chore: bump version to 1.3.12

- ManoAdamW now uses torch.optim.AdamW(fused=True, betas=(0.9, 0.95)) matching MuonAdamW, eliminating a confounding variable in optimizer comparison experiments
- only NoraNAdamW retains NAdamW, which is correct per the Nora paper design

.dockerignore

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

.github/workflows/release.yml

@@ -23,24 +23,33 @@ jobs:
         with:
           name: pure-wheel
           path: dist/*.whl
+          if-no-files-found: error
 
   build-cuda-linux:
-    name: Build CUDA wheel (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 (CUDA 12.8)
+      - name: Install torch (${{ matrix.cuda_tag }})
         run: |
-          pip install torch --index-url https://download.pytorch.org/whl/cu128
+          pip install torch --index-url https://download.pytorch.org/whl/${{ matrix.cuda_tag }}
 
-      - name: Setup CUDA
+      - name: Setup CUDA (${{ matrix.cuda_ver }})
         uses: Jimver/cuda-toolkit@v0.2.35
         with:
-          cuda: "12.8.0"
+          cuda: "${{ matrix.cuda_ver }}"
 
       - name: Build wheel (with CUDA kernels)
         run: |
@@ -48,8 +57,9 @@ jobs:
 
       - uses: actions/upload-artifact@v4
         with:
-          name: cuda-wheel-linux
+          name: cuda-wheel-linux-${{ matrix.cuda_tag }}
           path: dist/*.whl
+          if-no-files-found: error
 
   release:
     name: Attach wheels to release
@@ -58,14 +68,33 @@ jobs:
     permissions:
       contents: write
     steps:
-      - uses: actions/download-artifact@v4
+      - name: Download pure-Python wheel
+        uses: actions/download-artifact@v4
         with:
-          pattern: "*-wheel"
+          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: ./*.whl
+          files: |
+            release-assets/pure/*.whl
+            release-assets/cuda/*.whl
           tag_name: ${{ github.ref_name }}
           generate_release_notes: true

.gitignore

@@ -24,7 +24,7 @@
 !/.dockerignore
 !/Dockerfile
 !/docker-compose.yml
-!/assets/**
+!/docs/**
 !/CONTRIBUTING.md
 !/LICENSE
 !/pyproject.toml

