fix: stabilize paged decode attention kernels

- zero-fill split partials so combine skips unwritten splits deterministically
- skip loading masked KV in paged decode kernels to avoid 0*NaN output poisoning
- zero-fill shared memory tile buffers to prevent stale NaN leaking into softmax

.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

@@ -5,8 +5,16 @@
 !*/
 
 # Allow specific file types and root files
-!*.py
+!astrai/**/*.py
-!*.sh
+!scripts/**/*.py
+!tests/**/*.py
+!csrc/**/*.py
+
+!csrc/**/*.cu
+!csrc/**/*.h
+!csrc/**/*.cuh
+
+!scripts/**/*.sh
 
 # Allow GitHub files
 !/.github/**
@@ -16,8 +24,13 @@
 !/.dockerignore
 !/Dockerfile
 !/docker-compose.yml
-!/assets/**
+!/docs/**
 !/CONTRIBUTING.md
 !/LICENSE
 !/pyproject.toml
-!/README.md
+!/README.md
+# Allow extension modules (only source .py)
+!/astrai/extension/**/*.py
+
+# Allow build files
+!/setup.py

CONTRIBUTING.md

@@ -5,7 +5,7 @@ Thank you for your interest in contributing! This document provides step-by-step
 ## Quick Start
 
 ```bash
-git clone https://github.com/your-username/AstrAI.git
+git clone https://github.com/ViperEkura/AstrAI.git
 cd AstrAI
 pip install -e ".[dev]"     # install with dev dependencies (pytest, ruff)
 ```

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/cu126
+    --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 .
 

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>
@@ -9,18 +9,18 @@
 <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?color=76bad9" alt="release">
+  <img src="https://img.shields.io/github/v/tag/ViperEkura/AstrAI?label=Release&color=76bad9" alt="release">
-  <img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fapi.github.com%2Frepos%2FViperEkura%2FAstrAI&query=%24.stargazers_count&label=stars&suffix=%20stars&color=76bad9" alt="stars">
+  <img src="https://img.shields.io/github/stars/ViperEkura/AstrAI?style=flat&label=Stars&color=76bad9" alt="stars">
-  <img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fapi.github.com%2Frepos%2FViperEkura%2FAstrAI&query=%24.forks_count&label=forks&suffix=%20forks&color=76bad9" alt="forks">
+  <img src="https://img.shields.io/github/forks/ViperEkura/AstrAI?style=flat&label=Forks&color=76bad9" alt="forks">
 </div>
 <br>
 
 <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>
@@ -28,7 +28,8 @@
 ## 📖 Table of Contents
 
 - [Features](#features)
-- [Quick Start](#quick-start)
+- [Getting Started](#getting-started)
+- [Demo](#demo)
 - [Documentation](#documentation)
 - [Contributing](#contributing)
 - [Community](#community)
@@ -49,39 +50,51 @@
 - 🤗 **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.
 
-### Quick Start
+### Getting Started
 
-#### Installation
+End-to-end walkthrough in 5 steps:
+
+**1. Install**
 
 ```bash
 git clone https://github.com/ViperEkura/AstrAI.git
 cd AstrAI
-pip install -e .
+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)
 ```
 
-For development dependencies:
+**2. Download model**
 
 ```bash
-pip install -e ".[dev]"
+python scripts/demo/download.py    # downloads 1B checkpoint to params/
 ```
 
-#### Download Pre-trained Model
+**3. Preprocess data**
 
-Download pre-trained model weights (1B bilingual checkpoint) to `params/`:
+Create `pretrain.json` (preprocessing config for `seq` strategy):
+
+```json
+{
+    "version": 1,
+    "input": {"sections": [{"field": "text", "action": "train"}]},
+    "preprocessing": {"max_seq_len": 2048},
+    "output": {"storage_format": "bin"}
+}
+```
 
 ```bash
-python scripts/demo/download.py
+python scripts/tools/preprocess.py data/*.jsonl -o output/ -c pretrain.json
 ```
 
-Or download manually from [HuggingFace](https://huggingface.co/ViperEk/KHAOSZ) into `params/`.
+**4. Train**
-
-#### Train a Model
 
 ```bash
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 nohup python scripts/tools/train.py \
     --nprocs=4 \
+    --parallel_mode=ddp \
     --train_type=seq \
     --data_root_path=/path/to/dataset \
     --param_path=/path/to/model \
@@ -90,9 +103,7 @@ nohup python scripts/tools/train.py \
     --warmup_ratio=0.05 \
     --max_lr=1e-4 \
     --max_grad_norm=1.0 \
-    --adamw_beta1=0.9 \
+    --weight_decay=0.1 \
-    --adamw_beta2=0.95 \
-    --adamw_weight_decay=0.01 \
     --window_size=2048 \
     --ckpt_interval=10000 \
     --ckpt_dir=./checkpoint \
@@ -101,15 +112,54 @@ nohup python scripts/tools/train.py \
     > out.log 2> err.log &
 ```
 
-Full reference at [Parameter Guide](assets/docs/params.md).
+**5. Serve & query**
 
-#### Generate Text
+```bash
+# Terminal 1: start server
+python scripts/tools/server.py --param_path ./params --device cuda
+
+# Terminal 2: query
+curl http://localhost:8000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{"messages":[{"role":"user","content":"Hello"}],"max_tokens":512}'
+```
+
+### Demo
+
+Check out the demos in the `scripts/demo/` folder:
+
+```bash
+# Download model weights (required before running demos)
+python scripts/demo/download.py                      # model → params/
+
+# Interactive streaming chat (multi-turn, maintains history)
+python scripts/demo/stream_chat.py
+# Type your message after >>, type !exit to quit
+
+# Batch generation (5 hardcoded prompts, non-streaming)
+python scripts/demo/generate_batch.py
+
+# Single-prompt autoregressive streaming
+python scripts/demo/generate_ar.py
+```
+
+All generation demos use `temperature=0.8`, `top_p=0.95`, `top_k=50`, `max_tokens=2048` by default and require `params/` to contain model weights (run `download.py` first).
+
+Watch a video walkthrough on [bilibili](https://www.bilibili.com/video/BV1fuLB6yEj6).
+
+---
+
+See [Documentation](#documentation) for full references beyond the examples above.
+
+#### Text Generation
+
+Batch generation from a JSONL file:
 
 ```bash
 python scripts/tools/generate.py \
-    --param_path /path/to/model \
+    --param_path ./params \
-    --input_json_file /path/to/input.json \
+    --input_json_file input.jsonl \
-    --output_json_file /path/to/output.json
+    --output_json_file output.jsonl
 ```
 
 #### Docker
@@ -123,9 +173,6 @@ docker build -t astrai:latest .
 # Run with GPU support
 docker run --gpus all -it astrai:latest
 
-# Run with specific GPUs
-docker run --gpus '"device=0,1"' -it astrai:latest
-
 # Run inference server
 docker run --gpus all -p 8000:8000 astrai:latest \
   python -m scripts.tools.server --port 8000 --device cuda
@@ -142,88 +189,47 @@ docker compose --profile cpu up -d
 
 > **Note**: `--gpus all` is required for CUDA support. Without it, `torch.cuda.is_available()` will return `False`.
 
-#### Start HTTP Server
+#### HTTP API Examples
 
-Start the inference server with OpenAI and Anthropic-compatible HTTP API:
+Additional request examples beyond the [Getting Started](#getting-started) flow:
 
 ```bash
-python -m scripts.tools.server --port 8000 --device cuda
-```
-
-Make requests:
-
-```bash
-# OpenAI-compatible
-curl -X POST http://localhost:8000/v1/chat/completions \
-  -H "Content-Type: application/json" \
-  -d '{
-    "messages": [{"role": "user", "content": "Hello"}],
-    "max_tokens": 512
-  }'
-
 # OpenAI-compatible streaming
 curl -X POST http://localhost:8000/v1/chat/completions \
   -H "Content-Type: application/json" \
-  -d '{
+  -d '{"messages":[{"role":"user","content":"Tell a story"}],"stream":true,"max_tokens":500}'
-    "messages": [{"role": "user", "content": "Tell a story"}],
-    "stream": true,
-    "max_tokens": 500
-  }'
 
 # Anthropic-compatible
 curl -X POST http://localhost:8000/v1/messages \
   -H "Content-Type: application/json" \
-  -d '{
+  -d '{"model":"astrai","system":"You are a helpful assistant.","messages":[{"role":"user","content":"Hello"}],"max_tokens":512}'
-    "model": "astrai",
-    "system": "You are a helpful assistant.",
-    "messages": [{"role": "user", "content": "Hello"}],
-    "max_tokens": 512
-  }'
 
 # Anthropic-compatible streaming with stop sequences
 curl -X POST http://localhost:8000/v1/messages \
   -H "Content-Type: application/json" \
-  -d '{
+  -d '{"model":"astrai","messages":[{"role":"user","content":"Write a story"}],"max_tokens":500,"stream":true,"stop_sequences":["The end"]}'
-    "model": "astrai",
-    "messages": [{"role": "user", "content": "Write a story"}],
-    "max_tokens": 500,
-    "stream": true,
-    "stop_sequences": ["The end"]
-  }'
 
 # Health check
 curl http://localhost:8000/health
 ```
 
-#### Demo
+See [Inference Guide](docs/guides/inference.md) for SSE streaming format, error codes, and stats endpoint.
-
-Check out the demos in the `scripts/demo/` folder:
-
-```bash
-# Download pre‑processed data (required before running demos)
-python scripts/demo/download.py
-
-# Interactive streaming chat
-python scripts/demo/stream_chat.py
-
-# Batch generation
-python scripts/demo/generate_batch.py
-
-# Auto‑regressive generation
-python scripts/demo/generate_ar.py
-```
-
-Watch a video walkthrough on [bilibili](https://www.bilibili.com/video/BV1fuLB6yEj6).
 
 ### Documentation
 
 | Document | Description |
 |----------|-------------|
-| [Parameter Guide](./assets/docs/params.md) | Training & inference parameters |
+| [Get Started](./docs/get-started.md) | Installation and quickstart |
-| [Architecture](./assets/docs/architecture.md) | System architecture, class diagram & design patterns |
+| [CLI Reference](./docs/guides/params.md) | Parameters for all CLI tools (train, server, generate, preprocess) |
-| [Training](./assets/docs/training.md) | Training loop, strategies & formulas |
+| [Preprocessing](./docs/guides/preprocessing.md) | Declarative JSON-driven data preprocessing |
-| [Inference](./assets/docs/inference.md) | KVCache, continuous batching, sampling & HTTP API |
+| [Training](./docs/guides/training.md) | Training loop, strategies & formulas |
-| [Data Flow](./assets/docs/dataflow.md) | Data pipeline, storage backends & dataset architecture |
+| [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
 
@@ -240,7 +246,7 @@ 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
 

assets/docs/dataflow.md

@@ -1,57 +0,0 @@
-# Data Flow
-
-This document describes the data pipeline: from raw text to model input tensors.
-
-## Overview
-
-```
-Raw Text → AutoTokenizer → Token IDs → .h5/.bin → Dataset → 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.
-
-Storage format is auto-detected by `detect_format()`; backends are dispatched via registry:
-
-```
-StoreFactory.create("h5")  → H5Store
-StoreFactory.create("bin") → MmapStore
-```
-
-H5 backend supports shared memory via `.share_memory_()`. Bin (mmap) uses OS page-cache sharing natively.
-
-## Data Keys by Training Type
-
-| Type | Storage Keys |
-|------|-------------|
-| `seq` | `sequence` (→ input_ids, target_ids via offset-by-1) |
-| `sft` | `sequence`, `loss_mask` |
-| `dpo` | `chosen`, `rejected`, `chosen_mask`, `rejected_mask` |
-| `grpo` | `prompts`, `responses`, `masks`, `rewards` |
-
-## Dataset Architecture
-
-```
-DatasetFactory.load(train_type, load_path, window_size, stride, storage_type)
-  → StoreFactory.create(detect_format(path))
-    → Store._data[Dict[str, List[Tensor]]] + _cum[Dict[str, List[int]]]
-      → BaseDataset.__getitem__(idx)
-        → sliding window [begin, end) via get_index(idx)
-```
-
-`window_size` = max input length, `stride` = step between consecutive samples (defaults to `window_size`).
-
-## 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: 2026-05-28

assets/docs/inference.md

@@ -1,148 +0,0 @@
-# Inference
-
-## 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
-
-Six 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` 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
-```
-
-`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) |
-| `temperature` | float | 1.0 | Sampling temperature (>= 0.0) |
-| `top_p` | float | 1.0 | Nucleus threshold |
-| `top_k` | int | 50 | Top-k count |
-| `max_tokens` | Optional[int] | None | Max generation length |
-| `stream` | bool | False | Stream output |
-
-## 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
-await engine.generate_async("Hello", ...)       # -> AsyncGenerator[str]
-```
-
-> Document Update Time: 2026-05-28

assets/docs/params.md

@@ -1,99 +0,0 @@
-# Parameter Documentation
-
-## Training Parameters
-
-### Basic Parameters
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `--train_type` | Training type (`seq`, `sft`, `dpo`, `grpo`) | required |
-| `--data_root_path` | Dataset root directory | required |
-| `--param_path` | Model parameters or checkpoint path | required |
-| `--n_epoch` | Total training epochs | 1 |
-| `--batch_per_device` | Batch size per device | 1 |
-| `--grad_accum_steps` | Gradient accumulation steps between optimizer steps | 1 |
-
-### Learning Rate Scheduling
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `--warmup_ratio` | Fraction of total steps used for LR warmup | 0.05 |
-| `--max_lr` | Maximum learning rate (cosine decay after warmup) | 3e-4 |
-| `--max_grad_norm` | Maximum gradient norm for clipping | 1.0 |
-
-### Optimizer (AdamW)
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `--adamw_beta1` | AdamW beta1 | 0.9 |
-| `--adamw_beta2` | AdamW beta2 | 0.95 |
-| `--adamw_weight_decay` | AdamW weight decay | 0.01 |
-
-### Data Loading
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `--window_size` | Max input sequence length | model config `max_len` |
-| `--stride` | Stride for sliding window over sequences | None |
-| `--random_seed` | Random seed for reproducibility | 3407 |
-| `--num_workers` | DataLoader worker processes | 4 |
-| `--no_pin_memory` | Disable pin_memory (enabled by default) | (flag) |
-
-### Checkpoint & Resume
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `--ckpt_interval` | Iterations between checkpoints | 5000 |
-| `--ckpt_dir` | Checkpoint save directory | checkpoint |
-| `--start_epoch` | Resume from epoch (0 = from scratch) | 0 |
-| `--start_batch` | Resume from batch iteration | 0 |
-
-### Distributed Training
-
-| Parameter | Description | Default |
-|-----------|-------------|---------|
-| `--nprocs` | Number of GPUs / processes | 1 |
-| `--parallel_mode` | Parallel strategy (`none`, `ddp`, or `fsdp`) | none |
-| `--device_type` | Device type | cuda |
-| `--start_method` | Multiprocessing start method (`spawn`, `fork`, `forkserver`) | spawn |
-
-### Strategy-specific
-
-| Parameter | Description | Default | Used by |
-|-----------|-------------|---------|---------|
-| `--dpo_beta` | DPO beta value | 0.1 | `dpo` |
-| `--label_smoothing` | Label smoothing for cross-entropy loss | 0.05 | `seq`, `sft` |
-| `--group_size` | GRPO group size | 4 | `grpo` |
-| `--grpo_clip_eps` | GRPO clipping epsilon | 0.2 | `grpo` |
-| `--grpo_kl_coef` | GRPO KL penalty coefficient | 0.01 | `grpo` |
-| `--grpo_sync_interval` | GRPO ref_model sync interval (steps) | 200 | `grpo` |
-
-### Usage Example
-
-```bash
-export CUDA_VISIBLE_DEVICES=0,1,2,3
-
-nohup python scripts/tools/train.py \
-    --nprocs=4 \
-    --train_type=seq \
-    --data_root_path=/path/to/dataset \
-    --param_path=/path/to/model \
-    --batch_per_device=4 \
-    --grad_accum_steps=8 \
-    --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 \
-    --window_size=2048 \
-    --ckpt_interval=10000 \
-    --ckpt_dir=./checkpoint \
-    --random_seed=3407 \
-    --label_smoothing=0.05 \
-    > out.log 2> err.log &
-```
-
----
-
-> Document Update Time: 2026-05-24

assets/docs/training.md

@@ -1,227 +0,0 @@
-# Training
-
-## Model Architecture
-
-The model uses a decoder-only Transformer with **GQA** (Grouped Query Attention) and optional **MLA** (Multi-head Latent Attention). 1.0 billion parameters, Chinese–English bilingual.
-
-```mermaid
-flowchart TB
-    subgraph Layers["Transformer Layers"]
-        direction TB
-        A[Input Embedding] --> B[Transformer Block\nLayer 1]
-        B --> C[Transformer Block\nLayer ...]
-        C --> D[Transformer Block\nLayer ...]
-        D --> E[RMSNorm]
-        E --> F[Linear]
-        F --> G[SoftMax]
-    end
-
-    subgraph TransformerBlock["Transformer Block"]
-        direction TB
-        H[x] --> I[RMSNorm]
-        I --> J[Linear → Q/K/V]
-        J --> K[Q]; J --> L[K]; J --> M[V]
-        K --> N[RoPE]; L --> O[RoPE]
-        N --> P["Q @ K^T / sqrt(d)"]; O --> P
-        P --> Q[Masked SoftMax]; Q --> R[S @ V]; M --> R
-        R --> S[Linear]; S --> T[+]; H --> T
-        T --> U[RMSNorm]
-        U --> V["Linear (gate)"]; U --> W["Linear (up)"]
-        V --> X[SiLU]; X --> Y[×]; W --> Y
-        Y --> Z["Linear (down)"]; Z --> AA[+]; T --> AA
-        AA --> BB[x']
-    end
-```
-
-### Autoregression
-
-Given a token sequence, the model predicts the probability of the next token. Each generated token is appended to the input and fed back, repeating until an end-of-sequence token or max length.
-
-### Causal Mask
-
-```
-sequence : [[1, 2, 3, 4, 5, 6]]
-input_ids: [[1, 2, 3, 4, 5]]
-target_ids: [[2, 3, 4, 5, 6]]
-```
-
-Lower-triangular mask prevents attending to future positions:
-
-```
-[[0, -inf, -inf, -inf, -inf],
- [0,    0, -inf, -inf, -inf],
- [0,    0,    0, -inf, -inf],
- [0,    0,    0,    0, -inf],
- [0,    0,    0,    0,    0]]
-```
-
-### Rotary Position Embedding (RoPE)
-
-RoPE embeds position into Q/K vectors via complex rotation:
-
-$$ q_i = R_i W_q x_i, \quad k_j = R_j W_k x_j, \quad q_i^T k_j = x_i^T W_q^T R_{i-j} W_k x_j $$
-
-The complex rotation `freqs_cis` is pre-computed once (`cos, sin` pairs per position). `apply_rotary_emb` multiplies Q/K as complex numbers.
-
-## Training Loop
-
-Two-level loop: **epoch** → **batch**. Optimizer step fires every `grad_accum_steps` batches.
-
-```
-on_train_begin
-  on_epoch_begin
-    for batch in dataloader:
-      on_batch_begin
-      with executor.accumulate(model):
-        loss = strategy(batch)
-        stand_loss = loss / executor.grad_accum_steps
-        executor.backward(stand_loss)
-        iteration += 1
-        on_batch_end
-
-        if executor.sync_gradients:
-          on_optimizer_step
-          optimizer.step()
-          optimizer.zero_grad()
-          if scheduler:
-            scheduler.step()
-    on_epoch_end
-on_train_end
-```
-
-### Callback Lifecycle
-
-| Hook | Fires | Default callback |
-|------|-------|-----------------|
-| `on_train_begin` | Before training starts | `GradientCheckpointingCallback` |
-| `on_optimizer_step` | Every accumulation window | `GradientClippingCallback`, `ValidationCallback` |
-| `on_batch_end` | Every batch | `CheckpointCallback`, `MetricLoggerCallback`, `ProgressBarCallback` |
-| `on_train_end` | Training ends | `CheckpointCallback`, `MetricLoggerCallback` (final save) |
-
-Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `metric_logger` (JSONL, rank-0), `progress_bar` (tqdm), `gradient_clipping`, `validation` (periodic validation on val_dataset).
-
-## Strategies
-
-### SEQ (Pre-training)
-
-Next-token cross-entropy with optional label smoothing:
-
-$$
-L_{\text{PT}} = -\sum_{t=1}^{T} \log P(x_t \mid x_{\lt t}; \theta)
-$$
-
-Keys: `input_ids`, `target_ids`
-
-### SFT (Supervised Fine-Tuning)
-
-Masked cross-entropy (`ignore_index=-100`) over response tokens:
-
-$$
-L_{\text{SFT}} = -\sum_{t=P+1}^{P+L} \log P(s_t \mid s_{\lt t}; \theta)
-$$
-
-Keys: `input_ids`, `target_ids`, `loss_mask`
-
-### DPO (Direct Preference Optimization)
-
-Frozen reference model, preference margin via log-ratio:
-
-$$
-L_{\text{DPO}} = -\mathbb{E}\left[\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\text{ref}}(y_w\mid x)} - \beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\text{ref}}(y_l\mid x)}\right)\right]
-$$
-
-Parameters: `beta=0.1`. Keys: `chosen`, `rejected`, `chosen_mask`, `rejected_mask`.
-
-### GRPO (Group Relative Policy Optimization)
-
-On-policy PPO with group-normalized advantages:
-
-$$
-\text{Advantage}_i = \frac{r_i - \mu}{\sigma + \epsilon}
-$$
-
-$$
-L_{\text{GRPO}} = -\mathbb{E}\left[\min\left(\frac{\pi_\theta}{\pi_{\text{ref}}}A,\; \text{clip}\left(\frac{\pi_\theta}{\pi_{\text{ref}}}, 1-\epsilon, 1+\epsilon\right)A\right)\right] + \lambda \cdot \mathbb{E}\left[(\log\pi_\theta - \log\pi_{\text{ref}})^2\right]
-$$
-
-Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`, `sync_interval=200`.
-
-Keys: `prompts`, `responses`, `masks`, `rewards`.
-
-## LR Schedulers
-
-| Type | Class | Description |
-|------|-------|-------------|
-| Cosine | `CosineScheduler` | Linear warmup → cosine decay to `min_rate` |
-| SGDR | `SGDRScheduler` | Cosine annealing with warm restarts (`t_mult=2`) |
-
-Created by `SchedulerFactory.create(optimizer, schedule_type, **kwargs)`.
-
-## Gradient Checkpointing
-
-Trades compute for memory by recomputing activations during backward pass. Specify module types via `gradient_checkpointing_modules`:
-
-```python
-from astrai.model.components.decoder_block import DecoderBlock
-config = TrainConfig(..., gradient_checkpointing_modules=[DecoderBlock])
-```
-
-Callback wraps each `DecoderBlock.forward` with `torch.utils.checkpoint.checkpoint(use_reentrant=False)`, compatible with `torch.compile`. Uses `nn.Module.apply()` for traversal — works through DDP wrappers without manual unwrap. Empty list (default) means no-op.
-
-## Checkpoint
-
-```
-Checkpoint(state_dict, epoch, iteration, extra, meta, config)
-  ├── save(save_dir)    rank-0 only: meta.json (epoch/iteration/timestamp) + config.json (model config) + state_dict.safetensors + optional {key}.pt (optimizer.pt, scheduler.pt)
-  └── load(save_dir)    broadcasts metadata from rank-0
-```
-
-Optimizer/scheduler state persisted by default via `Checkpoint.extra`.  
-Model config (`context.model_config`) saved into `config.json` during training via `CheckpointCallback`.
-
-## TrainContextBuilder (Builder Pattern)
-
-```python
-context = (
-    TrainContextBuilder(config)
-        .with_resume_dir(resume_dir)
-        .build()
-)
-# Returns TrainContext with model, strategy, optimizer, scheduler, dataloader, checkpoint
-```
-
-- Loads checkpoint weights if provided
-- Creates executor via `ExecutorFactory.create(cfg.parallel_mode, grad_accum_steps=cfg.grad_accum_steps, **cfg.executor_kwargs)`
-- Calls `executor.prepare(model, optimizer, dataloader, scheduler)` for model distribution (e.g. DDP) + gradient accumulation wrappers
-- Creates `ResumableDistributedSampler` for shuffle+resume
-- Builds strategy via `StrategyFactory.create(train_type, ...)`
-
-## Training CLI
-
-```bash
-export CUDA_VISIBLE_DEVICES=0,1,2,3
-
-nohup python scripts/tools/train.py \
-    --nprocs=4 \
-    --train_type=seq \
-    --data_root_path=/path/to/dataset \
-    --param_path=/path/to/model \
-    --batch_per_device=4 \
-    --grad_accum_steps=8 \
-    --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 \
-    --window_size=2048 \
-    --ckpt_interval=10000 \
-    --ckpt_dir=./checkpoint \
-    --random_seed=3407 \
-    --label_smoothing=0.05 \
-    > out.log 2> err.log &
-```
-
-Full parameter reference at [params.md](params.md).
-
-> Document Update Time: 2026-05-28

astrai/__init__.py

@@ -1,34 +1,126 @@
-__version__ = "1.3.7"
+__version__ = "1.3.11"
 __author__ = "ViperEkura"
 
+import logging
+import os
+
 from astrai.config import (
     AutoRegressiveLMConfig,
+    BaseModelConfig,
+    ConfigFactory,
     EncoderConfig,
+    PipelineConfig,
     TrainConfig,
 )
-from astrai.dataset import DatasetFactory
+from astrai.dataset import (
+    BaseDataset,
+    DatasetFactory,
+    RDSampler,
+    Store,
+    StoreFactory,
+)
 from astrai.factory import BaseFactory
 from astrai.inference import (
     GenerationRequest,
     InferenceEngine,
+    ProtocolHandler,
+    SamplingPipeline,
+    get_app,
+    run_server,
+    sample,
 )
-from astrai.model import AutoModel, AutoRegressiveLM
+from astrai.model import (
-from astrai.tokenize import AutoTokenizer
+    AutoModel,
-from astrai.trainer import CallbackFactory, SchedulerFactory, StrategyFactory, Trainer
+    AutoRegressiveLM,
+    EmbeddingEncoder,
+    LoRAConfig,
+    inject_lora,
+)
+from astrai.parallel import (
+    ExecutorFactory,
+    get_rank,
+    get_world_size,
+    only_on_rank,
+    spawn_parallel_fn,
+)
+from astrai.preprocessing import Pipeline, filter_by_length
+from astrai.serialization import Checkpoint
+from astrai.tokenize import AutoTokenizer, ChatTemplate
+from astrai.trainer import (
+    BaseScheduler,
+    BaseStrategy,
+    CallbackFactory,
+    SchedulerFactory,
+    StrategyFactory,
+    TrainCallback,
+    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",
-    "EncoderConfig",
+    "AutoModel",
-    "TrainConfig",
-    "DatasetFactory",
     "AutoTokenizer",
+    "BaseDataset",
+    "BaseFactory",
+    "BaseModelConfig",
+    "BaseScheduler",
+    "BaseStrategy",
+    "CallbackFactory",
+    "ChatTemplate",
+    "Checkpoint",
+    "ConfigFactory",
+    "DatasetFactory",
+    "EmbeddingEncoder",
+    "EncoderConfig",
+    "ExecutorFactory",
     "GenerationRequest",
     "InferenceEngine",
-    "Trainer",
+    "LoRAConfig",
-    "CallbackFactory",
+    "Pipeline",
-    "StrategyFactory",
+    "PipelineConfig",
+    "ProtocolHandler",
+    "RDSampler",
+    "SamplingPipeline",
     "SchedulerFactory",
-    "BaseFactory",
+    "Store",
-    "AutoModel",
+    "StoreFactory",
+    "StrategyFactory",
+    "TrainCallback",
+    "TrainConfig",
+    "Trainer",
+    "filter_by_length",
+    "get_app",
+    "get_rank",
+    "get_world_size",
+    "inject_lora",
+    "only_on_rank",
+    "run_server",
+    "sample",
+    "setup_logging",
+    "spawn_parallel_fn",
 ]

astrai/config/__init__.py

@@ -4,13 +4,22 @@ from astrai.config.model_config import (
     ConfigFactory,
     EncoderConfig,
 )
+from astrai.config.preprocess_config import (
+    InputConfig,
+    OutputConfig,
+    PipelineConfig,
+    ProcessingConfig,
+)
 from astrai.config.train_config import TrainConfig
 
 __all__ = [
-    # Model configuration
     "BaseModelConfig",
     "AutoRegressiveLMConfig",
     "EncoderConfig",
     "ConfigFactory",
     "TrainConfig",
+    "InputConfig",
+    "OutputConfig",
+    "PipelineConfig",
+    "ProcessingConfig",
 ]

astrai/config/base.py

@@ -1,86 +1,38 @@
 import json
-from dataclasses import MISSING, dataclass, fields
+from dataclasses import asdict
-from typing import Any, Dict, Optional, Self, get_type_hints
+from pathlib import Path
+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 = {}
+        result = {}
-        for fld in fields(self):
+        for k, v in asdict(self).items():
-            v = getattr(self, fld.name)
+            if isinstance(v, tuple):
-            if isinstance(v, (str, int, float, bool)):
+                v = list(v)
-                d[fld.name] = v
+            try:
-            elif v is None:
+                json.dumps(v)
-                d[fld.name] = None
+                result[k] = v
-            elif isinstance(v, (dict, list, tuple)):
+            except (TypeError, ValueError):
-                try:
+                # Skip non-serializable runtime objects (e.g. model_fn, dataset).
-                    val = list(v) if isinstance(v, tuple) else v
+                # TrainConfig mixes hyperparams with callables/datasets; only the
-                    json.dumps(val)
+                # JSON-serializable subset is written to checkpoint meta.
-                    d[fld.name] = val
+                pass
-                except (TypeError, ValueError):
+        return result
-                    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
 
     @classmethod
     def from_dict(cls, d: Dict[str, Any]) -> Self:
-        hints = get_type_hints(cls)
+        return cls(**d)
-        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
+    @classmethod
-    def _unwrap_optional(tp) -> Optional[type]:
+    def from_file(cls, path: Union[str, Path]) -> Self:
-        if tp is None:
+        with open(path, "r", encoding="utf-8") as f:
-            return None
+            return cls.from_dict(json.load(f))
-        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 to_file(self, path: Union[str, Path]):
-    def _coerce(value: Any, target_type: type) -> Any:
+        with open(path, "w", encoding="utf-8") as f:
-        if target_type is bool and isinstance(value, bool):
+            json.dump(self.to_dict(), f, indent=2, ensure_ascii=False)
-            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
-        raise TypeError

astrai/config/model_config.py

@@ -1,10 +1,14 @@
-import json
+from typing import Any, Dict, Optional
-from dataclasses import dataclass
+
-from typing import Any, Dict, Optional, Self
+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``."""
@@ -18,75 +22,140 @@ 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
-    @classmethod
-    def from_file(cls, config_path: str) -> Self:
-        with open(config_path, "r") as f:
-            raw: Dict[str, Any] = json.load(f)
-        return cls.from_dict(raw)
-
-    def to_file(self, config_path: str):
-        d = self.to_dict()
-        config_dict = {k: v for k, v in d.items() if v is not None}
-        with open(config_path, "w") as f:
-            json.dump(config_dict, f, indent=4)
 
 
 @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.
+    """
 
     vocab_size: Optional[int] = None
-    dim: Optional[int] = None
+    hidden_size: Optional[int] = None
-    n_layers: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
+    rms_norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
+    intermediate_size: Optional[int] = None
-    tie_weight: Optional[bool] = None
+    tie_word_embeddings: Optional[bool] = None
-
+    max_position_embeddings: Optional[int] = None
-    max_len: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
-
     attn_type: str = "gqa"
-    n_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
-    n_kv_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
 
+    @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
+
 
 @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
+    hidden_size: Optional[int] = None
-    n_layers: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
+    rms_norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
+    intermediate_size: Optional[int] = None
-
+    max_position_embeddings: Optional[int] = None
-    max_len: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
-
+    attn_type: str = "gqa"
-    n_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
-    n_kv_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

