;
+    p.kv_stride_h = p.kv_len * p.head_dim;
+    p.kv_stride_l = p.head_dim;
+    p.kv_stride_d = 1;
+    p.mask_b_stride = p.kv_len;
+    p.mask_h_stride = 0;
+    p.mask_q_stride = 0;
+}
+
+// Set default Q strides for a paged decode params struct.
+template<typename P>
+inline void set_default_paged_strides(P& p) {
+    p.q_stride_b  = p.q_head * p.q_len * p.head_dim;
+    p.q_stride_h  = p.q_len * p.head_dim;
+    p.q_stride_l  = p.head_dim;
+    p.q_stride_d  = 1;
+    p.mask_b_stride = p.kv_len;
+    p.mask_h_stride = 0;
+    p.mask_q_stride = 0;
+}
+
+// Generic CPU reference for multi-query / grouped-query attention.
+// Tensor shapes (all float*):
+//   Q : [B, Hq, q_len, D]
+//   K : [B, Hk, kv_len, D]
+//   V : [B, Hk, kv_len, D]
+//   O : [B, Hq, q_len, D]
+// mask: if q_len == 1, shape is [B, kv_len]; otherwise mask is not supported.
+// causal_offset: -1 = non-causal; >=0 = absolute position of first Q token.
+static void cpu_attention_ref(
+    const float* Q, const float* K, const float* V, const bool* mask,
+    float* O, int B, int Hq, int Hk, int q_len, int kv_len, int D,
+    int causal_offset
+) {
+    float scale = 1.0f / sqrtf((float)D);
+    int n_rep = Hq / Hk;
+    for (int b = 0; b < B; b++) {
+        #pragma omp parallel for collapse(2) schedule(dynamic)
+        for (int h = 0; h < Hq; h++) {
+            for (int qi = 0; qi < q_len; qi++) {
+                int kv_h = h / n_rep;
+                float mv = -INFINITY, sv = 0.0f;
+                float accum[256] = {0.0f};
+                int lim = kv_len;
+                if (causal_offset >= 0) {
+                    int c = qi + causal_offset + 1;
+                    lim = (c < kv_len) ? c : kv_len;
+                }
+                for (int kj = 0; kj < lim; kj++) {
+                    if (mask != nullptr && q_len == 1) {
+                        if (!mask[b * kv_len + kj]) continue;
+                    }
+                    float dot = 0.0f;
+                    size_t q_idx = ((size_t)b * Hq + h) * q_len + qi;
+                    size_t kv_idx = ((size_t)b * Hk + kv_h) * kv_len + kj;
+                    for (int d = 0; d < D; d++)
+                        dot += Q[q_idx * D + d] * K[kv_idx * D + d];
+                    dot *= scale;
+                    float nm = fmaxf(mv, dot);
+                    float a = expf(mv - nm);
+                    float b_exp = expf(dot - nm);
+                    sv = sv * a + b_exp;
+                    for (int d = 0; d < D; d++)
+                        accum[d] = accum[d] * a + V[kv_idx * D + d] * b_exp;
+                    mv = nm;
+                }
+                float inv = 1.0f / sv;
+                size_t o_idx = ((size_t)b * Hq + h) * q_len + qi;
+                for (int d = 0; d < D; d++)
+                    O[o_idx * D + d] = accum[d] * inv;
+            }
+        }
+    }
+}

docker-compose.yml

@@ -3,6 +3,8 @@ services:
     build:
       context: .
       dockerfile: Dockerfile
+      args:
+        CUDA_TAG: ${CUDA_TAG:-cu128}
     user: "${UID:-1000}:${GID:-1000}"
     ports:
       - "8000:8000"
@@ -29,6 +31,8 @@ services:
     build:
       context: .
       dockerfile: Dockerfile
+      args:
+        CUDA_TAG: ${CUDA_TAG:-cu128}
     user: "${UID:-1000}:${GID:-1000}"
     ports:
       - "8000:8000"