Dockerfile

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

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>
@@ -17,7 +17,7 @@
 
 <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/ViperEkura">HuggingFace</a>
@@ -213,18 +213,23 @@ curl -X POST http://localhost:8000/v1/messages \
 curl http://localhost:8000/health
 ```
 
-See [Inference Guide](assets/docs/inference.md) for SSE streaming format, error codes, and stats endpoint.
+See [Inference Guide](docs/guides/inference.md) for SSE streaming format, error codes, and stats endpoint.
 
 ### Documentation
 
 | Document | Description |
 |----------|-------------|
-| [CLI Reference](./assets/docs/params.md) | Parameters for all CLI tools (train, server, generate, preprocess) |
-| [Architecture](./assets/docs/architecture.md) | System architecture, class diagram & design patterns |
-| [Training](./assets/docs/training.md) | Training loop, strategies & formulas |
-| [Inference](./assets/docs/inference.md) | KVCache, continuous batching, sampling & HTTP API |
-| [Data Flow](./assets/docs/dataflow.md) | Data pipeline, storage backends & dataset architecture |
-| [Preprocessing](./assets/docs/preprocessing.md) | Declarative JSON-driven data preprocessing |
+| [Get Started](./docs/get-started.md) | Installation and quickstart |
+| [CLI Reference](./docs/guides/params.md) | Parameters for all CLI tools (train, server, generate, preprocess) |
+| [Preprocessing](./docs/guides/preprocessing.md) | Declarative JSON-driven data preprocessing |
+| [Training](./docs/guides/training.md) | Training loop, strategies & formulas |
+| [Inference](./docs/guides/inference.md) | KVCache, continuous batching, sampling & HTTP API |
+| [Evaluation](./docs/guides/evaluation.md) | HumanEval, MMLU, PPL, ROUGE, IFD, IFEval |
+| [Distributed](./docs/guides/distributed.md) | Multi-GPU DDP / FSDP training |
+| [Architecture](./docs/developer/architecture.md) | System architecture, class diagram & design patterns |
+| [Data Flow](./docs/developer/dataflow.md) | Data pipeline, storage backends & dataset architecture |
+| [Internals](./docs/developer/internals.md) | Training internals: loss formulas, callback lifecycle, KV cache |
+| [CUDA Kernels](./docs/developer/cuda_kernels.md) | Custom CUDA attention kernels & benchmarks |
 
 ### Contributing
 

astrai/__init__.py

@@ -1,6 +1,9 @@
-__version__ = "1.3.10"
+__version__ = "1.3.12"
 __author__ = "ViperEkura"
 
+import logging
+import os
+
 from astrai.config import (
     AutoRegressiveLMConfig,
     BaseModelConfig,
@@ -53,6 +56,30 @@ from astrai.trainer import (
     Trainer,
 )
 
+
+def setup_logging(level: str = "INFO"):
+    """Attach a handler to the ``astrai`` logger (only, not root).
+
+    Call once per process, e.g. at the top of CLI scripts.
+    Set ``ASTR_LOG_LEVEL`` to override the default ``INFO``.
+    """
+    _logger = logging.getLogger("astrai")
+    if _logger.handlers:
+        return
+    _level = getattr(
+        logging, os.environ.get("ASTR_LOG_LEVEL", level).upper(), logging.INFO
+    )
+    _logger.setLevel(_level)
+    _handler = logging.StreamHandler()
+    _handler.setFormatter(
+        logging.Formatter(
+            "%(asctime)s | %(levelname)-7s | %(name)s | %(message)s",
+            datefmt="%Y-%m-%d %H:%M:%S",
+        )
+    )
+    _logger.addHandler(_handler)
+
+
 __all__ = [
     "AutoRegressiveLM",
     "AutoRegressiveLMConfig",
@@ -94,5 +121,6 @@ __all__ = [
     "only_on_rank",
     "run_server",
     "sample",
+    "setup_logging",
     "spawn_parallel_fn",
 ]

astrai/config/base.py

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

astrai/config/model_config.py

@@ -1,9 +1,14 @@
-from dataclasses import dataclass
 from typing import Any, Dict, Optional
 
+from pydantic import field_validator
+from pydantic.dataclasses import dataclass
+
 from astrai.config.base import BaseConfig
 from astrai.factory import BaseFactory
 
+_ATTN_TYPES = frozenset({"gqa", "mla"})
+_FFN_TYPES = frozenset({"mlp", "moe"})
+
 
 class ConfigFactory(BaseFactory[BaseConfig]):
     """Factory that dispatches config classes by ``model_type``."""
@@ -17,7 +22,12 @@ class ConfigFactory(BaseFactory[BaseConfig]):
 
 @dataclass
 class BaseModelConfig(BaseConfig):
-    """Base config with ``model_type`` dispatch and file I/O."""
+    """Base config with ``model_type`` dispatch and file I/O.
+
+    Args:
+        model_type (Optional[str]): Model type identifier for AutoModel dispatch. Defaults to None.
+        neftune_alpha (float): NEFTune noise alpha, 0=disabled, typical: 5.0. Defaults to 0.0.
+    """
 
     model_type: Optional[str] = None
     neftune_alpha: float = 0.0
@@ -26,57 +36,126 @@ class BaseModelConfig(BaseConfig):
 @dataclass
 @ConfigFactory.register("autoregressive_lm")
 class AutoRegressiveLMConfig(BaseModelConfig):
-    """Configuration for autoregressive language model."""
+    """Configuration for autoregressive language model.
+
+    Args:
+        model_type (Optional[str]): Model type identifier for AutoModel dispatch. Defaults to None.
+        neftune_alpha (float): NEFTune noise alpha, 0=disabled, typical: 5.0. Defaults to 0.0.
+        vocab_size (Optional[int]): Vocabulary size. Defaults to None.
+        hidden_size (Optional[int]): Hidden dimension size. Defaults to None.
+        num_hidden_layers (Optional[int]): Number of transformer layers. Defaults to None.
+        rms_norm_eps (Optional[float]): Epsilon for RMSNorm. Defaults to None.
+        intermediate_size (Optional[int]): Intermediate size in FFN. Defaults to None.
+        tie_word_embeddings (Optional[bool]): Whether to tie embedding and lm_head weights. Defaults to None.
+        max_position_embeddings (Optional[int]): Maximum sequence length the model was trained with. Defaults to None.
+        rope_theta (Optional[float]): Base frequency for RoPE. Defaults to None.
+        rope_scaling (Optional[dict]): RoPE scaling config, e.g. {"type": "linear", "factor": 4.0}. Defaults to None.
+        attn_type (str): Attention type: 'gqa' or 'mla'. Defaults to "gqa".
+        num_attention_heads (Optional[int]): Number of query attention heads. Defaults to None.
+        num_key_value_heads (Optional[int]): Number of key/value heads for GQA. Defaults to None.
+        use_qk_norm (Optional[bool]): Whether to apply RMSNorm to Q/K. Defaults to None.
+        use_gated_attention (Optional[bool]): Whether to use gated attention. Defaults to None.
+        kv_lora_rank (Optional[int]): KV compression rank, MLA only. Defaults to None.
+        qk_nope_head_dim (Optional[int]): Non-RoPE head dimension, MLA only. Defaults to None.
+        qk_rope_head_dim (Optional[int]): RoPE head dimension, MLA only. Defaults to None.
+        ffn_type (str): FFN type: 'mlp' or 'moe'. Defaults to "mlp".
+        n_routed_experts (Optional[int]): Number of routed experts, MoE only. Defaults to None.
+        n_shared_experts (Optional[int]): Number of shared experts, MoE only. Defaults to None.
+        n_activated_experts (Optional[int]): Number of activated experts per token, MoE only. Defaults to None.
+        topk_method (Optional[str]): Top-k routing method, MoE only. Defaults to None.
+    """
 
     vocab_size: Optional[int] = None
-    dim: Optional[int] = None
-    n_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
-    tie_weight: Optional[bool] = None
-
-    max_len: Optional[int] = None
+    hidden_size: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
+    rms_norm_eps: Optional[float] = None
+    intermediate_size: Optional[int] = None
+    tie_word_embeddings: Optional[bool] = None
+    max_position_embeddings: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
-
     attn_type: str = "gqa"
-    n_heads: Optional[int] = None
-    n_kv_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
+    num_key_value_heads: Optional[int] = None
     use_qk_norm: Optional[bool] = None
     use_gated_attention: Optional[bool] = None
-
     kv_lora_rank: Optional[int] = None
     qk_nope_head_dim: Optional[int] = None
     qk_rope_head_dim: Optional[int] = None
-
     ffn_type: str = "mlp"
     n_routed_experts: Optional[int] = None
     n_shared_experts: Optional[int] = None
     n_activated_experts: Optional[int] = None
     topk_method: Optional[str] = None
 
+    @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
-    n_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
-
-    max_len: Optional[int] = None
+    hidden_size: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
+    rms_norm_eps: Optional[float] = None
+    intermediate_size: Optional[int] = None
+    max_position_embeddings: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
-
     attn_type: str = "gqa"
-    n_heads: Optional[int] = None
-    n_kv_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
+    num_key_value_heads: Optional[int] = None
     use_qk_norm: Optional[bool] = None
     use_gated_attention: Optional[bool] = None
-
     ffn_type: str = "mlp"
     pooling_type: Optional[str] = None
     normalize_embeddings: Optional[bool] = None
+
+    @field_validator("attn_type")
+    def _validate_attn_type(cls, v: str) -> str:
+        if v not in _ATTN_TYPES:
+            raise ValueError(
+                f"attn_type must be one of {sorted(_ATTN_TYPES)}, got {v!r}"
+            )
+        return v
+
+    @field_validator("ffn_type")
+    def _validate_ffn_type(cls, v: str) -> str:
+        if v not in _FFN_TYPES:
+            raise ValueError(f"ffn_type must be one of {sorted(_FFN_TYPES)}, got {v!r}")
+        return v

astrai/config/preprocess_config.py

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

astrai/config/train_config.py

@@ -1,7 +1,9 @@
-from dataclasses import dataclass, field, fields
+from dataclasses import field
 from typing import Any, Callable, Dict, List, Optional
 
 import torch.nn as nn
+from pydantic import ConfigDict, field_validator, model_validator
+from pydantic.dataclasses import dataclass
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
 from torch.utils.data import Dataset
@@ -9,147 +11,204 @@ from torch.utils.data import Dataset
 from astrai.config.base import BaseConfig
 from astrai.model.components.lora import LoRAConfig
 
-
-def required(**kw):
-    return {"required": True, **kw}
+_TRAIN_TYPES = frozenset({"seq", "sft", "dpo", "grpo", "online_grpo", "online_dpo"})
+_PARALLEL_MODES = frozenset({"none", "ddp", "fsdp"})
+_BACKENDS = frozenset({"nccl", "gloo"})
+_START_METHODS = frozenset({"spawn", "fork", "forkserver"})
+_COMPILE_MODES = frozenset({"default", "reduce-overhead", "max-autotune"})
 
 
-@dataclass
+@dataclass(config=ConfigDict(arbitrary_types_allowed=True))
 class TrainConfig(BaseConfig):
-    # basic setting
-    model_fn: Callable[[], nn.Module] = field(
-        default=None, metadata=required(help="Model factory for training.")
-    )
-    strategy: str = field(default=None, metadata=required(help="Training strategy."))
-    dataset: Dataset = field(
-        default=None, metadata=required(help="Dataset for training.")
-    )
-    optimizer_fn: Callable[[nn.Module], Optimizer] = field(
-        default=None, metadata=required(help="Optimizer factory for training.")
-    )
-    scheduler_fn: Callable[[Optimizer], LRScheduler] = field(
-        default=None, metadata=required(help="Scheduler factory for training.")
-    )
-    n_epoch: int = field(default=1, metadata={"help": "Number of epochs for training."})
-    batch_per_device: int = field(
-        default=4, metadata={"help": "Batch size per device."}
-    )
-    grad_accum_steps: int = field(
-        default=1, metadata={"help": "Number of iterations between steps."}
-    )
-    max_grad_norm: Optional[float] = field(
-        default=None,
-        metadata={"help": "Maximum gradient norm. None disables clipping."},
-    )
-    gradient_checkpointing_modules: List[str] = field(
-        default_factory=list,
-        metadata={"help": "Module types to enable activation checkpointing for."},
-    )
+    """Training configuration.
 
-    # checkpoint setting
-    start_epoch: int = field(default=0, metadata={"help": "Start epoch for training."})
-    start_samples: int = field(
-        default=0,
-        metadata={
-            "help": "Start samples count (per rank). Superseded by checkpoint consumed_samples."
-        },
-    )
-    ckpt_dir: str = field(
-        default="./checkpoint", metadata={"help": "Checkpoint directory."}
-    )
-    ckpt_interval: int = field(
-        default=5000,
-        metadata={"help": "Number of optimizer steps between checkpoints."},
-    )
+    Combines hyperparameters with runtime objects (model_fn, dataset, etc.).
+    Only JSON-serializable fields are written to checkpoint meta via to_dict().
 
-    # lora setting
-    lora: Optional[LoRAConfig] = field(
-        default=None,
-        metadata={"help": "LoRA config. None means full fine-tuning."},
-    )
+    Args:
+        model_fn (Callable[[], nn.Module]): Model factory for training.
+        strategy (str): Training strategy (seq, sft, dpo, grpo, online_*).
+        dataset (Dataset): Dataset for training.
+        optimizer_fn (Callable[[nn.Module], Optimizer]): Optimizer factory for training.
+        optimizer_name (Optional[str]): Serializable built-in optimizer identifier. Defaults to None.
+        optimizer_hyperparameters (Dict[str, Any]): Serializable optimizer settings. Defaults to {}.
+        scheduler_fn (Callable[[Optimizer], LRScheduler]): Scheduler factory for training.
+        n_epoch (int): Number of epochs for training. Defaults to 1.
+        batch_per_device (int): Batch size per device. Defaults to 4.
+        grad_accum_steps (int): Number of iterations between optimizer steps. Defaults to 1.
+        max_grad_norm (Optional[float]): Maximum gradient norm. None disables clipping. Defaults to 1.0.
+        gradient_checkpointing_modules (List[type]): Module types to enable activation checkpointing for. Defaults to [].
+        compile_mode (Optional[str]): torch.compile mode: 'default', 'reduce-overhead', 'max-autotune', or None. Defaults to None.
+        start_epoch (int): Start epoch for training. Defaults to 0.
+        start_samples (int): Start samples count (per rank). Superseded by checkpoint consumed_samples. Defaults to 0.
+        ckpt_dir (str): Checkpoint directory. Defaults to "./checkpoint".
+        ckpt_interval (int): Number of optimizer steps between checkpoints. Defaults to 5000.
+        lora (Optional[LoRAConfig]): LoRA config. None means full fine-tuning. Defaults to None.
+        metrics (List[str]): Metrics to record during training. Defaults to ["loss", "lr", "grad_norm"].
+        random_seed (int): Random seed. Defaults to 3407.
+        num_workers (int): Number of workers for dataloader. Defaults to 0.
+        prefetch_factor (Optional[int]): Prefetch factor for dataloader. Defaults to None.
+        pin_memory (bool): Pin memory for dataloader. Defaults to False.
+        collate_fn (Optional[Callable[[List[Any]], Any]]): Collate function for dataloader (e.g. dpo_collate_fn). Defaults to None.
+        nprocs (int): Number of processes for distributed training. Defaults to 1.
+        backend (str): Distributed training backend. Defaults to "nccl".
+        master_addr (str): Master address for distributed training. Defaults to "localhost".
+        master_port (str): Master port for distributed training. Defaults to "29500".
+        parallel_mode (str): Parallel strategy: none, ddp, fsdp. Defaults to "none".
+        start_method (str): Multiprocessing start method: spawn/fork/forkserver. Defaults to "spawn".
+        device_type (str): Device type for distributed training. Defaults to "cuda".
+        val_dataset (Optional[Dataset]): Dataset for validation. Defaults to None.
+        val_split (Optional[float]): Ratio to split from training dataset for validation, e.g. 0.05. Defaults to None.
+        val_step (int): Number of optimizer steps between validation runs. Defaults to 1000.
+        neftune_alpha (float): NEFTune noise alpha, 0=disabled, typical: 5.0. Defaults to 0.0.
+        rollout_interval (int): Number of optimizer steps between online rollouts. Defaults to 512.
+        rollout_temperature (float): Sampling temperature for online rollout. Defaults to 0.7.
+        rollout_top_k (int): Top-k filtering for online rollout, 0=disable. Defaults to 0.
+        rollout_top_p (float): Top-p (nucleus) filtering for online rollout. Defaults to 0.9.
+        rollout_max_tokens (int): Maximum generated tokens per response in rollout. Defaults to 1024.
+        reward_model_fn (Optional[Callable]): Factory for reward model, required for online RL strategies. Defaults to None.
+        executor_kwargs (Dict[str, Any]): Extra kwargs passed to ExecutorFactory.create(). Defaults to {}.
+        extra_kwargs (Dict[str, Any]): Other arguments. Defaults to {}.
+    """
 
-    # metric setting
-    log_dir: str = field(
-        default="./checkpoint/logs", metadata={"help": "Directory for metric logs."}
-    )
-    metrics: List[str] = field(
-        default_factory=lambda: ["loss", "lr", "grad_norm"],
-        metadata={"help": "Metrics to record during training."},
-    )
+    model_fn: Callable[[], nn.Module]
+    strategy: str
+    dataset: Dataset
+    optimizer_fn: Callable[[nn.Module], Optimizer]
+    scheduler_fn: Callable[[Optimizer], LRScheduler]
+    optimizer_name: Optional[str] = None
+    optimizer_hyperparameters: Dict[str, Any] = field(default_factory=dict)
+    n_epoch: int = 1
+    batch_per_device: int = 4
+    grad_accum_steps: int = 1
+    max_grad_norm: Optional[float] = 1.0
+    gradient_checkpointing_modules: List[type] = field(default_factory=list)
+    compile_mode: Optional[str] = None
 
-    # dataloader setting
-    random_seed: int = field(default=3407, metadata={"help": "Random seed."})
-    num_workers: int = field(
-        default=0, metadata={"help": "Number of workers for dataloader."}
-    )
-    prefetch_factor: Optional[int] = field(
-        default=None, metadata={"help": "Prefetch factor for dataloader."}
-    )
-    pin_memory: bool = field(
-        default=False, metadata={"help": "Pin memory for dataloader."}
-    )
-    collate_fn: Optional[Callable[[List[Any]], Any]] = field(
-        default=None,
-        metadata={"help": "Collate function for dataloader (e.g. dpo_collate_fn)."},
-    )
+    start_epoch: int = 0
+    start_samples: int = 0
+    ckpt_dir: str = "./checkpoint"
+    ckpt_interval: int = 5000
 
-    # distributed training
-    nprocs: int = field(
-        default=1, metadata={"help": "Number of processes for distributed training."}
-    )
-    backend: str = field(
-        default="nccl", metadata={"help": "Distributed training backend."}
-    )
-    master_addr: str = field(
-        default="localhost",
-        metadata={"help": "Master address for distributed training."},
-    )
-    master_port: str = field(
-        default="29500", metadata={"help": "Master port for distributed training."}
-    )
-    parallel_mode: str = field(
-        default="none",
-        metadata={"help": "Parallel strategy: none, ddp, fsdp."},
-    )
-    start_method: str = field(
-        default="spawn",
-        metadata={"help": "Multiprocessing start method (spawn/fork/forkserver)."},
-    )
+    lora: Optional[LoRAConfig] = None
 
-    # others
-    device_type: str = field(
-        default="cuda", metadata={"help": "Device type for distributed training."}
-    )
-    val_dataset: Optional[Dataset] = field(
-        default=None, metadata={"help": "Dataset for validation."}
-    )
-    val_split: Optional[float] = field(
-        default=None,
-        metadata={
-            "help": "Ratio to split from training dataset for validation (e.g. 0.05). Ignored if val_dataset is set."
-        },
-    )
-    val_step: int = field(
-        default=1000,
-        metadata={"help": "Number of optimizer steps between validation runs."},
-    )
-    neftune_alpha: float = field(
-        default=0.0,
-        metadata={"help": "NEFTune noise alpha (0=disabled, typical: 5.0)."},
-    )
+    metrics: List[str] = field(default_factory=lambda: ["loss", "lr", "grad_norm"])
 
-    executor_kwargs: Dict[str, Any] = field(
-        default_factory=dict,
-        metadata={"help": "Extra kwargs passed to ExecutorFactory.create()."},
-    )
-    extra_kwargs: Dict[str, Any] = field(
-        default_factory=dict, metadata={"help": "Other arguments."}
-    )
+    random_seed: int = 3407
+    num_workers: int = 0
+    prefetch_factor: Optional[int] = None
+    pin_memory: bool = False
+    collate_fn: Optional[Callable[[List[Any]], Any]] = None
 
-    def __post_init__(self):
-        self.validate()
+    nprocs: int = 1
+    backend: str = "nccl"
+    master_addr: str = "localhost"
+    master_port: str = "29500"
+    parallel_mode: str = "none"
+    start_method: str = "spawn"
 
-    def validate(self):
-        for fld in fields(self):
-            if fld.metadata.get("required") and getattr(self, fld.name) is None:
-                raise ValueError(f"TrainConfig.{fld.name} is required but got None.")
+    device_type: str = "cuda"
+    val_dataset: Optional[Dataset] = None
+    val_split: Optional[float] = None
+    val_step: int = 1000
+    neftune_alpha: float = 0.0
+
+    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

@@ -6,7 +6,6 @@ from astrai.dataset.dataset import (
 )
 from astrai.dataset.sampler import RDSampler
 from astrai.dataset.storage import (
-    H5Store,
     JsonlStore,
     MmapStore,
     Recordable,
@@ -17,9 +16,7 @@ from astrai.dataset.storage import (
 )
 from astrai.serialization import (
     load_bin,
-    load_h5,
     save_bin,
-    save_h5,
 )
 
 __all__ = [
@@ -31,12 +28,9 @@ __all__ = [
     "Streamable",
     "Recordable",
     "StoreFactory",
-    "H5Store",
     "MmapStore",
     "JsonlStore",
     "detect_format",
-    "save_h5",
-    "load_h5",
     "save_bin",
     "load_bin",
     "RDSampler",

astrai/dataset/dataset.py

@@ -190,7 +190,8 @@ def grpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
       - rewards:  [G]
 
     Output:
-      - prompts:   [B, P_max]
+      - prompts:   [B, P_max], left-padded
+      - prompt_mask: [B, P_max]
       - responses: [B, G, R_max]
       - masks:     [B, G, R_max]
       - rewards:   [B, G]
@@ -201,13 +202,15 @@ def grpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
     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"]
+        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)
@@ -217,6 +220,7 @@ def grpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
 
     return {
         "prompts": prompts,
+        "prompt_mask": prompt_mask,
         "responses": responses,
         "masks": masks,
         "rewards": rewards,
@@ -310,7 +314,7 @@ class DatasetFactory(BaseFactory["BaseDataset"]):
                 stream datasets (SEQ/SFT).  Record datasets ignore it.
             stride: Stride between consecutive stream samples
                 (default: same as *window_size*).
-            storage_type: Storage backend ("h5", "bin", "jsonl") or
+            storage_type: Storage backend ("bin", "jsonl") or
                 None for auto-detection.
             tokenizer_path: Path to tokenizer for lazy JSONL
                 tokenisation (record datasets only).
@@ -346,7 +350,15 @@ class DatasetFactory(BaseFactory["BaseDataset"]):
         if processor is not None:
             store.load(load_path, processor=processor, **kwargs)
         else:
-            store.load(load_path, **kwargs)
+            load_kwargs = dict(kwargs)
+            if (
+                tokenizer_path is not None
+                and storage_type == "jsonl"
+                and train_type in ("seq", "sft")
+                and "tokenizer_path" not in load_kwargs
+            ):
+                load_kwargs["tokenizer_path"] = tokenizer_path