@@ -0,0 +1,152 @@
+"""Pipeline configuration for JSONL preprocessing.
+
+Supports single-sequence (SFT/pretrain) and multi-output (DPO/GRPO)
+modes, both driven declaratively through ``input.sections`` or
+``input.sources``.
+"""
+
+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):
+    """Declarative input mapping.
+
+    Single-output mode (backward-compatible)::
+
+        {"input": {"sections": [{"field": "messages", ...}]}}
+
+    Multi-output mode (DPO / GRPO)::
+
+        {"input": {"sources": {
+            "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
+    sources: Optional[Dict[str, Dict]] = None
+
+
+@dataclass
+class ProcessingConfig(BaseConfig):
+    """Processing configuration for tokenization and packing.
+
+    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 for storage.
+
+    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
+    storage_format: str = "bin"
+    max_tokens_per_shard: int = 100_000_000
+    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)
+    mask_default: str = "mask"
+    preprocessing: ProcessingConfig = field(default_factory=ProcessingConfig)
+    output: OutputConfig = field(default_factory=OutputConfig)

astrai/config/train_config.py

@@ -1,7 +1,9 @@
-from dataclasses import dataclass, field, fields
+from dataclasses import field
-from typing import Callable, List, Optional
+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
@@ -9,132 +11,200 @@ from torch.utils.data import Dataset
 from astrai.config.base import BaseConfig
 from astrai.model.components.lora import LoRAConfig
 
-
+_TRAIN_TYPES = frozenset({"seq", "sft", "dpo", "grpo", "online_grpo", "online_dpo"})
-def required(**kw):
+_PARALLEL_MODES = frozenset({"none", "ddp", "fsdp"})
-    return {"required": True, **kw}
+_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
+    """Training configuration.
-    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 = field(
-        default_factory=list,
-        metadata={"help": "Module types to enable activation checkpointing for."},
-    )
 
-    # checkpoint setting
+    Combines hyperparameters with runtime objects (model_fn, dataset, etc.).
-    start_epoch: int = field(default=0, metadata={"help": "Start epoch for training."})
+    Only JSON-serializable fields are written to checkpoint meta via to_dict().
-    start_batch: int = field(
-        default=0, metadata={"help": "Start batch iteration for training."}
-    )
-    ckpt_dir: str = field(
-        default="./checkpoint", metadata={"help": "Checkpoint directory."}
-    )
-    ckpt_interval: int = field(
-        default=5000, metadata={"help": "Number of iterations between checkpoints."}
-    )
 
-    # lora setting
+    Args:
-    lora: Optional[LoRAConfig] = field(
+        model_fn (Callable[[], nn.Module]): Model factory for training.
-        default=None,
+        strategy (str): Training strategy (seq, sft, dpo, grpo, online_*).
-        metadata={"help": "LoRA config. None means full fine-tuning."},
+        dataset (Dataset): Dataset for training.
-    )
+        optimizer_fn (Callable[[nn.Module], Optimizer]): Optimizer factory for training.
+        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.
+        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
+    model_fn: Callable[[], nn.Module]
-    log_dir: str = field(
+    strategy: str
-        default="./checkpoint/logs", metadata={"help": "Directory for metric logs."}
+    dataset: Dataset
-    )
+    optimizer_fn: Callable[[nn.Module], Optimizer]
-    log_interval: int = field(
+    scheduler_fn: Callable[[Optimizer], LRScheduler]
-        default=100,
+    n_epoch: int = 1
-        metadata={"help": "Number of batch iterations between metric logs."},
+    batch_per_device: int = 4
-    )
+    grad_accum_steps: int = 1
-    metrics: List[str] = field(
+    max_grad_norm: Optional[float] = 1.0
-        default_factory=lambda: ["loss", "lr"],
+    gradient_checkpointing_modules: List[type] = field(default_factory=list)
-        metadata={"help": "Metrics to record during training."},
+    compile_mode: Optional[str] = None
-    )
 
-    # dataloader setting
+    start_epoch: int = 0
-    random_seed: int = field(default=3407, metadata={"help": "Random seed."})
+    start_samples: int = 0
-    num_workers: int = field(
+    ckpt_dir: str = "./checkpoint"
-        default=0, metadata={"help": "Number of workers for dataloader."}
+    ckpt_interval: int = 5000
-    )
-    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."}
-    )
 
-    # distributed training
+    lora: Optional[LoRAConfig] = None
-    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)."},
-    )
 
-    # others
+    metrics: List[str] = field(default_factory=lambda: ["loss", "lr", "grad_norm"])
-    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_step: int = field(
-        default=1000,
-        metadata={"help": "Number of optimizer steps between validation runs."},
-    )
 
-    executor_kwargs: dict = field(
+    random_seed: int = 3407
-        default_factory=dict,
+    num_workers: int = 0
-        metadata={"help": "Extra kwargs passed to ExecutorFactory.create()."},
+    prefetch_factor: Optional[int] = None
-    )
+    pin_memory: bool = False
-    extra_kwargs: dict = field(
+    collate_fn: Optional[Callable[[List[Any]], Any]] = None
-        default_factory=dict, metadata={"help": "Other arguments."}
-    )
 
-    def __post_init__(self):
+    nprocs: int = 1
-        self.validate()
+    backend: str = "nccl"
+    master_addr: str = "localhost"
+    master_port: str = "29500"
+    parallel_mode: str = "none"
+    start_method: str = "spawn"
 
-    def validate(self):
+    device_type: str = "cuda"
-        for fld in fields(self):
+    val_dataset: Optional[Dataset] = None
-            if fld.metadata.get("required") and getattr(self, fld.name) is None:
+    val_split: Optional[float] = None
-                raise ValueError(f"TrainConfig.{fld.name} is required but got None.")
+    val_step: int = 1000
+    neftune_alpha: float = 0.0
+
+    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")
+    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

@@ -1,31 +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
@@ -13,296 +37,470 @@ 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.
+    Holds a :class:`Store`.  All sample-id indexing is delegated to the
-    Uses a storage abstraction for format-agnostic data loading.
+    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.store: Store = store
-        self.stride = stride
+        validate_keys(store, self.required_keys)
-        self.storage: Optional[Store] = None
 
-    @property
+    def __len__(self) -> int:
-    def required_keys(self) -> List[str]:
+        return len(self.store)
-        """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):
-        """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"),
-                          or None for auto-detection
-
-        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)
-        self._load_path = load_path
-        self.storage.load(load_path)
-        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)
 
     @property
     def keys(self) -> List[str]:
-        """Return the available data keys."""
+        return self.store.keys
-        if self.storage is None:
-            return []
-        return self.storage.keys
 
-    def get_index(self, index: int) -> tuple:
+    @property
-        """Calculate begin and end indices for a sample.
+    def token_count(self) -> int:
-
+        return self.store.token_count
-        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
 
     @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.
+    Use :meth:`DatasetFactory.register("custom")` to register new
-    All default dataset types (seq, sft, dpo, grpo) are registered automatically
+    dataset classes; they must inherit from :class:`BaseDataset`.
-    when their classes are defined with the decorator.
-
-    Example usage:
-        @DatasetFactory.register("custom")
-        class CustomDataset(BaseDataset):
-            ...
-
-        dataset = DatasetFactory.create("custom", window_size, stride)
     """
 
-    @classmethod
-    def _validate_component(cls, dataset_cls: type):
-        """Validate that the dataset class inherits from BaseDataset."""
-        if not issubclass(dataset_cls, BaseDataset):
-            raise TypeError(f"{dataset_cls.__name__} must inherit from BaseDataset")
-
-    @classmethod
-    def create(cls, train_type: str, window_size: int, stride: int) -> "BaseDataset":
-        """Create a dataset instance.
-
-        Args:
-            train_type: Type of training ("seq", "sft", "dpo", "grpo")
-            window_size: Window size for data sampling
-            stride: Stride between consecutive samples
-
-        Returns:
-            Dataset instance
-        """
-        return super().create(train_type, window_size, stride)
-
     @classmethod
     def load(
         cls,
         train_type: str,
-        load_path: str,
+        load_path: Optional[str] = None,
-        window_size: int,
+        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
+            train_type: Registered dataset name ("seq", "sft", "dpo",
-            load_path: Path to the data file
+                "grpo", …).
-            window_size: Window size for data sampling
+            load_path: Path to the data file or directory (ignored if
-            stride: Stride between consecutive samples (default: same as window_size)
+                *store* is given).
-            storage_type: Storage type ("h5", "bin") or None for auto-detection
+            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)
+        processor = cls._maybe_build_processor(
-        dataset.load(load_path, storage_type=storage_type)
+            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)
 
-    @classmethod
+        return cls.create(train_type, store=store)
-    def available_types(cls) -> list:
+
-        """Return list of registered dataset type names."""
+    @staticmethod
-        return cls.list_registered()
+    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.
 
-    def __init__(self, window_size: int, stride: int):
+    Stream mode: ``store.fetch(begin, end, "sequence")`` returns the
-        super().__init__(window_size, stride)
+    input window; the +1 shifted call returns the next-token target.
+    """
 
-    @property
+    required_keys = ["sequence"]
-    def required_keys(self) -> List[str]:
-        return ["sequence"]
 
-    def _fetch_data(self, begin_idx: int, end_idx: int) -> Tensor:
+    def __getitem__(self, index: int):
-        return self.storage.fetch(begin_idx, end_idx, "sequence")
+        begin, end = self.store.sample_window(index)
-
+        x = self.store.fetch(begin, end, "sequence")
-    def __getitem__(self, index):
+        y = self.store.fetch(begin + 1, end + 1, "sequence")
-        begin_idx, end_idx = self.get_index(index)
+        return {
-
+            "input_ids": x.to(dtype=torch.long),
-        x = self._fetch_data(begin_idx, end_idx).to(dtype=torch.long)
+            "target_ids": y.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}
 
 
 @DatasetFactory.register("sft")
 class SFTDataset(BaseDataset):
-    """Dataset for supervised fine-tuning with loss masking."""
+    """Dataset for supervised fine-tuning with loss masking.
 
-    def __init__(self, window_size: int, stride: int):
+    Stream mode: ``sequence``/``loss_mask``/``position_ids`` are sliced
-        super().__init__(window_size, stride)
+    to the window.  ``loss_mask`` and ``target_ids`` use the +1 shifted
+    slice so they align with the predicted positions.
+    """
 
-    @property
+    required_keys = ["sequence", "loss_mask", "position_ids"]
-    def required_keys(self) -> List[str]:
-        return ["sequence", "loss_mask"]
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
+    def __getitem__(self, index: int):
-        return self.storage.fetch(begin_idx, end_idx, key)
+        begin, end = self.store.sample_window(index)
-
+        x = self.store.fetch(begin, end, "sequence")
-    def __getitem__(self, index):
+        y = self.store.fetch(begin + 1, end + 1, "sequence")
-        begin_idx, end_idx = self.get_index(index)
+        position_ids = self.store.fetch(begin, end, "position_ids")
-
+        loss_mask = self.store.fetch(begin + 1, end + 1, "loss_mask")
-        x = self._fetch_data(begin_idx, end_idx, "sequence").to(dtype=torch.long)
+        return {
-        y = self._fetch_data(begin_idx + 1, end_idx + 1, "sequence").to(
+            "input_ids": x.to(dtype=torch.long),
-            dtype=torch.long
+            "target_ids": y.to(dtype=torch.long),
-        )
+            "position_ids": position_ids.to(dtype=torch.long),
-        loss_mask = self._fetch_data(begin_idx + 1, end_idx + 1, "loss_mask").to(
+            "loss_mask": loss_mask.to(dtype=torch.bool),
-            dtype=torch.bool
+        }
-        )
-
-        return {"input_ids": x, "target_ids": y, "loss_mask": loss_mask}
 
 
 @DatasetFactory.register("dpo")
 class DPODataset(BaseDataset):
-    """Dataset for Direct Preference Optimization training."""
+    """Record-structured dataset for Direct Preference Optimization.
 
-    def __init__(self, window_size: int, stride: int):
+    Each sample is one preference pair (chosen + rejected) and is an
-        super().__init__(window_size, stride)
+    independent training unit — no windowing, stride, or cross-record
+    concatenation.  This keeps each sequence self-contained so attention
+    never leaks across preference pairs.
 
-    @property
+    Two loading paths (handled by :class:`DatasetFactory`):
-    def required_keys(self) -> List[str]:
-        return ["chosen", "rejected", "chosen_mask", "rejected_mask"]
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
+    - **Pre-tokenized** (bin): ``store.load(path)`` reads per-record
-        return self.storage.fetch(begin_idx, end_idx, key)
+      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``.
+    """
 
-    def __getitem__(self, index: int):
+    required_keys = ["chosen", "rejected", "chosen_mask", "rejected_mask"]
-        begin_idx, end_idx = self.get_index(index)
 
-        chosen = self._fetch_data(begin_idx, end_idx, "chosen").to(dtype=torch.long)
+    def make_processor(self, tokenizer, max_len: int):
-        rejected = self._fetch_data(begin_idx, end_idx, "rejected").to(dtype=torch.long)
+        return partial(dpo_processor, tokenizer=tokenizer, max_len=max_len)
-        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
-        )
 
+    def __getitem__(self, index: int) -> Dict[str, Tensor]:
         return {
-            "chosen": chosen,
+            "chosen": self.store.fetch_record(index, "chosen").to(dtype=torch.long),
-            "rejected": rejected,
+            "rejected": self.store.fetch_record(index, "rejected").to(dtype=torch.long),
-            "chosen_mask": chosen_mask,
+            "chosen_mask": self.store.fetch_record(index, "chosen_mask").to(
-            "rejected_mask": rejected_mask,
+                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.
 
-    def __init__(self, window_size: int, stride: int):
+    Each sample is one prompt with its group of responses and scalar
-        super().__init__(window_size, stride)
+    rewards — an independent training unit with no windowing or stride.
 
-    @property
+    Expected storage layout (produced by JsonlStore or pre-tokenized):
-    def required_keys(self) -> List[str]:
-        return ["prompts", "responses", "masks", "rewards"]
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
+    - ``prompts``:   List[Tensor]  — one 1-D token tensor per record
-        return self.storage.fetch(begin_idx, end_idx, key)
+    - ``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.store.fetch_record(index, "prompts")
-
+        responses = self.store.fetch_record(index, "responses")
-        prompts = self._fetch_data(begin_idx, end_idx, "prompts").to(dtype=torch.long)
+        masks = self.store.fetch_record(index, "masks")
-        responses = self._fetch_data(begin_idx, end_idx, "responses").to(
+        rewards = self.store.fetch_record(index, "rewards")
-            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")
-
         return {
-            "prompts": prompts,
+            "prompts": prompts.to(dtype=torch.long),
-            "responses": responses,
+            "responses": [r.to(dtype=torch.long) for r in responses],
-            "masks": masks,
+            "masks": [m.to(dtype=torch.bool) for m in masks],
-            "rewards": rewards,
+            "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,
@@ -74,6 +82,7 @@ class ResumableDistributedSampler(Sampler[int]):
 
         self.epoch += 1
         self._indices = None
+        self.iter = self.iter % self.num_samples_per_replica
 
     @property
     def _remaining(self):

astrai/dataset/storage.py

@@ -1,97 +1,69 @@
 """Storage backends for different data formats.
 
-Layers:
+Architecture (composition over inheritance):
-  - 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)
 
-Key properties:
+    Store (ABC)               — owns _data/_cum/_offsets bookkeeping
-  - Multi-segment:   segments kept as-is, no forced concatenation — safe for
+                                  + window_size/stride for sample-id
-                      datasets larger than RAM
+                                  indexing.  __getitem__/__len__ produce
-  - Explicit length: _length = min(total elements across keys), set at load,
+                                  the smallest iterable unit so Dataset
-                      __len__ returns O(1)
+                                  classes are pure delegators.
-  - Zero-copy mmap:  MmapStore wraps np.memmap(mode="r"), all DataLoader
+    Streamable (mixin)        — raw token slice  fetch(begin, end, keys)
-                      workers share OS page-cache pages
+    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
+import glob
 import json
-import os
+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 h5py
-import numpy as np
 import torch
 from torch import Tensor
 
+from astrai.config.preprocess_config import PipelineConfig
 from astrai.factory import BaseFactory
+from astrai.preprocessing.transform import TokenizeTransform
+from astrai.serialization import (
+    load_bin,
+    load_bin_offsets,
+)
 
-
+logger = logging.getLogger(__name__)
-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]]):
-    os.makedirs(file_path, exist_ok=True)
-    meta = {}
-    for key, tensors in tensor_group.items():
-        cat = torch.cat(tensors, dim=0)
-        meta[key] = {"shape": list(cat.shape), "dtype": str(cat.dtype).split(".")[-1]}
-        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)
-
-
-def load_bin(file_path: str) -> Dict[str, List[Tensor]]:
-    with open(os.path.join(file_path, "meta.json"), "r") as f:
-        meta = json.load(f)
-    segments: Dict[str, List[Tensor]] = {}
-    for key, info in meta.items():
-        arr = np.memmap(
-            os.path.join(file_path, f"{key}.bin"),
-            dtype=info["dtype"],
-            mode="r+",
-            shape=tuple(info["shape"]),
-        )
-        segments[key] = [torch.from_numpy(arr)]
-    return segments
 
 
 def detect_format(load_path: str) -> str:
@@ -101,7 +73,7 @@ def detect_format(load_path: str) -> str:
         load_path: Directory or file path
 
     Returns:
-        Format string ("h5" or "bin")
+        Format string ("h5", "bin", "jsonl", or "processed")
 
     Raises:
         FileNotFoundError: If no supported data files are found
@@ -109,142 +81,546 @@ def detect_format(load_path: str) -> str:
     root = Path(load_path)
     if root.is_file():
         suffix = root.suffix.lower()
-        if suffix in (".h5", ".hdf5"):
+        if suffix == ".jsonl":
-            return "h5"
+            return "jsonl"
         raise ValueError(f"Unsupported file format: {suffix}")
 
-    h5_files = list(root.rglob("*.h5")) + list(root.rglob("*.hdf5"))
+    bin_files = [Path(p) for p in glob.glob(str(root / "**" / "*.bin"), recursive=True)]
-    if h5_files:
+    if bin_files:
-        return "h5"
+        has_meta = (root / "meta.json").exists() or len(
-    bin_files = list(root.rglob("*.bin"))
+            [Path(p) for p in glob.glob(str(root / "**" / "meta.json"), recursive=True)]
-    if bin_files and (root / "meta.json").exists():
+        ) > 0
-        return "bin"
+        if has_meta:
+            return "bin"
+    jsonl_files = [
+        Path(p) for p in glob.glob(str(root / "**" / "*.jsonl"), recursive=True)
+    ]
+    if jsonl_files:
+        return "jsonl"
     raise FileNotFoundError(f"No supported data files found at {load_path}")
 
 
 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).
+    A Store owns both its data layout AND its sample-id → token/record
-    ``len(store)`` returns ``self._length`` (explicit, O(1)), the minimum
+    index translation.  Datasets are thin wrappers that bind a Store
-    total element count across all keys.
+    to a particular train-type's key mapping; they never know about
+    window/stride math.
 
-    Subclasses fill ``self._data`` and ``self._cum`` during ``load()``
+    Two iteration modes:
-    via ``_normalize()``.
+
+    - **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:
+        return _stream_fetch(self, begin, end, keys)
-            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}
 
-    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 = []
+def _stream_fetch(self, begin: int, end: int, keys: Union[str, List[str]]):
-        for i in range(seg_start, seg_end + 1):
+    if not getattr(self, "_data", None):
-            prev = cum[i - 1] if i > 0 else 0
+        raise RuntimeError("Store not loaded")
-            s = max(begin - prev, 0)
+    if not (0 <= begin < self._length and 0 <= end <= self._length):
-            e = min(end - prev, segments[i].shape[0])
+        raise ValueError(
-            results.append(segments[i][s:e])
+            f"Index out of bounds: begin={begin}, end={end}, length={self._length}"
-
-        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
         )
+    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.
+    """Factory for creating Store instances by type name."""
-
-    Example::
-
-        @StoreFactory.register("custom")
-        class CustomStore(Store):
-            ...
-    """
-
-    @classmethod
-    def _validate_component(cls, store_cls: type):
-        if not issubclass(store_cls, Store):
-            raise TypeError(f"{store_cls.__name__} must inherit from Store")
-
-
-@StoreFactory.register("h5")
-class H5Store(Store):
-    """HDF5-based storage backend (pre-tokenized data)."""
-
-    def load(self, path: str):
-        self._normalize(load_h5(path))
 
 
 @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/
+    - **Stream**: always available via :meth:`fetch`.
-          meta.json          # {key: {shape, dtype}, ...}
+    - **Record** (``fetch_record(i, key)``): only when ``meta.json``
-          <key>.bin          # raw numpy array, one per key
+      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 = []
-        raw = load_bin(path)
+        root = Path(path)
-        self._normalize(raw)
+        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, offsets=all_offsets or None)
         for tensors in self._data.values():
             self._mmap_refs.extend(tensors)
+
+
+class JsonlSource:
+    """Read raw JSON records from a ``.jsonl`` file or directory.
+
+    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
+
+    _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)
+
+    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 +1,49 @@
+"""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,
+    CudaBackend,
+    TorchNativeBackend,
+    attention,
+    attn_backend,
+    get_backend,
+)
+from astrai.extension.attention_ops import (
+    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",
+    "CudaBackend",
+    "TorchNativeBackend",
+    "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 +1,423 @@
+"""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 ``TorchNativeBackend`` singleton.
+
+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 math
+from abc import ABC, abstractmethod
+from contextlib import contextmanager
+from typing import Optional, Union
+
+import torch
+import torch.nn.functional as F
+from torch import Tensor
+
+from astrai.extension.attention_ops import attn_paged_decode, attn_prefill
+from astrai.extension.loader import is_available
+from astrai.inference.core.cache import KVCache
+
+_current_backend: contextvars.ContextVar["AttentionBackend"] = contextvars.ContextVar(
+    "attn_backend"
+)
+
+
+class ATTN_BACKEND(enum.Enum):
+    """Backend selector enum, mirroring ``torch.nn.attention.SDPBackend``."""
+
+    TORCH_NATIVE = "torch_native"
+    CUDA = "cuda"
+
+
+def get_backend() -> "AttentionBackend":
+    """Return the active backend for the current thread/context.
+
+    Falls back to a ``TorchNativeBackend`` singleton when no backend
+    has been activated via ``with``.
+    """
+    try:
+        return _current_backend.get()
+    except LookupError:
+        return _default_backend
+
+
+@contextmanager
+def attn_backend(backend: Union[ATTN_BACKEND, "AttentionBackend", type]):
+    """Context manager to select an attention backend.
+
+    Mirrors ``torch.nn.attention.sdpa_kernel``. Accepts an
+    ``ATTN_BACKEND`` enum value, a backend class, or a backend instance.
+
+    Examples::
+
+        with attn_backend(ATTN_BACKEND.TORCH_NATIVE):
+            ...
+        with attn_backend(TorchNativeBackend):
+            ...
+        with attn_backend(TorchNativeBackend()):
+            ...
+    """
+    if isinstance(backend, ATTN_BACKEND):
+        instance = _BACKEND_REGISTRY[backend]()
+    elif isinstance(backend, type) and issubclass(backend, AttentionBackend):
+        instance = backend()
+    elif isinstance(backend, AttentionBackend):
+        instance = backend
+    else:
+        raise TypeError(
+            f"expected ATTN_BACKEND, AttentionBackend type, 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 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()
+    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."""
+
+
+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.
+    """
+
+    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:
+            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
+            if kv_cache.page_table is not None:
+                indices = kv_cache.page_table
+            else:
+                indices = kv_cache.req_to_token[kv_cache.req_pool_indices, :max_len]
+            if kv_cache.decode_mask is not None:
+                pos_mask = kv_cache.decode_mask
+            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))
+            k = kv_cache.k_buffer[layer_id, indices]
+            v = kv_cache.v_buffer[layer_id, indices]
+
+        n_rep = q.size(2) // k.size(2)
+        if n_rep > 1:
+            k = repeat_kv(k, n_rep)
+            v = repeat_kv(v, n_rep)
+
+        q = q.permute(0, 2, 1, 3)
+        k = k.permute(0, 2, 1, 3)
+        v = v.permute(0, 2, 1, 3)
+
+        out = F.scaled_dot_product_attention(q, k, v, attn_mask, is_causal=is_causal)
+        out = out.permute(0, 2, 1, 3).contiguous().flatten(2)
+        return out
+
+
+_default_backend = TorchNativeBackend()
+
+
+class CudaBackend(AttentionBackend):
+    """CUDA kernel backend with direct KV cache access.
+
+    Decode path: writes K/V to cache, then calls ``attn_paged_decode``
+    with ``page_size=1`` (each token slot is a single-token "page").
+    The ``req_to_token`` table serves directly as the page table.
+
+    Prefill path: writes K/V to cache, gathers full-sequence K/V via
+    indirect indexing (same as TorchNativeBackend), then calls
+    ``attn_prefill``.
+
+    Training path (``kv_cache is None``): calls ``attn_prefill`` directly
+    on the projected q/k/v.
+
+    Falls back to ``TorchNativeBackend`` for any path where the
+    corresponding CUDA kernel is not available.
+    """
+
+    def __init__(self):
+        self._fallback = TorchNativeBackend()
+
+    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 or not is_available("attn_paged_decode"):
+            return self._fallback.fwd_decode(
+                q, k, v, kv_cache, layer_id, attn_mask, is_causal
+            )
+
+        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
+
+        if kv_cache.page_table is not None:
+            page_table = kv_cache.page_table
+        else:
+            page_table = kv_cache.req_to_token[kv_cache.req_pool_indices, :max_len]
+
+        k_cache = kv_cache.k_buffer[layer_id].unsqueeze(1)
+        v_cache = kv_cache.v_buffer[layer_id].unsqueeze(1)
+
+        if q.size(0) == 1:
+            mask = None
+        elif kv_cache.decode_mask is not None:
+            mask = kv_cache.decode_mask
+        else:
+            mask = (
+                torch.arange(max_len, device=q.device)[None, :]
+                < kv_cache.seq_lens[:, None]
+            )
+
+        out = attn_paged_decode(
+            q,
+            page_table,
+            k_cache,
+            v_cache,
+            page_size=1,
+            kv_len=max_len,
+            mask=mask,
+            is_causal=is_causal,
+        )
+
+        out = out.flatten(2)
+        return out
+
+    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:
+            if is_available("attn_prefill"):
+                out = attn_prefill(q, k, v, mask=attn_mask, is_causal=is_causal)
+                return out.flatten(2)
+            return self._fallback.fwd_prefill(
+                q, k, v, kv_cache, layer_id, attn_mask, is_causal
+            )
+
+        if not is_available("attn_prefill"):
+            return self._fallback.fwd_prefill(
+                q, k, v, kv_cache, layer_id, attn_mask, is_causal
+            )
+
+        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]
+        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))
+        k_full = kv_cache.k_buffer[layer_id, indices]
+        v_full = kv_cache.v_buffer[layer_id, indices]
+
+        out = attn_prefill(q, k_full, v_full, mask=attn_mask, is_causal=is_causal)
+        return out.flatten(2)
+
+
+_BACKEND_REGISTRY: dict[ATTN_BACKEND, type[AttentionBackend]] = {
+    ATTN_BACKEND.TORCH_NATIVE: TorchNativeBackend,
+    ATTN_BACKEND.CUDA: CudaBackend,
+}

astrai/extension/attention_ops.py

@@ -0,0 +1,117 @@
+"""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 torch
+
+from astrai.extension.loader import _available, _modules
+
+
+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: torch.Tensor | None = 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=1
+    )
+
+
+def attn_prefill(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    mask: torch.Tensor | None = 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=1
+    )
+
+
+def attn_paged_decode(
+    q: torch.Tensor,
+    page_table: torch.Tensor,
+    k_cache: torch.Tensor,
+    v_cache: torch.Tensor,
+    page_size: int,
+    kv_len: int,
+    mask: torch.Tensor | None = None,
+    is_causal: bool = False,
+) -> torch.Tensor:
+    """Paged GQA decode attention (q_len == 1, direct page-table access).
+
+    Args:
+        q: [batch, 1, n_heads, head_dim] (blhd, bf16)
+        page_table: [batch, max_pages] (int64)
+        k_cache: [n_pages, page_size, n_kv_heads, head_dim] (bf16)
+        v_cache: same as k_cache
+        page_size: tokens per page
+        kv_len: actual sequence length per request
+        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_paged_decode")
+    causal_offset = (kv_len - 1) if is_causal else -1
+    return _modules["attn_paged_decode"].attn_paged_decode(
+        q,
+        page_table,
+        k_cache,
+        v_cache,
+        page_size,
+        kv_len,
+        mask=mask,
+        causal_offset=causal_offset,
+        layout=1,
+    )

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 +1,36 @@
+"""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", "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

@@ -1,149 +1,124 @@
-"""Base factory class for extensible component registration."""
+"""Base factory with decorator-based registration and kwarg-filtered instantiation."""
 
 import inspect
+import sys
 from abc import ABC
-from typing import Callable, Dict, Generic, List, Optional, Tuple, Type, TypeVar
+from typing import (
+    Callable,
+    Dict,
+    ForwardRef,
+    Generic,
+    List,
+    Optional,
+    Type,
+    TypeVar,
+    Union,
+    get_args,
+    get_origin,
+)
 
 T = TypeVar("T")
 
 
-class Registry:
+def _resolve_base_type(
-    """Flexible registry for component classes with category and priority support.
+    arg: Union[Type, str, ForwardRef], factory_cls: type
+) -> Optional[Type]:
+    """Resolve the generic type-arg T to a concrete class.
 
-    This registry stores component classes with optional metadata (category, priority).
+    - Concrete class (``BaseFactory[MyBase]``): returned directly.
-    It provides methods for registration, retrieval, and listing with filtering.
+    - 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
 
-    def __init__(self):
+    if isinstance(arg, str):
-        self._entries = {}  # name -> (component_cls, category, priority)
+        name = arg
+    elif isinstance(arg, ForwardRef):
+        name = arg.__forward_arg__
+    else:
+        return None
 
-    def register(
+    mod = sys.modules.get(factory_cls.__module__)
-        self,
+    if mod is None:
-        name: str,
+        return None
-        component_cls: Type,
+    try:
-        category: Optional[str] = None,
+        return eval(name, vars(mod))  # noqa: S307
-        priority: int = 0,
+    except NameError:
-    ):
+        return None
-        """Register a component class with optional category and priority."""
-        if name in self._entries:
-            raise ValueError(f"Component '{name}' is already registered")
-        self._entries[name] = (component_cls, category, priority)
 
-    def get(self, name: str) -> Type:
-        """Get component class by name."""
-        if name not in self._entries:
-            raise KeyError(f"Component '{name}' not found in registry")
-        return self._entries[name][0]
 
-    def get_with_metadata(self, name: str) -> Tuple[Type, Optional[str], int]:
+def _validate_component(component_cls: Type, base: Optional[Type]) -> None:
-        """Get component class with its metadata."""
+    """Validate that *component_cls* inherits from *base*.
-        entry = self._entries.get(name)
-        if entry is None:
-            raise KeyError(f"Component '{name}' not found in registry")
-        return entry
 
-    def contains(self, name: str) -> bool:
+    No-op when *base* is ``None`` (e.g. forward-ref resolution failed).
-        """Check if a name is registered."""
+    """
-        return name in self._entries
+    if base is not None and not issubclass(component_cls, base):
-
+        raise TypeError(f"{component_cls.__name__} must inherit from {base.__name__}")
-    def list_names(self) -> List[str]:
-        """Return list of registered component names."""
-        return sorted(self._entries.keys())
-
-    def list_by_category(self, category: str) -> List[str]:
-        """Return names of components belonging to a specific category."""
-        return sorted(
-            name for name, (_, cat, _) in self._entries.items() if cat == category
-        )
-
-    def list_by_priority(self, reverse: bool = False) -> List[str]:
-        """Return names sorted by priority (default ascending)."""
-        return sorted(
-            self._entries.keys(),
-            key=lambda name: self._entries[name][2],
-            reverse=reverse,
-        )
-
-    def entries(self) -> Dict[str, Tuple[Type, Optional[str], int]]:
-        """Return raw entries dictionary."""
-        return self._entries.copy()
 
 
 class BaseFactory(ABC, Generic[T]):
-    """Generic factory class for component registration and creation.
+    """Generic factory with decorator-based registration.
 
-    This base class provides a decorator-based registration pattern
+    Create a factory by subclassing with the desired base type::
-    for creating extensible component factories.
 
-    Example usage:
+        class MyFactory(BaseFactory[MyBase]):
-        class MyFactory(BaseFactory[MyBaseClass]):
             pass
 
+    Register components with the ``register`` decorator::
+
         @MyFactory.register("custom")
-        class CustomComponent(MyBaseClass):
+        class CustomComponent(MyBase):
             ...
 