+            store.load(load_path, **load_kwargs)
 
         return cls.create(train_type, store=store)
 
@@ -372,7 +384,7 @@ class DatasetFactory(BaseFactory["BaseDataset"]):
         """Build an on-the-fly tokenisation processor if applicable.
 
         Only raw JSONL + record datasets (DPO/GRPO) need a processor;
-        pre-tokenised backends (H5/bin) and stream datasets (SEQ/SFT)
+        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":
@@ -439,7 +451,7 @@ class DPODataset(BaseDataset):
 
     Two loading paths (handled by :class:`DatasetFactory`):
 
-    - **Pre-tokenized** (H5/bin): ``store.load(path)`` reads per-record
+    - **Pre-tokenized** (bin): ``store.load(path)`` reads per-record
       tensors; ``__getitem__`` returns them directly.
     - **Raw JSONL** (``tokenizer_path=...``): builds a lazy processor
       via :func:`dpo_processor` that tokenises on the fly — no packing,

astrai/dataset/storage.py

@@ -10,7 +10,6 @@ Architecture (composition over inheritance):
     Streamable (mixin)        — raw token slice  fetch(begin, end, keys)
     Recordable (mixin)        — raw record slice fetch_record(idx, keys)
 
-    H5Store(Store, Streamable, Recordable)
     MmapStore(Store, Streamable, Recordable)
     JsonlStore(Store, Streamable, Recordable)
 
@@ -36,9 +35,9 @@ 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 (H5/
-JSONL) or opaque shards (bin).  Record access for bin relies on
-``_offsets`` instead.
+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
@@ -56,12 +55,12 @@ from typing import Callable, Dict, List, Optional, Tuple, Union
 import torch
 from torch import Tensor
 
+from astrai.config.preprocess_config import PipelineConfig
 from astrai.factory import BaseFactory
 from astrai.preprocessing.transform import TokenizeTransform
 from astrai.serialization import (
     load_bin,
     load_bin_offsets,
-    load_h5,
 )
 
 logger = logging.getLogger(__name__)
@@ -82,19 +81,10 @@ def detect_format(load_path: str) -> str:
     root = Path(load_path)
     if root.is_file():
         suffix = root.suffix.lower()
-        if suffix in (".h5", ".hdf5"):
-            return "h5"
         if suffix == ".jsonl":
             return "jsonl"
         raise ValueError(f"Unsupported file format: {suffix}")
 
-    h5_files = [
-        Path(p)
-        for pattern in ("*.h5", "*.hdf5")
-        for p in glob.glob(str(root / "**" / pattern), recursive=True)
-    ]
-    if h5_files:
-        return "h5"
     bin_files = [Path(p) for p in glob.glob(str(root / "**" / "*.bin"), recursive=True)]
     if bin_files:
         has_meta = (root / "meta.json").exists() or len(
@@ -184,7 +174,7 @@ class Store(ABC):
         """Number of records available via :meth:`fetch_record`.
 
         Non-zero only when the backing layout provides per-record
-        indexing (H5/JSONL segments or bin ``_offsets``).
+        indexing (JSONL segments or bin ``_offsets``).
         """
         return self._num_records
 
@@ -268,7 +258,7 @@ class Store(ABC):
         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 (H5/JSONL), ``_data[key]`` is
+        ``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
@@ -304,7 +294,7 @@ class Store(ABC):
                     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 H5/JSONL.",
+                        "supported). Merge shards or use JSONL.",
                         key,
                         len(segs),
                     )
@@ -329,7 +319,7 @@ class Streamable:
     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 (H5/JSONL/bin+offsets), the
+    that can also serve record access (JSONL/bin+offsets), the
     ``fetch_record`` API from :class:`Recordable` is used instead.
     """
 
@@ -414,33 +404,6 @@ class StoreFactory(BaseFactory["Store"]):
     """Factory for creating Store instances by type name."""
 
 
-@StoreFactory.register("h5")
-class H5Store(Store, Streamable, Recordable):
-    """HDF5-based storage backend (pre-tokenized data).
-
-    Each key is stored as a group of per-record datasets (``data_0``,
-    ``data_1``, …).  Supports both access modes:
-
-    - **Stream**: ``fetch(begin, end, key)`` and ``store[i]`` slice
-      across concatenated records via ``_cum`` — used by SEQ/SFT.
-    - **Record**: ``fetch_record(i, key)`` and ``store[i]`` (when
-      ``window_size == 0``) index ``_data[key]`` directly — used by
-      DPO/GRPO.
-    """
-
-    segments_are_records = True
-
-    def __init__(
-        self,
-        window_size: int = 0,
-        stride: Optional[int] = None,
-    ):
-        super().__init__(window_size=window_size, stride=stride)
-
-    def load(self, path: str, **kwargs):
-        self._normalize(load_h5(path))
-
-
 @StoreFactory.register("bin")
 class MmapStore(Store, Streamable, Recordable):
     """Memory-mapped binary storage backend.
@@ -545,18 +508,29 @@ class JsonlSource:
 
 @StoreFactory.register("jsonl")
 class JsonlStore(Store, Streamable, Recordable):
-    """JSONL reader with two tokenisation modes.
+    """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.
 
-    Two modes, selected at :meth:`load` time:
+    Three ways to supply an eager transform (first match wins):
 
-    - **Eager** (default): applies a :class:`TokenizeTransform` to every
-      record at load time and registers per-key tensors via
-      ``_normalize``.  Both ``fetch`` (stream) and ``fetch_record``
-      (record) work.
+    - **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.
@@ -565,6 +539,16 @@ class JsonlStore(Store, Streamable, Recordable):
     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,
@@ -587,14 +571,20 @@ class JsonlStore(Store, Streamable, Recordable):
         if transform is None:
             root = Path(path)
             config_path = root / self.CONFIG_NAME if root.is_dir() else None
-            if config_path is None or not config_path.exists():
+            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, or pass processor= for lazy "
-                    f"on-the-fly tokenisation."
+                        f"explicit transform, pass processor= for lazy "
+                        f"on-the-fly tokenisation, or pass tokenizer_path= to "
+                        f"use the built-in messages config."
                     )
-            transform = TokenizeTransform.from_config_file(str(config_path))
+                config = PipelineConfig.from_dict(self._DEFAULT_MESSAGES_CONFIG)
+                transform = TokenizeTransform(config, tokenizer_path)
 
         transformed = transform.apply(records)
         self._normalize(transformed)

astrai/extension/__init__.py

@@ -4,26 +4,46 @@ 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
 
-Interface (shared by all wrappers):
-    causal_offset: -1 = non-causal; >=0 = absolute position of first Q token
-    mask:       2D [batch, kv_len] or 3D [batch, q_len, kv_len] (bool, True = keep)
-    scale:      0.0 = auto (1/sqrt(head_dim)); >0 = explicit
-    layout:     "bhld" (default) or "blhd"
+Layout convention: all q/k/v are ``[batch, seq_len, n_heads, head_dim]``
+(blhd). Scale is always ``1/sqrt(head_dim)``.
 
-Causal and mask can coexist — both are applied simultaneously.
-
-Each wrapper dispatches to its compiled CUDA kernel (``astrai.extension.attn_*``)
-when available, otherwise falls back to ``torch.nn.functional.scaled_dot_product_attention``.
+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.ops import attn_decode, attn_paged_decode, attn_prefill
+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 +1,422 @@
+"""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
+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

@@ -11,14 +11,14 @@ import logging
 
 logger = logging.getLogger(__name__)
 
-KERNEL_NAMES = ["attn_decode", "attn_prefill", "attn_paged_decode"]
+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".{_name}", package=__package__)
+        _mod = importlib.import_module(f".lib.{_name}", package=__package__)
         _available[_name] = True
         _modules[_name] = _mod
     except ImportError:

astrai/extension/ops.py

@@ -1,246 +0,0 @@
-"""GQA attention wrapper functions — one entry point per compiled kernel.
-
-Each wrapper dispatches to its CUDA kernel (loaded in ``loader.py``) when
-available, otherwise falls back to ``torch`` SDPA.
-
-Interface (all functions):
-    causal_offset: -1 = non-causal; >=0 = absolute position of first Q token
-    mask:       2D [batch, kv_len] or 3D [batch, q_len, kv_len] (bool)
-    scale:      0.0 = auto (1/sqrt(head_dim)); >0 = explicit
-    layout:     "bhld" (default) or "blhd"
-
-Add new kernel wrappers here; split into per-variant files only if this file
-grows large.
-"""
-
-import math
-
-import torch
-import torch.nn.functional as F
-
-from astrai.extension.loader import _available, _modules
-
-_LAYOUT_CODES: dict[str, int] = {"bhld": 0, "blhd": 1}
-
-
-def _parse_layout(layout: str | int) -> int:
-    if isinstance(layout, int):
-        return layout
-    code = _LAYOUT_CODES.get(layout.lower())
-    if code is None:
-        raise ValueError(
-            f"unknown layout '{layout}', expected one of {list(_LAYOUT_CODES)}"
-        )
-    return code
-
-
-def _to_bhld(t: torch.Tensor, layout: int) -> torch.Tensor:
-    """Normalize to b h l d view. Zero-copy transpose if layout==1 (b l h d)."""
-    if layout == 1:
-        return t.transpose(1, 2)
-    return t
-
-
-def _expand_kv_heads(
-    k: torch.Tensor, v: torch.Tensor, q_head: int
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """Expand K/V heads to match Q heads for GQA fallback."""
-    kv_head = k.size(1)
-    if kv_head == q_head:
-        return k, v
-    group = q_head // kv_head
-    k = k.repeat_interleave(group, dim=1)
-    v = v.repeat_interleave(group, dim=1)
-    return k, v
-
-
-def _build_attn_mask(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    mask: torch.Tensor | None,
-    causal_offset: int,
-    scale: float,
-) -> tuple[torch.Tensor | None, float]:
-    """Build SDPA-compatible attn_mask + resolved scale.
-
-    q and k must already be in b h l d layout.
-    Causal and mask can coexist: causal sets -inf above the diagonal, mask
-    sets -inf for padded positions. Both are OR'd into a single bool mask.
-    """
-    q_len = q.size(2)
-    kv_len = k.size(2)
-    head_dim = q.size(3)
-    resolved_scale = scale if scale and scale > 0 else 1.0 / math.sqrt(head_dim)
-
-    attn_mask = None
-
-    if mask is not None:
-        if mask.dim() == 2:
-            # [batch, kv_len] → [batch, 1, 1, kv_len]
-            attn_mask = mask[:, None, None, :]
-        elif mask.dim() == 3:
-            # [batch, q_len, kv_len] → [batch, 1, q_len, kv_len]
-            attn_mask = mask[:, None, :, :]
-        else:
-            raise ValueError(f"mask must be 2D or 3D, got {mask.dim()}D")
-
-    if causal_offset >= 0:
-        batch = q.size(0)
-        # q row i attends to kv cols 0..(causal_offset + i)
-        q_idx = torch.arange(q_len, device=q.device).unsqueeze(1)  # [q_len, 1]
-        kv_idx = torch.arange(kv_len, device=q.device).unsqueeze(0)  # [1, kv_len]
-        causal_bool = kv_idx > (causal_offset + q_idx)  # True = masked out
-        causal_mask = causal_bool.unsqueeze(0).expand(
-            batch, -1, -1
-        )  # [batch, q_len, kv_len]
-        causal_mask = causal_mask[:, None, :, :]  # [batch, 1, q_len, kv_len]
-
-        if attn_mask is not None:
-            attn_mask = attn_mask | causal_mask
-        else:
-            attn_mask = causal_mask
-
-    return attn_mask, resolved_scale
-
-
-def _torch_fallback(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    mask: torch.Tensor | None,
-    causal_offset: int,
-    scale: float,
-    q_layout: int,
-    kv_layout: int | None = None,
-) -> torch.Tensor:
-    """Reference attention via ``scaled_dot_product_attention``.
-
-    q_layout / kv_layout: 0 = b h l d, 1 = b l h d.
-    If kv_layout is None, uses q_layout (Q and K/V share the same layout).
-    """
-    if kv_layout is None:
-        kv_layout = q_layout
-    q = _to_bhld(q, q_layout)
-    k = _to_bhld(k, kv_layout)
-    v = _to_bhld(v, kv_layout)
-    k, v = _expand_kv_heads(k, v, q.size(1))
-    attn_mask, resolved_scale = _build_attn_mask(q, k, mask, causal_offset, scale)
-    out = F.scaled_dot_product_attention(
-        q, k, v, attn_mask=attn_mask, is_causal=False, scale=resolved_scale
-    )
-    # Restore Q's original layout
-    if q_layout == 1:
-        out = out.transpose(1, 2)
-    return out
-
-
-def _gather_kv_from_pages(
-    page_table: torch.Tensor,
-    k_cache: torch.Tensor,
-    v_cache: torch.Tensor,
-    page_size: int,
-    kv_len: int,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """Gather contiguous K/V from paged cache for torch SDPA fallback.
-
-    Shapes:
-        page_table : [batch, max_pages] (int64)
-        k_cache    : [n_pages, page_size, n_kv_heads, head_dim]
-        v_cache    : same as k_cache
-    Returns:
-        k, v : [batch, kv_len, n_kv_heads, head_dim]  (b l h d)
-    """
-    batch, max_pages = page_table.shape
-    _, ps, n_kv_heads, head_dim = k_cache.shape
-    if ps != page_size:
-        raise ValueError(f"k_cache page_size mismatch: {ps} vs {page_size}")
-
-    # Vectorized gather: build physical page + offset indices, then advanced-index
-    positions = torch.arange(kv_len, device=page_table.device)
-    logical_pages = positions // page_size  # [kv_len]
-    page_offsets = positions % page_size  # [kv_len]
-
-    phys_pages = page_table[:, logical_pages]  # [batch, kv_len]
-    # k_cache[phys_pages, page_offsets] → [batch, kv_len, n_kv_heads, head_dim] (b l h d)
-    k = k_cache[phys_pages, page_offsets]
-    v = v_cache[phys_pages, page_offsets]
-    return k, v
-
-
-def attn_decode(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    mask: torch.Tensor | None = None,
-    causal_offset: int = -1,
-    scale: float = 0.0,
-    layout: str = "bhld",
-) -> torch.Tensor:
-    li = _parse_layout(layout)
-    if _available["attn_decode"]:
-        return _modules["attn_decode"].attn_decode(
-            q,
-            k,
-            v,
-            mask=mask,
-            causal_offset=causal_offset,
-            scale=scale,
-            layout=li,
-        )
-    return _torch_fallback(q, k, v, mask, causal_offset, scale, q_layout=li)
-
-
-def attn_prefill(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    mask: torch.Tensor | None = None,
-    causal_offset: int = -1,
-    scale: float = 0.0,
-    layout: str = "bhld",
-) -> torch.Tensor:
-    li = _parse_layout(layout)
-    if _available["attn_prefill"]:
-        return _modules["attn_prefill"].attn_prefill(
-            q,
-            k,
-            v,
-            mask=mask,
-            causal_offset=causal_offset,
-            scale=scale,
-            layout=li,
-        )
-    return _torch_fallback(q, k, v, mask, causal_offset, scale, q_layout=li)
-
-
-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,
-    causal_offset: int = -1,
-    scale: float = 0.0,
-    layout: str = "bhld",
-) -> torch.Tensor:
-    li = _parse_layout(layout)
-    if _available["attn_paged_decode"]:
-        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,
-            scale=scale,
-            layout=li,
-        )
-    # Gathered K/V are always b l h d
-    k, v = _gather_kv_from_pages(page_table, k_cache, v_cache, page_size, kv_len)
-    return _torch_fallback(
-        q, k, v, mask, causal_offset, scale, q_layout=li, kv_layout=1
-    )

astrai/extension/rotary_backend.py

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

astrai/extension/rotary_ops.py

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

astrai/factory.py

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

astrai/inference/__init__.py

@@ -30,21 +30,16 @@ from astrai.inference.api.openai import OpenAIResponseBuilder
 from astrai.inference.core import (
     STOP,
     Allocator,
-    CacheView,
-    ContiguousCache,
-    ContiguousCacheView,
     Executor,
     InferenceScheduler,
     KVCache,
-    PageCache,
-    PageCacheView,
+    KVStorage,
     PagePool,
     PrefixCache,
-    Storage,
+    ReqToTokenPool,
     Task,
     TaskManager,
     TaskStatus,
-    TaskTable,
     page_hash,
 )
 from astrai.inference.engine import GenerationRequest, InferenceEngine
@@ -68,16 +63,11 @@ __all__ = [
     "TaskManager",
     "TaskStatus",
     "Allocator",
-    "CacheView",
     "KVCache",
-    "ContiguousCache",
-    "ContiguousCacheView",
-    "PageCache",
-    "PageCacheView",
+    "KVStorage",
     "PagePool",
     "PrefixCache",
-    "Storage",
-    "TaskTable",
+    "ReqToTokenPool",
     "page_hash",
     "sample",
     "BaseSamplingStrategy",

astrai/inference/api/server.py

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

astrai/inference/api/tool_parser.py

@@ -22,13 +22,10 @@ class BaseToolParser(ABC):
     Maintains streaming state internally so that each call to :meth:`feed`
     can diff against previously emitted content.
 
-    Parameters
-    ----------
-    tools : list of dict, optional
-        Tool definitions from the request.
-    tool_choice : str
-        ``"auto"`` / ``"required"`` / ``"none"`` or a named tool choice
-        dict.
+    Args:
+        tools (list of dict, optional): Tool definitions from the request.
+        tool_choice (str): ``"auto"`` / ``"required"`` / ``"none"`` or a named
+            tool choice dict.
     """
 
     def __init__(self, tools: Optional[List[Dict]] = None, tool_choice: str = "auto"):
@@ -51,14 +48,12 @@ class BaseToolParser(ABC):
 
         Returns an empty list when nothing new should be emitted.
 
-        Parameters
-        ----------
-        body : str
-            The complete accumulated generated text so far.
-        current_token_ids : list of int, optional
-            All token IDs decoded into *body* (cumulative).
-        delta_token_ids : list of int, optional
-            Only the token IDs for this chunk.
+        Args:
+            body (str): The complete accumulated generated text so far.
+            current_token_ids (list of int, optional): All token IDs decoded
+                into *body* (cumulative).
+            delta_token_ids (list of int, optional): Only the token IDs for
+                this chunk.
         """
 
     @abstractmethod

astrai/inference/core/__init__.py

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

astrai/inference/core/cache.py

@@ -1,7 +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 abc import ABC, abstractmethod
 from collections import OrderedDict
-from typing import Callable, Dict, List, Optional, Tuple
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional
 
 import torch
 from torch import Tensor
@@ -108,426 +122,380 @@ class PrefixCache:
             self._hash_to_page[h] = page_idx
 
 
-class PagePool:
-    """Orchestrates allocator (page management) and PrefixCache (content addressing)."""
+class ReqToTokenPool:
+    """Maps [req_idx, pos] -> physical token slot in KV storage.
 
-    def __init__(self, allocator: Allocator, prefix: PrefixCache):
-        self._alloc = allocator
-        self._prefix = prefix
-        self._alloc.on_evict = prefix.evict
+    Each row is one request; each column is a sequence position. The value
+    at [req_idx, pos] is the flat index into the KV storage buffers.
+    """
 
-    @property
-    def allocator(self) -> Allocator:
-        return self._alloc
-
-    @property
-    def prefix(self) -> PrefixCache:
-        return self._prefix
-
-    def alloc(self) -> int:
-        return self._alloc.alloc()
-
-    def free(self, idx: int):
-        keep = self._prefix.has_page(idx)
-        self._alloc.free(idx, keep_cached=keep)
-        if not keep:
-            self._prefix.evict(idx)
-
-    def inc_ref(self, idx: int):
-        self._alloc.inc_ref(idx)
-
-    def lookup(self, token_ids: List[int]) -> List[int]:
-        hits = self._prefix.lookup(token_ids)
-        for p in hits:
-            self._alloc.touch(p)
-        return hits
-
-    def record(self, page_idx: int, token_ids: List[int], logical_page_idx: int):
-        self._prefix.record(page_idx, token_ids, logical_page_idx)
-
-
-class TaskTable:
-    """Maps task_ids to page tables and cached token counts."""
-
-    def __init__(self, page_size: int):
-        self._page_size = page_size
-        self._pages: Dict[str, List[int]] = {}
-        self._cached: Dict[str, int] = {}
+    def __init__(self, size: int, max_context_len: int, device: torch.device):
+        self.size = size
+        self.max_context_len = max_context_len
+        self.req_to_token = torch.zeros(
+            (size, max_context_len), dtype=torch.long, device=device
+        )
+        self.free_slots = list(range(size))
         self._lock = threading.Lock()
 
-    def 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
-            self._cached[task_id] = cached
+            if num_reqs > len(self.free_slots):
+                return None
+            slots = self.free_slots[:num_reqs]
+            self.free_slots = self.free_slots[num_reqs:]
+            return slots
 
-    def 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:
-        with self._lock:
-            return self._cached.get(task_id, 0)
-
-    def pop(self, task_id: str) -> Tuple[List[int], int]:
-        with self._lock:
-            pages = self._pages.pop(task_id, [])
-            cached = self._cached.pop(task_id, 0)
-            return pages, cached
-
-    def get_ref(self, task_id: str) -> List[int]:
-        with self._lock:
-            return self._pages.setdefault(task_id, [])
-
-    def table_tensor(self, task_ids: List[str], device: torch.device) -> Tensor:
-        with self._lock:
-            states = [self._pages.get(tid, []) for tid in task_ids]
-            max_pages = max((len(s) for s in states), default=0)
-            rows = [s + [-1] * (max_pages - len(s)) for s in states]
-            return torch.tensor(rows, dtype=torch.long, device=device)
+    def write(self, indices, values):
+        self.req_to_token[indices] = values
 
 
-class Storage:
-    """KV-cache tensor storage with paged write/gather."""
+class KVStorage:
+    """Token-level KV cache storage.
+
+    Buffers: [n_layers, size, n_kv_heads, head_dim]. Each token occupies
+    one slot indexed by ReqToTokenPool.
+    """
 
     def __init__(
         self,
+        size: int,
         n_layers: int,
-        n_pages: int,
-        page_size: int,
         n_kv_heads: int,
         head_dim: int,
         device: torch.device,
         dtype: torch.dtype,
     ):
-        self.page_size = page_size
-        self.k_cache = torch.empty(
-            (n_layers, n_pages, page_size, n_kv_heads, head_dim),
-            device=device,
-            dtype=dtype,
+        self.size = size
+        self.k_buffer = torch.empty(
+            (n_layers, size, n_kv_heads, head_dim), device=device, dtype=dtype
         )
-        self.v_cache = torch.empty(
-            (n_layers, n_pages, page_size, n_kv_heads, head_dim),
-            device=device,
-            dtype=dtype,
+        self.v_buffer = torch.empty(
+            (n_layers, size, n_kv_heads, head_dim), device=device, dtype=dtype
         )
 
-    def write(
-        self,
-        layer_id: int,
-        page_table: Tensor,
-        start_pos: int,
-        k: Tensor,
-        v: Tensor,
-    ):
-        seq_len = k.size(1)
-        if seq_len == 0:
-            return
-        page_size = self.page_size
-        written = 0
-        first_page = start_pos // page_size
-        last_page = (start_pos + seq_len - 1) // page_size
-        for pi in range(first_page, last_page + 1):
-            phys_pages = page_table[:, pi]
-            page_start = pi * page_size
-            write_start = max(page_start, start_pos)
-            write_end = min(page_start + page_size, start_pos + seq_len)