-        component = MyFactory.create("custom", *args, **kwargs)
+        obj = MyFactory.create("custom", *args, **kwargs)
+
+    ``create()`` filters kwargs to match the component's ``__init__``
+    signature so components don't need ``**kwargs`` just to absorb
+    unrelated parameters.
     """
 
-    _registry: Registry
+    _entries: Dict[str, Type[T]]
 
     def __init_subclass__(cls, **kwargs):
         super().__init_subclass__(**kwargs)
-        cls._registry = Registry()
+        for orig_base in getattr(cls, "__orig_bases__", ()):
+            if get_origin(orig_base) is BaseFactory:
+                (arg,) = get_args(orig_base)
+                cls._entries = {}
+                cls._component_base = _resolve_base_type(arg, cls)
+                return
 
     @classmethod
-    def register(
+    def register(cls, name: str) -> Callable[[Type[T]], Type[T]]:
-        cls, name: str, category: Optional[str] = None, priority: int = 0
+        """Decorator to register a component class.
-    ) -> Callable[[Type[T]], Type[T]]:
-        """Decorator to register a component class with optional category and priority.
 
-        Args:
+        Validates that the decorated class inherits from the generic
-            name: Registration name for the component
+        type parameter ``T`` declared on the factory.
-            category: Optional category for grouping components
-            priority: Priority for ordering (default 0)
-
-        Returns:
-            Decorator function that registers the component class
-
-        Raises:
-            TypeError: If the decorated class doesn't inherit from the base type
         """
 
         def decorator(component_cls: Type[T]) -> Type[T]:
-            cls._validate_component(component_cls)
+            _validate_component(component_cls, cls._component_base)
-            cls._registry.register(
+            if name in cls._entries:
-                name, component_cls, category=category, priority=priority
+                raise ValueError(f"Component '{name}' is already registered")
-            )
+            cls._entries[name] = component_cls
             return component_cls
 
         return decorator
 
     @classmethod
     def create(cls, name: str, *args, **kwargs) -> T:
-        """Create a component instance by name.
+        """Create a component instance by name, filtering kwargs to match
-
+        the component's ``__init__`` signature.
-        Filters kwargs to match the component's __init__ signature,
-        so components don't need to declare **kwargs just to absorb
-        parameters meant for other components.
-
-        Args:
-            name: Registered name of the component
-            *args: Positional arguments passed to component constructor
-            **kwargs: Keyword arguments passed to component constructor
-
-        Returns:
-            Component instance
-
-        Raises:
-            ValueError: If the component name is not registered
         """
-        if not cls._registry.contains(name):
+        component_cls = cls._entries.get(name)
+        if component_cls is None:
             raise ValueError(
-                f"Unknown component: '{name}'. "
+                f"Unknown component: '{name}'. Supported types: {sorted(cls._entries)}"
-                f"Supported types: {sorted(cls._registry.list_names())}"
             )
-        component_cls = cls._registry.get(name)
         sig = inspect.signature(component_cls.__init__)
         has_var_kwargs = any(
             p.kind == inspect.Parameter.VAR_KEYWORD for p in sig.parameters.values()
@@ -157,70 +132,22 @@ 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 that the component class is valid for this factory.
-
-        Override this method in subclasses to add custom validation.
-
-        Args:
-            component_cls: Component class to validate
-
-        Raises:
-            TypeError: If the component class is invalid
-        """
-        pass
-
     @classmethod
     def get_component_class(cls, name: str) -> Type[T]:
-        """Get the registered component class by name without instantiating it.
+        """Get the registered component class without instantiating it."""
-
+        entry = cls._entries.get(name)
-        Args:
+        if entry is None:
-            name: Registered name of the component
-
-        Returns:
-            The component class itself
-
-        Raises:
-            ValueError: If the component name is not registered
-        """
-        if not cls._registry.contains(name):
             raise ValueError(
-                f"Unknown component: '{name}'. "
+                f"Unknown component: '{name}'. Supported types: {sorted(cls._entries)}"
-                f"Supported types: {sorted(cls._registry.list_names())}"
             )
-        return cls._registry.get(name)
+        return entry
 
     @classmethod
-    def list_registered(cls) -> list:
+    def list_registered(cls) -> List[str]:
-        """List all registered component names.
+        """List all registered component names."""
-
+        return sorted(cls._entries)
-        Returns:
-            List of registered component names
-        """
-        return cls._registry.list_names()
 
     @classmethod
     def is_registered(cls, name: str) -> bool:
-        """Check if a component name is registered.
+        """Check if a component name is registered."""
-
+        return name in cls._entries
-        Args:
-            name: Component name to check
-
-        Returns:
-            True if registered, False otherwise
-        """
-        return cls._registry.contains(name)
-
-    @classmethod
-    def list_by_category(cls, category: str) -> List[str]:
-        """List registered component names in a category."""
-        return cls._registry.list_by_category(category)
-
-    @classmethod
-    def list_by_priority(cls, reverse: bool = False) -> List[str]:
-        """List registered component names sorted by priority."""
-        return cls._registry.list_by_priority(reverse)
-
-
-__all__ = ["Registry", "BaseFactory"]

astrai/inference/__init__.py

@@ -6,18 +6,23 @@ Layers:
   - protocols/:   Response builders (OpenAI, Anthropic)
   - transport/:   SSE transport utilities
   - engine.py:    Facade (InferenceEngine), Value Object (GenerationRequest)
-  - sample.py:    Strategy pattern (TemperatureStrategy, TopKStrategy, TopPStrategy)
+  - sample.py:    Strategy pattern (TemperatureStrategy, TopKStrategy, TopPStrategy, FrequencyPenaltyStrategy)
 """
 
 from astrai.inference.api import (
     AnthropicMessage,
+    BaseToolParser,
     ChatCompletionRequest,
     ChatMessage,
+    FunctionDef,
     GenContext,
     MessagesRequest,
     ProtocolHandler,
+    SimpleJsonToolParser,
     StopChecker,
-    app,
+    ToolDef,
+    ToolParserFactory,
+    get_app,
     run_server,
 )
 from astrai.inference.api.anthropic import AnthropicResponseBuilder
@@ -28,19 +33,19 @@ from astrai.inference.core import (
     Executor,
     InferenceScheduler,
     KVCache,
-    KvcacheView,
+    KVStorage,
     PagePool,
     PrefixCache,
-    Storage,
+    ReqToTokenPool,
     Task,
     TaskManager,
     TaskStatus,
-    TaskTable,
     page_hash,
 )
 from astrai.inference.engine import GenerationRequest, InferenceEngine
 from astrai.inference.sample import (
     BaseSamplingStrategy,
+    FrequencyPenaltyStrategy,
     SamplingPipeline,
     TemperatureStrategy,
     TopKStrategy,
@@ -59,27 +64,32 @@ __all__ = [
     "TaskStatus",
     "Allocator",
     "KVCache",
-    "KvcacheView",
+    "KVStorage",
     "PagePool",
     "PrefixCache",
-    "Storage",
+    "ReqToTokenPool",
-    "TaskTable",
     "page_hash",
     "sample",
     "BaseSamplingStrategy",
     "TemperatureStrategy",
     "TopKStrategy",
     "TopPStrategy",
+    "FrequencyPenaltyStrategy",
     "SamplingPipeline",
     "ProtocolHandler",
     "StopChecker",
     "GenContext",
+    "BaseToolParser",
+    "SimpleJsonToolParser",
+    "ToolParserFactory",
     "OpenAIResponseBuilder",
     "AnthropicResponseBuilder",
     "ChatMessage",
     "ChatCompletionRequest",
+    "FunctionDef",
+    "ToolDef",
     "AnthropicMessage",
     "MessagesRequest",
-    "app",
+    "get_app",
     "run_server",
 ]

astrai/inference/api/__init__.py

@@ -1,23 +1,39 @@
-"""Inference API: protocol handler, stop checker, and FastAPI server."""
+"""Inference API: protocol handler, stop checker, tool parsers, and FastAPI server.
+
+``app`` is no longer a module-level global. Use :func:`get_app` to access the
+lazy singleton FastAPI instance.
+"""
 
 from astrai.inference.api.protocol import GenContext, ProtocolHandler, StopChecker
 from astrai.inference.api.server import (
     AnthropicMessage,
     ChatCompletionRequest,
     ChatMessage,
+    FunctionDef,
     MessagesRequest,
-    app,
+    ToolDef,
+    get_app,
     run_server,
 )
+from astrai.inference.api.tool_parser import (
+    BaseToolParser,
+    SimpleJsonToolParser,
+    ToolParserFactory,
+)
 
 __all__ = [
     "ProtocolHandler",
     "StopChecker",
     "GenContext",
+    "BaseToolParser",
+    "SimpleJsonToolParser",
+    "ToolParserFactory",
     "AnthropicMessage",
     "ChatCompletionRequest",
     "ChatMessage",
+    "FunctionDef",
+    "ToolDef",
     "MessagesRequest",
-    "app",
+    "get_app",
     "run_server",
 ]

astrai/inference/api/anthropic.py

@@ -1,5 +1,6 @@
 """Anthropic message completion response builder."""
 
+import time
 import uuid
 from typing import Any, Dict, List, Tuple, Union
 
@@ -39,9 +40,8 @@ class AnthropicResponseBuilder(ResponseBuilder):
         prompt = engine.tokenizer.apply_chat_template(messages, tokenize=False)
         ctx = GenContext(
             resp_id=f"msg_{uuid.uuid4().hex[:24]}",
-            created=0,
+            created=int(time.time()),
             model=request.model,
-            prompt_tokens=0,
         )
         stop_sequences = getattr(request, "stop_sequences", None) or []
         return prompt, ctx, stop_sequences
@@ -72,15 +72,17 @@ class AnthropicResponseBuilder(ResponseBuilder):
             ),
         ]
 
-    def format_chunk(self, token: str) -> str:
+    def format_chunk(self, token: str, **kwargs) -> List[str]:
-        return sse_event(
+        return [
-            {
+            sse_event(
-                "type": "content_block_delta",
+                {
-                "index": 0,
+                    "type": "content_block_delta",
-                "delta": {"type": "text_delta", "text": token},
+                    "index": 0,
-            },
+                    "delta": {"type": "text_delta", "text": token},
-            event="content_block_delta",
+                },
-        )
+                event="content_block_delta",
+            )
+        ]
 
     def format_stream_end(self, ctx: GenContext, stop: StopInfo) -> List[str]:
         events: List[str] = []

astrai/inference/api/openai.py

@@ -1,7 +1,9 @@
 """OpenAI chat completion response builder."""
 
+import logging
+import time
 import uuid
-from typing import Any, Dict, List, Tuple
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 from pydantic import BaseModel
 
@@ -11,24 +13,78 @@ from astrai.inference.api.protocol import (
     StopInfo,
     sse_event,
 )
+from astrai.inference.api.tool_parser import BaseToolParser, ToolParserFactory
 from astrai.inference.engine import InferenceEngine
 
+logger = logging.getLogger(__name__)
+
+_UNSUPPORTED_PARAMS = (
+    "n",
+    "presence_penalty",
+    "logit_bias",
+    "user",
+)
+
+
+def _resolve_tool_choice(
+    request: BaseModel,
+) -> Union[str, Dict[str, Any]]:
+    tc = getattr(request, "tool_choice", None)
+    if tc is None:
+        return "auto"
+    if isinstance(tc, str):
+        return tc
+    if isinstance(tc, dict):
+        return tc
+    return "auto"
+
+
+def _resolve_tools(request: BaseModel) -> Optional[List[Dict[str, Any]]]:
+    raw = getattr(request, "tools", None)
+    if not raw:
+        return None
+    if isinstance(raw, list):
+        return [t.model_dump() if hasattr(t, "model_dump") else t for t in raw]
+    return None
+
 
 class OpenAIResponseBuilder(ResponseBuilder):
     def prepare(
         self, request: BaseModel, engine: InferenceEngine
     ) -> Tuple[str, GenContext, List[str]]:
         messages = [{"role": m.role, "content": m.content} for m in request.messages]
-        prompt = engine.tokenizer.apply_chat_template(messages, tokenize=False)
+        tools = _resolve_tools(request)
+        prompt = engine.tokenizer.apply_chat_template(
+            messages, tokenize=False, tools=tools or []
+        )
 
         self._resp_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
         self._model = request.model
 
+        for param in _UNSUPPORTED_PARAMS:
+            value = getattr(request, param, None)
+            fields = getattr(type(request), "model_fields", {})
+            default = fields[param].default if param in fields else None
+            if value is not None and value != default:
+                logger.warning(
+                    "ChatCompletionRequest param '%s'=%r is not supported"
+                    " and will be ignored",
+                    param,
+                    value,
+                )
+
+        self._parser: Optional[BaseToolParser] = None
+        if tools:
+            tool_choice = _resolve_tool_choice(request)
+            self._parser = ToolParserFactory.create(
+                "simple_json", tools=tools, tool_choice=tool_choice
+            )
+        self._content_started = False
+
         ctx = GenContext(
             resp_id=self._resp_id,
-            created=0,
+            created=int(time.time()),
             model=self._model,
-            prompt_tokens=0,
         )
         stop = request.stop
         stop_sequences = (
@@ -55,7 +111,82 @@ class OpenAIResponseBuilder(ResponseBuilder):
             )
         ]
 
-    def format_chunk(self, token: str) -> str:
+    def format_chunk(self, token: str, **kwargs) -> List[str]:
+        body = kwargs.get("body", "")
+        if self._parser is not None:
+            return self._format_tool_chunk(body, **kwargs)
+
+        return [
+            sse_event(
+                {
+                    "id": self._resp_id,
+                    "object": "chat.completion.chunk",
+                    "created": 0,
+                    "model": self._model,
+                    "choices": [
+                        {
+                            "index": 0,
+                            "delta": {"content": token},
+                            "finish_reason": None,
+                        }
+                    ],
+                }
+            )
+        ]
+
+    def _format_tool_chunk(self, body: str, **kwargs) -> List[str]:
+        deltas = self._parser.feed(
+            body,
+            current_token_ids=kwargs.get("current_token_ids"),
+            delta_token_ids=kwargs.get("delta_token_ids"),
+        )
+        events: List[str] = []
+        for d in deltas:
+            if "content" in d:
+                if not self._content_started:
+                    events.append(self._role_chunk())
+                    self._content_started = True
+                events.append(
+                    sse_event(
+                        {
+                            "id": self._resp_id,
+                            "object": "chat.completion.chunk",
+                            "created": 0,
+                            "model": self._model,
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "delta": {"content": d["content"]},
+                                    "finish_reason": None,
+                                }
+                            ],
+                        }
+                    )
+                )
+            elif "tool_calls" in d:
+                if not self._content_started:
+                    events.append(self._role_chunk())
+                    self._content_started = True
+                events.append(
+                    sse_event(
+                        {
+                            "id": self._resp_id,
+                            "object": "chat.completion.chunk",
+                            "created": 0,
+                            "model": self._model,
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "delta": {"tool_calls": d["tool_calls"]},
+                                    "finish_reason": None,
+                                }
+                            ],
+                        }
+                    )
+                )
+        return events
+
+    def _role_chunk(self) -> str:
         return sse_event(
             {
                 "id": self._resp_id,
@@ -63,12 +194,19 @@ class OpenAIResponseBuilder(ResponseBuilder):
                 "created": 0,
                 "model": self._model,
                 "choices": [
-                    {"index": 0, "delta": {"content": token}, "finish_reason": None}
+                    {
+                        "index": 0,
+                        "delta": {"role": "assistant"},
+                        "finish_reason": None,
+                    }
                 ],
             }
         )
 
     def format_stream_end(self, ctx: GenContext, stop: StopInfo) -> List[str]:
+        finish_reason = "stop"
+        if self._parser is not None and self._parser.has_tool_calls:
+            finish_reason = "tool_calls"
         return [
             sse_event(
                 {
@@ -76,7 +214,9 @@ class OpenAIResponseBuilder(ResponseBuilder):
                     "object": "chat.completion.chunk",
                     "created": ctx.created,
                     "model": self._model,
-                    "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}],
+                    "choices": [
+                        {"index": 0, "delta": {}, "finish_reason": finish_reason}
+                    ],
                 }
             ),
             sse_event(
@@ -91,6 +231,32 @@ class OpenAIResponseBuilder(ResponseBuilder):
     def format_response(
         self, ctx: GenContext, content: str, stop: StopInfo
     ) -> Dict[str, Any]:
+        if self._parser is not None:
+            parsed = self._parser.parse_complete(content)
+            if parsed and parsed.get("tool_calls"):
+                return {
+                    "id": self._resp_id,
+                    "object": "chat.completion",
+                    "created": ctx.created,
+                    "model": self._model,
+                    "choices": [
+                        {
+                            "index": 0,
+                            "message": {
+                                "role": "assistant",
+                                "content": parsed.get("content"),
+                                "tool_calls": parsed["tool_calls"],
+                            },
+                            "finish_reason": "tool_calls",
+                        }
+                    ],
+                    "usage": {
+                        "prompt_tokens": ctx.prompt_tokens,
+                        "completion_tokens": ctx.completion_tokens,
+                        "total_tokens": ctx.prompt_tokens + ctx.completion_tokens,
+                    },
+                }
+
         return {
             "id": self._resp_id,
             "object": "chat.completion",

astrai/inference/api/protocol.py

@@ -35,7 +35,7 @@ class GenContext:
     resp_id: str
     created: int
     model: str
-    prompt_tokens: int
+    prompt_tokens: int = 0
     completion_tokens: int = 0
 
 
@@ -64,7 +64,7 @@ class StopChecker:
 class ResponseBuilder(ABC):
     """Interface for protocol-specific response formatting.
 
-    A new protocol requires one concrete builder implementing 6 methods.
+    A new protocol requires one concrete builder implementing 5 methods.
     """
 
     @abstractmethod
@@ -78,8 +78,15 @@ class ResponseBuilder(ABC):
         """SSE events that open the stream."""
 
     @abstractmethod
-    def format_chunk(self, token: str) -> str:
+    def format_chunk(self, token: str, **kwargs) -> List[str]:
-        """SSE event for a single generated token."""
+        """SSE events for a single generated token.
+
+        ``body`` (the full accumulated text so far) is always provided
+        as a keyword argument. Additional keyword arguments such as
+        ``current_token_ids`` and ``delta_token_ids`` may be included
+        for tool parsers that need token-level information.
+        Returns a list of SSE event strings (may be empty).
+        """
 
     @abstractmethod
     def format_stream_end(self, ctx: GenContext, stop: StopInfo) -> List[str]:
@@ -118,6 +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:
@@ -137,15 +145,25 @@ class ProtocolHandler:
             body = ""
             yielded = ""
             matched = None
+            token_ids: List[int] = []
             async for token in agen:
-                ctx.completion_tokens += 1
                 body += token
 
+                new_ids = self.engine.tokenizer.encode(token)
+                token_ids.extend(new_ids)
+
                 matched = checker.check(body)
                 if matched:
                     break
 
-                yield self.builder.format_chunk(token)
+                ctx.completion_tokens += 1
+                for event in self.builder.format_chunk(
+                    token,
+                    body=body,
+                    current_token_ids=token_ids,
+                    delta_token_ids=new_ids,
+                ):
+                    yield event
                 yielded += token
 
             stop = StopInfo(matched=matched, body=body, yielded=yielded)
@@ -168,7 +186,6 @@ class ProtocolHandler:
         matched = None
 
         async for token in agen:
-            ctx.completion_tokens += 1
             chunks.append(token)
             body += token
 
@@ -176,6 +193,8 @@ class ProtocolHandler:
             if matched:
                 break
 
+            ctx.completion_tokens += 1
+
         content = "".join(chunks)
         stop = StopInfo(matched=matched, body=body)
         return self.builder.format_response(ctx, content, stop)

astrai/inference/api/server.py

@@ -3,6 +3,9 @@ OpenAI / Anthropic-compatible chat completion server backed by continuous-batchi
 
 Protocol-specific formatting is delegated to ``astrai.inference.protocol``.
 This module owns the FastAPI app, request/response schemas, and dependency wiring.
+
+``app`` is lazily constructed — importing this module does NOT create a FastAPI instance.
+Use :func:`get_app` to access the singleton.
 """
 
 import logging
@@ -12,7 +15,7 @@ from typing import Any, Dict, List, Optional, Union
 
 import torch
 import uvicorn
-from fastapi import FastAPI, HTTPException
+from fastapi import APIRouter, FastAPI, HTTPException
 from pydantic import BaseModel, Field
 
 from astrai.inference.api.anthropic import AnthropicResponseBuilder
@@ -24,12 +27,25 @@ from astrai.tokenize import AutoTokenizer
 
 logger = logging.getLogger(__name__)
 
-_project_root = Path(__file__).parent.parent.parent
+_app_instance: Optional[FastAPI] = None
 
 
 class ChatMessage(BaseModel):
     role: str
-    content: str
+    content: Optional[str] = None
+    tool_calls: Optional[List[Dict[str, Any]]] = None
+    tool_call_id: Optional[str] = None
+
+
+class FunctionDef(BaseModel):
+    name: str
+    description: Optional[str] = None
+    parameters: Optional[Dict[str, Any]] = None
+
+
+class ToolDef(BaseModel):
+    type: str = "function"
+    function: FunctionDef
 
 
 class ChatCompletionRequest(BaseModel):
@@ -48,6 +64,8 @@ class ChatCompletionRequest(BaseModel):
     frequency_penalty: Optional[float] = Field(default=0.0, ge=-2.0, le=2.0)
     logit_bias: Optional[Dict[int, float]] = None
     user: Optional[str] = None
+    tools: Optional[List[ToolDef]] = None
+    tool_choice: Optional[Union[str, Dict[str, Any]]] = "auto"
 
 
 class AnthropicMessage(BaseModel):
@@ -84,17 +102,16 @@ async def lifespan(app: FastAPI):
         logger.info("Inference engine shutdown complete")
 
 
-app = FastAPI(title="AstrAI Inference Server", version="0.2.0", lifespan=lifespan)
+router = APIRouter()
 
 
 def _create_engine(
-    param_path: Optional[Path] = None,
+    param_path: Path,
     device: str = "cuda",
     dtype: torch.dtype = torch.bfloat16,
     max_batch_size: int = 16,
+    max_seq_len: Optional[int] = None,
 ) -> InferenceEngine:
-    if param_path is None:
-        param_path = _project_root / "params"
     if not param_path.exists():
         raise FileNotFoundError(f"Parameter directory not found: {param_path}")
 
@@ -107,39 +124,56 @@ 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
 
 
+def get_app() -> FastAPI:
+    """Return the singleton FastAPI instance (lazily created on first call)."""
+    global _app_instance
+    if _app_instance is None:
+        _app_instance = FastAPI(
+            title="AstrAI Inference Server",
+            version="0.2.0",
+            lifespan=lifespan,
+        )
+        _app_instance.include_router(router)
+        _app_instance.state.server_config = {}
+        _app_instance.state.engine = None
+    return _app_instance
+
+
 def _get_engine() -> InferenceEngine:
-    engine = app.state.engine
+    engine = get_app().state.engine
     if engine is None:
         raise HTTPException(status_code=503, detail="Engine not initialized")
     return engine
 
 
-@app.get("/health")
+@router.get("/health")
 async def health():
+    app = get_app()
     return {
         "status": "ok",
         "model_loaded": app.state.engine is not None,
     }
 
 
-@app.get("/stats")
+@router.get("/stats")
 async def get_stats():
     return _get_engine().get_stats()
 
 
-@app.post("/v1/chat/completions")
+@router.post("/v1/chat/completions")
 async def chat_completion(request: ChatCompletionRequest):
     engine = _get_engine()
     handler = ProtocolHandler(request, engine, OpenAIResponseBuilder())
     return await handler.handle()
 
 
-@app.post("/v1/messages")
+@router.post("/v1/messages")
 async def create_message(request: MessagesRequest):
     engine = _get_engine()
     handler = ProtocolHandler(request, engine, AnthropicResponseBuilder())
@@ -147,19 +181,22 @@ async def create_message(request: MessagesRequest):
 
 
 def run_server(
+    param_path: Path,
     host: str = "0.0.0.0",
     port: int = 8000,
     reload: bool = False,
     device: str = "cuda",
     dtype: torch.dtype = torch.bfloat16,
-    param_path: Optional[Path] = None,
     max_batch_size: int = 16,
+    max_seq_len: Optional[int] = None,
 ):
+    app = get_app()
     app.state.server_config = {
         "device": device,
         "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

@@ -0,0 +1,339 @@
+"""Tool call parsers for extracting structured tool calls from model output.
+
+Patterned after vLLM's ToolParser abstraction. Each parser knows how to
+detect and incrementally extract tool calls from raw generated text.
+
+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
+from typing import Dict, List, Optional
+
+from astrai.factory import BaseFactory
+
+
+class BaseToolParser(ABC):
+    """Abstract tool call parser — one instance per request.
+
+    Maintains streaming state internally so that each call to :meth:`feed`
+    can diff against previously emitted content.
+
+    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"):
+        self.tools = tools or []
+        self.tool_choice = tool_choice
+
+    @abstractmethod
+    def feed(
+        self,
+        body: str,
+        current_token_ids: Optional[List[int]] = None,
+        delta_token_ids: Optional[List[int]] = None,
+    ) -> List[Dict]:
+        """Feed the *full* accumulated text each step.
+
+        Returns a list of delta dicts to emit. Each delta is one of:
+
+        - ``{"content": "text"}``      — plain text delta
+        - ``{"tool_calls": [...]}``    — tool-call delta (OpenAI format)
+
+        Returns an empty list when nothing new should be emitted.
+
+        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
+    def parse_complete(self, body: str) -> Optional[Dict]:
+        """Parse the *complete* generated text after generation ends.
+
+        Returns ``None`` when no tool calls were found, otherwise a dict
+        with ``content`` (str or None) and ``tool_calls`` (list of dicts).
+        """
+
+    @property
+    @abstractmethod
+    def has_tool_calls(self) -> bool:
+        """True if the parser detected at least one tool call in the stream."""
+
+
+class ToolParserFactory(BaseFactory["BaseToolParser"]):
+    pass
+
+
+_TOOL_CALL_HEAD_RE = re.compile(r'\{\s*"name"\s*:')
+
+
+def _scan_json(text: str, start: int = 0):
+    """Scan for a complete JSON object starting at *start*.
+
+    Returns ``(end, complete)`` where *end* is one-past the closing
+    brace (or ``len(text)`` if unclosed), and *complete* is a bool.
+    """
+    depth = 0
+    in_string = False
+    escape = False
+    for i in range(start, len(text)):
+        c = text[i]
+        if escape:
+            escape = False
+            continue
+        if c == "\\":
+            escape = True
+            continue
+        if c == '"':
+            in_string = not in_string
+            continue
+        if in_string:
+            continue
+        if c == "{":
+            depth += 1
+        elif c == "}":
+            depth -= 1
+            if depth == 0:
+                return i + 1, True
+    return len(text), False
+
+
+def _parse_tool_call_json(json_str: str, complete: bool):
+    """Extract *name* and *arguments* from a tool-call JSON string.
+
+    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
+    name = name_match.group(1)
+
+    args_match = re.search(r'"arguments"\s*:\s*(.*)', json_str, re.DOTALL)
+    if not args_match:
+        return name, "", True
+
+    raw = args_match.group(1).rstrip()
+    if raw.startswith("{"):
+        inner = raw[1:].rstrip()
+        if inner.endswith("}"):
+            inner = inner[:-1].rstrip()
+        raw = inner
+    return name, raw, True
+
+
+def _find_tool_calls(text: str, start_pos: int = 0):
+    """Find all complete ``{...}`` tool-call objects in *text*.
+
+    Returns a list of dicts with keys *start*, *end*, *name*, *args*,
+    *complete*.
+    """
+    results = []
+    pos = start_pos
+
+    while True:
+        brace = text.find("{", pos)
+        if brace == -1:
+            break
+
+        end, complete = _scan_json(text, brace)
+        if not complete:
+            break
+
+        json_str = text[brace:end]
+
+        name, args, valid = _parse_tool_call_json(json_str, complete=True)
+        if not valid or name is None:
+            pos = end
+            continue
+
+        results.append(
+            {
+                "start": brace,
+                "end": end,
+                "name": name,
+                "args": args,
+                "complete": True,
+            }
+        )
+        pos = end
+
+    return results
+
+
+def _find_partial_tool_call(text: str, start_pos: int = 0):
+    """Find one incomplete (still-generating) tool-call JSON object."""
+    brace = text.find("{", start_pos)
+    if brace == -1:
+        return None
+
+    json_str = text[brace:]
+    if '"name"' not in json_str:
+        return None
+
+    name, args, valid = _parse_tool_call_json(json_str, complete=False)
+    if not valid or name is None:
+        return None
+
+    return {
+        "start": brace,
+        "name": name,
+        "args": args,
+        "complete": False,
+    }
+
+
+@ToolParserFactory.register("simple_json")
+class SimpleJsonToolParser(BaseToolParser):
+    """Parser for models that output tool calls as plain JSON objects.
+
+    Detects ``{"name": "<func>", "arguments": {...}}`` anywhere in the
+    generated text.  Handles single and (non-overlapping) multiple tool
+    calls.  Text preceding the first tool call is emitted as plain
+    ``content`` deltas.
+    """
+
+    def __init__(self, tools=None, tool_choice="auto"):
+        super().__init__(tools, tool_choice)
+        self._emitted_content_len = 0
+        self._tc_state: List[Dict] = []
+        self._has_tool_calls = False
+
+    # -------------------------------------------------------------- feed
+
+    def feed(
+        self,
+        body: str,
+        current_token_ids: Optional[List[int]] = None,
+        delta_token_ids: Optional[List[int]] = None,
+    ) -> List[Dict]:
+        deltas: List[Dict] = []
+
+        completed = _find_tool_calls(body)
+
+        if not completed:
+            partial = _find_partial_tool_call(body)
+            if not partial:
+                return self._emit_plain_content(body, deltas)
+            all_tcs = [partial]
+        else:
+            all_tcs = completed
+            partial = _find_partial_tool_call(body, completed[-1]["end"])
+            if partial:
+                all_tcs = completed + [partial]
+
+        first_start = all_tcs[0]["start"]
+        if first_start > self._emitted_content_len:
+            content = body[self._emitted_content_len : first_start]
+            self._emitted_content_len = first_start
+            if content:
+                deltas.append({"content": content})
+
+        for i, tc in enumerate(all_tcs):
+            if i >= len(self._tc_state):
+                self._tc_state.append(
+                    {
+                        "id": f"call_{uuid.uuid4().hex[:12]}",
+                        "name_emitted": False,
+                        "args_emitted_len": 0,
+                    }
+                )
+                self._has_tool_calls = True
+            st = self._tc_state[i]
+
+            if not st["name_emitted"]:
+                st["name_emitted"] = True
+                deltas.append(
+                    {
+                        "tool_calls": [
+                            {
+                                "index": i,
+                                "id": st["id"],
+                                "type": "function",
+                                "function": {"name": tc["name"], "arguments": ""},
+                            }
+                        ]
+                    }
+                )
+
+            new_args = tc["args"]
+            if len(new_args) > st["args_emitted_len"]:
+                diff = new_args[st["args_emitted_len"] :]
+                st["args_emitted_len"] = len(new_args)
+                deltas.append(
+                    {
+                        "tool_calls": [
+                            {
+                                "index": i,
+                                "function": {"arguments": diff},
+                            }
+                        ]
+                    }
+                )
+
+        return deltas
+
+    def _emit_plain_content(self, body: str, deltas: List[Dict]) -> List[Dict]:
+        new_content = body[self._emitted_content_len :]
+        if new_content:
+            self._emitted_content_len = len(body)
+            deltas.append({"content": new_content})
+        return deltas
+
+    # -------------------------------------------------------- complete
+
+    def parse_complete(self, body: str) -> Optional[Dict]:
+        completed = _find_tool_calls(body)
+        if not completed:
+            return None
+
+        content = body[: completed[0]["start"]].strip() or None
+        tool_calls = []
+        for i, tc in enumerate(completed):
+            tool_calls.append(
+                {
+                    "id": f"call_{uuid.uuid4().hex[:12]}",
+                    "type": "function",
+                    "function": {
+                        "name": tc["name"],
+                        "arguments": tc["args"],
+                    },
+                }
+            )
+        return {"content": content, "tool_calls": tool_calls}
+
+    @property
+    def has_tool_calls(self) -> bool:
+        return self._has_tool_calls

astrai/inference/core/__init__.py

@@ -3,11 +3,10 @@
 from astrai.inference.core.cache import (
     Allocator,
     KVCache,
-    KvcacheView,
+    KVStorage,
     PagePool,
     PrefixCache,
-    Storage,
+    ReqToTokenPool,
-    TaskTable,
     page_hash,
 )
 from astrai.inference.core.executor import Executor
@@ -17,11 +16,10 @@ from astrai.inference.core.task import STOP, Task, TaskManager, TaskStatus
 __all__ = [
     "Allocator",
     "KVCache",
-    "KvcacheView",
+    "KVStorage",
     "PagePool",
     "PrefixCache",
-    "Storage",
+    "ReqToTokenPool",
-    "TaskTable",
     "page_hash",
     "Executor",
     "InferenceScheduler",

astrai/inference/core/cache.py

@@ -1,6 +1,21 @@
+"""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 PrefixCache (content 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 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
@@ -62,7 +77,8 @@ class Allocator:
 
     def touch(self, idx: int):
         with self._lock:
-            self._lru.move_to_end(idx)
+            if idx in self._lru:
+                self._lru.move_to_end(idx)
 
 
 class PrefixCache:
@@ -106,263 +122,380 @@ class PrefixCache:
             self._hash_to_page[h] = page_idx
 
 
-class PagePool:
+class ReqToTokenPool:
-    """Orchestrates allocator (page management) and PrefixCache (content addressing)."""
+    """Maps [req_idx, pos] -> physical token slot in KV storage.
 