-            offset = write_start - page_start
-            chunk = write_end - write_start
-            valid = phys_pages >= 0
-            if not valid.all():
-                if valid.any():
-                    valid_pages = phys_pages[valid]
-                    self.k_cache[layer_id, valid_pages, offset : offset + chunk] = k[
-                        valid, written : written + chunk
-                    ]
-                    self.v_cache[layer_id, valid_pages, offset : offset + chunk] = v[
-                        valid, written : written + chunk
-                    ]
-                written += chunk
-                continue
-            self.k_cache[layer_id, phys_pages, offset : offset + chunk] = k[
-                :, written : written + chunk
-            ]
-            self.v_cache[layer_id, phys_pages, offset : offset + chunk] = v[
-                :, written : written + chunk
-            ]
-            written += chunk
+    def get_key_buffer(self, layer_id: int) -> Tensor:
+        return self.k_buffer[layer_id]
 
-    def gather(
-        self, layer_id: int, page_table: Tensor, total_len: int
-    ) -> Tuple[Tensor, Tensor]:
-        safe = page_table.clamp(min=0)
-        k = self.k_cache[layer_id, safe]
-        v = self.v_cache[layer_id, safe]
-        k = k.flatten(1, 2)
-        v = v.flatten(1, 2)
-        if (page_table < 0).any():
-            invalid = (
-                (page_table < 0)
-                .unsqueeze(-1)
-                .expand(-1, -1, self.page_size)
-                .flatten(1, 2)
-            )
-            invalid = invalid[:, :, None, None].expand_as(k)
-            k = k.masked_fill(invalid, 0.0)
-            v = v.masked_fill(invalid, 0.0)
-        k = k[:, :total_len]
-        v = v[:, :total_len]
-        return k, v
+    def get_value_buffer(self, layer_id: int) -> Tensor:
+        return self.v_buffer[layer_id]
+
+    def set_kv_buffer(self, layer_id: int, loc: Tensor, k: Tensor, v: Tensor) -> None:
+        self.k_buffer[layer_id, loc] = k
+        self.v_buffer[layer_id, loc] = v
 
 
-class CacheView(ABC):
-    """Abstract view passed to attention layers for KV-cache I/O."""
+@dataclass
+class KVCache:
+    """Pure data struct passed to model for KV cache I/O.
 
-    @abstractmethod
-    def write(self, layer_id: int, k: Tensor, v: Tensor): ...
+    The attention layer does raw buffer indexing — no methods, no abstraction.
 
-    @abstractmethod
-    def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]: ...
+    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 KVCache(ABC):
-    """Abstract KV-cache facade for scheduler/executor."""
+class PagePool:
+    """Top-level KV cache manager.
 
-    @abstractmethod
-    def task_alloc(self, task_id: str, prompt_ids: List[int]) -> bool: ...
+    Combines KVStorage + ReqToTokenPool + Allocator + PrefixCache.
 
-    @abstractmethod
-    def task_free(self, task_id: str): ...
-
-    @abstractmethod
-    def task_extend(self, task_id: str, pos: int) -> bool: ...
-
-    @abstractmethod
-    def bind_tasks(
-        self,
-        task_ids: List[str],
-        total_len: int,
-        device: torch.device,
-        write_positions: Optional[Tensor] = None,
-    ) -> CacheView: ...
-
-    def task_cached(self, task_id: str) -> int:
-        return 0
-
-    def task_record_hashes(
-        self, task_id: str, prompt_ids: List[int], start_logical_page: int = 0
-    ): ...
-
-
-class PageCacheView(CacheView):
-    """Bundles Storage + page_table + total_len for attention layers."""
-
-    def __init__(self, storage: Storage, page_table: Tensor, total_len: int = 0):
-        self._storage = storage
-        self._page_table = page_table
-        self._total_len = total_len
-
-    def write(self, layer_id: int, k: Tensor, v: Tensor):
-        start_pos = self._total_len - k.size(1)
-        self._storage.write(layer_id, self._page_table, start_pos, k, v)
-
-    def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]:
-        return self._storage.gather(layer_id, self._page_table, self._total_len)
-
-
-class PageCache(KVCache):
-    """Paged KV-cache with prefix sharing."""
+    Args:
+        n_layers: Number of transformer layers.
+        n_kv_heads: Number of KV attention heads.
+        head_dim: Dimension per head.
+        max_batch_size: Maximum concurrent requests.
+        max_seq_len: Maximum sequence length per request.
+        device, dtype: Tensor device and dtype.
+        page_size: Page size for paged mode (1 = token-level).
+        n_tokens: Total token slots for paged mode. None = contiguous mode
+            (pre-allocates max_batch_size * max_seq_len).
+    """
 
     def __init__(
         self,
         n_layers: int,
-        n_pages: int,
-        page_size: int,
         n_kv_heads: int,
         head_dim: int,
-        device: torch.device,
-        dtype: torch.dtype,
-    ):
-        self.page_size = page_size
-        self._pool = PagePool(Allocator(n_pages), PrefixCache(page_size))
-        self._table = TaskTable(page_size)
-        self._storage = Storage(
-            n_layers, n_pages, page_size, n_kv_heads, head_dim, device, dtype
-        )
-
-    def task_alloc(self, task_id: str, prompt_ids: List[int]) -> bool:
-        hits = self._pool.lookup(prompt_ids)
-        cached = len(hits) * self.page_size
-        for p in hits:
-            self._pool.inc_ref(p)
-
-        remaining = len(prompt_ids) - cached
-        n_new = (
-            (remaining + self.page_size - 1) // self.page_size if remaining > 0 else 0
-        )
-        new_pages: List[int] = []
-        if n_new > 0:
-            for _ in range(n_new):
-                p = self._pool.alloc()
-                if p < 0:
-                    for hp in hits:
-                        self._pool.free(hp)
-                    for np in new_pages:
-                        self._pool.free(np)
-                    return False
-                new_pages.append(p)
-
-        self._table.set(task_id, hits + new_pages, cached)
-        return True
-
-    def task_free(self, task_id: str):
-        page_table, _ = self._table.pop(task_id)
-        for idx in page_table:
-            self._pool.free(idx)
-
-    def task_extend(self, task_id: str, pos: int) -> bool:
-        page_table = self._table.get(task_id)
-        needed = (pos + 1 + self.page_size - 1) // self.page_size
-        while len(page_table) < needed:
-            p = self._pool.alloc()
-            if p < 0:
-                return False
-            page_table.append(p)
-        return True
-
-    def task_cached(self, task_id: str) -> int:
-        return self._table.get_cached(task_id)
-
-    def task_record_hashes(
-        self, task_id: str, prompt_ids: List[int], start_logical_page: int = 0
-    ):
-        page_table = self._table.get(task_id)
-        full_pages = len(prompt_ids) // self.page_size
-        for i in range(start_logical_page, full_pages):
-            self._pool.record(page_table[i], prompt_ids, i)
-
-    def bind_tasks(
-        self,
-        task_ids: List[str],
-        total_len: int,
-        device: torch.device,
-        write_positions: Optional[Tensor] = None,
-    ) -> PageCacheView:
-        page_table = self._table.table_tensor(task_ids, device)
-        return PageCacheView(self._storage, page_table, total_len)
-
-
-class ContiguousCacheView(CacheView):
-    """Contiguous KV-cache view for attention layers."""
-
-    def __init__(
-        self,
-        cache: "ContiguousCache",
-        batch_indices: Tensor,
-        total_len: int = 0,
-        write_positions: Optional[Tensor] = None,
-    ):
-        self._cache = cache
-        self._batch_indices = batch_indices
-        self._total_len = total_len
-        self._write_positions = write_positions
-
-    def write(self, layer_id: int, k: Tensor, v: Tensor):
-        seq_len = k.size(1)
-        indices = self._batch_indices
-        if self._write_positions is not None and seq_len == 1:
-            pos = self._write_positions
-            self._cache.k[layer_id, indices, pos] = k.squeeze(1)
-            self._cache.v[layer_id, indices, pos] = v.squeeze(1)
-            for s, p in zip(indices.tolist(), pos.tolist()):
-                cur = self._cache._slot_len.get(s, 0)
-                if p + 1 > cur:
-                    self._cache._slot_len[s] = p + 1
-        else:
-            start_pos = self._total_len - seq_len
-            self._cache.k[layer_id, indices, start_pos : start_pos + seq_len] = k
-            self._cache.v[layer_id, indices, start_pos : start_pos + seq_len] = v
-            new_len = start_pos + seq_len
-            for s in indices.tolist():
-                cur = self._cache._slot_len.get(s, 0)
-                if new_len > cur:
-                    self._cache._slot_len[s] = new_len
-
-    def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]:
-        max_len = max(
-            self._cache._slot_len.get(int(s), 0) for s in self._batch_indices.tolist()
-        )
-        indices = self._batch_indices
-        k = self._cache.k[layer_id, indices, :max_len]
-        v = self._cache.v[layer_id, indices, :max_len]
-        return k, v
-
-
-class ContiguousCache(KVCache):
-    """Contiguous per-slot KV cache (default implementation)."""
-
-    def __init__(
-        self,
-        n_layers: int,
         max_batch_size: int,
         max_seq_len: int,
-        n_kv_heads: int,
-        head_dim: int,
         device: torch.device,
         dtype: torch.dtype,
+        page_size: int = 1,
+        n_tokens: Optional[int] = None,
     ):
+        self.page_size = page_size
+        self.max_batch_size = max_batch_size
         self.max_seq_len = max_seq_len
-        self.k = torch.zeros(
-            n_layers,
-            max_batch_size,
-            max_seq_len,
-            n_kv_heads,
-            head_dim,
-            device=device,
-            dtype=dtype,
+        self.device = device
+        self.dtype = dtype
+        self.n_layers = n_layers
+        self.n_kv_heads = n_kv_heads
+        self.head_dim = head_dim
+
+        self.contiguous = n_tokens is None
+        if self.contiguous:
+            self.n_tokens = max_batch_size * max_seq_len
+        else:
+            self.n_tokens = n_tokens
+
+        self._storage = KVStorage(
+            self.n_tokens, n_layers, n_kv_heads, head_dim, device, dtype
         )
-        self.v = torch.zeros(
-            n_layers,
-            max_batch_size,
-            max_seq_len,
-            n_kv_heads,
-            head_dim,
-            device=device,
-            dtype=dtype,
+        self._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._slot_len: Dict[int, int] = {}
-        self._task_slot: Dict[str, int] = {}
-        self._free_slots = list(range(max_batch_size))
-        self._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:
-        if not self._free_slots:
+        req_slots = self._req_pool.alloc(1)
+        if req_slots is None:
             return False
-        slot = self._free_slots.pop(0)
-        self._task_slot[task_id] = slot
-        self._slot_len[slot] = 0
+        req_idx = req_slots[0]
+        self._task_req[task_id] = req_idx
+
+        if self.contiguous:
+            self._task_len[req_idx] = len(prompt_ids)
+            self._task_cached[task_id] = 0
+            return True
+
+        n_tokens_needed = len(prompt_ids)
+        cached = 0
+
+        if self._prefix is not None:
+            hits = self._prefix.lookup(prompt_ids)
+            cached = len(hits) * self.page_size
+            for p in hits:
+                self._alloc.inc_ref(p)
+            self._task_pages[task_id] = list(hits)
+            self._task_slots[task_id] = []
+        else:
+            self._task_pages[task_id] = []
+            self._task_slots[task_id] = []
+
+        remaining = n_tokens_needed - cached
+        if remaining > 0:
+            if self.page_size == 1:
+                slots = self._alloc_tokens(remaining)
+                if slots is None:
+                    for p in self._task_pages[task_id]:
+                        self._alloc.free(p)
+                    self._req_pool.free([req_idx])
+                    del self._task_req[task_id]
+                    return False
+                self._task_slots[task_id] = slots
+            else:
+                n_new_pages = (remaining + self.page_size - 1) // self.page_size
+                new_pages = []
+                for _ in range(n_new_pages):
+                    p = self._alloc.alloc()
+                    if p < 0:
+                        for hp in self._task_pages[task_id]:
+                            self._alloc.free(hp)
+                        for np_ in new_pages:
+                            self._alloc.free(np_)
+                        self._req_pool.free([req_idx])
+                        del self._task_req[task_id]
+                        return False
+                    new_pages.append(p)
+                self._task_pages[task_id].extend(new_pages)
+
+        self._write_req_to_token(task_id, prompt_ids, cached)
+        self._task_len[req_idx] = len(prompt_ids)
+        self._task_cached[task_id] = cached
         return True
 
     def task_free(self, task_id: str):
-        slot = self._task_slot.pop(task_id, None)
-        if slot is not None:
-            self._slot_len.pop(slot, None)
-            self._free_slots.append(slot)
+        req_idx = self._task_req.pop(task_id, None)
+        if req_idx is None:
+            return
+        self._task_len.pop(req_idx, None)
+        self._task_cached.pop(task_id, None)
+
+        if not self.contiguous:
+            if self._prefix is not None:
+                for p in self._task_pages.get(task_id, []):
+                    keep = self._prefix.has_page(p)
+                    self._alloc.free(p, keep_cached=keep)
+                    if not keep:
+                        self._prefix.evict(p)
+            else:
+                for p in self._task_pages.get(task_id, []):
+                    self._alloc.free(p)
+            self._task_pages.pop(task_id, None)
+            self._task_slots.pop(task_id, None)
+
+        self._req_pool.free([req_idx])
 
     def task_extend(self, task_id: str, pos: int) -> bool:
+        req_idx = self._task_req.get(task_id)
+        if req_idx is None:
+            return False
+
+        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
+            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:
-        slot = self._task_slot.get(task_id)
-        if slot is None:
-            return 0
-        return self._slot_len.get(slot, 0)
+        return self._task_cached.get(task_id, 0)
+
+    def task_record_hashes(
+        self, task_id: str, prompt_ids: List[int], start_logical_page: int = 0
+    ):
+        if self._prefix is None or self.contiguous:
+            return
+        pages = self._task_pages.get(task_id, [])
+        full_pages = len(prompt_ids) // self.page_size
+        for i in range(start_logical_page, min(full_pages, len(pages))):
+            self._prefix.record(pages[i], prompt_ids, i)
+
+    # ---- bind for forward ----
 
     def bind_tasks(
         self,
         task_ids: List[str],
-        total_len: int,
+        seq_lens: List[int],
         device: torch.device,
-        write_positions: Optional[Tensor] = None,
-    ) -> ContiguousCacheView:
-        slots = [self._task_slot[tid] for tid in task_ids]
-        batch_indices = torch.tensor(slots, dtype=torch.long, device=device)
-        return ContiguousCacheView(
-            self, batch_indices, total_len, write_positions=write_positions
+        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
@@ -19,7 +19,7 @@ class Executor:
         self,
         model: AutoModel,
         tokenizer: AutoTokenizer,
-        kv_cache: KVCache,
+        kv_cache: PagePool,
         device: Optional[str] = None,
         dtype: Optional[torch.dtype] = None,
     ):
@@ -43,19 +43,42 @@ class Executor:
         )
 
         task_ids = [t.task_id for t in tasks]
+        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(
-                    start_pos, prompt_len, dtype=torch.long, device=self.device
-                )
-                .unsqueeze(0)
-                .expand(batch_sz, -1),
-                paged_cache=self.kv_cache.bind_tasks(task_ids, prompt_len, self.device),
+                input_mask=input_mask,
+                position_ids=position_ids,
+                kv_cache=self.kv_cache.bind_tasks(
+                    task_ids, [prompt_len] * batch_sz, self.device, start_pos=start_pos
+                ),
             )
 
-    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 []
 
@@ -68,7 +91,10 @@ 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]
 
@@ -79,43 +105,64 @@ class Executor:
             [t.frequency_penalty for t in tasks], device=self.device
         )
 
+        has_freq = bool((freq_penalties != 0).any())
+        if has_freq:
             history_lists = []
-        mask_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)
-            mask_lists.append([True] * len(ids))
+                history_lens.append(len(ids))
 
-        max_len = max(len(h) for h in history_lists)
+            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, m) in enumerate(zip(history_lists, mask_lists)):
-            padded_ids[i, : len(h)] = torch.tensor(
+            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, : len(m)] = torch.tensor(
-                m, dtype=torch.bool, 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.kv_cache.bind_tasks(
+                input_mask=input_mask,
+                kv_cache=self.kv_cache.bind_tasks(
                     task_ids,
-                    total_len,
+                    [t.next_pos + 1 for t in tasks],
                     self.device,
-                    write_positions=position_ids,
                 ),
                 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,

astrai/inference/core/scheduler.py

@@ -1,10 +1,11 @@
 import logging
 import threading
+import uuid
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch
 
-from astrai.inference.core.cache import ContiguousCache, KVCache
+from astrai.inference.core.cache import PagePool
 from astrai.inference.core.executor import Executor
 from astrai.inference.core.task import STOP, Task, TaskManager, TaskStatus
 from astrai.model.automodel import AutoModel
@@ -22,45 +23,43 @@ class InferenceScheduler:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 16,
         max_seq_len: Optional[int] = None,
-        max_prompt_len: int = 2048,
         device: Optional[str] = None,
         dtype: Optional[torch.dtype] = None,
-        cache: Optional[KVCache] = None,
+        cache: Optional[PagePool] = None,
     ):
         config = model.config
 
         if max_seq_len is not None:
             self.max_seq_len = max_seq_len
-        elif config.max_len is not None:
-            self.max_seq_len = config.max_len
+        elif config.max_position_embeddings is not None:
+            self.max_seq_len = config.max_position_embeddings
         else:
             raise ValueError(
                 "max_seq_len must be provided either as argument "
-                "or in model config (config.max_len)"
+                "or in model config (config.max_position_embeddings)"
             )
         self.device = device or next(model.parameters()).device
         self.dtype = dtype or next(model.parameters()).dtype
 
-        head_dim = config.dim // config.n_heads
+        head_dim = config.hidden_size // config.num_attention_heads
 
         if cache is not None:
             self._cache = cache
         else:
-            self._cache = ContiguousCache(
-                config.n_layers,
-                max_batch_size,
-                self.max_seq_len,
-                config.n_kv_heads,
-                head_dim,
-                self.device,
-                self.dtype,
+            self._cache = PagePool(
+                n_layers=config.num_hidden_layers,
+                n_kv_heads=config.num_key_value_heads,
+                head_dim=head_dim,
+                max_batch_size=max_batch_size,
+                max_seq_len=self.max_seq_len,
+                device=self.device,
+                dtype=self.dtype,
             )
 
         self._task_mgr = TaskManager(
             tokenizer=tokenizer,
             max_batch_size=max_batch_size,
             max_seq_len=self.max_seq_len,
-            max_prompt_len=max_prompt_len,
         )
 
         self._executor = Executor(
@@ -110,9 +109,11 @@ 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 cache.task_cached(t.task_id) < len(t.prompt_ids)
                 ]
@@ -138,10 +139,10 @@ class InferenceScheduler:
                                 t.task_id, t.prompt_ids, start_logical_page
                             )
 
-                decode_tasks = self._task_mgr.get_active_tasks()
+                decode_tasks = active
 
                 valid: List[Task] = []
-                for t in sorted(decode_tasks, key=lambda t: t.task_id):
+                for t in decode_tasks:
                     if cache.task_extend(t.task_id, t.next_pos):
                         valid.append(t)
                     else:
@@ -194,6 +195,117 @@ class InferenceScheduler:
             self._cache.task_free(task.task_id)
         for task in self._task_mgr.get_waiting_tasks():
             self._task_mgr.invoke_callback(task.task_id, STOP)
+            self._cache.task_free(task.task_id)
         self._task_mgr.clear_queues()
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
+
+    def run_batch(
+        self,
+        prompt_ids_list: List[List[int]],
+        *,
+        max_tokens: Optional[int] = None,
+        temperature: float = 1.0,
+        top_p: float = 1.0,
+        top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
+        return_logprobs: bool = False,
+    ) -> List[List[int]]:
+        """Synchronous batch generation without the scheduler thread.
+
+        Accepts already-tokenized prompts (no string round-trip) and runs
+        prefill + decode to completion on the calling thread.  Designed for
+        RL rollout, where logprobs of the behaviour policy must be collected
+        alongside generated tokens.
+
+        Args:
+            prompt_ids_list: ``B`` prompts, each a list of token IDs.
+            max_tokens: Maximum tokens to generate per prompt.  ``None``
+                uses ``self.max_seq_len - len(prompt_ids)``.
+            temperature/top_p/top_k/frequency_penalty/rep_window: Sampling
+                parameters (uniform across the batch).
+            return_logprobs: If ``True``, return ``(token_ids, logprobs)``
+                tuples per prompt (logprobs aligned 1-to-1 with token_ids).
+
+        Returns:
+            ``List[List[int]]`` of generated token IDs per prompt, or —
+            when ``return_logprobs`` is ``True`` —
+            ``List[Tuple[List[int], List[float]]]``.
+        """
+        stop_ids = self._task_mgr.tokenizer.stop_ids
+        cache = self._cache
+        seq_cap = self.max_seq_len
+
+        tasks: List[Task] = []
+        for ids in prompt_ids_list:
+            if len(ids) >= seq_cap:
+                tasks.append(None)
+                continue
+            t_max = max_tokens
+            if t_max is None:
+                t_max = seq_cap - len(ids)
+            else:
+                t_max = min(t_max, seq_cap - len(ids))
+            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__)
@@ -14,37 +16,30 @@ STOP = object()
 
 
 class StreamDecoder:
-    """Incremental decoder for byte-level BPE streaming.
+    """Incremental decoder backed by the tokenizers library's DecodeStream.
 
-    Byte-level BPE may split a single Unicode character (e.g. em-dash,
-    smart quotes) across multiple tokens. Decoding such a token in