-    def __init__(self, allocator: Allocator, prefix: PrefixCache):
+    Each row is one request; each column is a sequence position. The value
-        self._alloc = allocator
+    at [req_idx, pos] is the flat index into the KV storage buffers.
-        self._prefix = prefix
+    """
-        self._alloc.on_evict = prefix.evict
 
-    @property
+    def __init__(self, size: int, max_context_len: int, device: torch.device):
-    def allocator(self) -> Allocator:
+        self.size = size
-        return self._alloc
+        self.max_context_len = max_context_len
-
+        self.req_to_token = torch.zeros(
-    @property
+            (size, max_context_len), dtype=torch.long, device=device
-    def prefix(self) -> PrefixCache:
+        )
-        return self._prefix
+        self.free_slots = list(range(size))
-
-    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):
+    def alloc(self, num_reqs: int) -> Optional[List[int]]:
         with self._lock:
-            self._pages[task_id] = page_table
+            if num_reqs > len(self.free_slots):
-            self._cached[task_id] = cached
+                return None
+            slots = self.free_slots[:num_reqs]
+            self.free_slots = self.free_slots[num_reqs:]
+            return slots
 
-    def get(self, task_id: str) -> List[int]:
+    def free(self, req_indices: List[int]):
         with self._lock:
-            return self._pages.get(task_id, [])
+            self.free_slots.extend(req_indices)
 
-    def get_cached(self, task_id: str) -> int:
+    def write(self, indices, values):
-        with self._lock:
+        self.req_to_token[indices] = values
-            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:
+class KVStorage:
-    """KV-cache tensor storage with paged write/gather."""
+    """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_pages: int,
-        page_size: int,
         n_kv_heads: int,
         head_dim: int,
         device: torch.device,
         dtype: torch.dtype,
     ):
-        self.page_size = page_size
+        self.size = size
-        self.k_cache = torch.empty(
+        self.k_buffer = torch.empty(
-            (n_layers, n_pages, page_size, n_kv_heads, head_dim),
+            (n_layers, size, n_kv_heads, head_dim), device=device, dtype=dtype
-            device=device,
-            dtype=dtype,
         )
-        self.v_cache = torch.empty(
+        self.v_buffer = torch.empty(
-            (n_layers, n_pages, page_size, n_kv_heads, head_dim),
+            (n_layers, size, n_kv_heads, head_dim), device=device, dtype=dtype
-            device=device,
-            dtype=dtype,
         )
 
-    def write(
+    def get_key_buffer(self, layer_id: int) -> Tensor:
-        self,
+        return self.k_buffer[layer_id]
-        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(
+    def get_value_buffer(self, layer_id: int) -> Tensor:
-        self, layer_id: int, page_table: Tensor, total_len: int
+        return self.v_buffer[layer_id]
-    ) -> 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
 
-
+    def set_kv_buffer(self, layer_id: int, loc: Tensor, k: Tensor, v: Tensor) -> None:
-class KvcacheView:
+        self.k_buffer[layer_id, loc] = k
-    """Bundles Storage + page_table + total_len for attention layers."""
+        self.v_buffer[layer_id, loc] = v
-
-    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)
 
 
+@dataclass
 class KVCache:
-    """Facade: page management + KV-cache I/O for continuous batching."""
+    """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
+        page_table: [batch, max_len] — precomputed gather indices for decode;
+            None for prefill or when not yet computed.
+        decode_mask: [batch, max_len] bool — precomputed position validity
+            mask for decode; None for prefill or single-batch decode.
+    """
+
+    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
+    page_table: Optional[Tensor] = None
+    decode_mask: Optional[Tensor] = None
+
+
+class PagePool:
+    """Top-level KV cache manager.
+
+    Combines KVStorage + ReqToTokenPool + Allocator + PrefixCache.
+
+    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,
+        max_batch_size: int,
+        max_seq_len: int,
         device: torch.device,
         dtype: torch.dtype,
+        page_size: int = 1,
+        n_tokens: Optional[int] = None,
     ):
         self.page_size = page_size
-        self._pool = PagePool(Allocator(n_pages), PrefixCache(page_size))
+        self.max_batch_size = max_batch_size
-        self._table = TaskTable(page_size)
+        self.max_seq_len = max_seq_len
-        self._storage = Storage(
+        self.device = device
-            n_layers, n_pages, page_size, n_kv_heads, head_dim, device, dtype
+        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._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[PrefixCache] = None
+        else:
+            n_pages = self.n_tokens // page_size
+            self._alloc = Allocator(n_pages)
+            self._prefix = PrefixCache(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()
+
+    # ---- task lifecycle ----
 
     def task_alloc(self, task_id: str, prompt_ids: List[int]) -> bool:
-        hits = self._pool.lookup(prompt_ids)
+        req_slots = self._req_pool.alloc(1)
-        cached = len(hits) * self.page_size
+        if req_slots is None:
-        for p in hits:
+            return False
-            self._pool.inc_ref(p)
+        req_idx = req_slots[0]
+        self._task_req[task_id] = req_idx
 
-        remaining = len(prompt_ids) - cached
+        if self.contiguous:
-        n_new = (
+            self._task_len[req_idx] = len(prompt_ids)
-            (remaining + self.page_size - 1) // self.page_size if remaining > 0 else 0
+            self._task_cached[task_id] = 0
-        )
+            return True
-        new_pages: List[int] = []
+
-        if n_new > 0:
+        n_tokens_needed = len(prompt_ids)
-            for _ in range(n_new):
+        cached = 0
-                p = self._pool.alloc()
+
-                if p < 0:
+        if self._prefix is not None:
-                    for hp in hits:
+            hits = self._prefix.lookup(prompt_ids)
-                        self._pool.free(hp)
+            cached = len(hits) * self.page_size
-                    for np in new_pages:
+            for p in hits:
-                        self._pool.free(np)
+                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
-                new_pages.append(p)
+                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._table.set(task_id, hits + new_pages, cached)
+        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):
-        page_table, _ = self._table.pop(task_id)
+        req_idx = self._task_req.pop(task_id, None)
-        for idx in page_table:
+        if req_idx is None:
-            self._pool.free(idx)
+            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:
-        page_table = self._table.get(task_id)
+        req_idx = self._task_req.get(task_id)
-        needed = (pos + 1 + self.page_size - 1) // self.page_size
+        if req_idx is None:
-        while len(page_table) < needed:
+            return False
-            p = self._pool.alloc()
+
-            if p < 0:
+        if self.contiguous:
+            return pos < self.max_seq_len
+
+        if self.page_size == 1:
+            slots = self._alloc_tokens(1)
+            if slots is None:
                 return False
-            page_table.append(p)
+            self._task_slots.setdefault(task_id, []).extend(slots)
+            self._req_pool.req_to_token[req_idx, pos] = slots[0]
+        else:
+            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
+
+        self._task_len[req_idx] = pos + 1
         return True
 
     def task_cached(self, task_id: str) -> int:
-        return self._table.get_cached(task_id)
+        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
     ):
-        page_table = self._table.get(task_id)
+        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, full_pages):
+        for i in range(start_logical_page, min(full_pages, len(pages))):
-            self._pool.record(page_table[i], prompt_ids, i)
+            self._prefix.record(pages[i], prompt_ids, i)
 
-    def make_table_tensor(self, task_ids: List[str], device: torch.device) -> Tensor:
+    # ---- bind for forward ----
-        return self._table.table_tensor(task_ids, device)
 
-    def bind(self, page_table: Tensor, total_len: int = 0) -> KvcacheView:
+    def bind_tasks(
-        return KvcacheView(self._storage, page_table, total_len)
+        self,
+        task_ids: List[str],
+        seq_lens: List[int],
+        device: torch.device,
+        start_pos: Optional[int] = None,
+    ) -> KVCache:
+        req_indices = [self._task_req[tid] for tid in task_ids]
+        req_pool_indices = torch.tensor(req_indices, dtype=torch.long, device=device)
+        seq_lens_t = torch.tensor(seq_lens, dtype=torch.long, device=device)
+
+        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
+            ]
+            page_table = None
+            decode_mask = None
+        else:
+            write_pos = seq_lens_t - 1
+            out_cache_loc = self._req_pool.req_to_token[
+                req_pool_indices, write_pos
+            ].unsqueeze(-1)
+            ml = max(seq_lens)
+            page_table = self._req_pool.req_to_token[req_pool_indices, :ml]
+            if len(task_ids) > 1:
+                decode_mask = (
+                    torch.arange(ml, device=device)[None, :] < seq_lens_t[:, None]
+                )
+            else:
+                decode_mask = 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),
+            page_table=page_table,
+            decode_mask=decode_mask,
+        )
+
+    # ---- 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

@@ -3,7 +3,7 @@ from typing import List, Optional
 
 import torch
 
-from astrai.inference.core.cache import KVCache
+from astrai.inference.core.cache import PagePool
 from astrai.inference.core.task import Task
 from astrai.inference.sample import sample
 from astrai.model.automodel import AutoModel
@@ -19,13 +19,13 @@ class Executor:
         self,
         model: AutoModel,
         tokenizer: AutoTokenizer,
-        page_cache: KVCache,
+        kv_cache: PagePool,
         device: Optional[str] = None,
         dtype: Optional[torch.dtype] = None,
     ):
         self.model = model
         self.tokenizer = tokenizer
-        self.page_cache = page_cache
+        self.kv_cache = kv_cache
         self.device = device or next(model.parameters()).device
         self.dtype = dtype or next(model.parameters()).dtype
 
@@ -43,20 +43,42 @@ class Executor:
         )
 
         task_ids = [t.task_id for t in tasks]
-        page_tables = self.page_cache.make_table_tensor(task_ids, 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
+        )
 
         with torch.inference_mode():
             self.model(
                 input_ids,
-                position_ids=torch.arange(
+                input_mask=input_mask,
-                    start_pos, prompt_len, dtype=torch.long, device=self.device
+                position_ids=position_ids,
-                )
+                kv_cache=self.kv_cache.bind_tasks(
-                .unsqueeze(0)
+                    task_ids, [prompt_len] * batch_sz, self.device, start_pos=start_pos
-                .expand(batch_sz, -1),
+                ),
-                paged_cache=self.page_cache.bind(page_tables, total_len=prompt_len),
             )
 
-    def execute_decode(self, tasks: List[Task]) -> List[int]:
+    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 []
 
@@ -69,26 +91,84 @@ class Executor:
         position_ids = torch.tensor(
             [t.next_pos for t in tasks], dtype=torch.long, device=self.device
         )
-        total_len = position_ids.max().item() + 1
+        total_len = max(t.next_pos for t in tasks) + 1
+        input_mask = position_ids[:, None, None] >= torch.arange(
+            total_len, device=self.device
+        )
 
         task_ids = [t.task_id for t in tasks]
-        page_tables = self.page_cache.make_table_tensor(task_ids, self.device)
 
         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)
+        freq_penalties = torch.tensor(
+            [t.frequency_penalty for t in tasks], device=self.device
+        )
+
+        has_freq = bool((freq_penalties != 0).any())
+        if has_freq:
+            history_lists = []
+            history_lens = []
+            for t in tasks:
+                window = t.rep_window
+                prompt_part = t.prompt_ids[-window:]
+                ids = prompt_part + t.output_ids
+                history_lists.append(ids)
+                history_lens.append(len(ids))
+
+            max_len = max(history_lens) if history_lens else 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, h in enumerate(history_lists):
+                L = history_lens[i]
+                padded_ids[i, :L] = torch.as_tensor(
+                    h, dtype=torch.long, device=self.device
+                )
+                padded_mask[i, :L] = True
+        else:
+            padded_ids = None
+            padded_mask = None
 
         with torch.inference_mode():
             outputs = self.model(
                 input_ids.unsqueeze(1),
-                paged_cache=self.page_cache.bind(page_tables, total_len=total_len),
+                input_mask=input_mask,
+                kv_cache=self.kv_cache.bind_tasks(
+                    task_ids,
+                    [t.next_pos + 1 for t in tasks],
+                    self.device,
+                ),
                 position_ids=position_ids.unsqueeze(1),
             )
             logits = outputs["logits"][:, -1, :]
 
+        if return_logprobs:
+            tokens, logprobs = sample(
+                logits,
+                temperature=temperatures,
+                top_k=top_ks,
+                top_p=top_ps,
+                frequency_penalty=freq_penalties,
+                input_ids=padded_ids,
+                input_mask=padded_mask,
+                return_logprobs=True,
+            )
+            tokens_list = tokens.tolist()
+            logprobs_list = logprobs.tolist()
+            for t, lp in zip(tasks, logprobs_list):
+                t.output_logprobs.append(float(lp))
+            return list(zip(tokens_list, logprobs_list))
+
         return sample(
             logits,
             temperature=temperatures,
             top_k=top_ks,
             top_p=top_ps,
+            frequency_penalty=freq_penalties,
+            input_ids=padded_ids,
+            input_mask=padded_mask,
         ).tolist()

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 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
@@ -14,7 +15,7 @@ logger = logging.getLogger(__name__)
 
 
 class InferenceScheduler:
-    """Four-phase continuous batching loop: cleanup -> refill -> prefill -> decode."""
+    """Continuous batching loop: cleanup -> refill -> prefill -> decode (all groups)."""
 
     def __init__(
         self,
@@ -22,73 +23,74 @@ class InferenceScheduler:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 16,
         max_seq_len: Optional[int] = None,
-        max_prompt_len: int = 2048,
-        page_size: int = 64,
         device: Optional[str] = None,
         dtype: Optional[torch.dtype] = 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:
+        elif config.max_position_embeddings is not None:
-            self.max_seq_len = config.max_len
+            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
 
-        n_pages = (
+        head_dim = config.hidden_size // config.num_attention_heads
-            max_batch_size * (self.max_seq_len + page_size) + page_size - 1
-        ) // page_size
 
-        self._page_cache = KVCache(
+        if cache is not None:
-            config.n_layers,
+            self._cache = cache
-            n_pages,
+        else:
-            page_size,
+            self._cache = PagePool(
-            config.n_kv_heads,
+                n_layers=config.num_hidden_layers,
-            config.dim // config.n_heads,
+                n_kv_heads=config.num_key_value_heads,
-            self.device,
+                head_dim=head_dim,
-            self.dtype,
+                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(
             model=model,
             tokenizer=tokenizer,
-            page_cache=self._page_cache,
+            kv_cache=self._cache,
             device=self.device,
             dtype=self.dtype,
         )
 
-        self._running = False
+        self._stop_event = threading.Event()
+        self._loop_thread: Optional[threading.Thread] = None
 
     def add_task(self, prompt: str, **kwargs) -> str:
         return self._task_mgr.add_task(prompt, **kwargs)
 
     def remove_task(self, task_id: str):
         for task in self._task_mgr.remove_task(task_id):
-            self._page_cache.task_free(task.task_id)
+            self._cache.task_free(task.task_id)
 
     def get_stats(self) -> Dict[str, Any]:
         return self._task_mgr.get_stats()
 
     def _run_generation_loop(self):
         stop_ids = self._task_mgr.tokenizer.stop_ids
+        cache = self._cache
         try:
-            while self._running:
+            while not self._stop_event.is_set():
                 finished = self._task_mgr.remove_finished_tasks(stop_ids)
                 for task in finished:
-                    self._page_cache.task_free(task.task_id)
+                    cache.task_free(task.task_id)
 
                 active = self._task_mgr.get_active_tasks()
                 available = self._task_mgr.max_batch_size - len(active)
@@ -96,7 +98,7 @@ class InferenceScheduler:
                     candidates = self._task_mgr.pull_candidates(available)
                     failed = []
                     for task in candidates:
-                        if self._page_cache.task_alloc(task.task_id, task.prompt_ids):
+                        if cache.task_alloc(task.task_id, task.prompt_ids):
                             self._task_mgr.activate(task)
                         else:
                             failed.append(task)
@@ -107,11 +109,13 @@ class InferenceScheduler:
                     self._task_mgr.wait_for_tasks(timeout=1.0)
                     continue
 
+                active = self._task_mgr.get_active_tasks()
+
                 to_prefill = [
                     t
-                    for t in self._task_mgr.get_active_tasks()
+                    for t in active
                     if t.output_tokens == 0
-                    and self._page_cache.task_cached(t.task_id) < len(t.prompt_ids)
+                    and cache.task_cached(t.task_id) < len(t.prompt_ids)
                 ]
                 if to_prefill:
                     for t in to_prefill:
@@ -121,85 +125,187 @@ class InferenceScheduler:
                     for t in to_prefill:
                         key = (
                             len(t.prompt_ids),
-                            self._page_cache.task_cached(t.task_id),
+                            cache.task_cached(t.task_id),
                         )
                         groups.setdefault(key, []).append(t)
 
                     for (prompt_len, start_pos), group in groups.items():
                         self._executor.execute_prefill(group, prompt_len, start_pos)
-                        start_logical_page = start_pos // self._page_cache.page_size
+                        start_logical_page = start_pos // getattr(
+                            cache, "page_size", 64
+                        )
                         for t in group:
-                            self._page_cache.task_record_hashes(
+                            cache.task_record_hashes(
-                                t.task_id,
+                                t.task_id, t.prompt_ids, start_logical_page
-                                t.prompt_ids,
-                                start_logical_page=start_logical_page,
                             )
 
-                pos_groups: Dict[int, List[Task]] = {}
+                decode_tasks = active
-                for t in self._task_mgr.get_active_tasks():
-                    pos_groups.setdefault(t.next_pos, []).append(t)
 
-                if pos_groups:
+                valid: List[Task] = []
-                    best_key = max(pos_groups, key=lambda k: len(pos_groups[k]))
+                for t in decode_tasks:
-                    group = sorted(pos_groups[best_key], key=lambda t: t.task_id)
+                    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)
 
-                    valid: List[Task] = []
+                if valid:
-                    for t in group:
+                    next_tokens = self._executor.execute_decode(valid)
-                        if self._page_cache.task_extend(t.task_id, t.next_pos):
-                            valid.append(t)
-                        else:
-                            t.status = TaskStatus.ABORTED
-                            if t.stream_callback:
-                                t.stream_callback(STOP)
 
-                    if valid:
+                    for t, ntok in zip(valid, next_tokens):
-                        next_tokens = self._executor.execute_decode(valid)
+                        t.output_ids.append(ntok)
+                        t.output_tokens += 1
+                        new_text = t.decode_new_token(self._task_mgr.tokenizer)
+                        if new_text:
+                            self._task_mgr.invoke_callback(t.task_id, new_text)
 
-                        for t, ntok in zip(valid, next_tokens):
+                    for t in valid:
-                            t.output_ids.append(ntok)
+                        if t.is_finished(stop_ids):
-                            t.output_tokens += 1
+                            remaining = t.flush_remaining(self._task_mgr.tokenizer)
-                            pos = t.input_tokens + t.output_tokens
+                            if remaining:
-                            extend_ok = self._page_cache.task_extend(t.task_id, pos)
+                                self._task_mgr.invoke_callback(t.task_id, remaining)
-                            if t.stream_callback:
+                            self._task_mgr.invoke_callback(t.task_id, STOP)
-                                t.stream_callback(
-                                    self._task_mgr.tokenizer.decode([ntok])
-                                )
-                            if not extend_ok:
-                                t.status = TaskStatus.ABORTED
-                                if t.stream_callback:
-                                    t.stream_callback(STOP)
-
-                        for t in valid:
-                            if t.is_finished(stop_ids):
-                                if t.stream_callback:
-                                    t.stream_callback(STOP)
 
         except Exception as e:
+            self._stop_event.set()
             logger.error(f"Scheduler loop crashed: {e}", exc_info=True)
             for task in self._task_mgr.get_active_tasks():
-                if task.stream_callback:
+                self._task_mgr.invoke_callback(task.task_id, STOP)
-                    task.stream_callback(STOP)
+                cache.task_free(task.task_id)
-                self._page_cache.task_free(task.task_id)
             for task in self._task_mgr.get_waiting_tasks():
-                if task.stream_callback:
+                self._task_mgr.invoke_callback(task.task_id, STOP)
-                    task.stream_callback(STOP)
             self._task_mgr.clear_queues()
-            raise
 
     def start(self):
-        if not self._running:
+        if self._loop_thread is not None and self._loop_thread.is_alive():
-            self._running = True
+            return
-            t = threading.Thread(target=self._run_generation_loop, daemon=True)
+        self._stop_event.clear()
-            t.start()
+        t = threading.Thread(target=self._run_generation_loop, daemon=True)
-            self._loop_thread = t
+        t.start()
+        self._loop_thread = t
 
     def stop(self):
-        self._running = False
+        self._stop_event.set()
         self._task_mgr.wake()
-        if hasattr(self, "_loop_thread"):
+        if self._loop_thread is not None:
             self._loop_thread.join(timeout=2.0)
+        self._loop_thread = None
         for task in self._task_mgr.get_active_tasks():
-            self._page_cache.task_free(task.task_id)
+            self._task_mgr.invoke_callback(task.task_id, STOP)
+            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))
+            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]
+            prefill_groups: Dict[Tuple[int, int], List[Task]] = {}
+            for t in live:
+                key = (len(t.prompt_ids), cache.task_cached(t.task_id))
+                prefill_groups.setdefault(key, []).append(t)
+            for (prompt_len, start_pos), group in prefill_groups.items():
+                self._executor.execute_prefill(group, prompt_len, start_pos)
+
+            while live:
+                valid: List[Task] = []
+                for t in sorted(live, key=lambda x: x.task_id):
+                    if cache.task_extend(t.task_id, t.next_pos):
+                        valid.append(t)
+                    else:
+                        t.status = TaskStatus.ABORTED
+                if not valid:
+                    break
+
+                step_out = self._executor.execute_decode(
+                    valid, return_logprobs=return_logprobs
+                )
+                if return_logprobs:
+                    for t, (ntok, _lp) in zip(valid, step_out):
+                        t.output_ids.append(ntok)
+                        t.output_tokens += 1
+                else:
+                    for t, ntok in zip(valid, step_out):
+                        t.output_ids.append(ntok)
+                        t.output_tokens += 1
+
+                live = [t for t in valid 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."""
 
@@ -33,7 +62,8 @@ class Task:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
-        stream_callback: Optional[Callable[[str], None]] = None,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
     ):
         self.task_id = task_id
         self.prompt_ids = prompt_ids
@@ -41,14 +71,37 @@ 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.stream_callback = stream_callback
+        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:
@@ -70,15 +123,14 @@ 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] = []
+        self._callbacks: Dict[str, Callable[[str], None]] = {}
 
         self._task_event = threading.Event()
         self._lock = threading.Lock()
@@ -93,14 +145,16 @@ 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:
+        if len(prompt_ids) > self.max_seq_len:
-            prompt_ids = prompt_ids[-self.max_prompt_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
@@ -117,12 +171,15 @@ class TaskManager:
             temperature=temperature,
             top_p=top_p,
             top_k=top_k,
-            stream_callback=stream_callback,
+            frequency_penalty=frequency_penalty,
+            rep_window=rep_window,
         )
 
         with self._lock:
             self.waiting_queue.append(task)
             self._total_tasks += 1
+            if stream_callback:
+                self._callbacks[task_id] = stream_callback
 
         self._task_event.set()
         return task_id
@@ -134,8 +191,14 @@ class TaskManager:
                 t for t in self.waiting_queue if t.task_id != task_id
             )
             self.active_tasks = [t for t in self.active_tasks if t.task_id != task_id]
+            self._callbacks.pop(task_id, None)
         return removed_active
 
+    def invoke_callback(self, task_id: str, token: str):
+        cb = self._callbacks.get(task_id)
+        if cb:
+            cb(token)
+
     def get_stats(self) -> Dict[str, Any]:
         return {
             "total_tasks": self._total_tasks,
@@ -186,7 +249,10 @@ class TaskManager:
         return bool(self.active_tasks or self.waiting_queue)
 
     def wait_for_tasks(self, timeout: float = 1.0):
-        self._task_event.clear()
+        with self._lock:
+            if self.waiting_queue or self.active_tasks:
+                return
+            self._task_event.clear()
         self._task_event.wait(timeout=timeout)
 
     def get_active_tasks(self) -> List[Task]:
@@ -201,6 +267,7 @@ class TaskManager:
         with self._lock:
             self.waiting_queue.clear()
             self.active_tasks.clear()
+            self._callbacks.clear()
 
     def wake(self):
         self._task_event.set()

astrai/inference/engine.py

@@ -8,6 +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 PagePool
 from astrai.inference.core.scheduler import InferenceScheduler
 from astrai.inference.core.task import STOP
 from astrai.tokenize import AutoTokenizer
@@ -73,6 +74,8 @@ class GenerationRequest:
         top_p: float = 1.0,
         temperature: float = 1.0,
         max_tokens: Optional[int] = None,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
         stream: bool = False,
     ):
         if not (isinstance(top_k, int) and top_k >= 0):
@@ -81,12 +84,21 @@ class GenerationRequest:
             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 non-negative number")
+        if not (
+            isinstance(frequency_penalty, (int, float))
+            and -2.0 <= frequency_penalty <= 2.0
+        ):
+            raise ValueError("frequency_penalty must be between -2.0 and 2.0")
+        if not (isinstance(rep_window, int) and rep_window > 0):
+            raise ValueError("rep_window must be a positive integer")
 
         self.messages = messages
         self.top_k = top_k
         self.top_p = top_p
         self.temperature = temperature
         self.max_tokens = max_tokens
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
         self.stream = stream
 
 
@@ -99,8 +111,7 @@ class InferenceEngine:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 1,
         max_seq_len: Optional[int] = None,
-        max_prompt_len: int = 2048,
+        cache: Optional[PagePool] = None,
-        page_size: int = 128,
     ):
         self.model = model
         self.tokenizer = tokenizer
@@ -109,8 +120,7 @@ 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,
-            page_size=page_size,
         )
 
         self.scheduler.start()
@@ -130,17 +140,33 @@ 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
+                prompts,
+                is_batch,
+                max_tokens,
+                temperature,
+                top_p,
+                top_k,
+                frequency_penalty,
+                rep_window,
             )
         else:
             return self._generate_non_streaming(
-                prompts, is_batch, max_tokens, temperature, top_p, top_k
+                prompts,
+                is_batch,
+                max_tokens,
+                temperature,
+                top_p,
+                top_k,
+                frequency_penalty,
+                rep_window,
             )
 
     def generate_async(
@@ -150,9 +176,18 @@ 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
+            [prompt],
+            False,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
         )
 
         async def _agen():
@@ -183,6 +218,8 @@ class InferenceEngine:
             temperature=request.temperature,
             top_p=request.top_p,
             top_k=request.top_k,
+            frequency_penalty=request.frequency_penalty,
+            rep_window=request.rep_window,
         )
 
     def _submit_tasks(
@@ -192,6 +229,8 @@ class InferenceEngine:
         temperature: float,
         top_p: float,
         top_k: int,
+        frequency_penalty: float,
+        rep_window: int,
     ) -> Tuple[GenerateResult, List[str]]:
         n = len(prompts)
         result = GenerateResult(count=n)
@@ -204,6 +243,8 @@ class InferenceEngine:
                 temperature=temperature,
                 top_p=top_p,
                 top_k=top_k,
+                frequency_penalty=frequency_penalty,
+                rep_window=rep_window,
                 stream_callback=cb,
             )
             task_ids.append(task_id)
@@ -224,9 +265,17 @@ class InferenceEngine:
         temperature: float,
         top_p: float,
         top_k: int,
+        frequency_penalty: float,
+        rep_window: int,
     ) -> Generator:
         result, task_ids = self._submit_tasks(
-            prompts, max_tokens, temperature, top_p, top_k
+            prompts,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
         )
         n = len(prompts)
         remaining = n
@@ -260,9 +309,17 @@ class InferenceEngine:
         temperature: float,
         top_p: float,
         top_k: int,
+        frequency_penalty: float,
+        rep_window: int,
     ) -> Union[str, List[str]]:
         result, task_ids = self._submit_tasks(
-            prompts, max_tokens, temperature, top_p, top_k
+            prompts,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
         )
 
         try:

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
+(temperature, top-k, top-p, frequency penalty) is a pluggable
-can be composed into a pipeline.
+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
@@ -19,16 +19,28 @@ 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.
         """
+        raise NotImplementedError
 
 
 class TemperatureStrategy(BaseSamplingStrategy):
@@ -41,13 +53,21 @@ class TemperatureStrategy(BaseSamplingStrategy):
     def __init__(self, temperature: Union[float, Tensor] = 1.0):
         self.temperature = temperature
 
-    def apply(self, logits, filter_value=-float("inf")):
+    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)
+            t = torch.clamp(t, min=1e-8)
             if (t != 1.0).any():
-                logits = logits / t.to(logits.device, non_blocking=True).view(-1, 1)
+                logits = logits / t
         elif t != 1.0:
-            logits = logits / t
+            logits = logits / max(t, 1e-8)
         return logits
 
 
@@ -61,7 +81,13 @@ class TopKStrategy(BaseSamplingStrategy):
     def __init__(self, top_k: Union[int, Tensor] = 0):
         self.top_k = top_k
 
-    def apply(self, logits, filter_value=-float("inf")):
+    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)
@@ -98,7 +124,9 @@ class TopPStrategy(BaseSamplingStrategy):
     def __init__(self, top_p: Union[float, Tensor] = 1.0):
         self.top_p = top_p
 
-    def _apply(self, logits, top_p, filter_value):
+    def _apply(
+        self, logits: Tensor, top_p: Union[float, Tensor], filter_value: float
+    ) -> Tensor:
         sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
         cum_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
         remove = cum_probs > top_p
@@ -109,7 +137,13 @@ class TopPStrategy(BaseSamplingStrategy):
         logits[mask] = filter_value
         return logits
 
-    def apply(self, logits, filter_value=-float("inf")):
+    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)
@@ -120,6 +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.
 