-    isolation produces U+FFFD (replacement char). This decoder
-    accumulates token IDs and only emits text once the trailing
-    characters are complete, buffering incomplete multi-byte sequences
-    until the next token arrives.
+    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__ = ("_tokenizer", "_ids", "_emitted")
+    __slots__ = ("_stream", "_tok")
 
     def __init__(self, tokenizer: AutoTokenizer):
-        self._tokenizer = tokenizer
-        self._ids: List[int] = []
-        self._emitted: str = ""
+        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.
         """
-        self._ids.append(token_id)
-        full = self._tokenizer.decode(self._ids, skip_special_tokens=True)
-        if full.endswith("\ufffd"):
-            return ""
-        if len(full) > len(self._emitted):
-            diff = full[len(self._emitted) :]
-            self._emitted = full
-            return diff
-        return ""
+        chunk = self._stream.step(self._tok, token_id)
+        return chunk or ""
 
 
 class TaskStatus(Enum):
@@ -81,6 +76,7 @@ class Task:
 
         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()
@@ -100,18 +96,11 @@ class Task:
     def flush_remaining(self, tokenizer: AutoTokenizer) -> str:
         """Emit any text still buffered in the decoder.
 
-        Called when generation terminates (max_tokens reached, stop
-        sequence, or external removal) to avoid dropping a final
-        incomplete-looking fragment that is actually complete when
-        adjacent to the stop token.
+        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.
         """
-        if self._decoder is None or not self.output_ids:
-            return ""
-        full = tokenizer.decode(self.output_ids, skip_special_tokens=True)
-        if len(full) > len(self._decoder._emitted):
-            diff = full[len(self._decoder._emitted) :]
-            self._decoder._emitted = full
-            return diff
         return ""
 
     @property
@@ -134,12 +123,10 @@ class TaskManager:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 16,
         max_seq_len: int = 8192,
-        max_prompt_len: int = 512,
     ):
         self.tokenizer = tokenizer
         self.max_batch_size = max_batch_size
         self.max_seq_len = max_seq_len
-        self.max_prompt_len = max_prompt_len
 
         self.waiting_queue: Deque[Task] = deque()
         self.active_tasks: List[Task] = []
@@ -164,10 +151,10 @@ class TaskManager:
     ) -> str:
         task_id = f"task_{int(time.time())}_{uuid.uuid4().hex[:8]}"
         prompt_ids = self.tokenizer.encode(prompt)
-        if len(prompt_ids) > self.max_prompt_len:
-            prompt_ids = prompt_ids[-self.max_prompt_len :]
+        if len(prompt_ids) > self.max_seq_len:
+            prompt_ids = prompt_ids[-self.max_seq_len :]
 
-        if len(prompt_ids) >= self.max_seq_len:
+        if len(prompt_ids) > self.max_seq_len:
             if stream_callback:
                 stream_callback(STOP)
             return task_id

astrai/inference/engine.py

@@ -8,7 +8,7 @@ from typing import Any, AsyncGenerator, Dict, Generator, List, Optional, Tuple,
 import torch
 import torch.nn as nn
 
-from astrai.inference.core.cache import KVCache
+from astrai.inference.core.cache import PagePool
 from astrai.inference.core.scheduler import InferenceScheduler
 from astrai.inference.core.task import STOP
 from astrai.tokenize import AutoTokenizer
@@ -111,9 +111,7 @@ class InferenceEngine:
         tokenizer: AutoTokenizer,
         max_batch_size: int = 1,
         max_seq_len: Optional[int] = None,
-        max_prompt_len: int = 2048,
-        page_size: int = 128,
-        cache: Optional[KVCache] = None,
+        cache: Optional[PagePool] = None,
     ):
         self.model = model
         self.tokenizer = tokenizer
@@ -122,7 +120,6 @@ class InferenceEngine:
             tokenizer=self.tokenizer,
             max_batch_size=max_batch_size,
             max_seq_len=max_seq_len,
-            max_prompt_len=max_prompt_len,
             cache=cache,
         )
 

astrai/inference/sample.py

@@ -276,7 +276,8 @@ class SamplingPipeline(BaseSamplingStrategy):
         filter_value: float = -float("inf"),
         input_ids: Optional[Tensor] = None,
         input_mask: Optional[Tensor] = None,
-    ) -> Tensor:
+        return_logprobs: bool = False,
+    ):
         """Apply strategies then sample (softmax + multinomial).
 
         Short-circuits to ``argmax`` when temperature is exactly 0
@@ -286,21 +287,41 @@ class SamplingPipeline(BaseSamplingStrategy):
             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.
         """
-        for s in self.strategies:
-            if isinstance(s, TemperatureStrategy) and self._is_greedy(s.temperature):
-                return logits.argmax(dim=-1)
-            break
+        if self._is_greedy_pipeline():
+            tokens = logits.argmax(dim=-1)
+            if not return_logprobs:
+                return tokens
+            log_probs = torch.log_softmax(logits.float(), dim=-1)
+            chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
+            return tokens, chosen
 
-        return torch.multinomial(
-            torch.softmax(
-                self.apply(logits, filter_value, input_ids, input_mask), dim=-1
-            ),
-            num_samples=1,
+        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()
@@ -313,31 +334,71 @@ def sample(
     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]``.
     """
-    if SamplingPipeline._is_greedy(temperature):
-        return logits.argmax(dim=-1)
-    return SamplingPipeline(
-        [
+    greedy = (
+        (
+            isinstance(temperature, Tensor)
+            and temperature.numel() == 1
+            and temperature.item() == 0
+        )
+        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),
-            FrequencyPenaltyStrategy(frequency_penalty),
     ]
-    ).sample(logits, filter_value, input_ids, input_mask)
+    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):
-    """
-    Autoregressive language model base class.
-    Provides model loading/saving, registration, and generation.
-    """
+class ModelFactory(BaseFactory[nn.Module]):
+    """Pure factory for model dispatch, separated from nn.Module state."""
+
+
+class AutoModel(nn.Module):
+    """Model base class with loading/saving and generation."""
 
     def __init__(self, config: BaseModelConfig):
         super().__init__()
@@ -68,7 +69,7 @@ class AutoModel(BaseFactory["AutoModel"], nn.Module):
         config = ConfigFactory.load(raw)
         model_type = config.model_type or "autoregressive_lm"
 
-        actual_cls = AutoModel.get_component_class(model_type)
+        actual_cls = ModelFactory.get_component_class(model_type)
 
         with _disable_random_init(enable=disable_random_init):
             model = actual_cls(config)

astrai/model/components/__init__.py

@@ -1,4 +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

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

astrai/model/components/decoder_block.py

@@ -4,7 +4,7 @@ from typing import Optional
 import torch.nn as nn
 from torch import Tensor
 
-from astrai.inference.core.cache import CacheView
+from astrai.inference.core.cache import KVCache
 from astrai.model.components.attention import AttnFactory
 from astrai.model.components.mlp import FFNFactory
 from astrai.model.components.norm import RMSNorm
@@ -14,10 +14,18 @@ class DecoderBlock(nn.Module):
     def __init__(self, config, layer_id: int):
         super().__init__()
         cfg = asdict(config)
-        cfg["down_init_std"] = 0.02 / (2 * config.n_layers) ** 0.5
+        cfg.update(
+            dim=config.hidden_size,
+            dim_ffn=config.intermediate_size,
+            n_layers=config.num_hidden_layers,
+            n_heads=config.num_attention_heads,
+            n_kv_heads=config.num_key_value_heads,
+            norm_eps=config.rms_norm_eps,
+            down_init_std=0.02 / (2 * config.num_hidden_layers) ** 0.5,
+        )
         self.attention = AttnFactory.create(config.attn_type, **cfg, layer_id=layer_id)
-        self.input_norm = RMSNorm(config.dim, config.norm_eps)
-        self.post_attention_norm = RMSNorm(config.dim, config.norm_eps)
+        self.input_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
         self.mlp = FFNFactory.create(config.ffn_type, **cfg)
 
     def forward(
@@ -25,13 +33,15 @@ class DecoderBlock(nn.Module):
         x: Tensor,
         rotary_emb: Tensor,
         attention_mask: Optional[Tensor] = None,
-        paged_cache: Optional[CacheView] = 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/lora.py

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

astrai/model/components/rope.py

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

astrai/model/encoder.py

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

astrai/model/transformer.py

@@ -5,8 +5,8 @@ import torch.nn as nn
 from torch import Tensor
 
 from astrai.config.model_config import AutoRegressiveLMConfig
-from astrai.inference.core.cache import CacheView
-from astrai.model.automodel import AutoModel
+from astrai.inference.core.cache import KVCache
+from astrai.model.automodel import AutoModel, ModelFactory
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.model.components.embedding import Embedding
 from astrai.model.components.linear import Linear
@@ -15,35 +15,18 @@ from astrai.model.components.rope import RotaryEmbedding
 
 
 def process_attention_mask(
-    input_tensor: Tensor,
-    position_ids: Optional[Tensor],
-    input_mask: Optional[Tensor] = None,
-    is_causal: bool = False,
+    input_mask: Optional[Tensor],
 ) -> Optional[Tensor]:
-    if position_ids is None:
+    if input_mask is None:
         return None
-    if input_mask is not None and input_mask.dim() > 2:
+    if input_mask.dim() == 2:
+        return input_mask[:, None, None, :]
+    if input_mask.dim() == 3:
+        return input_mask[:, None, :, :]
     return input_mask
 
-    device = input_tensor.device
-    B = input_tensor.size(0)
-    T = position_ids.max().item() + 1
 
-    if input_mask is None:
-        if position_ids.min().item() == 0 and is_causal:
-            return None
-        attend = torch.ones(B, 1, T, dtype=torch.bool, device=device)
-    else:
-        attend = input_mask[:, :T].to(device=device, dtype=torch.bool).unsqueeze(1)
-
-    if is_causal:
-        causal = position_ids.unsqueeze(-1) >= torch.arange(T, device=device)
-        attend = attend & causal
-
-    return attend.unsqueeze(1)
-
-
-@AutoModel.register("autoregressive_lm")
+@ModelFactory.register("autoregressive_lm")
 class AutoRegressiveLM(AutoModel):
     """Autoregressive language model with paged KV cache."""
 
@@ -53,24 +36,32 @@ class AutoRegressiveLM(AutoModel):
         rope_dim = (
             config.qk_rope_head_dim
             if config.attn_type == "mla"
-            else config.dim // config.n_heads
+            else config.hidden_size // config.num_attention_heads
         )
         rope_base = config.rope_theta if config.rope_theta is not None else 10000
         self.rotary_embedding = RotaryEmbedding(
-            rope_dim, config.max_len, rope_base, rope_scaling=config.rope_scaling
+            rope_dim,
+            config.max_position_embeddings,
+            rope_base,
+            rope_scaling=config.rope_scaling,
         )
         self.embed_tokens = Embedding(
-            config.vocab_size, config.dim, neftune_alpha=config.neftune_alpha
+            config.vocab_size,
+            config.hidden_size,
+            neftune_alpha=config.neftune_alpha,
         )
 
         self.layers = nn.ModuleList(
-            [DecoderBlock(config, layer_id) for layer_id in range(config.n_layers)]
+            [
+                DecoderBlock(config, layer_id)
+                for layer_id in range(config.num_hidden_layers)
+            ]
         )
 
-        self.norm = RMSNorm(config.dim, config.norm_eps)
-        self.lm_head = Linear(config.dim, config.vocab_size)
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.lm_head = Linear(config.hidden_size, config.vocab_size)
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             self.lm_head.weight = self.embed_tokens.weight
 
         self.apply(self._init_weights)
@@ -85,7 +76,7 @@ class AutoRegressiveLM(AutoModel):
 
         state_dict = dict(state_dict)
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             # same tensor for embed and lm_head
             if embed_key in state_dict:
                 state_dict[lm_head_key] = state_dict[embed_key]
@@ -101,7 +92,7 @@ class AutoRegressiveLM(AutoModel):
             destination=destination, prefix=prefix, keep_vars=keep_vars
         )
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             lm_head_key = prefix + "lm_head.weight"
             if lm_head_key in state_dict:
                 del state_dict[lm_head_key]
@@ -112,17 +103,18 @@ class AutoRegressiveLM(AutoModel):
         self,
         input_ids: Tensor,
         input_mask: Optional[Tensor] = None,
-        paged_cache: Optional[CacheView] = None,
+        kv_cache: Optional[KVCache] = None,
         position_ids: Optional[Tensor] = None,
     ) -> Dict[str, Tensor]:
         assert input_ids.ndim == 2
 
         x = self.embed_tokens(input_ids)
         rotary_emb = self.rotary_embedding(x, position_ids)
-        attn_mask = process_attention_mask(x, position_ids, input_mask, is_causal=True)
+        attn_mask = process_attention_mask(input_mask)
+        use_sdpa_causal_mask = attn_mask is None
 
         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/optim/__init__.py

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

astrai/optim/composite.py

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

astrai/optim/mano_adamw.py

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

astrai/optim/muon_adamw.py

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

astrai/optim/nora_nadamw.py

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

astrai/parallel/__init__.py

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

astrai/parallel/executor.py

@@ -4,17 +4,19 @@ import contextlib
 import logging
 import os
 from contextlib import contextmanager
-from typing import Optional, Tuple
+from typing import Any, Callable, Dict, Optional, Tuple
 
 import torch
 import torch.distributed as dist
 import torch.nn as nn
-from torch.distributed.fsdp import FullStateDictConfig, StateDictType
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.distributed.fsdp import (
+    FSDPModule,
+    fully_shard,
+)
+from torch.distributed.tensor import DTensor
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
-from torch.utils.data import DataLoader
 
 from astrai.factory import BaseFactory
 from astrai.parallel.setup import get_rank, get_world_size
@@ -22,6 +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)
@@ -86,19 +164,28 @@ class BaseExecutor:
 
     def prepare(
         self,
-        model: nn.Module,
-        optimizer: Optional[Optimizer] = None,
-        dataloader: Optional[DataLoader] = None,
-        scheduler: Optional[LRScheduler] = None,
-    ) -> Tuple[
-        nn.Module, Optional[Optimizer], Optional[DataLoader], Optional[LRScheduler]
-    ]:
+        model_fn: Callable[[], nn.Module],
+        optimizer_fn: Optional[Callable[[nn.Module], Optimizer]] = None,
+        scheduler_fn: Optional[Callable[[Optimizer], LRScheduler]] = None,
+        before_wrap: Optional[Callable[[nn.Module], nn.Module]] = None,
+        after_wrap: Optional[Callable[[nn.Module], nn.Module]] = None,
+    ) -> Tuple[nn.Module, Optional[Optimizer], Optional[LRScheduler]]:
+        model = model_fn()
+        if before_wrap is not None:
+            model = before_wrap(model)
         model = self._prepare_model(model)
-        if optimizer is not None:
+        if after_wrap is not None:
+            model = after_wrap(model)
+        optimizer = None
+        scheduler = None
+        if optimizer_fn is not None:
+            optimizer = optimizer_fn(model)
+            if scheduler_fn is not None:
+                scheduler = scheduler_fn(optimizer)
             optimizer = AccumOptimizer(optimizer, self.gradient_state)
             if scheduler is not None:
                 scheduler = AccumScheduler(scheduler, self.gradient_state)
-        return model, optimizer, dataloader, scheduler
+        return model, optimizer, scheduler
 
     def _prepare_model(self, model: nn.Module) -> nn.Module:
         return model
@@ -148,14 +235,7 @@ class BaseExecutor:
     def grad_accum_steps(self) -> int:
         return self.gradient_state.num_steps
 
-    def clip_grad_norm(self, model: nn.Module, max_norm: Optional[float]) -> float:
-        if max_norm is None:
-            total_norm = torch.norm(
-                torch.stack(
-                    [p.grad.norm(2) for p in model.parameters() if p.grad is not None]
-                )
-            )
-            return total_norm.item()
+    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()
@@ -234,78 +314,115 @@ class DDPExecutor(BaseExecutor):
 
 @ExecutorFactory.register("fsdp")
 class FSDPExecutor(BaseExecutor):
+    """FSDP executor using `torch.distributed.fsdp.fully_shard` (per-module API).
+
+    Wraps each child module individually via ``fully_shard``.
+    Skips the root model because ``ABC + Generic[T]`` in the MRO makes
+    ``fully_shard``'s dynamic ``__class__`` assignment fail at the CPython level.
+    Original ``Parameter`` objects are preserved (as DTensors) — no
+    ``FlatParameter``, no ``use_orig_params=True`` hack.
+    """
+
     def __init__(
         self,
         grad_accum_steps: int = 1,
-        process_group=None,
-        sharding_strategy=None,
-        cpu_offload=None,
-        auto_wrap_policy=None,
-        backward_prefetch=None,
-        mixed_precision=None,
-        ignored_modules=None,
-        param_init_fn=None,
-        sync_module_states: bool = False,
-        forward_prefetch: bool = False,
-        limit_all_gathers: bool = True,
-        ignored_states=None,
-        device_mesh=None,
+        mesh: Optional[Any] = None,
+        mp_policy: Optional[Any] = None,
+        reshard_after_forward: bool = False,
     ):
         super().__init__(grad_accum_steps=grad_accum_steps)
-        self._fsdp_kwargs = {
-            k: v
-            for k, v in dict(
-                process_group=process_group,
-                sharding_strategy=sharding_strategy,
-                cpu_offload=cpu_offload,
-                auto_wrap_policy=auto_wrap_policy,
-                backward_prefetch=backward_prefetch,
-                mixed_precision=mixed_precision,
-                ignored_modules=ignored_modules,
-                param_init_fn=param_init_fn,
-                sync_module_states=sync_module_states,
-                forward_prefetch=forward_prefetch,
-                limit_all_gathers=limit_all_gathers,
-                use_orig_params=True,
-                ignored_states=ignored_states,
-                device_mesh=device_mesh,
-            ).items()
-            if v is not None
-        }
-        self._original_model: Optional[nn.Module] = None
+        self._mesh = mesh
+        self._mp_policy = mp_policy
+        self._reshard_after_forward = reshard_after_forward
 
     def _prepare_model(self, model: nn.Module) -> nn.Module:
         if not self.use_distributed:
             logger.warning("FSDP backend selected but world_size=1, model not wrapped")
             return model
-        self._original_model = model
-        device_id = torch.device("cuda", get_rank())
-        model = FSDP(model, device_id=device_id, **self._fsdp_kwargs)
-        logger.info("Model wrapped with FSDP (world_size=%d)", get_world_size())
+
+        kwargs = dict(
+            mesh=self._mesh,
+            mp_policy=self._mp_policy,
+            reshard_after_forward=self._reshard_after_forward,
+        )
+        kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+        for child in model.children():
+            if isinstance(child, nn.ModuleList):
+                for sub in child:
+                    fully_shard(sub, **kwargs)
+            else:
+                fully_shard(child, **kwargs)
+
+        logger.info(
+            "FSDP wrapping applied to %d direct children (root skipped for ABC compat)",
+            len(list(model.children())),
+        )
         return model
 
+    @contextmanager
     def _no_sync(self, model: nn.Module):
-        if isinstance(model, FSDP):
-            return model.no_sync()
-        return contextlib.nullcontext()
+        fsdp_modules = [m for m in model.modules() if isinstance(m, FSDPModule)]
+        if fsdp_modules:
+            for m in fsdp_modules:
+                m.set_requires_gradient_sync(False, recurse=True)
+            try:
+                yield
+            finally:
+                for m in fsdp_modules:
+                    m.set_requires_gradient_sync(True, recurse=True)
+        else:
+            yield
 
-    def clip_grad_norm(self, model: nn.Module, max_norm: Optional[float]) -> float:
-        if max_norm is None:
+    def clip_grad_norm(self, model: nn.Module, max_norm: float) -> float:
+        if not self.use_distributed:
             return super().clip_grad_norm(model, max_norm)
-        if isinstance(model, FSDP) and self.use_distributed:
-            total_norm = model.clip_grad_norm_(max_norm)
-            if isinstance(total_norm, torch.Tensor):
+
+        # FSDP params are DTensors (sharded across ranks).
+        # torch.nn.utils.clip_grad_norm_ computes LOCAL norm per rank,
+        # so we must all-reduce to get the global norm before clipping.
+        local_norm = torch.nn.utils.get_total_norm(
+            [p.grad for p in model.parameters() if p.grad is not None],
+        )
+        if isinstance(local_norm, DTensor):
+            local_norm = local_norm.to_local()
+        total_norm_sq = local_norm**2
+        dist.all_reduce(total_norm_sq, op=dist.ReduceOp.SUM)
+        total_norm = total_norm_sq.sqrt()
+
+        clip_coef = max_norm / (total_norm + 1e-6)
+        clip_coef_clamped = torch.clamp(clip_coef, max=1.0)
+        for p in model.parameters():
+            if p.grad is not None:
+                p.grad.mul_(clip_coef_clamped)
+
         return total_norm.item()
-            return total_norm
-        return super().clip_grad_norm(model, max_norm)
 
     def unwrap_model(self, model: nn.Module):
-        if isinstance(model, FSDP) and self.use_distributed:
-            with FSDP.state_dict_type(
-                model,
-                StateDictType.FULL_STATE_DICT,
-                FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
-            ):
+        if not self.use_distributed:
             return model.state_dict()
 
-        return model.state_dict()
+        # unshard() and full_tensor() are collective ops — all ranks must
+        # participate. Non-rank-0 ranks still call them but discard results.
+        for module in model.modules():
+            if isinstance(module, FSDPModule):
+                module.unshard()
+
+        state_dict = model.state_dict()
+        result = {}
+        for k, v in state_dict.items():
+            if isinstance(v, DTensor):
+                full = v.full_tensor()
+                if get_rank() == 0:
+                    result[k] = full
+            elif get_rank() == 0:
+                result[k] = v
+
+        for module in model.modules():
+            if isinstance(module, FSDPModule):
+                module.reshard()
+
+        if get_rank() != 0:
+            return None
+
+        return result

astrai/parallel/module.py

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

astrai/parallel/setup.py

@@ -1,5 +1,8 @@
+import logging
 import os
+import signal
 import socket
+import threading
 from abc import ABC, abstractmethod
 from contextlib import contextmanager
 from functools import wraps
@@ -9,6 +12,10 @@ 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:
@@ -115,6 +122,7 @@ def _run_single_rank(
     func: Callable,
     kwargs: dict,
 ):
+    install_early_signal_handlers()
     with setup_parallel(
         rank=rank,
         world_size=world_size,
@@ -155,6 +163,7 @@ class TorchrunStrategy(LaunchStrategy):
     """External orchestrator (torchrun, SLURM, K8s) — env vars pre-set."""
 
     def launch(self, func: Callable, **kwargs):
+        install_early_signal_handlers()
         rank = int(os.environ["RANK"])
         world_size = int(os.environ["WORLD_SIZE"])
         local_rank = int(os.environ.get("LOCAL_RANK", rank))
@@ -188,6 +197,7 @@ class LocalStrategy(LaunchStrategy):
             _run_single_rank(0, *args)
             return
 
+        install_early_signal_handlers()
         ctx = mp.start_processes(
             _run_single_rank,
             args=args,
@@ -195,14 +205,46 @@ class LocalStrategy(LaunchStrategy):
             start_method=self.start_method,
             join=False,
         )
+
+        parent_stop = threading.Event()
+        original_handlers = {}
+
+        def _parent_handler(signum, frame):
+            sig = signal.Signals(signum)
+            logger.warning(
+                "Parent (pid=%d) received %s, forwarding to children...",
+                os.getpid(),
+                sig.name,
+            )
+            parent_stop.set()
+            for p in ctx.processes:
+                if p.is_alive():
+                    p.terminate()
+
+        for sig in (signal.SIGTERM, signal.SIGINT):
+            prev = signal.signal(sig, _parent_handler)
+            if prev not in (signal.SIG_DFL, signal.SIG_IGN, None, _parent_handler):
+                original_handlers[sig] = prev
+
         try:
-            while not ctx.join():
+            while not ctx.join() and not parent_stop.is_set():
                 pass
         except BaseException:
+            logger.warning(
+                "Parent received unexpected exception, terminating children..."
+            )
             for p in ctx.processes:
+                if p.is_alive():
                     p.terminate()
-            ctx.join()
             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:

astrai/preprocessing/builder.py

@@ -94,6 +94,97 @@ class SectionRenderer:
 
         return all_ids, loss_mask
 
+    def process_sections_batch(
+        self,
+        items: list[dict],
+        sections: list,
+        config,
+        tokenizer,
+        *,
+        is_top_level=False,
+        filter_text=True,
+    ):
+        """Render and tokenize a group of records with batched Rust tokenization."""
+        has_template = any(s.get("template") for s in sections)
+        is_text_config = not has_template and all(
+            s["action"] == "train" for s in sections
+        )
+        plans: list[list[tuple[str, str, bool]]] = []
+
+        for item in items:
+            plan: list[tuple[str, str, bool]] = []
+            first_section = True
+            for sec in sections:
+                field = sec["field"]
+                action = sec["action"]
+                use_template = sec.get("template", False)
+                add_special = sec.get(
+                    "add_special_tokens", not use_template and first_section
+                )
+
+                if use_template:
+                    messages = item.get(field)
+                    if not isinstance(messages, list) or not messages:
+                        continue
+                    for msg in messages:
+                        role = msg.get("role", "")
+                        rendered = tokenizer.apply_chat_template(
+                            [msg], tokenize=False, add_generation_prompt=False
+                        )
+                        plan.append(
+                            (rendered, _resolve_action(action, role, config), False)
+                        )
+                else:
+                    text = str(item.get(field, ""))
+                    if not text.strip():
+                        continue
+                    if is_text_config and filter_text:
+                        pp = config.preprocessing
+                        if pp.min_chars > 0 and len(text) < pp.min_chars:
+                            continue
+                        if len(text) > pp.max_chars:
+                            continue
+                    plan.append((text, action, add_special))
+
+                first_section = False
+            plans.append(plan)
+
+        encoded: dict[tuple[int, int], list[int]] = {}
+        for add_special in (False, True):
+            refs = [
+                (item_idx, unit_idx, text)
+                for item_idx, plan in enumerate(plans)
+                for unit_idx, (text, _, add) in enumerate(plan)
+                if add == add_special
+            ]
+            if not refs:
+                continue
+            ids_batch = tokenizer.encode(
+                [text for _, _, text in refs], add_special_tokens=add_special
+            )
+            for (item_idx, unit_idx, _), ids in zip(refs, ids_batch):
+                encoded[(item_idx, unit_idx)] = ids
+
+        outputs = []
+        max_len = config.preprocessing.max_seq_len
+        for item_idx, plan in enumerate(plans):
+            all_ids = []
+            loss_mask = []
+            if is_top_level and has_template and tokenizer.bos_token_id is not None:
+                all_ids.append(tokenizer.bos_token_id)
+                loss_mask.append(0)
+            for unit_idx, (_, action, _) in enumerate(plan):
+                ids = encoded[(item_idx, unit_idx)]
+                all_ids.extend(ids)
+                loss_mask.extend([1 if action == "train" else 0] * len(ids))
+            all_ids = all_ids[:max_len]
+            loss_mask = loss_mask[: len(all_ids)]
+            if not all_ids or (is_top_level and has_template and len(all_ids) <= 1):
+                outputs.append((None, None))
+            else:
+                outputs.append((all_ids, loss_mask))
+        return outputs
+
     def process_list_field(self, item: dict, sections: list, config, tokenizer):
         """Tokenize a list-valued field, preserving per-element boundaries.
 
@@ -147,6 +238,42 @@ class SectionRenderer:
             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"
@@ -214,6 +341,9 @@ class BaseMaskBuilder(ABC):
     @abstractmethod
     def build(self, item: dict, config, tokenizer) -> Optional[dict]: ...
 
+    def build_batch(self, items: list[dict], config, tokenizer) -> list[Optional[dict]]:
+        return [self.build(item, config, tokenizer) for item in items]
+
 
 class MaskBuilderFactory(BaseFactory["BaseMaskBuilder"]):
     pass
@@ -248,6 +378,27 @@ class SingleOutputMaskBuilder(BaseMaskBuilder):
             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):
@@ -317,6 +468,49 @@ class MultiOutputMaskBuilder(BaseMaskBuilder):
         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):
@@ -335,3 +529,9 @@ class SectionedMaskBuilder(BaseMaskBuilder):
         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/pipeline.py

@@ -1,7 +1,7 @@
 """Config-driven JSONL preprocessing pipeline.
 
 Composes a :class:`BaseMaskBuilder` (selected by ``input.type``) with
-sharding and flush to ``.h5`` / ``.bin`` storage.  Packing, position-id
+sharding and flush to ``.bin`` storage.  Packing, position-id
 generation and storage writing are each delegated to pluggable strategies,
 dispatched by configuration keys.
 
@@ -23,7 +23,6 @@ import tqdm
 from astrai.config.preprocess_config import PipelineConfig
 from astrai.preprocessing.core import (
     build_preprocessing_components,
-    iter_raw_records,
     primary_ids,
 )
 from astrai.preprocessing.packing import PackingStrategyFactory
@@ -81,6 +80,9 @@ class Pipeline:
     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
@@ -89,19 +91,31 @@ class Pipeline:
 
         pp = self.config.preprocessing
 
-        for item in tqdm.tqdm(
-            self._iter_items(), desc="Tokenizing", unit="docs", mininterval=0.5
-        ):
-            if pp.max_items and count >= pp.max_items:
-                break
-
+        progress = tqdm.tqdm(desc="Tokenizing", unit="docs", mininterval=0.5)
+        stop = False
+        for items in self._iter_batches(pp.batch_size):
+            progress.update(len(items))
             try:
-                result = self.transform(item)
+                results = self.transform_batch(items)
             except Exception:
                 logger.warning(
-                    "Failed to process item #%d, skipping", count + 1, exc_info=True
+                    "Failed to process batch, retrying records individually",
+                    exc_info=True,
                 )
-                continue
+                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
 
@@ -122,6 +136,10 @@ class Pipeline:
                     self._flush(domains, shard_idx)
                     domains.clear()
                     total_tokens = 0
+            if stop:
+                break
+
+        progress.close()
 
         if total_tokens > 0:
             self._flush(domains, shard_idx)
@@ -150,6 +168,17 @@ class Pipeline:
                             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]

astrai/preprocessing/writer.py

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

astrai/serialization/__init__.py

@@ -20,9 +20,7 @@ from astrai.serialization.checkpoint import (
 from astrai.serialization.dataset import (
     load_bin,
     load_bin_offsets,
-    load_h5,
     save_bin,
-    save_h5,
 )
 
 __all__ = [
@@ -39,7 +37,5 @@ __all__ = [
     "save_torch",
     "load_bin",
     "load_bin_offsets",
-    "load_h5",
     "save_bin",
-    "save_h5",
 ]

astrai/serialization/dataset.py

@@ -1,55 +1,14 @@
-"""Dataset storage serialization helpers (HDF5 / memory-mapped binary)."""
+"""Dataset storage serialization helpers (memory-mapped binary)."""
 
 import json
 import os
-from pathlib import Path
 from typing import Any, Dict, List, Optional
 
-import h5py
 import numpy as np
 import torch
 from torch import Tensor
 
 
-def save_h5(file_path: str, file_name: str, tensor_group: Dict[str, List[Tensor]]):
-    os.makedirs(file_path, exist_ok=True)
-    full_file_path = os.path.join(file_path, f"{file_name}.h5")
-    with h5py.File(full_file_path, "w") as f:
-        for key, tensors in tensor_group.items():
-            grp = f.create_group(key)
-            for idx, tensor in enumerate(tensors):
-                arr = tensor.cpu().numpy()
-                grp.create_dataset(f"data_{idx}", data=arr)
-
-
-def load_h5(file_path: str, share_memory=True) -> Dict[str, List[Tensor]]:
-    tensor_group: Dict[str, List[Tensor]] = {}
-
-    root_path = Path(file_path)
-    if root_path.is_file() and root_path.suffix in (".h5", ".hdf5"):
-        h5_files = [root_path]
-    else:
-        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]],
@@ -65,7 +24,7 @@ def save_bin(
     offsets, preserving backward compatibility.
 
     Nested keys (``List[List[Tensor]]`` such as GRPO ``responses``) are
-    not supported in bin format — use H5 for those.
+    not supported in bin format — use JSONL for those.
     """
     os.makedirs(file_path, exist_ok=True)
     record_keys = set(record_keys or [])
@@ -74,7 +33,7 @@ def save_bin(
         if tensors and isinstance(tensors[0], list):
             raise ValueError(
                 f"Nested key '{key}' (List[List[Tensor]]) is not supported "
-                f"in bin format. Use H5 or JSONL storage instead."
+                f"in bin format. Use JSONL storage instead."
             )
         cat = torch.cat(tensors, dim=0)
         entry: Dict[str, Any] = {
@@ -100,7 +59,7 @@ def load_bin(file_path: str) -> Dict[str, List[Tensor]]:
         arr = np.memmap(
             os.path.join(file_path, f"{key}.bin"),
             dtype=info["dtype"],
-            mode="r",
+            mode="c",
             shape=tuple(info["shape"]),
         )
         segments[key] = [torch.from_numpy(arr)]
@@ -112,7 +71,7 @@ def load_bin_offsets(file_path: str) -> Dict[str, List[int]]:
 
     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 (H5/JSONL layout).
+    indexing (JSONL layout).
     """
     with open(os.path.join(file_path, "meta.json"), "r") as f:
         meta = json.load(f)

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

@@ -38,12 +38,27 @@ class ChatTemplate:
         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.  By deferring compilation to first access, the
-        default pickle protocol serialises only ``template_str``; each
-        worker rebuilds the cache on first render.
+        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(
         cls,

astrai/tokenize/tokenizer.py

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

astrai/trainer/metric_util.py

@@ -22,6 +22,51 @@ def grad_norm(model: nn.Module, per_param: bool = False) -> float | Dict[str, fl
     return total_sq.sqrt().item()
 
 
+class GradSNRTracker:
+    """Track gradient signal-to-noise ratio via EMA of first/second moments.
+
+    SNR = E[g]^2 / Var(g) = E[g]^2 / (E[g^2] - E[g]^2)
+
+    The tracker accumulates per-parameter EMA moments across optimizer steps.
+    Call ``update`` after backward (before ``optimizer.step``) and read
+    ``snr`` to get the aggregate SNR across all parameters.
+    """
+
+    def __init__(self, beta: float = 0.999, eps: float = 1e-8):
+        self.beta = beta
+        self.eps = eps
+        self._first: Dict[int, torch.Tensor] = {}
+        self._second: Dict[int, torch.Tensor] = {}
+
+    @torch.no_grad()
+    def update(self, model: nn.Module) -> None:
+        beta = self.beta
+        for param in model.parameters():
+            if param.grad is None:
+                continue
+            pid = id(param)
+            g = param.grad.detach()
+            if pid not in self._first:
+                self._first[pid] = g.clone()
+                self._second[pid] = g.pow(2).clone()
+            else:
+                self._first[pid].mul_(beta).add_(g, alpha=1 - beta)
+                self._second[pid].mul_(beta).addcmul_(g, g, value=1 - beta)
+
+    @property
+    def snr(self) -> float:
+        if not self._first:
+            return 0.0
+        total_signal = 0.0
+        total_noise = 0.0
+        for m, v in zip(self._first.values(), self._second.values()):
+            signal = m.pow(2).sum().item()
+            noise = (v - m.pow(2)).clamp(min=0).sum().item()
+            total_signal += signal
+            total_noise += noise
+        return total_signal / (total_noise + self.eps)
+
+
 def ctx_get_loss(ctx):
     return ctx.loss
 
@@ -36,3 +81,10 @@ def ctx_get_val_loss(ctx):
 
 def ctx_get_grad_norm(ctx):
     return ctx.grad_norm
+
+
+def ctx_get_grad_snr(ctx):
+    tracker = getattr(ctx, "grad_snr_tracker", None)
+    if tracker is None:
+        return None
+    return tracker.snr

astrai/trainer/rollout.py

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

astrai/trainer/strategy.py

@@ -9,17 +9,8 @@ import torch.nn.functional as F
 from torch import Tensor
 
 from astrai.factory import BaseFactory
-
-
-def create_ref_model(
-    model_fn: Callable[[], nn.Module], state_dict: Dict[str, Tensor]
-) -> nn.Module:
-    """Create a frozen reference model from model_fn + full state dict."""
-    ref_model = model_fn()
-    ref_model.load_state_dict(state_dict)
-    ref_model.requires_grad_(False)
-    ref_model.eval()
-    return ref_model
+from astrai.parallel.executor import broadcast_state_dict
+from astrai.trainer.rollout import RolloutResult
 
 
 def move_to_device(batch: Dict[str, Tensor], device: str) -> Dict[str, Tensor]:
@@ -28,9 +19,10 @@ def move_to_device(batch: Dict[str, Tensor], device: str) -> Dict[str, Tensor]:
 
 
 def get_logprobs(
-    model: Union[nn.Module, Callable[..., Dict[str, Tensor]]],
+    model: nn.Module,
     input_ids: Tensor,
-    mask: Tensor,
+    attn_mask: Tensor,
+    loss_mask: Tensor,
     reduction: str,
 ) -> Tensor:
     """Compute token-wise log probabilities from model outputs.
@@ -38,7 +30,8 @@ def get_logprobs(
     Args:
         model: The language model
         input_ids: Input token IDs of shape [batch_size, seq_len]
-        mask: Attention mask of shape [batch_size, seq_len]
+        attn_mask: Attention mask passed to the model (may include causal).
+        loss_mask: Per-token mask for loss reduction.
         reduction: How to reduce over sequence dimension ("mean", "sum", "none")
 
     Returns:
@@ -51,9 +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(
@@ -61,13 +57,13 @@ 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:
@@ -87,7 +83,15 @@ def make_doc_boundary_mask(position_ids: Tensor) -> Tensor:
 
 
 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,
@@ -99,6 +103,7 @@ class BaseStrategy(ABC):
         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:
@@ -112,10 +117,54 @@ 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."""
+        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"]):
     """Factory class for creating training strategy instances.
@@ -238,13 +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]]
@@ -260,6 +327,29 @@ 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):
@@ -301,7 +391,11 @@ class GRPOStrategy(BaseStrategy):
 
     def sync_old_model(self):
         """Copy current policy weights to old model."""
-        self.old_model.load_state_dict(self.executor.unwrap_model(self.model))
+        state_dict = self.executor.unwrap_model(self.model)
+        if self.executor.use_distributed:
+            state_dict = broadcast_state_dict(state_dict)
+        if state_dict is not None:
+            self.old_model.load_state_dict(state_dict)
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
         batch = move_to_device(batch, self.device)
@@ -314,6 +408,12 @@ 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)
@@ -321,20 +421,32 @@ class GRPOStrategy(BaseStrategy):
         # response tokens.  get_logprobs shifts the mask by one position, so
         # the first response token's logprob (predicted from the last prompt
         # token) is correctly included.
-        full_masks = torch.cat([torch.zeros_like(prompt_expanded), masks_flat], dim=-1)
+        full_masks = torch.cat(
+            [torch.zeros_like(prompt_expanded, dtype=torch.bool), masks_flat], dim=-1
+        )
+
+        # 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, full_masks, "none"
+            self.model, full_sequences, attn_mask, full_masks, "none"
         )[:, prompt_len - 1 :]
         with torch.no_grad():
             token_log_probs_old = get_logprobs(
-                self.old_model, full_sequences, full_masks, "none"
+                self.old_model, full_sequences, attn_mask, full_masks, "none"
             )[:, prompt_len - 1 :]
             token_log_probs_ref = get_logprobs(
-                self.ref_model, full_sequences, full_masks, "none"
+                self.ref_model, full_sequences, attn_mask, full_masks, "none"
             )[:, prompt_len - 1 :]
 
         # Reshape to [B, G, response_len]
@@ -371,3 +483,26 @@ class GRPOStrategy(BaseStrategy):
         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

@@ -18,6 +18,7 @@ from astrai.parallel.setup import get_current_device
 from astrai.serialization import Checkpoint
 from astrai.trainer.metric_util import (
     ctx_get_grad_norm,
+    ctx_get_grad_snr,
     ctx_get_loss,
     ctx_get_lr,
     ctx_get_val_loss,
@@ -235,7 +236,7 @@ class ProgressBarCallback(TrainCallback):
 class MetricCallback(TrainCallback):
     def __init__(
         self,
-        log_dir: str,
+        ckpt_dir: str,
         save_interval: int,
         metrics: List[str] = None,
         val_step: int = 0,
@@ -246,8 +247,7 @@ class MetricCallback(TrainCallback):
         self.val_step = val_step
         self._next_val_step = 0
 
-        self.log_dir = Path(log_dir) if log_dir else Path.cwd() / "logs"
-        self.log_dir.mkdir(parents=True, exist_ok=True)
+        self.ckpt_dir = Path(ckpt_dir) if ckpt_dir else Path.cwd() / "checkpoint"
 
         self.log_cache = []
 
@@ -256,6 +256,7 @@ class MetricCallback(TrainCallback):
             "lr": ctx_get_lr,
             "val_loss": ctx_get_val_loss,
             "grad_norm": ctx_get_grad_norm,
+            "grad_snr": ctx_get_grad_snr,
         }
 
     def _metrics(self, context: TrainContext, names):
@@ -306,13 +307,15 @@ class MetricCallback(TrainCallback):
 
     @only_on_rank(0)
     def _flush(self, epoch, step):
-        log_file = self.log_dir / f"epoch_{epoch}_step_{step}_metric.jsonl"
+        log_file = self.ckpt_dir / f"epoch_{epoch}_step_{step}" / "metric.jsonl"
         log_file.parent.mkdir(parents=True, exist_ok=True)
         with open(log_file, "w") as f:
             for log in self.log_cache:
                 f.write(json.dumps(log) + "\n")
 
     def on_optimizer_step(self, context):
+        context.grad_snr_tracker.update(context.model)
+
         if (
             context.val_dataloader is not None
             and self.val_step > 0

astrai/trainer/train_context.py

@@ -1,3 +1,5 @@
+import logging
+import threading
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Dict, Optional, Self
@@ -8,12 +10,18 @@ from torch.utils.data import DataLoader, random_split
 
 from astrai.config.train_config import TrainConfig
 from astrai.dataset import RDSampler
+from astrai.inference.core.scheduler import InferenceScheduler
 from astrai.model.components.lora import inject_lora
-from astrai.parallel.executor import BaseExecutor, ExecutorFactory
+from astrai.parallel.executor import BaseExecutor, ExecutorFactory, create_ref_model
 from astrai.parallel.setup import get_current_device, get_rank, get_world_size
 from astrai.protocols import OptimizerProtocol, SchedulerProtocol
 from astrai.serialization import Checkpoint, load_json
-from astrai.trainer.strategy import BaseStrategy, StrategyFactory, create_ref_model
+from astrai.tokenize import AutoTokenizer
+from astrai.trainer.metric_util import GradSNRTracker
+from astrai.trainer.rollout import RolloutGenerator, RolloutRunner
+from astrai.trainer.strategy import BaseStrategy, StrategyFactory
+
+logger = logging.getLogger(__name__)
 
 
 @dataclass
@@ -27,11 +35,11 @@ class TrainContext:
     config: TrainConfig = field(default=None)
     model_config: dict = field(default_factory=dict)
     executor: BaseExecutor = field(default=None)
-
     epoch: int = field(default=0)
     consumed_samples: int = field(default=0)
     loss: float = field(default=0.0)
     grad_norm: Optional[float] = field(default=None)
+    grad_snr_tracker: GradSNRTracker = field(default_factory=GradSNRTracker)
     val_dataloader: Optional[DataLoader] = field(default=None)
     val_loss: Optional[float] = field(default=None)
 
@@ -39,6 +47,15 @@ class TrainContext:
     rank: int = field(default=0)
     kwargs: Dict[str, Any] = field(default_factory=dict)
 
+    _stop_event: threading.Event = field(default_factory=threading.Event)
+
+    @property
+    def stop_requested(self) -> bool:
+        return self._stop_event.is_set()
+
+    def request_stop(self) -> None:
+        self._stop_event.set()
+
     @property
     def optimizer_step(self) -> int:
         return self.consumed_samples // (
@@ -72,61 +89,72 @@ class TrainContextBuilder:
             **cfg.executor_kwargs,
         )
 
-        model = cfg.model_fn()
-        model = model.to(device=device)
-
         model_config = {}
         if self._param_path:
             config_path = Path(self._param_path) / "config.json"
             if config_path.exists():
                 model_config = load_json(config_path)
 