@@ -140,25 +252,76 @@ class SamplingPipeline(BaseSamplingStrategy):
     def __init__(self, strategies: List[BaseSamplingStrategy]):
         self.strategies = strategies
 
-    def apply(self, logits, filter_value=-float("inf")):
+    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()
+    @staticmethod
-    def sample(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def _is_greedy(temperature: Union[float, Tensor]) -> bool:
+        if isinstance(temperature, Tensor):
+            return temperature.numel() == 1 and temperature.item() == 0
+        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(
+        if self._is_greedy_pipeline():
-            torch.softmax(self.apply(logits, filter_value), dim=-1),
+            tokens = logits.argmax(dim=-1)
-            num_samples=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)
+        log_probs = torch.log_softmax(transformed.float(), dim=-1)
+        tokens = torch.multinomial(
+            torch.softmax(transformed, dim=-1), num_samples=1
         ).squeeze(-1)
+        if not return_logprobs:
+            return tokens
+        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()
@@ -167,22 +330,75 @@ 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(
+    greedy = (
-        [
+        (
-            TemperatureStrategy(temperature),
+            isinstance(temperature, Tensor)
-            TopKStrategy(top_k),
+            and temperature.numel() == 1
-            TopPStrategy(top_p),
+            and temperature.item() == 0
-        ]
+        )
-    ).sample(logits, filter_value)
+        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/automodel.py

@@ -40,11 +40,12 @@ def _disable_random_init(enable: bool = True):
             setattr(nn.init, n, fn)
 
 
-class AutoModel(BaseFactory["AutoModel"], nn.Module):
+class ModelFactory(BaseFactory[nn.Module]):
-    """
+    """Pure factory for model dispatch, separated from nn.Module state."""
-    Autoregressive language model base class.
+
-    Provides model loading/saving, registration, and generation.
+
-    """
+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

@@ -1,4 +1,5 @@
-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
@@ -6,7 +7,6 @@ from astrai.model.components.mlp import MLP
 from astrai.model.components.norm import RMSNorm
 from astrai.model.components.rope import (
     RotaryEmbedding,
-    apply_rotary_emb,
     get_rotary_emb,
 )
 
@@ -21,5 +21,4 @@ __all__ = [
     "RotaryEmbedding",
     "apply_rotary_emb",
     "get_rotary_emb",
-    "repeat_kv",
 ]

astrai/model/components/attention.py

@@ -5,28 +5,16 @@ 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 KvcacheView
+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]):
-    @classmethod
+    pass
-    def create(cls, attn_type: str, **kwargs) -> nn.Module:
-        return super().create(attn_type, **kwargs)
 
 
 @AttnFactory.register("gqa")
@@ -40,6 +28,7 @@ class GQA(nn.Module):
         norm_eps: float,
         use_gated_attention: bool,
         layer_id: int,
+        n_layers: int = 1,
     ):
         super().__init__()
         assert dim % n_heads == 0
@@ -57,7 +46,7 @@ class GQA(nn.Module):
         self.q_proj = Linear(dim, n_heads * self.head_dim)
         self.k_proj = Linear(dim, n_kv_heads * self.head_dim)
         self.v_proj = Linear(dim, n_kv_heads * self.head_dim)
-        self.o_proj = Linear(dim, dim)
+        self.o_proj = Linear(dim, dim, init_std=0.02 / (2 * n_layers) ** 0.5)
 
         if self.use_qk_norm:
             self.q_norm = RMSNorm(self.head_dim, norm_eps)
@@ -76,10 +65,9 @@ class GQA(nn.Module):
         x: Tensor,
         rotary_emb: Tensor,
         attn_mask: Tensor = None,
-        paged_cache: Optional[KvcacheView] = 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)
@@ -88,19 +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:
+        sdqa_out = attention(q, k, v, kv_cache, self.layer_id, attn_mask, is_causal)
-            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)
-        )
 
         if self.use_gated_attention:
             sdqa_out = sdqa_out * F.sigmoid(self.gate(x))
@@ -123,6 +99,7 @@ class MLA(nn.Module):
         use_qk_norm: bool,
         use_gated_attention: bool,
         layer_id: int,
+        n_layers: int = 1,
     ):
         super().__init__()
         self.dim = dim
@@ -150,7 +127,9 @@ class MLA(nn.Module):
             n_kv_heads * (2 * self.head_dim),
         )
 
-        self.o_proj = Linear(dim, dim, bias=False)
+        self.o_proj = Linear(
+            dim, dim, bias=False, init_std=0.02 / (2 * n_layers) ** 0.5
+        )
 
         if use_gated_attention:
             self.gate = Linear(dim, dim, bias=False)
@@ -160,10 +139,10 @@ class MLA(nn.Module):
         x: Tensor,
         rotary_emb: Tensor,
         attn_mask: Tensor = None,
-        paged_cache: Optional[KvcacheView] = 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)
@@ -192,18 +171,7 @@ class MLA(nn.Module):
             q = self.q_norm(q)
             k = self.k_norm(k)
 
-        if paged_cache is not None:
+        attn_out = attention(q, k, v, kv_cache, self.layer_id, attn_mask, is_causal)
-            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)
 
         if self.use_gated_attention:
             attn_out = attn_out * F.sigmoid(self.gate(x))

astrai/model/components/decoder_block.py

@@ -1,57 +1,47 @@
+from dataclasses import asdict
 from typing import Optional
 
 import torch.nn as nn
 from torch import Tensor
 
-from astrai.inference.core.cache import KvcacheView
+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.norm import RMSNorm
 
 
 class DecoderBlock(nn.Module):
-    def __init__(
+    def __init__(self, config, layer_id: int):
-        self,
-        dim: int,
-        n_heads: int,
-        dim_ffn: int,
-        n_kv_heads: int,
-        norm_eps: float,
-        use_qk_norm: bool,
-        use_gated_attention: bool,
-        layer_id: int,
-        attn_type: str = "gqa",
-        ffn_type: str = "mlp",
-        **kwargs,
-    ):
         super().__init__()
-        self.attention = AttnFactory.create(
+        cfg = asdict(config)
-            attn_type,
+        cfg.update(
-            dim=dim,
+            dim=config.hidden_size,
-            n_heads=n_heads,
+            dim_ffn=config.intermediate_size,
-            n_kv_heads=n_kv_heads,
+            n_layers=config.num_hidden_layers,
-            use_qk_norm=use_qk_norm,
+            n_heads=config.num_attention_heads,
-            norm_eps=norm_eps,
+            n_kv_heads=config.num_key_value_heads,
-            use_gated_attention=use_gated_attention,
+            norm_eps=config.rms_norm_eps,
-            layer_id=layer_id,
+            down_init_std=0.02 / (2 * config.num_hidden_layers) ** 0.5,
-            **kwargs,
         )
-        self.input_norm = RMSNorm(dim, norm_eps)
+        self.attention = AttnFactory.create(config.attn_type, **cfg, layer_id=layer_id)
-        self.post_attention_norm = RMSNorm(dim, norm_eps)
+        self.input_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
-        self.mlp = FFNFactory.create(ffn_type, dim, dim_ffn, **kwargs)
+        self.post_attention_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.mlp = FFNFactory.create(config.ffn_type, **cfg)
 
     def forward(
         self,
         x: Tensor,
         rotary_emb: Tensor,
         attention_mask: Optional[Tensor] = None,
-        paged_cache: Optional[KvcacheView] = None,
+        kv_cache: Optional[KVCache] = None,
+        is_causal: bool = False,
     ) -> Tensor:
         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

astrai/model/components/embedding.py

@@ -1,3 +1,5 @@
+import math
+
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -5,12 +7,20 @@ from torch import Tensor
 
 
 class Embedding(nn.Module):
-    def __init__(self, vocab_size: int, embedding_dim: int):
+    def __init__(self, vocab_size: int, embedding_dim: int, neftune_alpha: float = 0.0):
         super().__init__()
         self.weight = nn.Parameter(torch.empty((vocab_size, embedding_dim)))
+        self.neftune_noise_alpha = neftune_alpha
+
+    def set_neftune_alpha(self, alpha: float):
+        self.neftune_noise_alpha = alpha
 
     def reset_parameters(self):
         nn.init.normal_(self.weight, mean=0.0, std=0.02)
 
     def forward(self, x: Tensor) -> Tensor:
-        return F.embedding(x, self.weight)
+        out = F.embedding(x, self.weight)
+        if self.training and self.neftune_noise_alpha > 0.0:
+            eps = self.neftune_noise_alpha / math.sqrt(out.size(1))
+            out = out + eps * torch.randn_like(out)
+        return out

astrai/model/components/linear.py

@@ -5,13 +5,16 @@ from torch import Tensor
 
 
 class Linear(nn.Module):
-    def __init__(self, in_dim: int, out_dim: int, bias: bool = False):
+    def __init__(
+        self, in_dim: int, out_dim: int, bias: bool = False, init_std: float = 0.02
+    ):
         super().__init__()
         self.weight = nn.Parameter(torch.empty((out_dim, in_dim)))
         self.bias = nn.Parameter(torch.zeros(out_dim)) if bias else None
+        self.init_std = init_std
 
     def reset_parameters(self):
-        nn.init.kaiming_uniform_(self.weight, a=5**0.5)
+        nn.init.normal_(self.weight, mean=0.0, std=self.init_std)
         if self.bias is not None:
             fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight)
             bound = 1 / (fan_in**0.5)

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)
+        device = self.weight.device
-        self.lora_B = nn.Parameter(torch.zeros(self.weight.shape[0], r))
+        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

@@ -8,18 +8,16 @@ from astrai.model.components.linear import Linear
 
 
 class FFNFactory(BaseFactory[nn.Module]):
-    @classmethod
+    pass
-    def create(cls, ffn_type: str, dim: int, dim_ffn: int, **kwargs) -> nn.Module:
-        return super().create(ffn_type, dim, dim_ffn, **kwargs)
 
 
 @FFNFactory.register("mlp")
 class MLP(nn.Module):
-    def __init__(self, dim: int, dim_ffn: int):
+    def __init__(self, dim: int, dim_ffn: int, down_init_std: float = 0.02):
         super().__init__()
         self.up = Linear(dim, dim_ffn)
         self.gate = Linear(dim, dim_ffn)
-        self.down = Linear(dim_ffn, dim)
+        self.down = Linear(dim_ffn, dim, init_std=down_init_std)
 
     def forward(self, x: Tensor) -> Tensor:
         gated = self.up(x) * F.silu(self.gate(x))
@@ -37,6 +35,7 @@ class DeepSeekMoE(nn.Module):
         n_shared_experts: int = 1,
         n_activated_experts: int = 2,
         topk_method: str = "greedy",
+        n_layers: int = 1,
     ):
         super().__init__()
         self.dim = dim
@@ -46,12 +45,20 @@ class DeepSeekMoE(nn.Module):
         self.topk_method = topk_method
 
         self.router = Linear(dim, n_routed_experts, bias=False)
+        moe_scale = 1 / max(n_shared_experts, 1) + 1 / n_activated_experts
+        down_init_std = 0.02 / (2 * n_layers * moe_scale) ** 0.5
 
         self.shared_experts = nn.ModuleList(
-            [MLP(dim, dim_ffn) for _ in range(n_shared_experts)]
+            [
+                MLP(dim, dim_ffn, down_init_std=down_init_std)
+                for _ in range(n_shared_experts)
+            ]
         )
         self.routed_experts = nn.ModuleList(
-            [MLP(dim, dim_ffn) for _ in range(n_routed_experts)]
+            [
+                MLP(dim, dim_ffn, down_init_std=down_init_std)
+                for _ in range(n_routed_experts)
+            ]
         )
 
     def forward(self, x: Tensor) -> Tensor:

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 = get_rotary_emb(self.dim, max_len, self.base)
-        freqs_cis = torch.view_as_real(rotary_emb)
         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 self.freqs_cis[position_ids].float()
-        return torch.view_as_complex(position_freq_cis)

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
@@ -13,35 +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.hidden_size,
+            neftune_alpha=config.neftune_alpha,
         )
-        self.embed_tokens = Embedding(config.vocab_size, config.dim)
 
         self.layers = nn.ModuleList(
             [
-                DecoderBlock(
+                DecoderBlock(config, layer_id)
-                    config.dim,
+                for layer_id in range(config.num_hidden_layers)
-                    config.n_heads,
-                    config.dim_ffn,
-                    config.n_kv_heads,
-                    config.norm_eps,
-                    config.use_qk_norm,
-                    config.use_gated_attention,
-                    layer_id,
-                )
-                for layer_id in range(config.n_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
@@ -69,10 +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 = 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 KvcacheView
+from astrai.inference.core.cache import KVCache
-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.linear import Linear
@@ -15,38 +15,18 @@ from astrai.model.components.rope import RotaryEmbedding
 
 
 def process_attention_mask(
-    input_tensor: Tensor,
+    input_mask: Optional[Tensor],
-    position_ids: Optional[Tensor],
-    input_mask: Optional[Tensor] = None,
-    is_causal: bool = False,
 ) -> 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
-    dtype = input_tensor.dtype
-    B, S = input_tensor.size()[:2]
-    T = position_ids.max().item() + 1
-
     if input_mask is None:
-        if position_ids.min().item() == 0 and is_causal:
+        return None
-            return None
+    if input_mask.dim() == 2:
-        pad = torch.ones(B, T, dtype=torch.bool, device=device)
+        return input_mask[:, None, None, :]
-    else:
+    if input_mask.dim() == 3:
-        pad = input_mask[:, :T].to(device=device, dtype=torch.bool)
+        return input_mask[:, None, :, :]
-
+    return input_mask
-    attend = pad.view(B, 1, T).expand(B, S, T).clone()
-    if is_causal:
-        attend &= position_ids.unsqueeze(-1) >= torch.arange(T, device=device)
-
-    return torch.full(
-        (B, 1, S, T), -torch.finfo(dtype).max / 2, dtype=dtype, device=device
-    ).masked_fill_(attend.unsqueeze(1), 0.0)
 
 
-@AutoModel.register("autoregressive_lm")
+@ModelFactory.register("autoregressive_lm")
 class AutoRegressiveLM(AutoModel):
     """Autoregressive language model with paged KV cache."""
 
@@ -56,43 +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.hidden_size,
+            neftune_alpha=config.neftune_alpha,
         )
-        self.embed_tokens = Embedding(config.vocab_size, config.dim)
 
         self.layers = nn.ModuleList(
             [
-                DecoderBlock(
+                DecoderBlock(config, layer_id)
-                    config.dim,
+                for layer_id in range(config.num_hidden_layers)
-                    config.n_heads,
-                    config.dim_ffn,
-                    config.n_kv_heads,
-                    config.norm_eps,
-                    config.use_qk_norm,
-                    config.use_gated_attention,
-                    layer_id,
-                    attn_type=config.attn_type,
-                    ffn_type=config.ffn_type,
-                    n_routed_experts=config.n_routed_experts,
-                    n_shared_experts=config.n_shared_experts,
-                    n_activated_experts=config.n_activated_experts,
-                    topk_method=config.topk_method,
-                    kv_lora_rank=config.kv_lora_rank,
-                    qk_nope_head_dim=config.qk_nope_head_dim,
-                    qk_rope_head_dim=config.qk_rope_head_dim,
-                )
-                for layer_id in range(config.n_layers)
             ]
         )
 
-        self.norm = RMSNorm(config.dim, config.norm_eps)
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
-        self.lm_head = Linear(config.dim, config.vocab_size)
+        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)
@@ -107,7 +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]
@@ -123,7 +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]
@@ -134,17 +103,18 @@ class AutoRegressiveLM(AutoModel):
         self,
         input_ids: Tensor,
         input_mask: Optional[Tensor] = None,
-        paged_cache: Optional[KvcacheView] = 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
 
         for layer in self.layers:
-            x = layer(x, rotary_emb, attn_mask, paged_cache)
+            x = layer(x, rotary_emb, attn_mask, kv_cache, use_sdpa_causal_mask)
 
         hidden_states = self.norm(x)
         logits = self.lm_head(hidden_states)

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

@@ -2,16 +2,21 @@
 
 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 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
@@ -19,6 +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)
@@ -83,19 +164,28 @@ class BaseExecutor:
 
     def prepare(
         self,
-        model: nn.Module,
+        model_fn: Callable[[], nn.Module],
-        optimizer: Optional[Optimizer] = None,
+        optimizer_fn: Optional[Callable[[nn.Module], Optimizer]] = None,
-        dataloader: Optional[DataLoader] = None,
+        scheduler_fn: Optional[Callable[[Optimizer], LRScheduler]] = None,
-        scheduler: Optional[LRScheduler] = None,
+        before_wrap: Optional[Callable[[nn.Module], nn.Module]] = None,
-    ) -> Tuple[
+        after_wrap: Optional[Callable[[nn.Module], nn.Module]] = None,
-        nn.Module, Optional[Optimizer], Optional[DataLoader], Optional[LRScheduler]
+    ) -> 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:
+            if scheduler is not None:
-            scheduler = AccumScheduler(scheduler, self.gradient_state)
+                scheduler = AccumScheduler(scheduler, self.gradient_state)
-        return model, optimizer, dataloader, scheduler
+        return model, optimizer, scheduler
 
     def _prepare_model(self, model: nn.Module) -> nn.Module:
         return model
@@ -115,8 +205,23 @@ class BaseExecutor:
     def backward(self, loss: torch.Tensor):
         loss.backward()
 
-    def unwrap_model(self, model: nn.Module) -> nn.Module:
+    def unwrap_model(self, model: nn.Module):
-        return model
+        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:
@@ -130,6 +235,12 @@ class BaseExecutor:
     def grad_accum_steps(self) -> int:
         return self.gradient_state.num_steps
 
+    def clip_grad_norm(self, model: nn.Module, max_norm: float) -> float:
+        total_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
+        if isinstance(total_norm, torch.Tensor):
+            return total_norm.item()
+        return total_norm
+
 
 class ExecutorFactory(BaseFactory[BaseExecutor]):
     pass
@@ -180,7 +291,7 @@ class DDPExecutor(BaseExecutor):
         if not self.use_distributed:
             logger.warning("DDP backend selected but world_size=1, model not wrapped")
             return model
-        local_rank = get_rank()
+        local_rank = int(os.environ.get("LOCAL_RANK", get_rank()))
         model = DDP(
             model,
             device_ids=[local_rank],
@@ -195,73 +306,123 @@ class DDPExecutor(BaseExecutor):
             return model.no_sync()
         return contextlib.nullcontext()
 
-    def unwrap_model(self, model: nn.Module) -> nn.Module:
+    def unwrap_model(self, model: nn.Module):
         if isinstance(model, DDP):
-            return model.module
+            return model.module.state_dict()
-        return model
+        return model.state_dict()
 
 
 @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,
+        mesh: Optional[Any] = None,
-        sharding_strategy=None,
+        mp_policy: Optional[Any] = None,
-        cpu_offload=None,
+        reshard_after_forward: bool = False,
-        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,
-        use_orig_params: bool = False,
-        ignored_states=None,
-        device_mesh=None,
     ):
         super().__init__(grad_accum_steps=grad_accum_steps)
-        self._fsdp_kwargs = {
+        self._mesh = mesh
-            k: v
+        self._mp_policy = mp_policy
-            for k, v in dict(
+        self._reshard_after_forward = reshard_after_forward
-                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=use_orig_params,
-                ignored_states=ignored_states,
-                device_mesh=device_mesh,
-            ).items()
-            if v is not None
-        }
-        self._original_model: Optional[nn.Module] = None
 
     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())
+        kwargs = dict(
-        model = FSDP(model, device_id=device_id, **self._fsdp_kwargs)
+            mesh=self._mesh,
-        logger.info("Model wrapped with FSDP (world_size=%d)", get_world_size())
+            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):
+        fsdp_modules = [m for m in model.modules() if isinstance(m, FSDPModule)]
-            return model.no_sync()
+        if fsdp_modules:
-        return contextlib.nullcontext()
+            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 unwrap_model(self, model: nn.Module) -> nn.Module:
+    def clip_grad_norm(self, model: nn.Module, max_norm: float) -> float:
-        if self._original_model is not None:
+        if not self.use_distributed:
-            return self._original_model
+            return super().clip_grad_norm(model, max_norm)
-        if isinstance(model, FSDP):
+
-            return model._fsdp_wrapped_module
+        # FSDP params are DTensors (sharded across ranks).
-        return model
+        # 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 not self.use_distributed:
+            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

@@ -1,12 +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"]
@@ -30,6 +45,7 @@ def get_rank() -> int:
 def setup_parallel(
     rank: int,
     world_size: int,
+    local_rank: int,
     backend: str = "nccl",
     master_addr: str = "localhost",
     master_port: str = "29500",
@@ -41,20 +57,26 @@ def setup_parallel(
         return
 
     if world_size <= 1:
+        device_id = torch.device(device_type, local_rank)
+        os.environ["LOCAL_RANK"] = str(local_rank)
+        os.environ["WORLD_SIZE"] = "1"
+        os.environ["LOCAL_DEVICE"] = str(device_id)
         yield None
         return
 
-    device_id = torch.device(device_type, rank)
+    device_id = torch.device(device_type, local_rank)
 
     os.environ["MASTER_ADDR"] = master_addr
     os.environ["MASTER_PORT"] = master_port
-    os.environ["LOCAL_RANK"] = str(rank)
+    os.environ["LOCAL_RANK"] = str(local_rank)
     os.environ["WORLD_SIZE"] = str(world_size)
     os.environ["LOCAL_DEVICE"] = str(device_id)
 
-    dist.init_process_group(
+    pg_kwargs = dict(rank=rank, world_size=world_size, backend=backend)
-        rank=rank, world_size=world_size, backend=backend, device_id=device_id
+    if backend in ("nccl", "ccl"):
-    )
+        pg_kwargs["device_id"] = device_id
+
+    dist.init_process_group(**pg_kwargs)
 
     try:
         if backend == "nccl" and torch.cuda.is_available():
@@ -90,7 +112,7 @@ def only_on_rank(rank, sync=False):
     return decorator
 
 
-def wrapper_spawn_func(
+def _run_single_rank(
     rank: int,
     world_size: int,
     backend: str,
@@ -100,20 +122,143 @@ def wrapper_spawn_func(
     func: Callable,
     kwargs: dict,
 ):
-    try:
+    install_early_signal_handlers()
+    with setup_parallel(
+        rank=rank,
+        world_size=world_size,
+        local_rank=rank,
+        backend=backend,
+        master_addr=master_addr,
+        master_port=master_port,
+        device_type=device_type,
+    ):
+        func(**kwargs)
+
+
+class LaunchStrategy(ABC):
+    """Strategy for launching a function in a distributed context."""
+
+    def __init__(
+        self,
+        world_size: int,
+        backend: str,
+        master_addr: str,
+        master_port: str,
+        device_type: str,
+        start_method: str,
+    ):
+        self.world_size = world_size
+        self.backend = backend
+        self.master_addr = master_addr
+        self.master_port = master_port
+        self.device_type = device_type
+        self.start_method = start_method
+
+    @abstractmethod
+    def launch(self, func: Callable, **kwargs):
+        raise NotImplementedError
+
+
+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))
         with setup_parallel(
             rank=rank,
             world_size=world_size,
-            backend=backend,
+            local_rank=local_rank,
-            master_addr=master_addr,
+            backend=self.backend,
-            master_port=master_port,
+            master_addr=os.environ.get("MASTER_ADDR", self.master_addr),
-            device_type=device_type,
+            master_port=os.environ.get("MASTER_PORT", self.master_port),
+            device_type=self.device_type,
         ):
             func(**kwargs)
 
-    except Exception as e:
+
-        print(f"Error in rank {rank}: {e}")
+class LocalStrategy(LaunchStrategy):
-        raise
+    """Local launcher — single-process or mp.start_processes."""
+
+    def launch(self, func: Callable, **kwargs):
+        args = (
+            self.world_size,
+            self.backend,
+            self.master_addr,
+            self.master_port,
+            self.device_type,
+            func,
+            kwargs,
+        )
+
+        if self.world_size == 1:
+            _run_single_rank(0, *args)
+            return
+
+        install_early_signal_handlers()
+        ctx = mp.start_processes(
+            _run_single_rank,
+            args=args,
+            nprocs=self.world_size,
+            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() and not parent_stop.is_set():
+                pass
+        except BaseException:
+            logger.warning(
+                "Parent received unexpected exception, terminating children..."
+            )
+            for p in ctx.processes:
+                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:
+    """Detect the distributed launcher from environment.
+
+    Returns one of: "torchelastic", "torchrun", "external", "local".
+    """
+    if dist.is_torchelastic_launched():
+        return "torchelastic"
+    if "LOCAL_WORLD_SIZE" in os.environ:
+        return "torchrun"
+    if "RANK" in os.environ and "WORLD_SIZE" in os.environ:
+        return "external"
+    return "local"
 
 
 def spawn_parallel_fn(
@@ -121,46 +266,20 @@ 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,
 ):
-    # clear environment variables
+    if master_port is None:
-    for key in [
+        master_port = find_free_port()
-        "MASTER_ADDR",
+    launcher = _detect_launcher()
-        "MASTER_PORT",
+    if launcher in ("torchelastic", "torchrun", "external"):
-        "RANK",
+        strategy = TorchrunStrategy(
-        "WORLD_SIZE",
+            world_size, backend, master_addr, master_port, device_type, start_method
-        "LOCAL_RANK",
+        )
-        "LOCAL_DEVICE",
+    else:
-    ]:
+        strategy = LocalStrategy(
-        if key in os.environ:
+            world_size, backend, master_addr, master_port, device_type, start_method
-            del os.environ[key]
+        )
-
+    strategy.launch(func, **kwargs)
-    if world_size == 1:
-        device_id = torch.device(device_type, 0)
-        os.environ["LOCAL_RANK"] = "0"
-        os.environ["WORLD_SIZE"] = "1"
-        os.environ["LOCAL_DEVICE"] = str(device_id)
-
-        func(**kwargs)
-        return
-
-    wrapper_spawn_func_args = (
-        world_size,
-        backend,
-        master_addr,
-        master_port,
-        device_type,
-        func,
-        kwargs,
-    )
-
-    mp.start_processes(
-        wrapper_spawn_func,
-        args=wrapper_spawn_func_args,
-        nprocs=world_size,
-        start_method=start_method,
-        join=True,
-    )

astrai/preprocessing/__init__.py

@@ -0,0 +1,40 @@
+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,
+)
+
+__all__ = [
+    "BaseMaskBuilder",
+    "MaskBuilderFactory",
+    "MultiOutputMaskBuilder",
+    "PackingStrategy",
+    "PackingStrategyFactory",
+    "Pipeline",
+    "PositionIdStrategy",
+    "PositionIdStrategyFactory",
+    "SectionedMaskBuilder",
+    "SingleOutputMaskBuilder",
+    "StoreWriter",
+    "StoreWriterFactory",
+    "TokenizeTransform",
+    "filter_by_length",
+    "plan_bfd",
+]

astrai/preprocessing/builder.py

@@ -0,0 +1,537 @@
+"""Mask building for preprocessing pipeline.
+
+:class:`SectionRenderer` converts section specs into token ids and loss
+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
+from typing import Optional
+
+from astrai.factory import BaseFactory
+
+
+def _extract_domain(item: dict, domain_key: Optional[str]) -> str:
+    if not domain_key:
+        return "__default__"
+    val = item.get(domain_key, "__default__")
+    return val if isinstance(val, str) else "__default__"
+
+
+def _resolve_action(action: str, role: str, config) -> str:
+    if action == "$role":
+        return config.mask.get(role, config.mask_default)
+    return action
+
+
+class SectionRenderer:
+    """Render section specs into ``(ids, loss_mask)`` tuples."""
+
+    def process_sections(
+        self,
+        item: dict,
+        sections: list,
+        config,
+        tokenizer,
+        *,
+        is_top_level: bool = False,
+    ):
+        all_ids: list[int] = []
+        loss_mask: list[int] = []
+
+        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
+        )
+
+        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)
+
+        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:
+                success = self._append_template(
+                    item, field, action, tokenizer, config, all_ids, loss_mask
+                )
+                if not success:
+                    continue
+            else:
+                success = self._append_text(
+                    item,
+                    field,
+                    action,
+                    tokenizer,
+                    add_special,
+                    is_text_config,
+                    config,
+                    all_ids,
+                    loss_mask,
+                )
+                if not success:
+                    continue
+
+            first_section = False
+
+        max_len = config.preprocessing.max_seq_len
+        all_ids = all_ids[:max_len]
+        loss_mask = loss_mask[: len(all_ids)]
+
+        if not all_ids:
+            return None, None
+
+        if is_top_level and has_template and len(all_ids) <= 1:
+            return None, None
+
+        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):
+        """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"]
+            action = sec["action"]
+            use_template = sec.get("template", False)
+
+            values = item.get(field)
+            if not isinstance(values, list):
+                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, ids, mask
+                        )
+                else:
+                    wrapper = {field: str(val)}
+                    self._append_text(
+                        wrapper,
+                        field,
+                        action,
+                        tokenizer,
+                        False,
+                        False,
+                        config,
+                        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)
+
+        if not per_item_ids:
+            return None, None
+        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:
+        return len(sections) == 1 and sections[0].get("action") == "value"
+
+    @staticmethod
+    def extract_raw_value(item: dict, sections: list):
+        sec = sections[0]
+        field = sec["field"]
+        raw = item.get(field)
+        if raw is None:
+            return None
+        if isinstance(raw, list):
+            return [float(v) for v in raw]
+        return [float(raw)]
+
+    def _append_template(
+        self, item, field, action, tokenizer, config, all_ids, loss_mask
+    ):
+        messages = item.get(field)
+        if not isinstance(messages, list) or not messages:
+            return False
+        for msg in messages:
+            role = msg.get("role", "")
+            act = _resolve_action(action, role, config)
+            rendered = tokenizer.apply_chat_template(
+                [msg], tokenize=False, add_generation_prompt=False
+            )
+            ids = tokenizer.encode(rendered, add_special_tokens=False)
+            all_ids.extend(ids)
+            val = 1 if act == "train" else 0
+            loss_mask.extend([val] * len(ids))
+        return True
+
+    def _append_text(
+        self,
+        item,
+        field,
+        action,
+        tokenizer,
+        add_special,
+        is_text_config,
+        config,
+        all_ids,
+        loss_mask,
+    ):
+        text = str(item.get(field, ""))
+        if not text.strip():
+            return False
+        if is_text_config:
+            pp = config.preprocessing
+            if pp.min_chars > 0 and len(text) < pp.min_chars:
+                return False
+            if len(text) > pp.max_chars:
+                return False
+        ids = tokenizer.encode(text, add_special_tokens=add_special)
+        all_ids.extend(ids)
+        val = 1 if action == "train" else 0
+        loss_mask.extend([val] * len(ids))
+        return True
+
+
+class BaseMaskBuilder(ABC):
+    """Convert a JSONL item into token ids and optional loss_mask."""
+
+    @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("single")
+class SingleOutputMaskBuilder(BaseMaskBuilder):
+    """Build a single output sequence with optional loss mask.
+
+    Expects ``config.input.sections`` (list of section specs).
+    """
+
+    def __init__(self, renderer: Optional[SectionRenderer] = None):
+        self.renderer = renderer or SectionRenderer()
+
+    def build(self, item: dict, config, tokenizer) -> Optional[dict]:
+        sections = config.input.sections
+        if not sections:
+            return None
+
+        ids, mask = self.renderer.process_sections(
+            item, sections, config, tokenizer, is_top_level=True
+        )
+        if ids is None:
+            return None
+
+        result: dict = {
+            "sequence": ids,
+            "domain": _extract_domain(item, config.output.domain_key),
+        }
+        if not all(m == 1 for m in mask):
+            result["loss_mask"] = mask
+        return result
+
+    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
+
+        for output_key, spec in sources_spec.items():
+            sections = spec.get("sections", [])
+            if not sections:
+                continue
+
+            if self.renderer.is_value_section(sections):
+                ids = self.renderer.extract_raw_value(item, sections)
+                if ids is None:
+                    continue
+                result[output_key] = ids
+                any_output = True
+                continue
+
+            list_field = spec.get("list_field", False)
+            mask_key = spec.get("mask_key", f"{output_key}_mask")
+
+            if list_field:
+                ids, mask = self.renderer.process_list_field(
+                    item, sections, config, tokenizer
+                )
+                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
+                any_output = True
+                continue
+
+            ids, mask = self.renderer.process_sections(
+                item, sections, config, tokenizer, is_top_level=True
+            )
+
+            if ids is None:
+                continue
+
+            result[output_key] = ids
+            if not all(m == 1 for m in mask):
+                result[mask_key] = mask
+            elif "mask_key" in spec:
+                result[mask_key] = mask
+
+            any_output = True
+
+        if not any_output:
+            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]
+        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 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

@@ -0,0 +1,176 @@
+"""Sequence packing strategies for shard-level reordering and truncation.
+
+Each strategy receives the accumulated ``{key: [list of token lists]}``
+dict for a shard and returns a reordered / truncated version.  The
+pipeline later flattens the result into contiguous tensors.
+"""
+
+from abc import ABC, abstractmethod
+from typing import Dict, List
+
+from astrai.factory import BaseFactory
+
+
+def _truncate(seq: List[int], max_len: int, mode: str) -> List[int]:
+    if len(seq) <= max_len:
+        return seq
+    if mode == "keep_end":
+        return seq[-max_len:]
+    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."""
+
+    @abstractmethod
+    def apply(
+        self,
+        keys: Dict[str, List[List[int]]],
+        max_packed_len: int,
+        truncation_mode: str,
+    ) -> Dict[str, List[List[int]]]:
+        raise NotImplementedError
+
+
+class PackingStrategyFactory(BaseFactory["PackingStrategy"]):
+    pass
+
+
+@PackingStrategyFactory.register("simple")
+class SimplePacking(PackingStrategy):
+    def apply(
+        self,
+        keys: Dict[str, List[List[int]]],
+        max_packed_len: int,
+        truncation_mode: str,
+    ) -> Dict[str, List[List[int]]]:
+        return {
+            k: [_truncate(v, max_packed_len, truncation_mode) for v in vals]
+            for k, vals in keys.items()
+        }
+
+
+@PackingStrategyFactory.register("bfd")
+class BFDPacking(PackingStrategy):
+    """Best-Fit Decreasing bin packing.
+
+    Assigns sequences to bins using a best-fit heuristic (sorted by
+    decreasing length) and concatenates sequences within each bin into
+    a single packed sequence.  Packed sequences are truncated to
+    *max_packed_len* so that each packed bin fits within one context
+    window during training.
+    """
+
+    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
+        bins = plan_bfd(sequences, max_packed_len, truncation_mode)
+
+        packed: Dict[str, List[List[int]]] = {}
+        for k, vals in keys.items():
+            packed[k] = [
+                _truncate(
+                    self._concat_bin(vals, bin_indices),
+                    max_packed_len,
+                    truncation_mode,
+                )
+                for bin_indices in bins
+            ]
+        return packed
+
+    @staticmethod
+    def _concat_bin(vals: List[List[int]], indices: List[int]) -> List[int]:
+        result: List[int] = []
+        for i in indices:
+            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 _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