-        if not model_config and hasattr(model, "config"):
-            model_config = model.config.to_dict()
+        preloaded_state_dict = None
+        preloaded_epoch = cfg.start_epoch
+        preloaded_consumed = cfg.start_samples * get_world_size()
+        preloaded_checkpoint = None
+        if self._param_path:
+            checkpoint = Checkpoint.load_any(self._param_path)
+            if checkpoint is not None:
+                preloaded_state_dict = checkpoint.state_dict
+                if checkpoint.config:
+                    model_config = checkpoint.config
+                if self._resume:
+                    preloaded_epoch = checkpoint.epoch
+                    per_step = (
+                        cfg.batch_per_device * get_world_size() * cfg.grad_accum_steps
+                    )
+                    preloaded_consumed = (
+                        checkpoint.consumed_samples // per_step
+                    ) * per_step
+                    preloaded_checkpoint = checkpoint
+
+        if not model_config and hasattr(cfg.model_fn(), "config"):
+            model_config = cfg.model_fn().config.to_dict()
+
+        def _before_wrap(m):
+            m = m.to(device=device)
+            if cfg.lora is not None:
+                inject_lora(
+                    m,
+                    r=cfg.lora.r,
+                    alpha=cfg.lora.alpha,
+                    target_modules=set(cfg.lora.target_modules),
+                )
+            if preloaded_state_dict is not None:
+                m.load_state_dict(preloaded_state_dict, strict=False)
+            return m
+
+        def _after_wrap(m):
+            if cfg.compile_mode is not None:
+                logger.info("torch.compile enabled (mode=%s)", cfg.compile_mode)
+                m = torch.compile(m, mode=cfg.compile_mode)
+            return m
 
         context = TrainContext(
-            model=model,
             world_size=get_world_size(),
             rank=get_rank(),
             config=cfg,
             model_config=model_config,
             executor=executor,
+            epoch=preloaded_epoch,
+            consumed_samples=preloaded_consumed,
+            checkpoint=preloaded_checkpoint,
         )
 
-        if self._param_path:
-            checkpoint = Checkpoint.load_any(self._param_path)
-            if checkpoint is not None:
-                model.load_state_dict(checkpoint.state_dict, strict=False)
-                if checkpoint.config:
-                    context.model_config = checkpoint.config
-
-                if self._resume:
-                    context.epoch = checkpoint.epoch or cfg.start_epoch
-                    if checkpoint.consumed_samples > 0:
-                        per_step = (
-                            cfg.batch_per_device
-                            * context.world_size
-                            * cfg.grad_accum_steps
+        context.model, context.optimizer, context.scheduler = executor.prepare(
+            cfg.model_fn,
+            cfg.optimizer_fn,
+            cfg.scheduler_fn,
+            before_wrap=_before_wrap,
+            after_wrap=_after_wrap,
         )
-                        context.consumed_samples = (
-                            checkpoint.consumed_samples // per_step
-                        ) * per_step
-                    else:
-                        context.consumed_samples = (
-                            cfg.start_samples * context.world_size
-                        )
-                    context.checkpoint = checkpoint
-
-        if cfg.lora is not None:
-            inject_lora(
-                model,
-                r=cfg.lora.r,
-                alpha=cfg.lora.alpha,
-                target_modules=set(cfg.lora.target_modules),
-            )
-
-        context.optimizer = cfg.optimizer_fn(model)
-        context.scheduler = cfg.scheduler_fn(context.optimizer)
 
         train_dataset = cfg.dataset
         val_dataset = cfg.val_dataset
@@ -141,6 +169,15 @@ class TrainContextBuilder:
             )
 
         sampler_offset = context.consumed_samples // context.world_size
+
+        if self._resume and sampler_offset > 0:
+            offset = context.world_size - 1
+            num_samples_per_replica = (
+                len(train_dataset) + offset
+            ) // context.world_size
+            if num_samples_per_replica > 0:
+                context.epoch = sampler_offset // num_samples_per_replica
+
         sampler = RDSampler(
             data_source=train_dataset,
             start_epoch=context.epoch,
@@ -175,15 +212,6 @@ class TrainContextBuilder:
                 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"):
@@ -194,17 +222,23 @@ class TrainContextBuilder:
 
         strategy_kwargs = dict(cfg.extra_kwargs)
 
-        if cfg.strategy in ("dpo", "grpo"):
-            ref_model = create_ref_model(
-                cfg.model_fn, executor.unwrap_model(context.model)
-            ).to(device=device)
-            strategy_kwargs["ref_model"] = ref_model
+        needs_ref = cfg.strategy in (
+            "dpo",
+            "grpo",
+            "online_grpo",
+            "online_dpo",
+        )
+        needs_old = cfg.strategy in ("grpo", "online_grpo")
 
-        if cfg.strategy == "grpo":
-            old_model = create_ref_model(
-                cfg.model_fn, executor.unwrap_model(context.model)
-            ).to(device=device)
-            strategy_kwargs["old_model"] = old_model
+        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,
@@ -214,4 +248,44 @@ class TrainContextBuilder:
             **strategy_kwargs,
         )
 
+        # Enable online rollout when the train_type is an ``online_*`` variant.
+        is_online = cfg.strategy.startswith("online_")
+        if is_online:
+            if not context.strategy.supports_online():
+                raise ValueError(
+                    f"Strategy '{cfg.strategy}' does not support online rollout"
+                )
+            if cfg.reward_model_fn is None:
+                raise ValueError("reward_model_fn is required for online RL strategies")
+
+            tokenizer = AutoTokenizer.from_pretrained(self._param_path)
+            reward_model = cfg.reward_model_fn()
+
+            group_size = strategy_kwargs.get("group_size", 1)
+            rollout_batch_size = group_size * max(1, cfg.batch_per_device)
+            max_seq_len = getattr(context.model.config, "max_position_embeddings", None)
+
+            scheduler = InferenceScheduler(
+                model=context.model,
+                tokenizer=tokenizer,
+                max_batch_size=rollout_batch_size,
+                max_seq_len=max_seq_len,
+            )
+
+            generator = RolloutGenerator(
+                scheduler=scheduler,
+                tokenizer=tokenizer,
+                max_tokens=cfg.rollout_max_tokens,
+                group_size=group_size,
+                temperature=cfg.rollout_temperature,
+                top_k=cfg.rollout_top_k,
+                top_p=cfg.rollout_top_p,
+            )
+            runner = RolloutRunner(
+                generator=generator,
+                reward_model=reward_model,
+                rollout_interval=cfg.rollout_interval,
+            )
+            context.strategy.set_rollout_runner(runner)
+
         return context

astrai/trainer/trainer.py

@@ -1,8 +1,14 @@
 import logging
 from typing import List, Optional
 
+import torch.distributed as dist
+
 from astrai.config import TrainConfig
 from astrai.parallel.setup import spawn_parallel_fn
+from astrai.signal_handler import (
+    register_signal_handlers,
+    unregister_signal_handlers,
+)
 from astrai.trainer.train_callback import (
     CallbackFactory,
     TrainCallback,
@@ -36,7 +42,7 @@ class Trainer:
             ),
             CallbackFactory.create(
                 "metric",
-                log_dir=cfg.log_dir,
+                ckpt_dir=cfg.ckpt_dir,
                 save_interval=cfg.ckpt_interval,
                 metrics=cfg.metrics,
                 val_step=cfg.val_step,
@@ -58,6 +64,7 @@ class Trainer:
             .with_param_path(param_path, resume=resume)
             .build()
         )
+        register_signal_handlers(context)
         executor = context.executor
         self._call_callbacks("on_train_begin", context)
 
@@ -65,10 +72,14 @@ class Trainer:
             context.model.train()
 
             for epoch in range(context.epoch, context.config.n_epoch):
+                if context.stop_requested:
+                    break
                 context.epoch = epoch
                 self._call_callbacks("on_epoch_begin", context)
 
                 for batch in context.dataloader:
+                    if context.stop_requested:
+                        break
                     with executor.accumulate(context.model):
                         self._call_callbacks("on_batch_begin", context)
                         loss = context.strategy(batch)
@@ -83,6 +94,7 @@ class Trainer:
                         if executor.sync_gradients:
                             self._call_callbacks("on_optimizer_step", context)
                             context.optimizer.step()
+                            context.strategy.on_optimizer_step()
                             context.optimizer.zero_grad()
 
                             if context.scheduler:
@@ -90,12 +102,21 @@ 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, param_path: Optional[str] = None, resume: bool = False):
         cfg = self.train_config

csrc/build.py

@@ -1,6 +1,32 @@
 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
 
@@ -27,7 +53,7 @@ NVCC_FLAGS = [
     "--use_fast_math",
     "--ptxas-options=-O3,-v",
     "--extra-device-vectorization",
-    "--threads=8",
+    "--threads=16",
 ]
 
 
@@ -46,3 +72,4 @@ def register(name: str, sources: list[str] | None = None, **kwargs):
 register("attn_decode")
 register("attn_prefill")
 register("attn_paged_decode")
+register("rotary_emb")

csrc/kernels/attn_common.h

@@ -19,9 +19,11 @@ struct AttentionParams {
     // 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] (mask_q_stride=0) or 3D [batch, q_len, kv_len]
-    int mask_b_stride;   // = kv_len (both 2D and 3D)
-    int mask_q_stride;   // 2D: 0 (all q rows share); 3D: kv_len
+    // 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;
@@ -52,8 +54,9 @@ struct PagedAttentionParams {
     // Q strides (layout-agnostic)
     int q_stride_b, q_stride_h, q_stride_l, q_stride_d;
 
-    // Mask strides (2D or 3D)
+    // Mask strides (2D, 3D, or 4D)
     int mask_b_stride;
+    int mask_h_stride;
     int mask_q_stride;
 
     const T* __restrict__ q;

csrc/kernels/attn_decode.cu

@@ -1,51 +1,6 @@
-#include "attn_decode_split_kv.cuh"
+#include "attn_dispatchers.cuh"
 #include "attn_entry_utils.cuh"
 
-#ifndef ASTRAI_NO_MMA
-#include "attn_decode_split_kv_mma.cuh"
-#endif
-
-// Scalar fallback: one warp per query head, split-KV across grid.z.
-static void launch_scalar_decode(AttentionParams<bf16>& p) {
-    int group_size = p.q_head / p.kv_head;
-    int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
-    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
-    alloc_split_partials(p);
-
-    size_t smem = DC_CHUNK * p.head_dim * sizeof(bf16);
-    attn_decode_split_kv_kernel<<<dim3(p.batch * p.kv_head, 1, p.num_splits), dim3(32, group_size), smem>>>(p);
-    attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
-}
-
-#ifndef ASTRAI_NO_MMA
-// MMA head-packing requires G <= 16 (BR=16 rows). sm_80+ tensor-core
-// + cp.async wins even at G=1 (decode is memory-bound, not compute-bound).
-// STAGES=2 (double-buffer) for D<=128 (smem 16 KB); STAGES=1 for D=256
-// (double-buffer would be 32 KB, near the 48 KB static cap — keep single
-// to preserve occupancy).
-template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
-static void launch_mma_decode(AttentionParams<bf16>& p) {
-    int tiles_total = (p.kv_len + BC - 1) / BC;
-    p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
-    alloc_split_partials(p);
-
-    attn_decode_split_kv_mma_kernel<HEAD_DIM, BC, STAGES><<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
-    attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
-}
-#endif
-
-template <int HEAD_DIM>
-static void dispatch_decode(AttentionParams<bf16>& p) {
-#ifndef ASTRAI_NO_MMA
-    int G = p.q_head / p.kv_head;
-    if (G >= 1 && G <= 16) {
-        launch_mma_decode<HEAD_DIM, 32>(p);
-        return;
-    }
-#endif
-    launch_scalar_decode(p);
-}
-
 torch::Tensor attn_decode(
     torch::Tensor q,
     torch::Tensor k,
@@ -60,11 +15,11 @@ torch::Tensor attn_decode(
     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");
 
-    // O matches Q's original layout
     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;
 }

csrc/kernels/attn_decode_split_kv.cuh

@@ -2,16 +2,10 @@
 #include <cuda_bf16.h>
 #include <float.h>
 #include "attn_common.h"
-
-using bf16 = __nv_bfloat16;
+#include "attn_warp_utils.cuh"
 constexpr int DC_CHUNK = 64;
 
-__device__ inline float warp_reduce_sum(float val) {
-    for (int offset = 16; offset > 0; offset >>= 1)
-        val += __shfl_xor_sync(0xFFFFFFFF, val, offset);
-    return val;
-}
-
+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;
@@ -30,7 +24,7 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
 
     // 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;
+    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};
 
@@ -48,7 +42,8 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
 
         // 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) {
+        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;
@@ -60,24 +55,30 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
         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 += 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 (p.use_mask && p.mask && !p.mask[mask_base + kv_idx])
+            if constexpr (HasMask) {
+                if (!p.mask[mask_base + kv_idx])
                     partial = -FLT_MAX;
-            if (p.causal_offset >= 0 && kv_idx > p.causal_offset)
+            }
+            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);
+            float alpha = __expf(m - new_m);
+            float beta  = __expf(partial - new_m);
             d = d * alpha + beta;
 
-            // V: stride-based read
-            int v_off = kv_base + kv_idx * p.kv_stride_l + lane * hd_per_thread * p.kv_stride_d;
+            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] = acc_reg[i] * alpha + __bfloat162float(p.v[v_off + i * p.kv_stride_d]) * beta;
+                acc_reg[i] = fmaf(acc_reg[i], alpha,
+                                  __bfloat162float(p.v[v_off + i * p.kv_stride_d]) * beta);
             m = new_m;
         }
         __syncthreads();
@@ -85,7 +86,7 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
 
     // ---- write UN-normalised partials for this split ----
     size_t bh = (size_t)batch * p.q_head + q_head;
-    size_t slot = bh * p.num_splits + split;
+    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;
@@ -97,9 +98,6 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
     }
 }
 
-// Reduce split-K partials into the final bf16 output. One block per (batch,
-// q_head); each thread folds across all splits with a single-pass
-// online-rescale reduction (expf + FMA counts halved vs 3-pass original).
 __global__ void attn_decode_combine_kernel(AttentionParams<bf16> p) {
     int bh = blockIdx.x;
     int d = threadIdx.x;
@@ -108,7 +106,7 @@ __global__ void attn_decode_combine_kernel(AttentionParams<bf16> p) {
     int batch = bh / p.q_head;
     int q_head = bh % p.q_head;
 
-    size_t split_base = (size_t)bh * p.num_splits;
+    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;
 
@@ -120,13 +118,12 @@ __global__ void attn_decode_combine_kernel(AttentionParams<bf16> p) {
         float nm = fmaxf(m, mi);
         float corr = __expf(m - nm);
         float e = __expf(mi - nm);
-        acc = acc * corr + op[s * p.head_dim + d] * e;
-        l = l * corr + li * e;
+        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;
-    // Stride-based output write (q_len=1 for decode, so stride_l not needed)
     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

@@ -3,85 +3,72 @@
 #include <cuda_bf16.h>
 #include "attn_common.h"
 #include "attn_mma_utils.cuh"
-
-using bf16 = __nv_bfloat16;
+#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.
 //
-// Decode has q_len == 1, so S = q @ K^T is a GEMV per head — no tensor-core
-// work on its own. But GQA gives us G = q_head / kv_head query heads that all
-// share one kv_head. We pack those G 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 (K/V load is the decode
-// bottleneck, so the reuse is the win, not the flops). The KV sequence is
-// partitioned across gridDim.z blocks so that a decode with only
-// batch*kv_head independent tasks can fill all SMs. Each (batch, kv_head,
-// split) block computes an UN-normalised partial (Oacc, m, l) over its KV
-// slice; the combine kernel below reduces across splits. Fixes the "grid too
-// small" bottleneck (0.04 waves/SM → many blocks) for long-context,
-// small-batch decode.
-
-template <int HEAD_DIM, int BC, int STAGES = 2>
+// 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) {
-    constexpr int KD = HEAD_DIM / 16;
-    constexpr int NC8 = BC / 8;
-    constexpr int KT2 = BC / 16;
-    constexpr int DN8 = HEAD_DIM / 8;
-    constexpr int LD = HEAD_DIM;
-    constexpr int SWIZ_MASK = (HEAD_DIM >= 64) ? 7 : (HEAD_DIM / 8 - 1);
-    constexpr int VEC = 8;
-    constexpr int TOTAL = BC * HEAD_DIM;
-
     const int lane = threadIdx.x;
     const int gid = lane >> 2;
     const int tid4 = lane & 3;
 
-    const int kv_head = blockIdx.x;
+    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;
-    const int G = p.q_head / p.kv_head;
-    const int q_head0 = kv_head * G;
 
-    // Double-buffered shared memory for K/V (no sQ needed — Q goes direct
-    // from global to registers).
-    __shared__ __align__(16) bf16 sK[STAGES * BC * LD];
-    __shared__ __align__(16) bf16 sV[STAGES * BC * LD];
+    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;
 
-    // ---- Load Q directly from global into mma A-operand registers ----
+    // 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[KD][4];
-    load_q_mma_frags<KD>(p.q + q_base, p.q_stride_h, p.q_stride_d,
+    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[DN8][4];
+    float Oacc[Traits::DN8][4];
     #pragma unroll
-    for (int j = 0; j < DN8; j++)
+    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 — [batch, kv_head, kv_len, head_dim]
     const int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
-    const int tiles_total = (p.kv_len + BC - 1) / BC;
+    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 int has_mask = p.use_mask && p.mask;
 
-    // ---- Load tile lambda: predicated cp.async, unified full/partial ----
+    // ---- Load tile lambda: predicated cp.async ----
     auto load_tile = [&](int ti, int buf) {
-        int kv0 = ti * BC;
-        bf16* dK = sK + buf * BC * LD;
-        bf16* dV = sV + buf * BC * LD;
+        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 * VEC; i < TOTAL; i += 32 * VEC) {
-            int r = i / HEAD_DIM, d = i % HEAD_DIM;
+        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 * LD + swiz_col(d, r, SWIZ_MASK);
-            // KV stride-based: contiguous within head_dim (stride_d == 1 typically)
+            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);
@@ -89,73 +76,79 @@ __global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
         cp_async_commit();
     };
 
-    // ---- Prologue: issue first tile load ----
-    if (ti_begin < ti_end) {
-        load_tile(ti_begin, 0);
-    }
+    // ---- Multi-stage cp.async pipeline ----
+    // Prologue loads STAGES tiles; each loop iteration waits only for the
+    // oldest outstanding group (wait_group<STAGES-1>) so the STAGES-1 newer
+    // tile loads stay in flight and overlap with the current tile's compute.
+    constexpr int STAGES = Traits::STAGES;
+    const int ntiles = ti_end - ti_begin;
 
-    for (int ti = ti_begin; ti < ti_end; ti++) {
-        constexpr int BUF_MASK = (STAGES > 1) ? (STAGES - 1) : 0;
-        int buf = (ti - ti_begin) & BUF_MASK;
+    auto process_tile = [&](int it, int buf) {
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = (ti_begin + it) * Traits::BC;
 
-        // Wait for current tile, then issue next tile's prefetch (overlaps
-        // with this tile's compute). Single syncwarp covers both hazards.
-        // When STAGES==1, no prefetch — load happens at end of prior iter.
-        cp_async_wait_group<0>();
-        __syncwarp();
-        if constexpr (STAGES > 1) {
-            if (ti + 1 < ti_end)
-                load_tile(ti + 1, (ti + 1 - ti_begin) & BUF_MASK);
-        }
-
-        const bf16* bK = sK + buf * BC * LD;
-        const bf16* bV = sV + buf * BC * LD;
-        int kv0 = ti * BC;
-
-        float Sacc[NC8][4];
-        mma_compute_scores<KD, NC8>(Qa, bK, LD, SWIZ_MASK, lane, Sacc);
+        float Sacc[Traits::NC8][4];
+        mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
 
         #pragma unroll
-        for (int n8 = 0; n8 < NC8; n8++)
+        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, mask_q_stride irrelevant
-        int maxc = (p.causal_offset >= 0) ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
-        mma_softmax_tile<NC8, DN8>(kv0, maxc, maxc,
+        // 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,
-                                    batch,
-                                    p.mask, has_mask,
+                                           p.mask_b_stride, 0, 0,
+                                           batch, 0,
+                                           p.mask,
                                            Sacc, Oacc, m0, m1, l0, l1, lane);
 
-        mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);
-        __syncwarp();
+        mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+    };
 
-        if constexpr (STAGES == 1) {
-            if (ti + 1 < ti_end)
-                load_tile(ti + 1, 0);
+    if (ntiles >= STAGES) {
+        #pragma unroll
+        for (int i = 0; i < STAGES; i++)
+            load_tile(ti_begin + i, i);
+
+        for (int it = 0; it < ntiles; it++) {
+            cp_async_wait_group<STAGES - 1>();
+            __syncwarp();
+            process_tile(it, it & (STAGES - 1));
+            __syncwarp();
+            if (it + STAGES < ntiles)
+                load_tile(ti_begin + it + STAGES, (it + STAGES) & (STAGES - 1));
         }
+    } else {
+        // Fewer tiles than stages: load all, wait for all, process.
+        for (int i = 0; i < ntiles; i++)
+            load_tile(ti_begin + i, i);
+        cp_async_wait_group<0>();
+        __syncwarp();
+        for (int it = 0; it < ntiles; it++)
+            process_tile(it, it);
     }
 
     // ---- write UN-normalised partials for this split ----
     auto split_slot = [&](int h) -> size_t {
         size_t bh = (size_t)batch * p.q_head + h;
-        return bh * p.num_splits + split;
+        return bh * MAX_SPLITS + split;
     };
     #pragma unroll
-    for (int dn8 = 0; dn8 < DN8; dn8++) {
+    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) * HEAD_DIM;
+            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) * HEAD_DIM;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
             op[d] = Oacc[dn8][2];
             op[d + 1] = Oacc[dn8][3];
         }

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

@@ -1,17 +1,12 @@
 #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;
 
-inline int compute_num_splits(int base_blocks, int tiles_total) {
-    int sm_count = 0;