@@ -0,0 +1,277 @@
+"""Config-driven JSONL preprocessing pipeline.
+
+Composes a :class:`BaseMaskBuilder` (selected by ``input.type``) with
+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
+import logging
+import os
+from collections import defaultdict
+from itertools import chain
+from typing import Dict, List, Optional
+
+import torch
+import tqdm
+
+from astrai.config.preprocess_config import PipelineConfig
+from astrai.preprocessing.core import (
+    build_preprocessing_components,
+    primary_ids,
+)
+from astrai.preprocessing.packing import PackingStrategyFactory
+from astrai.preprocessing.writer import StoreWriterFactory
+
+logger = logging.getLogger(__name__)
+
+_STR_TO_DTYPE: dict[str, torch.dtype] = {
+    "bool": torch.bool,
+    "uint8": torch.uint8,
+    "int8": torch.int8,
+    "int16": torch.int16,
+    "int32": torch.int32,
+    "int64": torch.int64,
+    "float16": torch.float16,
+    "float32": torch.float32,
+    "float64": torch.float64,
+}
+
+
+def filter_by_length(text: str, min_len: int = 50, max_len: int = 2_000_000) -> bool:
+    return min_len <= len(text) <= max_len
+
+
+class Pipeline:
+    """Tokenization pipeline driven by a declarative :class:`PipelineConfig`.
+
+    Usage::
+
+        config = PipelineConfig.from_file("sft_pipeline.json")
+        Pipeline(config, ["data.jsonl"], output_dir="out", tokenizer_path="params").run()
+    """
+
+    def __init__(
+        self,
+        config: PipelineConfig,
+        input_paths: list[str],
+        output_dir: str,
+        tokenizer_path: str,
+    ):
+        os.makedirs(output_dir, exist_ok=True)
+        self.config = config
+        self.paths = input_paths
+        self.output_dir = output_dir
+        self.tokenizer_path = tokenizer_path
+
+        self.tokenizer, self.mask_builder, self._position_id = (
+            build_preprocessing_components(config, tokenizer_path)
+        )
+        self._packer = PackingStrategyFactory.create(
+            config.preprocessing.packing_strategy
+        )
+        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)
+
+    def transform_batch(self, items: list[dict]) -> list[Optional[dict]]:
+        return self.mask_builder.build_batch(items, self.config, self.tokenizer)
+
+    def run(self):
+        domains: dict = defaultdict(lambda: defaultdict(list))
+        total_tokens = 0
+        shard_idx: dict[str, int] = defaultdict(int)
+        count = 0
+
+        pp = self.config.preprocessing
+
+        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:
+                results = self.transform_batch(items)
+            except Exception:
+                logger.warning(
+                    "Failed to process batch, retrying records individually",
+                    exc_info=True,
+                )
+                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)
+
+            for result in results:
+                if pp.max_items and count >= pp.max_items:
+                    stop = True
+                    break
+                if result is None:
+                    continue
+
+                domain = result.pop("domain", "__default__")
+                ids = primary_ids(result)
+                if not ids:
+                    continue
+
+                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)
+
+                if total_tokens >= self.config.output.max_tokens_per_shard:
+                    self._flush(domains, shard_idx)
+                    domains.clear()
+                    total_tokens = 0
+            if stop:
+                break
+
+        progress.close()
+
+        if total_tokens > 0:
+            self._flush(domains, shard_idx)
+
+    @staticmethod
+    def _align_bucket(bucket: dict, result: dict, ids: list):
+        """Pad previously-accumulated keys that are missing from *result*."""
+        for key in list(bucket.keys()):
+            if key in result:
+                continue
+            bucket[key].append([0] * len(ids))
+
+    def _iter_items(self):
+        for path in self.paths:
+            with open(path, "r", encoding="utf-8") as f:
+                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():
+            idx = shard_idx[domain]
+
+            pp = self.config.preprocessing
+            original_sequences = keys.get("sequence", [])
+            mode = self.config.output.position_ids_mode
+
+            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 = 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
+
+            first_key = "sequence" if "sequence" in tensors else next(iter(tensors))
+            tqdm.tqdm.write(
+                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/position_id.py

@@ -0,0 +1,46 @@
+"""Position-id generation strategies for packed sequences.
+
+Each strategy takes the list of per-document token sequences after packing
+and returns a flat list of position ids (same total length as all
+sequences combined).  The pipeline wraps the result into a tensor and
+attaches it as ``position_ids``.
+"""
+
+from abc import ABC, abstractmethod
+from typing import List
+
+from astrai.factory import BaseFactory
+
+
+class PositionIdStrategy(ABC):
+    """Generate ``position_ids`` for packed sequences."""
+
+    @abstractmethod
+    def generate(self, sequences: List[List[int]]) -> List[int]:
+        raise NotImplementedError
+
+
+class PositionIdStrategyFactory(BaseFactory["PositionIdStrategy"]):
+    pass
+
+
+@PositionIdStrategyFactory.register("none")
+class NoPositionId(PositionIdStrategy):
+    def generate(self, sequences: List[List[int]]) -> List[int]:
+        return []
+
+
+@PositionIdStrategyFactory.register("doc_reset")
+class DocResetPositionId(PositionIdStrategy):
+    def generate(self, sequences: List[List[int]]) -> List[int]:
+        pos_ids = []
+        for seq in sequences:
+            pos_ids.extend(range(len(seq)))
+        return pos_ids
+
+
+@PositionIdStrategyFactory.register("continuous")
+class ContinuousPositionId(PositionIdStrategy):
+    def generate(self, sequences: List[List[int]]) -> List[int]:
+        total = sum(len(seq) for seq in sequences)
+        return list(range(total))

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

@@ -0,0 +1,56 @@
+"""Storage writer strategies for pipeline output.
+
+The :class:`StoreWriter` abstraction decouples the pipeline from the
+concrete storage format (bin).  The pipeline builds a ``{key:
+List[Tensor]}`` dict and delegates the write to the writer selected
+by ``output.storage_format``.
+"""
+
+import logging
+import os
+import shutil
+from abc import ABC, abstractmethod
+from typing import Dict, List
+
+import torch
+
+from astrai.factory import BaseFactory
+from astrai.serialization import save_bin
+
+logger = logging.getLogger(__name__)
+
+
+class StoreWriter(ABC):
+    """Write pre-tokenized tensors to disk in a format-specific way."""
+
+    @abstractmethod
+    def save(
+        self,
+        output_dir: str,
+        domain: str,
+        shard_idx: int,
+        tensors: Dict[str, List[torch.Tensor]],
+    ) -> None: ...
+
+
+class StoreWriterFactory(BaseFactory["StoreWriter"]):
+    pass
+
+
+@StoreWriterFactory.register("bin")
+class BinWriter(StoreWriter):
+    def save(self, output_dir, domain, shard_idx, tensors):
+        shard_path = os.path.join(output_dir, domain, f"shard_{shard_idx:04d}")
+        try:
+            save_bin(shard_path, tensors)
+        except Exception:
+            if os.path.exists(shard_path):
+                shutil.rmtree(shard_path, ignore_errors=True)
+            logger.error(
+                "Failed to write shard %s/%s_%04d, cleaned up partial output",
+                domain,
+                "shard",
+                shard_idx,
+                exc_info=True,
+            )
+            raise

astrai/serialization/__init__.py

@@ -0,0 +1,41 @@
+"""Serialization utilities for models and datasets.
+
+This package re-exports checkpoint helpers and dataset storage helpers so
+that existing imports from ``astrai.serialization`` continue to work.
+"""
+
+from astrai.serialization.checkpoint import (
+    Checkpoint,
+    load_json,
+    load_model_config,
+    load_model_weights,
+    load_safetensors,
+    load_state_dict,
+    load_torch,
+    save_json,
+    save_model,
+    save_safetensors,
+    save_torch,
+)
+from astrai.serialization.dataset import (
+    load_bin,
+    load_bin_offsets,
+    save_bin,
+)
+
+__all__ = [
+    "Checkpoint",
+    "load_json",
+    "load_model_config",
+    "load_model_weights",
+    "load_safetensors",
+    "load_state_dict",
+    "load_torch",
+    "save_json",
+    "save_model",
+    "save_safetensors",
+    "save_torch",
+    "load_bin",
+    "load_bin_offsets",
+    "save_bin",
+]

astrai/serialization/checkpoint.py

@@ -1,9 +1,11 @@
+"""Model checkpoint serialization helpers."""
+
 import io
 import json
 import time
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any, Dict, Union
+from typing import Any, Dict, Optional, Union
 
 import safetensors.torch as st
 import torch
@@ -136,7 +138,7 @@ def load_state_dict(path: Union[str, Path], broadcast: bool = False) -> dict:
 class Checkpoint:
     state_dict: Dict[str, Any] = field(default_factory=dict)
     epoch: int = 0
-    iteration: int = 0
+    consumed_samples: int = 0
     extra: Dict[str, Any] = field(default_factory=dict)
     meta: Dict[str, Any] = field(default_factory=dict)
     config: Dict[str, Any] = field(default_factory=dict)
@@ -145,12 +147,9 @@ class Checkpoint:
         save_path = Path(save_dir)
         save_path.mkdir(parents=True, exist_ok=True)
 
-        if get_rank() != 0:
-            return
-
         meta = {
             "epoch": self.epoch,
-            "iteration": self.iteration,
+            "consumed_samples": self.consumed_samples,
             "timestamp": time.strftime("%Y-%m-%dT%H:%M:%S"),
             **self.meta,
         }
@@ -176,7 +175,27 @@ class Checkpoint:
         return cls(
             state_dict=state_dict,
             epoch=meta.get("epoch", 0),
-            iteration=meta.get("iteration", 0),
+            consumed_samples=meta.get("consumed_samples", 0),
             extra=extra,
+            meta=meta,
             config=config,
         )
+
+    @classmethod
+    def load_any(cls, save_dir: str, broadcast: bool = False) -> Optional["Checkpoint"]:
+        save_path = Path(save_dir)
+        meta_path = save_path / _META_FILE
+        weights_path = save_path / _WEIGHTS_FILE
+
+        if meta_path.exists():
+            return cls.load(save_dir, broadcast=broadcast)
+
+        if weights_path.exists():
+            state_dict = load_state_dict(weights_path, broadcast=broadcast)
+            config = {}
+            config_path = save_path / _CONFIG_FILE
+            if config_path.exists():
+                config = load_json(config_path, broadcast)
+            return cls(state_dict=state_dict, config=config)
+
+        return None

astrai/serialization/dataset.py

@@ -0,0 +1,82 @@
+"""Dataset storage serialization helpers (memory-mapped binary)."""
+
+import json
+import os
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+import torch
+from torch import 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)
+        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)
+
+
+def load_bin(file_path: str) -> Dict[str, List[Tensor]]:
+    with open(os.path.join(file_path, "meta.json"), "r") as f:
+        meta = json.load(f)
+    segments: Dict[str, List[Tensor]] = {}
+    for key, info in meta.items():
+        arr = np.memmap(
+            os.path.join(file_path, f"{key}.bin"),
+            dtype=info["dtype"],
+            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

@@ -1,13 +1,11 @@
-from dataclasses import dataclass
+from functools import cached_property
 from typing import Any, Dict, List, Optional
 
 from jinja2 import Template
 
-# Message type for chat messages
 type MessageType = Dict[str, Any]
 
 
-@dataclass
 class ChatTemplate:
     """A chat template with Jinja2 rendering support.
 
@@ -15,23 +13,51 @@ class ChatTemplate:
         name: Unique identifier for the template.
         template_str: Jinja2 template string.
         description: Optional description.
-        default_variables: Optional dictionary of default variable values
+        default_variables: Optional dictionary of default variable values.
-            that will be passed to the template if not overridden during rendering.
         special_tokens: Optional dictionary mapping token names to their string values.
-            These tokens are automatically added to the template variables.
     """
 
-    name: str
+    def __init__(
-    template_str: str
+        self,
-    description: str = ""
+        name: str = "",
-    default_variables: Dict[str, Any] = None
+        template_str: str = "",
-    special_tokens: Dict[str, str] = None
+        description: str = "",
+        default_variables: Optional[Dict[str, Any]] = None,
+        special_tokens: Optional[Dict[str, str]] = None,
+    ):
+        self.name = name
+        self.template_str = template_str
+        self.description = description
+        self.default_variables = default_variables or {}
+        self.special_tokens = special_tokens or {}
 
-    def __post_init__(self):
+    @cached_property
-        if self.default_variables is None:
+    def _compiled(self) -> Template:
-            self.default_variables = {}
+        """Lazy-compiled Jinja2 template, cached on first access.
-        if self.special_tokens is None:
+
-            self.special_tokens = {}
+        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(
@@ -43,7 +69,7 @@ class ChatTemplate:
     ) -> "ChatTemplate":
         """Create a ChatTemplate instance directly from a template string."""
         return cls(
-            name="",  # empty name for ad‑hoc templates
+            name="",
             template_str=template_str,
             description=description,
             default_variables=default_variables,
@@ -73,5 +99,4 @@ class ChatTemplate:
         if system_prompt is not None:
             variables["system_prompt"] = system_prompt
 
-        jinja_template = Template(self.template_str)
+        return self._compiled.render(**variables)
-        return jinja_template.render(**variables)

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(
+        encoded_list = self._tokenizer.encode_batch(
-                tokens,
+            tokens,
-                is_pretokenized=is_pretokenized,
+            is_pretokenized=is_pretokenized,
-                add_special_tokens=add_special_tokens,
+            add_special_tokens=add_special_tokens,
-            )
+        )
-            return [
+        return [encoded.ids if out_ids else encoded.tokens for encoded in encoded_list]
-                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]]:
+    ) -> Union[str, List[int], List[str], List[List[int]]]:
-        """
+        """Apply the chat template and optionally tokenize.
-        Apply the chat template to messages and optionally tokenize the result.
+
+        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'.
+            messages: Single conversation (``Messages``) or batch of
-            system_prompt: Optional system prompt string (auto-converted to first message).
+                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).
+            add_generation_prompt: Whether to add the generation prompt.
-            **kwargs: Additional variables to pass to the template.
+            **kwargs: Additional template variables.
 
         Returns:
-            Either the rendered string or list of token IDs.
+            Single mode: ``str`` or ``List[int]``.
-
+            Batch mode: ``List[str]`` or ``List[List[int]]``.
-        Raises:
-            RuntimeError: If chat template is not set.
         """
         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/__init__.py

@@ -1,4 +1,3 @@
-from astrai.trainer.optim import Muon
 from astrai.trainer.schedule import BaseScheduler, SchedulerFactory
 from astrai.trainer.strategy import BaseStrategy, StrategyFactory
 from astrai.trainer.train_callback import (
@@ -10,8 +9,6 @@ from astrai.trainer.trainer import Trainer
 __all__ = [
     # Main trainer
     "Trainer",
-    # Optimizer
-    "Muon",
     # Strategy factory
     "StrategyFactory",
     "BaseStrategy",

astrai/trainer/metric_util.py

@@ -1,42 +1,25 @@
-from typing import Any, Callable, Dict
+from typing import Dict
 
 import torch
 import torch.nn as nn
 
 
-def _grad_stat(
+def grad_norm(model: nn.Module, per_param: bool = False) -> float | Dict[str, float]:
-    model: nn.Module, fn: Callable[[torch.Tensor], Any], default: Any
+    grads = [p.grad.detach() for p in model.parameters() if p.grad is not None]
-) -> dict:
+    if not grads:
-    results = {}
+        return 0.0
-    for name, param in model.named_parameters():
-        results[name] = default
-        if param.grad is not None:
-            results[name] = fn(param.grad.data)
-    return results
 
-
+    total_sq = torch.stack([g.pow(2).sum() for g in grads]).sum()
-def grad_norm(model: nn.Module, norm_type: int = 2) -> Dict[str, float]:
+    if per_param:
-    return _grad_stat(model, lambda g: g.norm(norm_type).item(), 0.0)
+        norms = {}
-
+        for name, param in model.named_parameters():
-
+            if param.grad is not None:
-def grad_std(model: nn.Module) -> Dict[str, float]:
+                norms[name] = param.grad.norm(2).item()
-    return _grad_stat(model, lambda g: g.std().item(), 0.0)
+            else:
-
+                norms[name] = 0.0
-
+        norms["total"] = total_sq.sqrt().item()
-def grad_max(model: nn.Module) -> Dict[str, float]:
+        return norms
-    return _grad_stat(model, lambda g: g.max().item(), -float("inf"))
+    return total_sq.sqrt().item()
-
-
-def grad_min(model: nn.Module) -> Dict[str, float]:
-    return _grad_stat(model, lambda g: g.min().item(), float("inf"))
-
-
-def grad_mean(model: nn.Module) -> Dict[str, float]:
-    return _grad_stat(model, lambda g: g.mean().item(), 0.0)
-
-
-def grad_nan_num(model: nn.Module) -> Dict[str, int]:
-    return _grad_stat(model, lambda g: g.isnan().sum().item(), 0)
 
 
 def ctx_get_loss(ctx):
@@ -52,24 +35,4 @@ def ctx_get_val_loss(ctx):
 
 
 def ctx_get_grad_norm(ctx):
-    return grad_norm(ctx.model)
+    return ctx.grad_norm
-
-
-def ctx_get_grad_std(ctx):
-    return grad_std(ctx.model)
-
-
-def ctx_get_grad_max(ctx):
-    return grad_max(ctx.model)
-
-
-def ctx_get_grad_min(ctx):
-    return grad_min(ctx.model)
-
-
-def ctx_get_grad_mean(ctx):
-    return grad_mean(ctx.model)
-
-
-def ctx_get_grad_nan_num(ctx):
-    return grad_nan_num(ctx.model)

astrai/trainer/optim.py

@@ -1,143 +0,0 @@
-import torch
-from torch.optim import Optimizer
-
-
-def _zeropower_via_newtonschulz(G: torch.Tensor, steps: int = 5):
-    assert G.ndim == 2
-    X = G
-    scale = max(1, G.size(0) / G.size(1)) ** 0.5
-    X = X / (X.norm() + 1e-7) * scale
-    if steps == 0:
-        return X
-    a, b, c = (3.4445, -4.7750, 2.0315)
-    for _ in range(steps):
-        A = X @ X.T
-        B = A @ X
-        X = a * X + b * B + c * (A @ B)
-    return X
-
-
-class Muon(Optimizer):
-    def __init__(
-        self,
-        params,
-        lr: float = 2e-3,
-        momentum: float = 0.95,
-        weight_decay: float = 0.0,
-        nesterov: bool = True,
-        ns_steps: int = 5,
-        adamw_lr: float = None,
-        adamw_betas: tuple = (0.9, 0.95),
-        adamw_eps: float = 1e-8,
-        adamw_wd: float = 0.0,
-    ):
-        defaults = dict(
-            lr=lr,
-            momentum=momentum,
-            weight_decay=weight_decay,
-            nesterov=nesterov,
-            ns_steps=ns_steps,
-            adamw_lr=adamw_lr if adamw_lr is not None else lr * 0.1,
-            adamw_betas=adamw_betas,
-            adamw_eps=adamw_eps,
-            adamw_wd=adamw_wd,
-        )
-        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:
-            params_2d, params_1d = [], []
-            grads_2d, grads_1d = [], []
-
-            for p in group["params"]:
-                if p.grad is None:
-                    continue
-                if p.grad.is_sparse:
-                    raise RuntimeError("Muon does not support sparse gradients")
-                if p.ndim >= 2:
-                    params_2d.append(p)
-                    grads_2d.append(p.grad)
-                else:
-                    params_1d.append(p)
-                    grads_1d.append(p.grad)
-
-            if params_2d:
-                self._muon_update_foreach(params_2d, grads_2d, group)
-            if params_1d:
-                self._adamw_update_foreach(params_1d, grads_1d, group)
-
-        return loss
-
-    def _muon_update_foreach(self, params_2d, grads_2d, group):
-        lr = group["lr"]
-        momentum = group["momentum"]
-        wd = group["weight_decay"]
-        nesterov = group["nesterov"]
-        ns_steps = group["ns_steps"]
-
-        if wd != 0:
-            torch._foreach_mul_(params_2d, 1 - lr * wd)
-
-        if nesterov:
-            grads_2d = torch._foreach_add(grads_2d, params_2d, alpha=wd)
-
-        bufs = []
-        for p, grad in zip(params_2d, grads_2d):
-            state = self.state[p]
-            if "momentum_buffer" not in state:
-                state["momentum_buffer"] = torch.zeros_like(grad)
-            bufs.append(state["momentum_buffer"])
-
-        torch._foreach_lerp_(bufs, grads_2d, 1 - momentum)
-
-        for p, buf in zip(params_2d, bufs):
-            update = _zeropower_via_newtonschulz(buf, steps=ns_steps)
-            scale = max(1, p.size(0) / p.size(1)) ** 0.5
-            p.add_(update, alpha=-lr * scale)
-
-    def _adamw_update_foreach(self, params_1d, grads_1d, group):
-        lr = group["adamw_lr"]
-        betas = group["adamw_betas"]
-        eps = group["adamw_eps"]
-        wd = group["adamw_wd"]
-
-        steps: list[int] = []
-        exp_avgs, exp_avg_sqs = [], []
-        has_state = []
-        for p in params_1d:
-            state = self.state[p]
-            if not state:
-                state["step"] = 0
-                state["exp_avg"] = torch.zeros_like(p)
-                state["exp_avg_sq"] = torch.zeros_like(p)
-                has_state.append(False)
-            else:
-                has_state.append(True)
-            state["step"] += 1
-            steps.append(state["step"])
-            exp_avgs.append(state["exp_avg"])
-            exp_avg_sqs.append(state["exp_avg_sq"])
-
-        beta1, beta2 = betas
-
-        torch._foreach_lerp_(exp_avgs, grads_1d, 1 - beta1)
-        grads_sq = torch._foreach_mul(grads_1d, grads_1d)
-        torch._foreach_lerp_(exp_avg_sqs, grads_sq, 1 - beta2)
-
-        bias_correction1 = [1 - beta1**s for s in steps]
-        bias_correction2 = [1 - beta2**s for s in steps]
-
-        if wd != 0:
-            torch._foreach_mul_(params_1d, 1 - lr * wd)
-
-        exp_avg_corrected = torch._foreach_div(exp_avgs, bias_correction1)
-        denom = torch._foreach_div(exp_avg_sqs, bias_correction2)
-        denom = torch._foreach_sqrt(denom)
-        torch._foreach_add_(denom, eps)
-        torch._foreach_addcdiv_(params_1d, exp_avg_corrected, denom, value=-lr)

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/schedule.py

@@ -2,7 +2,7 @@
 
 import math
 from abc import ABC, abstractmethod
-from typing import Any, Dict, List, Type
+from typing import Any, Dict, List
 
 from torch.optim.lr_scheduler import LRScheduler
 
@@ -31,7 +31,6 @@ class SchedulerFactory(BaseFactory["BaseScheduler"]):
     """Factory class for creating learning rate schedulers.
 
     Supports decorator-based registration for extensible scheduler types.
-    Also supports creation from ScheduleConfig objects.
 
     Example usage:
         @SchedulerFactory.register("custom")
@@ -41,33 +40,6 @@ class SchedulerFactory(BaseFactory["BaseScheduler"]):
         scheduler = SchedulerFactory.create("custom", optimizer, **kwargs)
     """
 
-    @classmethod
-    def _validate_component(cls, scheduler_cls: Type[BaseScheduler]):
-        """Validate that the scheduler class inherits from BaseScheduler."""
-        if not issubclass(scheduler_cls, BaseScheduler):
-            raise TypeError(f"{scheduler_cls.__name__} must inherit from BaseScheduler")
-
-    @classmethod
-    def create(
-        cls, optimizer, schedule_type: str = "none", **kwargs
-    ) -> "BaseScheduler":
-        """Create a scheduler instance by type name.
-
-        Args:
-            optimizer: PyTorch optimizer
-            schedule_type: Type of scheduler ("cosine", "sgdr")
-            **kwargs: Arguments passed to the scheduler constructor
-
-        Returns:
-            Scheduler instance
-        """
-        return super().create(schedule_type, optimizer, **kwargs)
-
-    @classmethod
-    def available_types(cls) -> list:
-        """Return list of registered scheduler type names."""
-        return cls.list_registered()
-
 
 # ----------- Scheduler implementations -----------
 
@@ -81,7 +53,7 @@ class CosineScheduler(BaseScheduler):
         optimizer,
         warmup_steps: int,
         lr_decay_steps: int,
-        min_rate: float = 0.05,
+        min_rate: float = 0.01,
         last_epoch: int = -1,
     ):
         self.warmup_steps = warmup_steps
@@ -93,11 +65,15 @@ class CosineScheduler(BaseScheduler):
     def get_lr(self) -> List[float]:
         # warmup
         if self.last_epoch < self.warmup_steps:
-            warmup_factor = max(self.min_rate, self.last_epoch / self.warmup_steps)
+            warmup_factor = max(
+                self.min_rate, self.last_epoch / max(self.warmup_steps, 1)
+            )
             return [base_lr * warmup_factor for base_lr in self.base_lrs]
 
         # cosine decay
-        decay_progress = (self.last_epoch - self.warmup_steps) / self.lr_decay_steps
+        decay_progress = (self.last_epoch - self.warmup_steps) / max(
+            self.lr_decay_steps, 1
+        )
         decay_progress = min(decay_progress, 1.0)
         cosine_decay = 0.5 * (1.0 + math.cos(math.pi * decay_progress))
         decay_factor = max(self.min_rate, cosine_decay)
@@ -132,7 +108,7 @@ class SGDRScheduler(BaseScheduler):
         optimizer,
         warmup_steps: int,
         cycle_length: int,
-        min_rate: float = 0.05,
+        min_rate: float = 0.01,
         t_mult: int = 2,
         last_epoch: int = -1,
     ):
@@ -146,7 +122,9 @@ class SGDRScheduler(BaseScheduler):
     def get_lr(self):
         # warmup
         if self.last_epoch < self.warmup_steps:
-            warmup_factor = max(self.min_rate, self.last_epoch / self.warmup_steps)
+            warmup_factor = max(
+                self.min_rate, self.last_epoch / max(self.warmup_steps, 1)
+            )
             return [base_lr * warmup_factor for base_lr in self.base_lrs]
 
         # SGDR
@@ -192,3 +170,66 @@ class SGDRScheduler(BaseScheduler):
         self.min_rate = state_dict.pop("min_rate")
         self.t_mult = state_dict.pop("t_mult")
         super().load_state_dict(state_dict)
+
+
+@SchedulerFactory.register("wsd")
+class WSDScheduler(BaseScheduler):
+    """WSD (Warmup-Stable-Decay) scheduler with sqrt cooldown.
+
+    warmup_steps:  linear warmup from min_rate to 1.0
+    stable_steps:   constant at base_lr
+    decay_steps:    sqrt decay from base_lr to min_rate
+    min_rate:       minimum lr as fraction of base_lr (default 0.0)
+    """
+
+    def __init__(
+        self,
+        optimizer,
+        warmup_steps: int,
+        stable_steps: int,
+        decay_steps: int,
+        min_rate: float = 0.01,
+        last_epoch: int = -1,
+    ):
+        self.warmup_steps = warmup_steps
+        self.stable_steps = stable_steps
+        self.decay_steps = decay_steps
+        self.min_rate = min_rate
+        self.total_steps = warmup_steps + stable_steps + decay_steps
+        super().__init__(optimizer, last_epoch)
+
+    def get_lr(self) -> List[float]:
+        if self.last_epoch < self.warmup_steps:
+            factor = max(self.min_rate, self.last_epoch / max(self.warmup_steps, 1))
+            return [base_lr * factor for base_lr in self.base_lrs]
+
+        offset = self.last_epoch - self.warmup_steps
+
+        if offset < self.stable_steps:
+            return list(self.base_lrs)
+
+        decay_ratio = (offset - self.stable_steps) / max(self.decay_steps, 1)
+        decay_ratio = min(decay_ratio, 1.0)
+        factor = (1.0 - self.min_rate) * (1.0 - decay_ratio) ** 2 + self.min_rate
+        return [base_lr * factor for base_lr in self.base_lrs]
+
+    def state_dict(self):
+        state = super().state_dict()
+        state.update(
+            {
+                "warmup_steps": self.warmup_steps,
+                "stable_steps": self.stable_steps,
+                "decay_steps": self.decay_steps,
+                "min_rate": self.min_rate,
+                "total_steps": self.total_steps,
+            }
+        )
+        return state
+
+    def load_state_dict(self, state_dict):
+        self.warmup_steps = state_dict.pop("warmup_steps")
+        self.stable_steps = state_dict.pop("stable_steps")
+        self.decay_steps = state_dict.pop("decay_steps")
+        self.min_rate = state_dict.pop("min_rate")
+        self.total_steps = state_dict.pop("total_steps")
+        super().load_state_dict(state_dict)

astrai/trainer/strategy.py

@@ -1,57 +1,37 @@
 """Training strategy implementations with factory pattern."""
 
-import copy
 from abc import ABC, abstractmethod
-from typing import Any, Callable, Dict, Union
+from typing import Callable, Dict, Union
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from torch import Tensor
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
-from torch.nn.parallel import DistributedDataParallel as DDP
 
 from astrai.factory import BaseFactory
+from astrai.parallel.executor import broadcast_state_dict
+from astrai.trainer.rollout import RolloutResult
 
 
-def unwrap_model(model: nn.Module) -> nn.Module:
+def move_to_device(batch: Dict[str, Tensor], device: str) -> Dict[str, Tensor]:
-    if isinstance(model, DDP):
-        return model.module
-    if isinstance(model, FSDP):
-        return model._fsdp_wrapped_module
-    return model
-
-
-def create_ref_model(model: nn.Module) -> nn.Module:
-    """Create a reference model for DPO/GRPO training.
-
-    Handles DDP-wrapped models safely by unwrapping first,
-    then creating a deep copy with frozen gradients.
-    """
-    original_model = unwrap_model(model)
-    ref_model = copy.deepcopy(original_model)
-    ref_model.requires_grad_(False)
-    ref_model.eval()
-    return ref_model
-
-
-def move_to_device(batch: Dict[str, Tensor], device: str) -> Any:
     """Move batch tensors to specified device with non-blocking transfer."""
     return {key: value.to(device, non_blocking=True) for key, value in batch.items()}
 
 
 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:
     """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:
@@ -64,9 +44,12 @@ 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"]
+    logits = model(
+        input_ids[:, :-1],
+        attn_mask[:, :, :-1, :-1] if attn_mask.dim() == 4 else attn_mask[:, :-1],
+    )["logits"]
     log_probs = torch.log_softmax(logits.float(), dim=-1)
 
     token_logprobs = torch.gather(
@@ -74,24 +57,53 @@ def get_logprobs(
     ).squeeze(-1)
 
     if reduction == "mean":
-        return (token_logprobs * shifted_mask).sum(dim=-1) / shifted_mask.sum(
+        return (token_logprobs * shifted_loss_mask).sum(dim=-1) / shifted_loss_mask.sum(
             dim=-1
         ).clamp(min=1.0)
     elif reduction == "sum":
-        return (token_logprobs * shifted_mask).sum(dim=-1)
+        return (token_logprobs * shifted_loss_mask).sum(dim=-1)
     else:
-        return token_logprobs * shifted_mask
+        return token_logprobs * shifted_loss_mask
+
+
+def make_doc_boundary_mask(position_ids: Tensor) -> Tensor:
+    S = position_ids.size(1)
+    device = position_ids.device
+    boundaries = position_ids[:, 1:] <= position_ids[:, :-1]
+    doc_ids = torch.cat(
+        [
+            torch.zeros(position_ids.size(0), 1, dtype=torch.long, device=device),
+            boundaries.long().cumsum(dim=1),
+        ],
+        dim=1,
+    )
+    same_doc = doc_ids.unsqueeze(-1) == doc_ids.unsqueeze(-2)
+    causal = torch.tril(torch.ones(S, S, dtype=torch.bool, device=device))
+    return (same_doc & causal).unsqueeze(1)
 
 
 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, model: Union[Callable[..., Dict[str, Tensor]]], device: str, **kwargs
+        self,
+        model: Union[nn.Module, Callable[..., Dict[str, Tensor]]],
+        device: str,
+        **kwargs,
     ):
         self.model = model
         self.device = device
+        self.executor = kwargs.pop("executor", None)
         self.extra_kwargs = kwargs
+        self._rollout_runner = None
 
     @abstractmethod
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
@@ -105,9 +117,53 @@ class BaseStrategy(ABC):
         """
         raise NotImplementedError
 
+    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 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]) -> Tensor:
-        """Allow calling strategy directly as a callable."""
+        """Run offline or online forward depending on runner injection."""
-        return self.compute_loss(batch)
+        if self._rollout_runner is None:
+            return self.compute_loss(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(train_batch)
 
 
 class StrategyFactory(BaseFactory["BaseStrategy"]):
@@ -124,32 +180,6 @@ class StrategyFactory(BaseFactory["BaseStrategy"]):
         strategy = StrategyFactory.create("custom", model, device)
     """
 
-    @classmethod
-    def _validate_component(cls, strategy_cls: type):
-        """Validate that the strategy class inherits from BaseStrategy."""
-        if not issubclass(strategy_cls, BaseStrategy):
-            raise TypeError(f"{strategy_cls.__name__} must inherit from BaseStrategy")
-
-    @classmethod
-    def create(cls, train_type: str, model, device: str, **kwargs) -> "BaseStrategy":
-        """Create a strategy instance based on training type.
-
-        Args:
-            train_type: Type of training ("seq", "sft", "dpo", "grpo")
-            model: Model instance for the strategy
-            device: Device to run the strategy on
-            **kwargs: Additional arguments passed to strategy constructor
-
-        Returns:
-            Strategy instance
-        """
-        return super().create(train_type, model, device, **kwargs)
-
-    @classmethod
-    def available_strategies(cls) -> list:
-        """Return list of registered strategy names."""
-        return cls.list_registered()
-
 
 # ============== Strategy Classes ==============
 # All strategies are registered at class definition time using the decorator
@@ -162,7 +192,13 @@ class SEQStrategy(BaseStrategy):
     Computes cross-entropy loss for next token prediction.
     """
 
-    def __init__(self, model, device, label_smoothing: float = 0.0, **kwargs):
+    def __init__(
+        self,
+        model: Union[nn.Module, Callable[..., Dict[str, Tensor]]],
+        device: str,
+        label_smoothing: float = 0.0,
+        **kwargs,
+    ):
         super().__init__(model, device, **kwargs)
         self.label_smoothing = label_smoothing
 
@@ -187,21 +223,31 @@ class SFTStrategy(BaseStrategy):
     Applies cross-entropy loss only to tokens where loss_mask is True.
     """
 
-    def __init__(self, model, device, label_smoothing: float = 0.0, **kwargs):
+    def __init__(
+        self,
+        model: Union[nn.Module, Callable[..., Dict[str, Tensor]]],
+        device: str,
+        label_smoothing: float = 0.0,
+        **kwargs,
+    ):
         super().__init__(model, device, **kwargs)
         self.label_smoothing = label_smoothing
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
         batch = move_to_device(batch, self.device)
-        input_ids, target_ids, loss_mask = (
+        input_ids, target_ids, position_ids, loss_mask = (
             batch["input_ids"],
             batch["target_ids"],
+            batch["position_ids"],
             batch["loss_mask"],
         )
 
         ignore_index = -100
-        logits = self.model(input_ids=input_ids)["logits"]
+        input_mask = make_doc_boundary_mask(position_ids)
-        target_ids = target_ids.masked_fill(loss_mask == 0, ignore_index)
+        target_ids = target_ids.masked_fill(~loss_mask, ignore_index)
+        logits = self.model(
+            input_ids=input_ids, position_ids=position_ids, input_mask=input_mask
+        )["logits"]
 
         loss = F.cross_entropy(
             input=logits.flatten(0, 1).float(),
@@ -225,12 +271,13 @@ 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(model)
+        self.ref_model = ref_model
         self.beta = beta
         self.reduction = reduction
 
@@ -240,13 +287,31 @@ class DPOStrategy(BaseStrategy):
         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
+
+        log_pi = get_logprobs(
+            self.model,
+            concat_ids,
+            full_mask,
+            concat_loss_mask,
+            self.reduction,
+        )
 
         with torch.no_grad():
             log_ref = get_logprobs(
-                self.ref_model, concat_ids, concat_mask, self.reduction
+                self.ref_model,
+                concat_ids,
+                full_mask,
+                concat_loss_mask,
+                self.reduction,
             )
 
         log_pi_chosen = log_pi[: chosen_ids.shape[0]]
@@ -262,46 +327,77 @@ class DPOStrategy(BaseStrategy):
 
         return dpo_loss
 
+    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
+    Implements GRPO following DeepSeek-R1 with token-level PPO clipping.
-    a frozen ref_model stores the old-policy log-probs.  ratio = exp(logπ_θ - logπ_ref),
+    Advantages are group-normalized from scalar per-response rewards and
-    clipped PPO objective.  Call ``sync_ref_model()`` after each data-generation round.
+    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(model)
+        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):
+    def sync_old_model(self):
-        """Copy current model weights to ref model."""
+        """Copy current policy weights to old model."""
-        ref_state = self.model.state_dict()
+        state_dict = self.executor.unwrap_model(self.model)
-        self.ref_model.load_state_dict(ref_state)
+        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()
-
         batch = move_to_device(batch, self.device)
         prompts = batch["prompts"]
         responses = batch["responses"]
@@ -312,33 +408,101 @@ 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)
+        # Prompt tokens are masked out (0) so logprobs are computed only for
-
+        # response tokens.  get_logprobs shifts the mask by one position, so
-        log_probs_policy = get_logprobs(
+        # the first response token's logprob (predicted from the last prompt
-            self.model, full_sequences, full_masks, self.reduction
+        # 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).
+        token_log_probs_policy = get_logprobs(
+            self.model, full_sequences, attn_mask, full_masks, "none"
+        )[:, prompt_len - 1 :]
         with torch.no_grad():
-            log_probs_ref = get_logprobs(
+            token_log_probs_old = get_logprobs(
-                self.ref_model, full_sequences, full_masks, self.reduction
+                self.old_model, full_sequences, attn_mask, full_masks, "none"
-            )
+            )[:, prompt_len - 1 :]
-            log_probs_ref = log_probs_ref.view(batch_size, group_size)
+            token_log_probs_ref = get_logprobs(
+                self.ref_model, full_sequences, attn_mask, full_masks, "none"
+            )[:, 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
+
+        # 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
 
-        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
 
         return total_loss
+
+    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

@@ -9,21 +9,15 @@ from typing import IO, Callable, List, Optional, Protocol, runtime_checkable
 import torch
 import torch.distributed as dist
 import torch.nn as nn
-from torch.nn.utils import clip_grad_norm_
 from torch.utils.checkpoint import checkpoint as torch_checkpoint
 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_grad_max,
-    ctx_get_grad_mean,
-    ctx_get_grad_min,
-    ctx_get_grad_nan_num,
     ctx_get_grad_norm,
-    ctx_get_grad_std,
     ctx_get_loss,
     ctx_get_lr,
     ctx_get_val_loss,
@@ -86,7 +80,9 @@ class GradientClippingCallback(TrainCallback):
         self.max_grad_norm = max_grad_norm
 
     def on_optimizer_step(self, context: TrainContext):
-        clip_grad_norm_(context.model.parameters(), self.max_grad_norm)
+        context.grad_norm = context.executor.clip_grad_norm(
+            context.model, self.max_grad_norm
+        )
 
 
 @CallbackFactory.register("gradient_checkpointing")
@@ -143,33 +139,38 @@ 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_iter = 0
+        self.last_ckpt_step = None
+
+    def on_train_begin(self, context: TrainContext):
+        self.last_ckpt_step = context.optimizer_step
 
     def _save_checkpoint(self, context: TrainContext):
-        unwrapped = context.executor.unwrap_model(context.model)
+        self.last_ckpt_step = context.optimizer_step
-        state_dict = unwrapped.state_dict()
-        self.last_ckpt_iter = context.iteration
 
-        if get_rank() == 0:
+        with context.executor.checkpoint_context(context.model) as state_dict:
-            save_path = os.path.join(
+            if state_dict is not None:
-                self.save_dir, f"epoch_{context.epoch}_iter_{context.iteration}"
+                save_path = os.path.join(
-            )
+                    self.save_dir,
-            extra = self.save_extra_fn(context)
+                    f"epoch_{context.epoch}_step_{context.optimizer_step}",
-            context.checkpoint = Checkpoint(
+                )
-                state_dict=state_dict,
+                extra = self.save_extra_fn(context)
-                epoch=context.epoch,
+                meta = context.config.to_dict()
-                iteration=context.iteration,
+                context.checkpoint = Checkpoint(
-                extra=extra,
+                    state_dict=state_dict,
-                config=context.model_config,
+                    epoch=context.epoch,
-            )
+                    consumed_samples=context.consumed_samples,
-            context.checkpoint.save(save_path)
+                    config=context.model_config,
+                    extra=extra,
+                    meta=meta,
+                )
+                context.checkpoint.save(save_path)
 
     def on_batch_end(self, context: TrainContext):
-        if context.iteration - self.last_ckpt_iter >= self.interval:
+        if context.optimizer_step - self.last_ckpt_step >= self.interval:
             self._save_checkpoint(context)
 
     def on_train_end(self, context: TrainContext):
-        if context.iteration != self.last_ckpt_iter:
+        if context.optimizer_step != self.last_ckpt_step:
             self._save_checkpoint(context)
 
     def on_error(self, context: TrainContext):
@@ -201,20 +202,24 @@ class ProgressBarCallback(TrainCallback):
 
     @only_on_rank(0)
     def on_epoch_begin(self, context: TrainContext):
+        total_steps = len(context.dataloader) // context.executor.grad_accum_steps
         self.progress_bar = tqdm(
-            context.dataloader,
+            total=total_steps,
             desc=f"Epoch {context.epoch + 1}/{self.num_epoch}",
             dynamic_ncols=True,
             file=self.file or sys.stdout,
         )
 
     @only_on_rank(0)
-    def on_batch_end(self, context: TrainContext):
+    def on_optimizer_step(self, context: TrainContext):
         postfix = {
+            "step": f"{context.optimizer_step:d}",
             "loss": f"{context.loss:.4f}",
             "lr": f"{context.optimizer.param_groups[-1]['lr']:.2e}",
         }
-        if context.val_loss > 0:
+        if context.grad_norm is not None:
+            postfix["grad_norm"] = f"{context.grad_norm:.2f}"
+        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)
@@ -226,22 +231,22 @@ class ProgressBarCallback(TrainCallback):
             self.progress_bar.close()
 
 
-@CallbackFactory.register("metric_logger")
+@CallbackFactory.register("metric")
-class MetricLoggerCallback(TrainCallback):
+class MetricCallback(TrainCallback):
     def __init__(
         self,
-        log_dir: str,
+        ckpt_dir: str,
         save_interval: int,
-        log_interval: int = 10,
         metrics: List[str] = None,
+        val_step: int = 0,
     ):
-        self.last_log_iter = 0
+        self.last_log_flush_step = None
         self.save_interval = save_interval
-        self.log_interval = log_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.ckpt_dir = Path(ckpt_dir) if ckpt_dir else Path.cwd() / "checkpoint"
-        self.log_dir.mkdir(parents=True, exist_ok=True)
 
         self.log_cache = []
 
@@ -250,53 +255,28 @@ class MetricLoggerCallback(TrainCallback):
             "lr": ctx_get_lr,
             "val_loss": ctx_get_val_loss,
             "grad_norm": ctx_get_grad_norm,
-            "grad_std": ctx_get_grad_std,
-            "grad_max": ctx_get_grad_max,
-            "grad_min": ctx_get_grad_min,
-            "grad_mean": ctx_get_grad_mean,
-            "grad_nan_num": ctx_get_grad_nan_num,
         }
 
-    def _get_log_data(self, context: TrainContext):
+    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
+        }
+
+    @only_on_rank(0)
+    def _append(self, event_type: str, context: TrainContext, **extra):
+        entry = {
+            "type": event_type,
             "timestamp": time.strftime("%Y-%m-%dT%H:%M:%S"),
             "epoch": context.epoch,
-            "iter": context.iteration,
+            "step": context.optimizer_step,
-            **{m: self._metric_funcs[m](context) for m in self.metrics},
+            "consumed_samples": context.consumed_samples,
+            **extra,
         }
+        self.log_cache.append(entry)
 
-    @only_on_rank(0)
+    def _run_validation(self, context: TrainContext) -> float:
-    def _add_log(self, log_data):
-        self.log_cache.append(log_data)
-
-    @only_on_rank(0)
-    def _save_log(self, epoch, iter):
-        log_file = self.log_dir / f"epoch_{epoch}_iter_{iter}_metric.jsonl"
-
-        with open(log_file, "w") as f:
-            for log in self.log_cache:
-                f.write(json.dumps(log) + "\n")
-
-    def on_batch_end(self, context):
-        if context.iteration % self.log_interval == 0:
-            log_data = self._get_log_data(context)
-            self._add_log(log_data)
-
-        if context.iteration - self.last_log_iter >= self.save_interval:
-            self._save_log(context.epoch, context.iteration)
-            self.last_log_iter = context.iteration
-
-    def on_train_end(self, context):
-        if context.iteration != self.last_log_iter:
-            self._save_log(context.epoch, context.iteration)
-
-    def on_error(self, context):
-        self._save_log(context.epoch, context.iteration)
-
-
-@CallbackFactory.register("validation")
-class ValidationCallback(TrainCallback):
-    def _run_validation(self, context: TrainContext):
         context.model.eval()
 
         total_loss = 0.0
@@ -308,27 +288,56 @@ class ValidationCallback(TrainCallback):
                 total_loss += loss.item()
                 num_batches += 1
 
-        avg_loss = total_loss / max(num_batches, 1)
-
         if context.world_size > 1 and dist.is_initialized():
-            loss_tensor = torch.tensor([avg_loss], device=get_current_device())
+            stats = torch.tensor(
-            dist.all_reduce(loss_tensor, op=dist.ReduceOp.AVG)
+                [total_loss, float(num_batches)], device=get_current_device()
-            avg_loss = loss_tensor.item()
+            )
+            dist.all_reduce(stats, op=dist.ReduceOp.SUM)
+            avg_loss = (stats[0] / stats[1]).item()
+        else:
+            avg_loss = total_loss / max(num_batches, 1)
 
-        context.val_loss = avg_loss
         context.model.train()
+        return avg_loss
 
-        step_count = context.iteration // context.config.grad_accum_steps
+    def on_train_begin(self, context: TrainContext):
-        logger.info(
+        self.last_log_flush_step = context.optimizer_step
-            f"Epoch {context.epoch + 1}, Step {step_count}, Val Loss: {avg_loss:.4f}"
-        )
 
-    def on_optimizer_step(self, context: TrainContext):
+    @only_on_rank(0)
-        if context.val_dataloader is None:
+    def _flush(self, epoch, step):
-            return
+        log_file = self.ckpt_dir / f"epoch_{epoch}_step_{step}" / "metric.jsonl"
-        cfg = context.config
+        log_file.parent.mkdir(parents=True, exist_ok=True)
-        if cfg.val_step <= 0:
+        with open(log_file, "w") as f:
-            return
+            for log in self.log_cache:
-        step_count = context.iteration // cfg.grad_accum_steps
+                f.write(json.dumps(log) + "\n")
-        if step_count % cfg.val_step == 0:
+
-            self._run_validation(context)
+    def on_optimizer_step(self, context):
+        if (
+            context.val_dataloader is not None
+            and self.val_step > 0
+            and context.optimizer_step >= self._next_val_step
+        ):
+            context.val_loss = self._run_validation(context)
+            self._next_val_step = context.optimizer_step + self.val_step
+            self._append("validation", context, val_loss=context.val_loss)
+
+        step_metrics = [m for m in self.metrics if m != "val_loss"]
+        self._append("step", context, **self._metrics(context, step_metrics))
+
+        if context.optimizer_step - self.last_log_flush_step >= self.save_interval:
+            self._flush(context.epoch, context.optimizer_step)
+            self.last_log_flush_step = context.optimizer_step
+
+    def on_epoch_end(self, context):
+        self._append("epoch", context)
+
+    def on_train_end(self, context):
+        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

@@ -1,19 +1,27 @@
+import logging
+import threading
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Optional, Self
+from typing import Any, Dict, Optional, Self
 
+import torch
 import torch.nn as nn
-from torch.utils.data import DataLoader
+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, load_model_weights
+from astrai.serialization import Checkpoint, load_json
+from astrai.tokenize import AutoTokenizer
+from astrai.trainer.rollout import RolloutGenerator, RolloutRunner
 from astrai.trainer.strategy import BaseStrategy, StrategyFactory
 
+logger = logging.getLogger(__name__)
+
 
 @dataclass
 class TrainContext:
@@ -26,16 +34,33 @@ class TrainContext:
     config: TrainConfig = field(default=None)
     model_config: dict = field(default_factory=dict)
     executor: BaseExecutor = field(default=None)
-
     epoch: int = field(default=0)
-    iteration: int = field(default=0)
+    consumed_samples: int = field(default=0)
     loss: float = field(default=0.0)
-    val_dataloader: DataLoader = field(default=None)
+    grad_norm: Optional[float] = field(default=None)
-    val_loss: float = field(default=0.0)
+    val_dataloader: Optional[DataLoader] = field(default=None)
+    val_loss: Optional[float] = field(default=None)
 
     world_size: int = field(default=1)
     rank: int = field(default=0)
-    kwargs: dict = field(default_factory=dict)
+    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 // (
+            self.config.batch_per_device
+            * self.world_size
+            * self.config.grad_accum_steps
+        )
 
 
 class TrainContextBuilder:
@@ -44,10 +69,12 @@ 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:
+    def with_param_path(self, param_path: Optional[str], resume: bool = False) -> Self:
-        self._resume_dir = resume_dir
+        self._param_path = param_path
+        self._resume = resume
         return self
 
     def build(self) -> TrainContext:
@@ -60,96 +87,129 @@ class TrainContextBuilder:
             **cfg.executor_kwargs,
         )
 
-        model = cfg.model_fn()
-        model = model.to(device=device)
-
         model_config = {}
-        if self._resume_dir:
+        if self._param_path:
-            config_path = Path(self._resume_dir) / "config.json"
+            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"):
+        preloaded_state_dict = None
-            model_config = model.config.to_dict()
+        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 is not None:
+        context.model, context.optimizer, context.scheduler = executor.prepare(
-            resume_path = Path(self._resume_dir)
+            cfg.model_fn,
-            if (resume_path / "meta.json").exists():
+            cfg.optimizer_fn,
-                checkpoint = Checkpoint.load(self._resume_dir)
+            cfg.scheduler_fn,
-                state_dict = checkpoint.state_dict
+            before_wrap=_before_wrap,
-                if checkpoint.config:
+            after_wrap=_after_wrap,
-                    context.model_config = checkpoint.config
+        )
-            else:
-                checkpoint = None
-                state_dict = load_model_weights(self._resume_dir)
-            model.load_state_dict(state_dict, strict=False)
-            if checkpoint is not None:
-                context.epoch = cfg.start_epoch
-                context.iteration = cfg.start_batch
-            context.checkpoint = checkpoint
 
-        if cfg.lora is not None:
+        train_dataset = cfg.dataset
-            inject_lora(
+        val_dataset = cfg.val_dataset
-                model,
+
-                r=cfg.lora.r,
+        if val_dataset is None and cfg.val_split is not None:
-                alpha=cfg.lora.alpha,
+            n_total = len(cfg.dataset)
-                target_modules=set(cfg.lora.target_modules),
+            n_val = max(1, int(n_total * cfg.val_split))
+            n_train = n_total - n_val
+            generator = torch.Generator().manual_seed(cfg.random_seed)
+            train_dataset, val_dataset = random_split(
+                cfg.dataset, [n_train, n_val], generator=generator
             )
 
-        context.optimizer = cfg.optimizer_fn(model)
+        sampler_offset = context.consumed_samples // context.world_size
-        context.scheduler = cfg.scheduler_fn(context.optimizer)
 
-        sampler_offset = context.iteration * cfg.batch_per_device
+        if self._resume and sampler_offset > 0:
-        sampler = ResumableDistributedSampler(
+            offset = context.world_size - 1
-            data_source=cfg.dataset,
+            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,
             seed=cfg.random_seed,
         )
         context.dataloader = DataLoader(
-            cfg.dataset,
+            train_dataset,
             batch_size=cfg.batch_per_device,
             sampler=sampler,
             num_workers=cfg.num_workers,
             pin_memory=cfg.pin_memory,
             prefetch_factor=cfg.prefetch_factor,
+            collate_fn=cfg.collate_fn,
         )
 
-        if cfg.val_dataset is not None:
+        if val_dataset is not None:
-            val_sampler = ResumableDistributedSampler(
+            val_sampler = RDSampler(
-                data_source=cfg.val_dataset,
+                data_source=val_dataset,
                 start_epoch=0,
                 start_iter=0,
                 seed=cfg.random_seed,
                 shuffle=False,
             )
             context.val_dataloader = DataLoader(
-                cfg.val_dataset,
+                val_dataset,
                 batch_size=cfg.batch_per_device,
                 sampler=val_sampler,
                 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"):
@@ -158,11 +218,72 @@ class TrainContextBuilder:
                     if obj is not None:
                         obj.load_state_dict(extra[name])
 
+        strategy_kwargs = dict(cfg.extra_kwargs)
+
+        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,
-            train_type=cfg.strategy,
             device=device,
-            **cfg.extra_kwargs,
+            executor=executor,
+            **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,
@@ -35,15 +41,14 @@ class Trainer:
                 cfg.ckpt_interval,
             ),
             CallbackFactory.create(
-                "metric_logger",
+                "metric",
-                log_dir=cfg.log_dir,
+                ckpt_dir=cfg.ckpt_dir,
                 save_interval=cfg.ckpt_interval,
-                log_interval=cfg.log_interval,
                 metrics=cfg.metrics,
+                val_step=cfg.val_step,
             ),
             CallbackFactory.create("progress_bar", cfg.n_epoch),
             CallbackFactory.create("gradient_clipping", cfg.max_grad_norm),
-            CallbackFactory.create("validation"),
         ]
         return callbacks
 
@@ -53,10 +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)
 
@@ -64,23 +72,29 @@ 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:
-                    self._call_callbacks("on_batch_begin", context)
+                    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
                         executor.backward(stand_loss)
-                        context.iteration += 1
+                        context.consumed_samples += (
+                            context.config.batch_per_device * context.world_size
+                        )
                         self._call_callbacks("on_batch_end", context)
 
                         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 +102,23 @@ 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 +128,6 @@ 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/__init__.py

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

csrc/build.py

@@ -0,0 +1,75 @@
+from pathlib import Path
+
+
+def cuda_toolkit_version() -> tuple[int, int] | None:
+    """Return ``(major, minor)`` of the nvcc on PATH, or ``None``.
+
+    Used by ``setup.py`` to detect nvcc/torch CUDA version mismatches
+    (e.g. nvcc 13.0 with a cu128 torch wheel) which cause cryptic ABI errors.
+    """
+    import shutil
+    import subprocess
+
+    nvcc = shutil.which("nvcc")
+    if nvcc is None:
+        return None
+    try:
+        out = subprocess.check_output(
+            [nvcc, "--version"], stderr=subprocess.STDOUT, text=True
+        )
+        for line in out.splitlines():
+            if "release" in line:
+                ver = line.split("release")[1].split(",")[0].strip()
+                major, minor = ver.split(".")
+                return (int(major), int(minor))
+    except Exception:
+        pass
+    return None
+
+
+def _arch_flags() -> list[str]:
+    import torch
+
+    if torch.cuda.is_available():
+        cap = torch.cuda.get_device_capability()
+    else:
+        cap = (8, 0)
+    ver = f"{cap[0]}{cap[1]}"
+    flags = [f"-gencode=arch=compute_{ver},code=sm_{ver}"]
+    # tensor-core mma path (mma.sync.m16n8k16.bf16) requires sm_80+; decide the
+    # kernel dispatch at build time via this define rather than at runtime.
+    if cap[0] < 8:
+        flags.append("-DASTRAI_NO_MMA")
+    return flags
+
+
+_kernels_dir = Path("csrc/kernels")
+REGISTRY: dict[str, dict] = {}
+
+CXX_FLAGS = ["-O3", "-funroll-loops"]
+NVCC_FLAGS = [
+    "-O3",
+    "--expt-relaxed-constexpr",
+    "--use_fast_math",
+    "--ptxas-options=-O3,-v",
+    "--extra-device-vectorization",
+    "--threads=16",
+]
+
+
+def register(name: str, sources: list[str] | None = None, **kwargs):
+    if sources is None:
+        sources = [str(_kernels_dir / f"{name}.cu")]
+    REGISTRY[name] = {
+        "sources": sources,
+        "cxx_flags": [*CXX_FLAGS],
+        "nvcc_flags": [*NVCC_FLAGS, *_arch_flags()],
+        "extra_link_args": kwargs.pop("extra_link_args", []),
+        **kwargs,
+    }
+
+
+register("attn_decode")
+register("attn_prefill")
+register("attn_paged_decode")
+register("rotary_emb")

csrc/kernels/attn_common.h

@@ -0,0 +1,71 @@
+#pragma once
+
+
+template<typename T, typename AT = float>
+struct AttentionParams {
+    int batch;
+    int q_head;
+    int kv_head;
+    int q_len;
+    int kv_len;
+    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;
+    // KV strides (K and V share the same layout — only base pointers differ)
+    int kv_stride_b, kv_stride_h, kv_stride_l, kv_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 T* __restrict__ q;
+    const T* __restrict__ k;
+    const T* __restrict__ v;
+    const bool* __restrict__ mask;
+
+    T* __restrict__ o;
+    AT* __restrict__ o_part;
+    AT* __restrict__ ml_part;
+};
+
+template<typename T, typename AT = float>
+struct PagedAttentionParams {
+    int batch;
+    int q_head;
+    int kv_head;
+    int q_len;
+    int kv_len;
+    int head_dim;
+    int use_mask;
+    int causal_offset;
+    float scale;
+
+    int num_splits;
+    int page_size;
+    int max_pages;
+
+    // Q strides (layout-agnostic)
+    int q_stride_b, q_stride_h, q_stride_l, q_stride_d;
+
+    // Mask strides (2D, 3D, or 4D)
+    int mask_b_stride;
+    int mask_h_stride;
+    int mask_q_stride;
+
+    const T* __restrict__ q;
+    const T* __restrict__ k_cache;
+    const T* __restrict__ v_cache;
+    const bool* __restrict__ mask;
+    const int64_t* __restrict__ page_table;
+
+    T* __restrict__ o;
+    AT* __restrict__ o_part;
+    AT* __restrict__ ml_part;
+};

csrc/kernels/attn_decode.cu

@@ -0,0 +1,37 @@
+#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
+) {
+    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 == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    alloc_split_partials(p);
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_decode, p);
+    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") = 0,
+        "GQA decode (tensor-core head-packing on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_decode_split_kv.cuh

@@ -0,0 +1,129 @@
+#pragma once
+#include <cuda_bf16.h>
+#include <float.h>
+#include "attn_common.h"
+#include "attn_warp_utils.cuh"
+constexpr int DC_CHUNK = 64;
+
+template <int HEAD_DIM, 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;
+
+    // Q: [batch, q_head, q_len=1, head_dim] — stride-based
+    float q_reg[8];
+    int q_off = batch * p.q_stride_b + q_head * p.q_stride_h
+              + 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]);
+
+    // KV: [batch, kv_head, kv_len, head_dim] — stride-based base
+    int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    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 k_smem[];
+
+    // Split-KV: each split processes a contiguous subset of chunks
+    int chunks_total = (p.kv_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, p.kv_len - chunk_start);
+
+        // Load K into shared memory (gather from strided global)
+        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 kv_idx = chunk_start + s;
+            int g_off = kv_base + kv_idx * p.kv_stride_l + d_dim * p.kv_stride_d;
+            k_smem[i] = p.k[g_off];
+        }
+        __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 > p.causal_offset)
+                    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;
+
+            int v_off = kv_base + kv_idx * p.kv_stride_l
+                        + lane * hd_per_thread * p.kv_stride_d;
+            for (int i = 0; i < hd_per_thread; i++)
+                acc_reg[i] = fmaf(acc_reg[i], alpha,
+                                  __bfloat162float(p.v[v_off + i * p.kv_stride_d]) * 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;
+    }
+}
+
+__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 = batch * p.q_stride_b + q_head * p.q_stride_h + d * p.q_stride_d;
+    p.o[o_off] = __float2bfloat16(acc * inv);
+}

csrc/kernels/attn_decode_split_kv_mma.cuh

@@ -0,0 +1,161 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+#include "attn_warp_utils.cuh"
+
+// Split-K (FlashDecoding) tensor-core decode via GQA head-packing.
+// 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.
+//
+// IsCausal and HasMask are compile-time bools — no runtime branch in the
+// inner compute loop.
+//
+// Traits = KernelTraits<HEAD_DIM, BC=32, WARPS=1, STAGES=<2 or 1>>.
+template <typename Traits, 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;
+
+    // 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.
+    // stride_row = p.q_stride_h for decode (q_len=1).
+    const int q_base = batch * p.q_stride_b + q_head0 * p.q_stride_h;
+    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 kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    const int tiles_total = (p.kv_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 ----
+    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 < p.kv_len;
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            int g_off = kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
+            cp_async_16_pred(&dK[off], &p.k[g_off], valid);
+            cp_async_16_pred(&dV[off], &p.v[g_off], valid);
+        }
+        cp_async_commit();
+    };
+
+    constexpr int BUF_MASK = (Traits::STAGES > 1) ? (Traits::STAGES - 1) : 0;
+
+    // Prologue
+    if (ti_begin < ti_end) {
+        load_tile(ti_begin, 0);
+    }
+
+    for (int ti = ti_begin; ti < ti_end; ti++) {
+        int buf = (ti - ti_begin) & BUF_MASK;
+
+        cp_async_wait_group<0>();
+        __syncwarp();
+        if constexpr (Traits::STAGES > 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, (ti + 1 - ti_begin) & BUF_MASK);
+        }
+
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = ti * 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
+        int maxc = IsCausal ? min(p.kv_len, p.causal_offset + 1) : p.kv_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);
+        __syncwarp();
+
+        if constexpr (Traits::STAGES == 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, 0);
+        }
+    }
+
+    // ---- 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 +1,195 @@
+#pragma once
+// Shared attention dispatchers — used by both production .cu and test .cu.
+// No torch dependency; pure CUDA.
+
+#include <cuda_runtime.h>
+#include <algorithm>
+#include "attn_warp_utils.cuh"
+#include "attn_prefill_split_q.cuh"
+#include "attn_decode_split_kv.cuh"
+#include "attn_paged_decode_split_kv.cuh"
+#ifndef ASTRAI_NO_MMA
+#include "attn_prefill_split_q_mma.cuh"
+#include "attn_decode_split_kv_mma.cuh"
+#include "attn_paged_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.
+inline int compute_num_splits(int base_blocks, int tiles_total,
+                              int min_tiles_per_split = 1) {
+    int sm_count = 0;
+    cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, 0);
+    int n = (2 * sm_count + 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)));
+}
+
+// ======================================================================
+// Prefill
+// ======================================================================
+
+#ifndef ASTRAI_NO_MMA
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_prefill_mma(AttentionParams<bf16>& p) {
+    constexpr int WARPS = 4;
+    constexpr int BC = (HEAD_DIM <= 128) ? 32 : 16;
+    using Traits = KernelTraits<HEAD_DIM, BC, WARPS, 2>;
+    dim3 grid((p.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, IsCausal, HasMask><<<grid, block>>>(p);
+}
+#endif
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_prefill_scalar(AttentionParams<bf16>& p) {
+    constexpr int G = 8, ROWS = 32, P_BC = 32;
+    dim3 grid((p.q_len + ROWS - 1) / ROWS, p.q_head, p.batch);
+    dim3 block(G, ROWS);
+    attn_prefill_split_q_kernel_t<HEAD_DIM, G, ROWS, P_BC, IsCausal, HasMask><<<grid, block>>>(p);
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_prefill(AttentionParams<bf16>& p) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+
+#ifndef ASTRAI_NO_MMA
+    if (is_causal) {
+        if (has_mask)      launch_prefill_mma<HEAD_DIM, true, true>(p);
+        else               launch_prefill_mma<HEAD_DIM, true, false>(p);
+    } else {
+        if (has_mask)      launch_prefill_mma<HEAD_DIM, false, true>(p);
+        else               launch_prefill_mma<HEAD_DIM, false, false>(p);
+    }
+#else
+    if (is_causal) {
+        if (has_mask)      launch_prefill_scalar<HEAD_DIM, true, true>(p);
+        else               launch_prefill_scalar<HEAD_DIM, true, false>(p);
+    } else {
+        if (has_mask)      launch_prefill_scalar<HEAD_DIM, false, true>(p);
+        else               launch_prefill_scalar<HEAD_DIM, false, false>(p);
+    }
+#endif
+}
+
+// ======================================================================
+// Decode
+// ======================================================================
+
+#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 <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_decode_mma(AttentionParams<bf16>& p, int group_size) {
+    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 tiles_total = (p.kv_len + BC - 1) / BC;
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, 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, IsCausal, HasMask><<<grid, 32>>>(p);
+}
+#endif
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_decode_scalar(AttentionParams<bf16>& p, int group_size) {
+    int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
+    size_t smem = 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, IsCausal, HasMask><<<grid, block, smem>>>(p);
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_decode(AttentionParams<bf16>& p) {
+    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
+    if (is_causal) {
+        if (has_mask)      launch_decode_mma<HEAD_DIM, true, true>(p, group_size);
+        else               launch_decode_mma<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_decode_mma<HEAD_DIM, false, true>(p, group_size);
+        else               launch_decode_mma<HEAD_DIM, false, false>(p, group_size);
+    }
+#else
+    if (is_causal) {
+        if (has_mask)      launch_decode_scalar<HEAD_DIM, true, true>(p, group_size);
+        else               launch_decode_scalar<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_decode_scalar<HEAD_DIM, false, true>(p, group_size);
+        else               launch_decode_scalar<HEAD_DIM, false, false>(p, group_size);
+    }
+#endif
+
+    attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
+}
+
+// ======================================================================
+// Paged Decode
+// ======================================================================
+
+#ifndef ASTRAI_NO_MMA
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_paged_decode_mma(PagedAttentionParams<bf16>& p, int group_size) {
+    int G = p.q_head / p.kv_head;
+    constexpr int MAX_G = 16;
+    constexpr int BC = 16;
+    int num_passes = (G + MAX_G - 1) / MAX_G;
+    int tiles_total = (p.kv_len + BC - 1) / BC;
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, 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);
+    paged_attn_decode_split_kv_mma_kernel<Traits, IsCausal, HasMask> <<<grid, 32>>>(p);
+}
+#endif
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_paged_decode_scalar(PagedAttentionParams<bf16>& p, int group_size) {
+    int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_CHUNK;
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
+    size_t smem = PDC_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);
+    paged_attn_decode_split_kv_kernel<HEAD_DIM, IsCausal, HasMask><<<grid, block, smem>>>(p);
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_paged_decode(PagedAttentionParams<bf16>& p) {
+    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
+    if (is_causal) {
+        if (has_mask)      launch_paged_decode_mma<HEAD_DIM, true, true>(p, group_size);
+        else               launch_paged_decode_mma<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_paged_decode_mma<HEAD_DIM, false, true>(p, group_size);
+        else               launch_paged_decode_mma<HEAD_DIM, false, false>(p, group_size);
+    }
+#else
+    if (is_causal) {
+        if (has_mask)      launch_paged_decode_scalar<HEAD_DIM, true, true>(p, group_size);
+        else               launch_paged_decode_scalar<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_paged_decode_scalar<HEAD_DIM, false, true>(p, group_size);
+        else               launch_paged_decode_scalar<HEAD_DIM, false, false>(p, group_size);
+    }
+#endif
+
+    paged_attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
+}

csrc/kernels/attn_entry_utils.cuh

@@ -0,0 +1,183 @@
+#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, arg)
+//   Expands to: fn<32>(arg); fn<64>(arg); etc.
+#define DISPATCH_HEAD_DIM(hd, fn, arg) \
+    switch (hd) { \
+        case 32:  fn<32>(arg); break; \
+        case 64:  fn<64>(arg); break; \
+        case 128: fn<128>(arg); break; \
+        case 256: fn<256>(arg); break; \
+        default: \
+            TORCH_CHECK(false, "unsupported head_dim ", hd, \
+                         " (supported: 32, 64, 128, 256)"); \
+    }
+
+template<typename P>
+inline void alloc_split_partials(P& p) {
+    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
+    auto o_part = torch::zeros(at::IntArrayRef{p.batch, p.q_head, MAX_SPLITS, p.head_dim}, fopt);
+    auto ml_part = torch::full(at::IntArrayRef{p.batch, p.q_head, MAX_SPLITS, 2}, -FLT_MAX, 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 == 1) 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 == 1) 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_params ----
+template<typename T>
+inline void attn_pack_paged_params(
+    torch::Tensor q,
+    torch::Tensor page_table,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    int64_t page_size,
+    int64_t kv_len,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout,
+    PagedAttentionParams<T>& p
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+
+    TORCH_CHECK(q.is_cuda() && page_table.is_cuda() && k_cache.is_cuda() && v_cache.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(page_table.dtype() == torch::kLong, "page_table must be int64");
+    TORCH_CHECK(k_cache.sizes() == v_cache.sizes(), "k_cache and v_cache must have identical shapes");
+
+    extract_q_dims_and_strides(q, layout, p);
+
+    p.kv_head = (int)k_cache.size(2);
+    p.kv_len = (int)kv_len;
+    p.page_size = (int)page_size;
+    p.max_pages = (int)page_table.size(1);
+
+    TORCH_CHECK(q.size(2) == 1, "Q seq_len must be 1 (decode)");
+    TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
+    TORCH_CHECK(k_cache.size(1) == page_size,
+                "k_cache dim 1 must equal page_size, got ",
+                k_cache.size(1), " vs ", page_size);
+
+    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);
+
+    p.page_table = page_table.data_ptr<int64_t>();
+    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.o = nullptr;
+    p.o_part = nullptr;
+    p.ml_part = nullptr;
+
+    pack_mask(mask, p);
+}

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 +1,42 @@
+#include "attn_dispatchers.cuh"
+#include "attn_entry_utils.cuh"
+
+torch::Tensor attn_paged_decode(
+    torch::Tensor q,
+    torch::Tensor page_table,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    int64_t page_size,
+    int64_t kv_len,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout
+) {
+    PagedAttentionParams<bf16> p;
+    attn_pack_paged_params(q, page_table, k_cache, v_cache,
+                           page_size, kv_len, mask, causal_offset, scale, layout, p);
+
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
+    auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    alloc_split_partials(p);
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_decode, p);
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_paged_decode", &attn_paged_decode,
+        py::arg("q"),
+        py::arg("page_table"),
+        py::arg("k_cache"),
+        py::arg("v_cache"),
+        py::arg("page_size"),
+        py::arg("kv_len"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = 0,
+        "Paged GQA decode — split-KV with direct page-table access.");
+}

csrc/kernels/attn_paged_decode_split_kv.cuh

@@ -0,0 +1,153 @@
+#pragma once
+#include <cuda_bf16.h>
+#include <float.h>
+#include "attn_common.h"
+#include "attn_warp_utils.cuh"
+constexpr int PDC_CHUNK = 64;
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+__global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<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;
+
+    float q_reg[8];
+    int q_off = batch * p.q_stride_b + q_head * p.q_stride_h
+              + lane * hd_per_thread * p.q_stride_d;
+    #pragma unroll
+    for (int i = 0; i < hd_per_thread; i++)
+        q_reg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
+
+    float m = -FLT_MAX, d = 0.0f, acc_reg[8] = {0.0f};
+
+    extern __shared__ __align__(16) bf16 k_smem[];
+
+    int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_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);
+
+    const int mask_base = batch * p.mask_b_stride + q_head * p.mask_h_stride;
+
+    for (int ci = ch_begin; ci < ch_end; ci++) {
+        int chunk_start = ci * PDC_CHUNK;
+        int this_chunk = min(PDC_CHUNK, p.kv_len - chunk_start);
+
+        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 pos = chunk_start + s;
+            int logical_page = pos / p.page_size;
+            int page_offset = pos % p.page_size;
+            int phys_page = p.page_table[batch * p.max_pages + logical_page];
+            if (phys_page >= 0) {
+                int64_t off = (int64_t)phys_page * p.page_size * p.kv_head * p.head_dim
+                            + (int64_t)page_offset * p.kv_head * p.head_dim
+                            + (int64_t)kv_head * p.head_dim
+                            + d_dim;
+                k_smem[i] = p.k_cache[off];
+            } else {
+                k_smem[i] = __float2bfloat16(0.0f);
+            }
+        }
+        __syncthreads();
+
+        for (int s = 0; s < this_chunk; s++) {
+            float partial = 0.0f;
+            #pragma unroll
+            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;
+            bool masked = false;
+            if constexpr (HasMask) {
+                if (!p.mask[mask_base + kv_idx])
+                    masked = true;
+            }
+            if constexpr (IsCausal) {
+                if (kv_idx > p.causal_offset)
+                    masked = true;
+            }
+            if (masked)
+                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;
+
+            int pos = chunk_start + s;
+            int logical_page = pos / p.page_size;
+            int page_offset = pos % p.page_size;
+            int phys_page = p.page_table[batch * p.max_pages + logical_page];
+            if (masked) {
+                #pragma unroll
+                for (int i = 0; i < hd_per_thread; i++)
+                    acc_reg[i] = fmaf(acc_reg[i], alpha, 0.0f);
+            } else if (phys_page >= 0) {
+                int64_t v_base = (int64_t)phys_page * p.page_size * p.kv_head * p.head_dim
+                               + (int64_t)page_offset * p.kv_head * p.head_dim
+                               + (int64_t)kv_head * p.head_dim;
+                #pragma unroll
+                for (int i = 0; i < hd_per_thread; i++)
+                    acc_reg[i] = fmaf(acc_reg[i], alpha,
+                                      __bfloat162float(p.v_cache[v_base + lane * hd_per_thread + i]) * beta);
+            } else {
+                #pragma unroll
+                for (int i = 0; i < hd_per_thread; i++)
+                    acc_reg[i] = fmaf(acc_reg[i], alpha, 0.0f);
+            }
+            m = new_m;
+        }
+        __syncthreads();
+    }
+
+    size_t bh = (size_t)batch * p.q_head + q_head;
+    size_t slot = bh * MAX_SPLITS + split;
+    int d0 = lane * hd_per_thread;
+    #pragma unroll
+    for (int i = 0; i < hd_per_thread; i++)
+        p.o_part[slot * p.head_dim + (d0 + i)] = acc_reg[i];
+    if (lane == 0) {
+        p.ml_part[slot * 2] = m;
+        p.ml_part[slot * 2 + 1] = d;
+    }
+}
+
+__global__ void paged_attn_decode_combine_kernel(PagedAttentionParams<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 = batch * p.q_stride_b + q_head * p.q_stride_h + d * p.q_stride_d;
+    p.o[o_off] = __float2bfloat16(acc * inv);
+}

csrc/kernels/attn_paged_decode_split_kv_mma.cuh

@@ -0,0 +1,173 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+#include "attn_warp_utils.cuh"
+
+// Paged split-KV tensor-core decode via GQA head-packing.
+// Reads K/V directly from the page pool through a page table — one tile
+// (BC=32) fits within a single page (page_size >= 32), so the page-table
+// lookup happens once per tile for cp.async.
+//
+// IsCausal and HasMask are compile-time bools.
+template <typename Traits, bool IsCausal, bool HasMask>
+__global__ void paged_attn_decode_split_kv_mma_kernel(PagedAttentionParams<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;
+
+    __shared__ __align__(16) bf16 sK[Traits::STAGES * Traits::BC * Traits::LD];
+    __shared__ __align__(16) bf16 sV[Traits::STAGES * Traits::BC * Traits::LD];
+
+    #pragma unroll
+    for (int i = lane; i < Traits::STAGES * Traits::BC * Traits::LD; i += 32) {
+        sK[i] = __float2bfloat16(0.0f);
+        sV[i] = __float2bfloat16(0.0f);
+    }
+    __syncwarp();
+
+    const int q_base = batch * p.q_stride_b + q_head0 * p.q_stride_h;
+    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 = (p.kv_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);
+
+    const int64_t page_stride = (int64_t)p.page_size * p.kv_head * Traits::HEAD_DIM;
+    const int64_t pos_stride  = (int64_t)p.kv_head * Traits::HEAD_DIM;
+    const int64_t head_off    = (int64_t)kv_head * Traits::HEAD_DIM;
+
+    // ---- Load tile lambda: paged addressing ----
+    // Unified per-element page-table lookup. When page_size >= BC, all
+    // elements in a tile share the same page, so the lookup is redundant
+    // but harmless (L1-cached). This avoids a branch on page_size.
+    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 < p.kv_len);
+            if constexpr (HasMask) {
+                valid = valid && p.mask[batch * p.mask_b_stride + kc];
+            }
+            int phys_page = valid ? p.page_table[batch * p.max_pages + kc] : 0;
+            valid = valid && (phys_page >= 0);
+            int page_off = kc % p.page_size;
+            int64_t gmem_base = (int64_t)phys_page * page_stride
+                              + (int64_t)page_off * pos_stride
+                              + head_off;
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            cp_async_16_pred(&dK[off], &p.k_cache[gmem_base + d], valid);
+            cp_async_16_pred(&dV[off], &p.v_cache[gmem_base + d], valid);
+        }
+        cp_async_commit();
+    };
+
+    constexpr int BUF_MASK = (Traits::STAGES > 1) ? (Traits::STAGES - 1) : 0;
+
+    if (ti_begin < ti_end) {
+        load_tile(ti_begin, 0);
+    }
+
+    for (int ti = ti_begin; ti < ti_end; ti++) {
+        int buf = (ti - ti_begin) & BUF_MASK;
+
+        cp_async_wait_group<0>();
+        __syncwarp();
+        if constexpr (Traits::STAGES > 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, (ti + 1 - ti_begin) & BUF_MASK);
+        }
+
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = ti * 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;
+
+        int maxc = IsCausal ? min(p.kv_len, p.causal_offset + 1) : p.kv_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);
+        __syncwarp();
+
+        if constexpr (Traits::STAGES == 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, 0);
+        }
+    }
+
+    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_prefill.cu

@@ -0,0 +1,35 @@
+#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
+) {
+    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 == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_prefill, p);
+    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") = 0,
+        "GQA prefill (tensor-core mma on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_prefill_split_q.cuh

@@ -0,0 +1,149 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+
+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.
+// Templated on <HEAD_DIM, 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, 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;
+
+    int kv_head = q_head / (p.q_head / p.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]
+    float qreg[DPT];
+    if (q_row < p.q_len) {
+        int q_off = batch * p.q_stride_b + q_head * p.q_stride_h
+                  + 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;
+
+    // KV: stride-based base
+    int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    int mask_batch_base = batch * p.mask_b_stride + q_head * p.mask_h_stride;
+    int tiles   = (p.kv_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, p.kv_len - kv0);
+
+        // Load K/V into shared memory from strided global
+        for (int i = lid; i < tlen * HEAD_DIM; i += tt) {
+            int s = i / HEAD_DIM;
+            int d_dim = i % HEAD_DIM;
+            int kv_idx = kv0 + s;
+            int g_off = kv_base + kv_idx * p.kv_stride_l + d_dim * p.kv_stride_d;
+            sK[i] = p.k[g_off];
+            sV[i] = p.v[g_off];
+        }
+        __syncthreads();
+
+        int lim = tlen;
+        if constexpr (IsCausal) {
+            if (q_row < p.q_len) {
+                int ep = q_row + p.causal_offset + 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 < p.q_len) {
+        int o_off = batch * p.q_stride_b + q_head * p.q_stride_h
+                  + 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 +1,146 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+
+// Tensor-core prefill flash attention (raw mma.sync PTX).
+// 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).
+//
+// 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, 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;
+
+    // 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 = batch * p.q_stride_b + q_head * p.q_stride_h;
+    const int qra = qrow0 + gid;
+    const int qrb = qrow0 + gid + 8;
+    const bool va = qra < p.q_len, vb = qrb < p.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;
+
+    // KV: stride-based base
+    const int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    const int tiles = (p.kv_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 + p.causal_offset;
+    const int block_max_kv =
+        blockIdx.x * Traits::WARPS * Traits::BR + Traits::WARPS * Traits::BR - 1
+        + p.causal_offset;
+
+    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 ----
+    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 < p.kv_len;
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            int g_off = kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
+            cp_async_16_pred(&dK[off], &p.k[g_off], valid);
+            cp_async_16_pred(&dV[off], &p.v[g_off], 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(p.kv_len, qr0 + p.causal_offset + 1)
+                                 : p.kv_len;
+            int maxc1 = IsCausal ? min(p.kv_len, qr1 + p.causal_offset + 1)
+                                 : p.kv_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 = batch * p.q_stride_b + q_head * p.q_stride_h;
+    #pragma unroll
+    for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+        int d = dn8 * 8 + 2 * tid4;
+        if (qr0 < p.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 < p.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 +1,87 @@
+#include <torch/extension.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
+) {
+    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");
+
+    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");
+
+    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>>>(
+        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
+    );
+
+    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_decode_test.cu

@@ -0,0 +1,185 @@
+/*
+Pure-C test — uses shared dispatcher.
+nvcc -I csrc -arch=sm_89 -O3 \
+    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
+    csrc/tests/attn_decode_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);
+}
+
+// Warmed-up, CUDA-event timed sweep over the production decode MMA path.
+static void bench() {
+    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 = 10, ITERS = 100;
+    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); }); };
+        double flops = 4.0 * B * Hq * (double)sl * D;
+        double bytes = 2.0 * (2.0 * nKV * sizeof(bf16));
+        BenchResult r = bench_kernel(launch, WARMUP, ITERS, flops, bytes);
+
+        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);
+    }
+}
+
+static int run_test(int B, int Hq, int Hk, int sl, int D, int causal) {
+    int gs = Hq / Hk;
+    printf("=== B=%d Hq=%d Hk=%d seq=%d D=%d gs=%d causal=%d ===\n",
+           B,Hq,Hk,sl,D,gs,causal);
+
+    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); });
+    cudaDeviceSynchronize();
+    double kms=now_ms()-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-8f);
+        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; }
+    }
+    printf("kernel: %.3f ms  max_abs_err: %.6e  max_rel_err: %.6e  %s\n\n",
+           kms, max_abs_err, max_rel_err, pass?"PASS":"FAIL");
+
+    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;
+}
+
+int main() {
+    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]);
+    int fail = 0;
+
+    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_test(B, Hq, Hk, sl, D, causal);
+        if (fail) break;
+    }
+
+    if (fail) {
+        printf("FAILED\n");
+        return fail;
+    }
+    printf("All tests passed!\n");
+    bench();
+    return 0;
+}

csrc/tests/attn_paged_decode_test.cu

@@ -0,0 +1,308 @@
+// Compile:
+//   nvcc -I csrc -arch=sm_89 -O3 --use_fast_math --ptxas-options=-O3 \
+//        --extra-device-vectorization csrc/tests/attn_paged_decode_test.cu \
+//        -o /tmp/test_paged && /tmp/test_paged
+
+#include <cstring>
+#include "test_utils.cuh"
+#include "../kernels/attn_dispatchers.cuh"
+
+static void gather_kv_cpu(
+    const bf16* h_k_pool, const bf16* h_v_pool,
+    const int64_t* h_pt, int B, int Hkv, int kv_len,
+    int page_size, int head_dim,
+    bf16* h_k, bf16* h_v)
+{
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    size_t page_stride = (size_t)page_size * Hkv * head_dim;
+    for (int b = 0; b < B; b++) {
+        for (int pos = 0; pos < kv_len; pos++) {
+            int log_pg = pos / page_size;
+            int pg_off = pos % page_size;
+            int phys = (int)h_pt[b * max_pages + log_pg];
+            for (int h = 0; h < Hkv; h++) {
+                size_t src_base = (size_t)phys * page_stride
+                                + (size_t)pg_off * Hkv * head_dim
+                                + h * head_dim;
+                size_t dst_base = ((size_t)b * Hkv + h) * kv_len * head_dim
+                                  + (size_t)pos * head_dim;
+                memcpy(h_k + dst_base, h_k_pool + src_base, head_dim * sizeof(bf16));
+                memcpy(h_v + dst_base, h_v_pool + src_base, head_dim * sizeof(bf16));
+            }
+        }
+    }
+}
+
+template <int HEAD_DIM>
+static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int causal, int seed) {
+    printf("B=%d Hq=%d Hkv=%d kv_len=%d page_sz=%d head_dim=%d causal=%d ... ",
+           B, Hq, Hkv, kv_len, page_size, HEAD_DIM, causal);
+    fflush(stdout);
+
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    int n_phys_pages = B * max_pages;
+    int max_splits = 32;
+
+    size_t sz_q  = (size_t)B * Hq * 1 * HEAD_DIM * sizeof(bf16);
+    size_t sz_o  = sz_q;
+    size_t sz_kv = (size_t)n_phys_pages * page_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_pt = (size_t)B * max_pages * sizeof(int64_t);
+    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_paged;
+    bf16 *d_k_pool, *d_v_pool;
+    int64_t* d_pt;
+    float *d_op, *d_ml;
+
+    cudaMalloc(&d_q, sz_q);
+    cudaMalloc(&d_o_paged, sz_o);
+    cudaMalloc(&d_k_pool, sz_kv);
+    cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_pt, sz_pt);
+    cudaMalloc(&d_op, sz_op);
+    cudaMalloc(&d_ml, sz_ml);
+
+    srand(seed);
+    auto rnd = [&]() { return (rand() / (float)RAND_MAX) * 2.0f - 1.0f; };
+
+    bf16* h_q = (bf16*)malloc(sz_q);
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++)
+        h_q[i] = __float2bfloat16(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);
+    size_t ps = (size_t)page_size * Hkv * HEAD_DIM;
+    for (int pg = 0; pg < n_phys_pages; pg++) {
+        for (int off = 0; off < page_size; off++) {
+            for (int h = 0; h < Hkv; h++) {
+                for (int d = 0; d < HEAD_DIM; d++) {
+                    float v = sinf((float)(pg * 7919 + off * 1049 + h * 331 + d));
+                    size_t idx = (size_t)pg * ps + (size_t)off * Hkv * HEAD_DIM
+                                 + h * HEAD_DIM + d;
+                    h_k_pool[idx] = __float2bfloat16(v);
+                    h_v_pool[idx] = __float2bfloat16(v * 0.3f);
+                }
+            }
+        }
+    }
+    cudaMemcpy(d_k_pool, h_k_pool, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, h_v_pool, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_pt = (int64_t*)malloc(sz_pt);
+    int next_pg = 0;
+    for (int b = 0; b < B; b++)
+        for (int p = 0; p < max_pages; p++)
+            h_pt[b * max_pages + p] = next_pg++;
+    cudaMemcpy(d_pt, h_pt, sz_pt, cudaMemcpyHostToDevice);
+
+    bf16* h_k_cont = (bf16*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(bf16));
+    bf16* h_v_cont = (bf16*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(bf16));
+    gather_kv_cpu(h_k_pool, h_v_pool, h_pt, B, Hkv, kv_len, page_size, HEAD_DIM, h_k_cont, h_v_cont);
+
+    float* h_q_f = (float*)malloc((size_t)B * Hq * HEAD_DIM * sizeof(float));
+    float* h_k_f = (float*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(float));
+    float* h_v_f = (float*)malloc((size_t)B * kv_len * 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 < B * kv_len * Hkv * HEAD_DIM; i++) {
+        h_k_f[i] = bf2f(h_k_cont[i]);
+        h_v_f[i] = bf2f(h_v_cont[i]);
+    }
+
+    float* h_o_ref = (float*)calloc(B * Hq * HEAD_DIM, sizeof(float));
+    cpu_attention_ref(h_q_f, h_k_f, h_v_f, nullptr, h_o_ref, B, Hq, Hkv,
+                      1, kv_len, HEAD_DIM, causal ? 0 : -1);
+
+    PagedAttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv; p.q_len = 1;
+    p.kv_len = kv_len; p.head_dim = HEAD_DIM;
+    p.use_mask = 0; p.causal_offset = causal ? 0 : -1;
+    set_default_paged_strides(p);
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.page_size = page_size; p.max_pages = max_pages;
+    p.page_table = d_pt;
+    p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.q = d_q; p.mask = nullptr; p.o = d_o_paged;
+    p.o_part = d_op; p.ml_part = d_ml;
+
+    dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(p); });
+    cudaDeviceSynchronize();
+
+    bf16* h_o_bf16 = (bf16*)malloc(sz_o);
+    cudaMemcpy(h_o_bf16, d_o_paged, sz_o, cudaMemcpyDeviceToHost);
+    float* h_o_paged = (float*)malloc(B * Hq * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++)
+        h_o_paged[i] = __bfloat162float(h_o_bf16[i]);
+
+    float max_abs_err = 0.0f, max_rel_err = 0.0f;
+    int bad_idx = -1;
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_paged[i] - h_o_ref[i]);
+        if (e > max_abs_err) { max_abs_err = e; bad_idx = i; }
+        float rel = e / fmaxf(fabsf(h_o_ref[i]), 1e-8f);
+        if (rel > max_rel_err) max_rel_err = rel;
+    }
+
+    const float atol = 0.01f, rtol = 0.01f;
+    bool pass = true;
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_paged[i] - h_o_ref[i]);
+        if (e > atol + rtol * fabsf(h_o_ref[i])) { pass = false; break; }
+    }
+
+    if (pass) {
+        printf("PASS (max_abs_err=%.4e max_rel_err=%.4e)\n", max_abs_err, max_rel_err);
+    } else {
+        int b = bad_idx / (Hq * HEAD_DIM);
+        int h = (bad_idx / HEAD_DIM) % Hq;
+        int d = bad_idx % HEAD_DIM;
+        printf("FAIL (max_abs_err=%.4e max_rel_err=%.4e at [%d,%d,%d]: ref=%.4f got=%.4f)\n",
+               max_abs_err, max_rel_err, b, h, d, h_o_ref[bad_idx], h_o_paged[bad_idx]);
+        printf("  ref[0..7]:");
+        for (int i = 0; i < 8 && i < HEAD_DIM; i++)
+            printf(" %.4f", h_o_ref[i]);
+        printf("\n  got[0..7]:");
+        for (int i = 0; i < 8 && i < HEAD_DIM; i++)
+            printf(" %.4f", h_o_paged[i]);
+        printf("\n");
+    }
+
+    free(h_q); free(h_k_pool); free(h_v_pool); free(h_pt);
+    free(h_k_cont); free(h_v_cont);
+    free(h_q_f); free(h_k_f); free(h_v_f);
+    free(h_o_ref); free(h_o_bf16); free(h_o_paged);
+    cudaFree(d_q); cudaFree(d_o_paged);
+    cudaFree(d_k_pool); cudaFree(d_v_pool); cudaFree(d_pt);
+    cudaFree(d_op); cudaFree(d_ml);
+
+    return pass ? 0 : 1;
+}
+
+struct TestCase {
+    int head_dim;
+    int B, Hq, Hkv, kv_len, page_size, causal, seed;
+};
+
+static const TestCase TESTS[] = {
+    {128, 1, 1, 1, 8, 128, 0, 1},
+    {128, 1, 4, 4, 128, 128, 0, 2},
+    {128, 2, 4, 4, 256, 128, 0, 3},
+    {128, 1, 4, 1, 64, 64, 0, 4},
+    {128, 1, 8, 2, 64, 128, 0, 5},
+    {128, 2, 16, 4, 128, 128, 0, 6},
+    {64, 1, 4, 2, 32, 128, 0, 7},
+    {256, 1, 2, 1, 16, 128, 0, 8},
+    {32, 1, 4, 2, 32, 64, 0, 9},
+    {128, 3, 8, 2, 256, 128, 0, 10},
+    {128, 2, 32, 8, 512, 128, 0, 11},
+    {128, 1, 16, 2, 256, 128, 0, 12},
+    {128, 2, 32, 4, 512, 128, 0, 13},
+    {128, 2, 8, 2, 128, 128, 1, 14},   // causal
+};
+
+static int dispatch_test(const TestCase& tc) {
+    int r = 0;
+    dispatch_by_head_dim(tc.head_dim, [&]<int D>() {
+        r = run_test<D>(tc.B, tc.Hq, tc.Hkv, tc.kv_len, tc.page_size, tc.causal, tc.seed);
+    });
+    return r;
+}
+
+template <int HEAD_DIM>
+static void bench_config(int B, int Hq, int Hkv, int kv_len, int page_size) {
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    int n_phys_pages = B * max_pages;
+    int max_splits = 32;
+
+    size_t sz_q  = (size_t)B * Hq * 1 * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)n_phys_pages * page_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_pt = (size_t)B * max_pages * sizeof(int64_t);
+    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_pt;
+    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_pt, sz_pt);
+    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_pt = (int64_t*)malloc(sz_pt);
+    int next_pg = 0;
+    for (int b = 0; b < B; b++)
+        for (int p = 0; p < max_pages; p++)
+            h_pt[b * max_pages + p] = next_pg++;
+    cudaMemcpy(d_pt, h_pt, sz_pt, cudaMemcpyHostToDevice);
+    free(h_pt);
+
+    PagedAttentionParams<bf16> pa;
+    pa.batch = B; pa.q_head = Hq; pa.kv_head = Hkv; pa.q_len = 1;
+    pa.kv_len = kv_len; pa.head_dim = HEAD_DIM;
+    pa.use_mask = 0; pa.causal_offset = -1;
+    set_default_paged_strides(pa);
+    pa.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    pa.page_size = page_size; pa.max_pages = max_pages;
+    pa.page_table = d_pt;
+    pa.k_cache = d_k_pool; pa.v_cache = d_v_pool;
+    pa.q = d_q; pa.mask = nullptr; pa.o = d_o;
+    pa.o_part = d_op; pa.ml_part = d_ml;
+
+    const int WARMUP = 10, ITERS = 100;
+    auto launch = [&]() {
+        dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(pa); });
+    };
+    double flops = 4.0 * B * Hq * (double)kv_len * HEAD_DIM;
+    size_t nKV = (size_t)B * Hkv * kv_len * HEAD_DIM;
+    double bytes = 2.0 * (2.0 * nKV * sizeof(bf16));
+    BenchResult r = bench_kernel(launch, WARMUP, ITERS, flops, bytes);
+
+    char cfg[64];
+    snprintf(cfg, sizeof(cfg),
+             "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d page=%3d",
+             B, Hq, Hkv, 1, kv_len, HEAD_DIM, page_size);
+    print_bench_row(cfg, r);
+
+    free(tmp);
+    cudaFree(d_q); cudaFree(d_o);
+    cudaFree(d_k_pool); cudaFree(d_v_pool); cudaFree(d_pt);
+    cudaFree(d_op); cudaFree(d_ml);
+}
+
+static void bench() {
+    printf("\n===== PAGED DECODE BENCH =====\n");
+    print_bench_header();
+    bench_config<128>(1, 32, 4, 512, 128);
+    bench_config<128>(1, 32, 4, 1024, 128);
+    bench_config<128>(1, 32, 4, 2048, 128);
+    bench_config<128>(1, 32, 4, 4096, 128);
+    bench_config<128>(16, 32, 4, 2048, 128);
+    bench_config<128>(32, 32, 4, 1024, 128);
+}
+
+int main() {
+    int n = sizeof(TESTS) / sizeof(TESTS[0]);
+    int fail = 0;
+    printf("=== Paged Decode vs CPU reference (%d cases) ===\n\n", n);
+
+    for (int i = 0; i < n; i++) {
+        fail += dispatch_test(TESTS[i]);
+        if (fail) break;
+    }
+
+    if (fail) {
+        printf("\nFAILED (%d/%d tests failed)\n", fail, n);
+        return fail;
+    }
+    printf("\nAll %d tests passed!\n", n);
+    bench();
+    return 0;
+}