-    cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, 0);
-    int n = (2 * sm_count + base_blocks - 1) / base_blocks;
-    return std::max(1, std::min(n, std::min(tiles_total, 32)));
-}
-
 // 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.
@@ -26,11 +21,15 @@ inline int compute_num_splits(int base_blocks, int tiles_total) {
                          " (supported: 32, 64, 128, 256)"); \
     }
 
+// The split kernel unconditionally writes every (batch, q_head, split) slot it
+// owns — including empty split ranges, which store m = -FLT_MAX so the combine
+// skips them. Allocators are therefore left uninitialized (torch::empty); the
+// per-call memset (torch::zeros / torch::full) was pure overhead.
 template<typename P>
 inline void alloc_split_partials(P& p) {
     auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
-    auto o_part = torch::empty({p.batch, p.q_head, p.num_splits, p.head_dim}, fopt);
-    auto ml_part = torch::empty({p.batch, p.q_head, p.num_splits, 2}, fopt);
+    auto o_part = torch::empty(at::IntArrayRef{p.batch, p.q_head, MAX_SPLITS, p.head_dim}, fopt);
+    auto ml_part = torch::empty(at::IntArrayRef{p.batch, p.q_head, MAX_SPLITS, 2}, fopt);
     p.o_part = (float*)o_part.data_ptr();
     p.ml_part = (float*)ml_part.data_ptr();
 }
@@ -50,6 +49,9 @@ inline void extract_q_dims_and_strides(torch::Tensor& q, int64_t layout, P& p) {
 }
 
 // ---- 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) {
@@ -60,18 +62,26 @@ inline void pack_mask(const c10::optional<torch::Tensor>& mask, P& p) {
         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) == p.q_len, "mask q_len mismatch");
+            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_q_stride = (int)m.stride(1);
+            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 [batch, kv_len] or 3D [batch, q_len, kv_len]");
+            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;
     }
 }

csrc/kernels/attn_mma_utils.cuh

@@ -3,10 +3,47 @@
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 
-// Shared MMA utilities for tensor-core GQA kernels.
-// mma.sync.m16n8k16 PTX wrappers, ldmatrix helpers, and bf16 packing.
+// 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;
 
-// mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32
 __device__ __forceinline__ void mma16816(float* d, const unsigned* a,
                                           const unsigned* b, const float* c) {
     asm volatile(
@@ -37,9 +74,7 @@ __device__ __forceinline__ unsigned pkb(bf16 a, bf16 b) {
 }
 
 // ldmatrix: cooperatively load mma fragments from smem (one instruction per
-// 16x16 / 16x8 tile) with the exact register layout mma expects — replaces the
-// scalar per-thread fragment packing, cutting shared-load instructions and bank
-// conflicts. Each lane supplies the shared address of one 8-wide row.
+// 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];"
@@ -60,29 +95,19 @@ __device__ __forceinline__ void ldmatrix_x2_trans(unsigned* r, const bf16* p) {
 }
 
 // XOR swizzle for shared-memory column at 8-bf16 chunk granularity.
-// Eliminates ldmatrix bank conflicts without LD padding: consecutive rows
-// land in distinct bank groups. swiz_col(d, r, mask) = ((d>>3)^(r&mask))<<3 | (d&7).
-// mask must cover log2(HEAD_DIM/8) chunk bits but stay within LD: use 7 for
-// HEAD_DIM>=64 (8+ chunks), 3 for HEAD_DIM=32 (4 chunks). Default 7 keeps
-// existing HEAD_DIM>=64 call sites working unchanged.
 __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,
-// bypassing registers. Eliminates shared-store bank conflicts and cuts
-// load-loop instruction count in half (1 cp.async vs 1 LDG + 1 STS).
-// Requires sm_80+.
+// 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 the
-// destination (src-size operand = 0 → no bytes read from src, so an
-// out-of-bounds src address is never dereferenced). Lets full and partial
-// tiles share one uniform async load path — no scalar fallback branch.
+// 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) {
@@ -100,9 +125,6 @@ __device__ __forceinline__ void cp_async_wait_all() {
     asm volatile("cp.async.wait_all;");
 }
 
-// Wait until at most N commit groups are still in flight. Used for
-// double-buffered pipelining: wait_group<1> lets the next tile's cp.async
-// continue while ensuring the current tile's data is ready.
 template <int N>
 __device__ __forceinline__ void cp_async_wait_group() {
     asm volatile("cp.async.wait_group %0;" :: "n"(N));
@@ -139,78 +161,65 @@ __device__ inline void load_q_mma_frags(
 }
 
 // ---------------------------------------------------------------------------
-// Shared MMA compute functions — used by both decode and prefill MMA kernels.
-// Extracted because S=Q@K^T, online softmax, and P@V are structurally identical
-// between the two kernels; only the per-row causal/mask bounds differ.
-// ---------------------------------------------------------------------------
-
 // S = Q @ K^T  (Qa pre-loaded by the caller; scale applied post-mma in the
 // caller to avoid bf16 precision loss).
-// LD and SWIZ_MASK are constexpr in the calling kernel — passing them as
-// runtime ints lets the compiler fold them while keeping the signature clean.
-template <int KD, int NC8>
+// Traits provides KD, NC8, LD, and SWIZ_MASK.
+// ---------------------------------------------------------------------------
+template <typename Traits>
 __device__ inline void mma_compute_scores(
-    const unsigned Qa[KD][4],
+    const unsigned Qa[Traits::KD][4],
     const bf16* __restrict__ sK,
-    int LD, 
-    int SWIZ_MASK, 
     int lane,
-    float Sacc[NC8][4])
+    float Sacc[Traits::NC8][4])
 {
     #pragma unroll
-    for (int n8 = 0; n8 < NC8; n8++) {
+    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 < KD; kt++) {
+        for (int kt = 0; kt < Traits::KD; kt++) {
             unsigned b[2];
-            ldmatrix_x2(b, &sK[krow_l * LD + swiz_col(kt * 16 + kcol_h, krow_l, SWIZ_MASK)]);
+            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.
-//   maxc0/maxc1: per-row KV column bounds (prefill: per-query-row causal limits;
-//     decode: same value for both rows since q_len==1).
-//   qrow0/qrow1: query row indices (for 3D mask indexing; decode passes 0).
-//   mask_b_stride/mask_q_stride: mask layout (2D: mask_q_stride=0; 3D: =kv_len).
-// Reads Sacc (Q@K^T scores), applies causal/mask, computes P = exp(S - nm),
-// rescales Oacc by exp(m_old - nm), and updates m/l — all in place.
-template <int NC8, int DN8>
+//
+// 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_q_stride,
-    int mask_batch,
+    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,
-    bool has_mask,
-    float Sacc[NC8][4],
-    float Oacc[DN8][4],
+    float Sacc[Traits::NC8][4],
+    float Oacc[Traits::DN8][4],
     float& m0, float& m1,
     float& l0, float& l1,
     int lane)
 {
     int tid4 = lane & 3;
 
-    // Mask out-of-bounds / masked columns: set -FLT_MAX so expf → 0 downstream
-    // without per-element sentinel checks. Compute tile-local row maxima.
     float rmax0 = -FLT_MAX, rmax1 = -FLT_MAX;
-    int mask_base0 = mask_batch * mask_b_stride + qrow0 * mask_q_stride;
-    int mask_base1 = mask_batch * mask_b_stride + qrow1 * mask_q_stride;
+    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 < NC8; n8++) {
+    for (int n8 = 0; n8 < Traits::NC8; n8++) {
         int cc = kv0 + n8 * 8 + 2 * tid4;
         int c1 = cc + 1;
-        bool b0 = (cc >= maxc0) || (has_mask && !mask[mask_base0 + cc]);
-        bool b1 = (c1 >= maxc0) || (has_mask && !mask[mask_base0 + c1]);
-        bool b2 = (cc >= maxc1) || (has_mask && !mask[mask_base1 + cc]);
-        bool b3 = (c1 >= maxc1) || (has_mask && !mask[mask_base1 + c1]);
+        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];
@@ -220,29 +229,20 @@ __device__ inline void mma_softmax_tile(
         rmax0 = fmaxf(rmax0, fmaxf(s0, s1));
         rmax1 = fmaxf(rmax1, fmaxf(s2, s3));
     }
-    // Warp-reduce row maxima across the 4-lane thread group (xor 1, xor 2).
     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));
 
-    // nm = max(running max m, tile-local max rmax) — updated running maximum.
     float nm0 = fmaxf(m0, rmax0), nm1 = fmaxf(m1, rmax1);
-    // corr rescales Oacc and l by exp(m_old - nm). When all-masked (m == nm ==
-    // -FLT_MAX), exp(0) = 1 — correct, no guard needed.
     float corr0 = __expf(m0 - nm0);
     float corr1 = __expf(m1 - nm1);
-    // pn guards only the all-masked-row edge: if nm == -FLT_MAX, exp(S - nm)
-    // gives 1 not 0 for masked entries. Two scalar masks replace 4*NC8
-    // per-element comparisons.
     float pn0 = (nm0 == -FLT_MAX) ? 0.0f : 1.0f;
     float pn1 = (nm1 == -FLT_MAX) ? 0.0f : 1.0f;
 
-    // P = exp(S - nm) for each element. Masked entries (Sacc = -FLT_MAX) give
-    // exp(-inf) ≈ 0 naturally; pn zero-fills the all-masked-row edge.
     float rsum0 = 0.0f, rsum1 = 0.0f;
     #pragma unroll
-    for (int n8 = 0; n8 < NC8; n8++) {
+    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);
@@ -261,22 +261,25 @@ __device__ inline void mma_softmax_tile(
     m0 = nm0; m1 = nm1;
 
     #pragma unroll
-    for (int j = 0; j < DN8; j++) {
+    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).
-template <int DN8, int KT2>
+// 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 LD, int SWIZ_MASK, int lane,
-    float Oacc[DN8][4])
+    int lane,
+    float Oacc[Traits::DN8][4])
 {
     #pragma unroll
-    for (int kt2 = 0; kt2 < KT2; kt2++) {
+    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]);
@@ -284,9 +287,10 @@ __device__ inline void mma_pv_accumulate(
         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 < DN8; dn8++) {
+        for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
             unsigned b[2];
-            ldmatrix_x2_trans(b, &sV[vrow_l * LD + swiz_col(dn8 * 8, vrow_l, SWIZ_MASK)]);
+            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

@@ -1,47 +1,6 @@
-#include "attn_paged_decode_split_kv.cuh"
-#ifndef ASTRAI_NO_MMA
-#include "attn_paged_decode_split_kv_mma.cuh"
-#endif
-
+#include "attn_dispatchers.cuh"
 #include "attn_entry_utils.cuh"
 
-static void launch_paged_scalar_decode(PagedAttentionParams<bf16>& p) {
-    int group_size = p.q_head / p.kv_head;
-    int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_CHUNK;
-    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
-    alloc_split_partials(p);
-
-    size_t smem = PDC_CHUNK * p.head_dim * sizeof(bf16);
-    dim3 grid = dim3(p.batch * p.kv_head, 1, p.num_splits);
-    dim3 block = dim3(32, group_size);
-    paged_attn_decode_split_kv_kernel<<<grid, block, smem>>>(p);
-    paged_attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
-}
-
-#ifndef ASTRAI_NO_MMA
-template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
-static void launch_paged_mma_decode(PagedAttentionParams<bf16>& p) {
-    int tiles_total = (p.kv_len + BC - 1) / BC;
-    p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
-    alloc_split_partials(p);
-
-    paged_attn_decode_split_kv_mma_kernel<HEAD_DIM, BC, STAGES><<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
-    paged_attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
-}
-#endif
-
-template <int HEAD_DIM>
-static void dispatch_paged_decode(PagedAttentionParams<bf16>& p) {
-#ifndef ASTRAI_NO_MMA
-    int G = p.q_head / p.kv_head;
-    if (G >= 1 && G <= 16 && p.page_size >= 32) {
-        launch_paged_mma_decode<HEAD_DIM, 32>(p);
-        return;
-    }
-#endif
-    launch_paged_scalar_decode(p);
-}
-
 torch::Tensor attn_paged_decode(
     torch::Tensor q,
     torch::Tensor page_table,
@@ -62,6 +21,7 @@ torch::Tensor attn_paged_decode(
     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;
 }

csrc/kernels/attn_paged_decode_split_kv.cuh

@@ -2,17 +2,10 @@
 #include <cuda_bf16.h>
 #include <float.h>
 #include "attn_common.h"
-
-using bf16 = __nv_bfloat16;
+#include "attn_warp_utils.cuh"
 constexpr int PDC_CHUNK = 64;
 
-__device__ inline float paged_warp_reduce_sum(float val) {
-    for (int offset = 16; offset > 0; offset >>= 1)
-        val += __shfl_xor_sync(0xFFFFFFFF, val, offset);
-    return val;
-}
-
-// Split-KV scalar decode: one warp per query head, grid.z partitions KV.
+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;
@@ -22,7 +15,6 @@ __global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p)
     int lane = threadIdx.x;
     int hd_per_thread = p.head_dim / 32;
 
-    // Q: stride-based [batch, q_head, q_len=1, head_dim]
     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;
@@ -39,14 +31,15 @@ __global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p)
     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;
+    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) {
+        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;
@@ -69,35 +62,48 @@ __global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p)
             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 = paged_warp_reduce_sum(partial) * p.scale;
+                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 (p.use_mask && p.mask && !p.mask[mask_base + kv_idx])
-                partial = -FLT_MAX;
-            if (p.causal_offset >= 0 && kv_idx > p.causal_offset)
+            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);
+            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 (phys_page >= 0) {
+            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] = acc_reg[i] * alpha + __bfloat162float(p.v_cache[v_base + lane * hd_per_thread + i]) * beta;
+                    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] = acc_reg[i] * alpha + 0.0f * beta;
+                    acc_reg[i] = fmaf(acc_reg[i], alpha, 0.0f);
             }
             m = new_m;
         }
@@ -105,7 +111,7 @@ __global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p)
     }
 
     size_t bh = (size_t)batch * p.q_head + q_head;
-    size_t slot = bh * p.num_splits + split;
+    size_t slot = bh * MAX_SPLITS + split;
     int d0 = lane * hd_per_thread;
     #pragma unroll
     for (int i = 0; i < hd_per_thread; i++)
@@ -124,7 +130,7 @@ __global__ void paged_attn_decode_combine_kernel(PagedAttentionParams<bf16> p) {
     int batch = bh / p.q_head;
     int q_head = bh % p.q_head;
 
-    size_t split_base = (size_t)bh * p.num_splits;
+    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;
 
@@ -136,8 +142,8 @@ __global__ void paged_attn_decode_combine_kernel(PagedAttentionParams<bf16> p) {
         float nm = fmaxf(m, mi);
         float corr = __expf(m - nm);
         float e = __expf(mi - nm);
-        acc = acc * corr + op[s * p.head_dim + d] * e;
-        l = l * corr + li * e;
+        acc = fmaf(acc, corr, op[s * p.head_dim + d] * e);
+        l = fmaf(l, corr, li * e);
         m = nm;
     }
 

csrc/kernels/attn_paged_decode_split_kv_mma.cuh

@@ -3,153 +3,165 @@
 #include <cuda_bf16.h>
 #include "attn_common.h"
 #include "attn_mma_utils.cuh"
-
-using bf16 = __nv_bfloat16;
+#include "attn_warp_utils.cuh"
 
 // Paged split-KV tensor-core decode via GQA head-packing.
-// Identical algorithm to attn_decode_split_kv_mma_kernel but reads K/V
-// directly from the page pool through a page table, eliminating the gather
-// copy.  Each tile (BC=32) fits within a single page (page_size >= 32), so
-// the page-table lookup happens once per tile for cp.async.
-
-template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
+// 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) {
-    constexpr int KD = HEAD_DIM / 16;
-    constexpr int NC8 = BC / 8;
-    constexpr int KT2 = BC / 16;
-    constexpr int DN8 = HEAD_DIM / 8;
-    constexpr int LD = HEAD_DIM;
-    constexpr int SWIZ_MASK = (HEAD_DIM >= 64) ? 7 : (HEAD_DIM / 8 - 1);
-    constexpr int VEC = 8;
-    constexpr int TOTAL = BC * HEAD_DIM;
-
     const int lane = threadIdx.x;
     const int gid = lane >> 2;
     const int tid4 = lane & 3;
 
-    const int kv_head = blockIdx.x;
+    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;
-    const int G = p.q_head / p.kv_head;
-    const int q_head0 = kv_head * G;
 
-    __shared__ __align__(16) bf16 sK[STAGES * BC * LD];
-    __shared__ __align__(16) bf16 sV[STAGES * BC * LD];
+    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();
 
-    // ---- Load Q directly from global into mma A-operand registers ----
     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[KD][4];
-    load_q_mma_frags<KD>(p.q + q_base, p.q_stride_h, p.q_stride_d,
+    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[DN8][4];
+    float Oacc[Traits::DN8][4];
     #pragma unroll
-    for (int j = 0; j < DN8; j++)
+    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 + BC - 1) / BC;
+    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 int has_mask = p.use_mask && p.mask;
 
-    // Paged strides (constant for the block)
-    const int64_t page_stride = (int64_t)p.page_size * p.kv_head * HEAD_DIM;
-    const int64_t pos_stride  = (int64_t)p.kv_head * HEAD_DIM;
-    const int64_t head_off    = (int64_t)kv_head * HEAD_DIM;
+    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: predicated cp.async, paged addressing ----
+    // ---- 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 * BC;
-        bf16* dK = sK + buf * BC * LD;
-        bf16* dV = sV + buf * BC * LD;
-        int logical_page = kv0 / p.page_size;
-        int phys_page = p.page_table[batch * p.max_pages + logical_page];
-        bool page_valid = (phys_page >= 0);
+        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 * VEC; i < TOTAL; i += 32 * VEC) {
-            int r = i / HEAD_DIM, d = i % HEAD_DIM;
+        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) && page_valid;
+            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 * LD + swiz_col(d, r, SWIZ_MASK);
+            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();
     };
 
-    // ---- Prologue: issue first tile load ----
-    if (ti_begin < ti_end) {
-        load_tile(ti_begin, 0);
-    }
+    // ---- Multi-stage cp.async pipeline ----
+    // Prologue loads STAGES tiles; each loop iteration waits only for the
+    // oldest outstanding group (wait_group<STAGES-1>) so the STAGES-1 newer
+    // tile loads stay in flight and overlap with the current tile's compute.
+    constexpr int STAGES = Traits::STAGES;
+    const int ntiles = ti_end - ti_begin;
 
-    for (int ti = ti_begin; ti < ti_end; ti++) {
-        constexpr int BUF_MASK = (STAGES > 1) ? (STAGES - 1) : 0;
-        int buf = (ti - ti_begin) & BUF_MASK;
+    auto process_tile = [&](int it, int buf) {
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = (ti_begin + it) * Traits::BC;
 
-        cp_async_wait_group<0>();
-        __syncwarp();
-        if constexpr (STAGES > 1) {
-            if (ti + 1 < ti_end)
-                load_tile(ti + 1, (ti + 1 - ti_begin) & BUF_MASK);
-        }
-
-        const bf16* bK = sK + buf * BC * LD;
-        const bf16* bV = sV + buf * BC * LD;
-        int kv0 = ti * BC;
-
-        float Sacc[NC8][4];
-        mma_compute_scores<KD, NC8>(Qa, bK, LD, SWIZ_MASK, lane, Sacc);
+        float Sacc[Traits::NC8][4];
+        mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
 
         #pragma unroll
-        for (int n8 = 0; n8 < NC8; n8++)
+        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, mask_q_stride irrelevant
-        int maxc = (p.causal_offset >= 0) ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
-        mma_softmax_tile<NC8, DN8>(kv0, maxc, maxc,
+        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,
-                                    batch,
-                                    p.mask, has_mask,
+                                           p.mask_b_stride, 0, 0,
+                                           batch, 0,
+                                           p.mask,
                                            Sacc, Oacc, m0, m1, l0, l1, lane);
 
-        mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);
+        mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+    };
+
+    if (ntiles >= STAGES) {
+        #pragma unroll
+        for (int i = 0; i < STAGES; i++)
+            load_tile(ti_begin + i, i);
+
+        for (int it = 0; it < ntiles; it++) {
+            cp_async_wait_group<STAGES - 1>();
             __syncwarp();
-
-        if constexpr (STAGES == 1) {
-            if (ti + 1 < ti_end)
-                load_tile(ti + 1, 0);
+            process_tile(it, it & (STAGES - 1));
+            __syncwarp();
+            if (it + STAGES < ntiles)
+                load_tile(ti_begin + it + STAGES, (it + STAGES) & (STAGES - 1));
         }
+    } else {
+        // Fewer tiles than stages: load all, wait for all, process.
+        for (int i = 0; i < ntiles; i++)
+            load_tile(ti_begin + i, i);
+        cp_async_wait_group<0>();
+        __syncwarp();
+        for (int it = 0; it < ntiles; it++)
+            process_tile(it, it);
     }
 
-    // ---- write UN-normalised partials for this split ----
     auto split_slot = [&](int h) -> size_t {
         size_t bh = (size_t)batch * p.q_head + h;
-        return bh * p.num_splits + split;
+        return bh * MAX_SPLITS + split;
     };
     #pragma unroll
-    for (int dn8 = 0; dn8 < DN8; dn8++) {
+    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) * HEAD_DIM;
+            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) * HEAD_DIM;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
             op[d] = Oacc[dn8][2];
             op[d + 1] = Oacc[dn8][3];
         }

csrc/kernels/attn_prefill.cu

@@ -1,35 +1,6 @@
-#include "attn_prefill_split_q.cuh"
+#include "attn_dispatchers.cuh"
 #include "attn_entry_utils.cuh"
 
-#ifndef ASTRAI_NO_MMA
-#include "attn_prefill_split_q_mma.cuh"
-#endif
-
-template <int HEAD_DIM>
-static void dispatch_prefill(AttentionParams<bf16>& p) {
-#ifndef ASTRAI_NO_MMA
-    constexpr int WARPS = 4, BR = 16;
-    // KV tile: bigger tiles amortize the per-tile cp.async wait + barrier +
-    // loop overhead over more tensor-core work (this kernel is latency-bound,
-    // not compute/bandwidth-bound), so BC=32 wins ~6-8% over BC=16 for
-    // D<=128. D=256 stays at 16: BC=32 double-buffered would need 64KB smem,
-    // over the 48KB static cap.
-    constexpr int BC = (HEAD_DIM <= 128) ? 32 : 16;
-    dim3 grid((p.q_len + BR * WARPS - 1) / (BR * WARPS), p.q_head, p.batch);
-    dim3 block(WARPS * 32, 1, 1);
-    // Static shared memory — double-buffered K/V only (no sQ: Q goes direct
-    // to registers). 2*BC*LD bf16 each for sK and sV → 4*BC*HEAD_DIM*2 bytes.
-    // Occupancy is smem-capped: D=64→3 blocks/SM (16KB), D=128→1 (32KB),
-    // D=256→1 (32KB, BC=16).
-    attn_prefill_split_q_mma_kernel<HEAD_DIM, WARPS, BC><<<grid, block>>>(p);
-#else
-    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, 1);
-    attn_prefill_split_q_kernel_t<HEAD_DIM, G, ROWS, P_BC><<<grid, block>>>(p);
-#endif
-}
-
 torch::Tensor attn_prefill(
     torch::Tensor q,
     torch::Tensor k,

csrc/kernels/attn_prefill_split_q.cuh

@@ -6,12 +6,9 @@
 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[]. Small per-thread footprint keeps
-// occupancy high; the S dot product is reduced across the G-lane group with a
-// short shuffle chain (log2(G) shuffles) instead of a full 32-lane warp reduce.
-// Online (per-kv) softmax — cheap because acc[] is only HEAD_DIM/G long.
-// Templated on <HEAD_DIM, G, ROWS, P_BC>. Block = (G, ROWS). G power-of-two,
-// G*ROWS a multiple of 32 with groups warp-aligned.
+// 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) {
@@ -21,8 +18,7 @@ __device__ __forceinline__ float group_reduce_sum(float v, unsigned mask) {
     return v;
 }
 
-// load 8 contiguous bf16 from (16-byte aligned) smem as one float4, unpack to
-// 8 floats — cuts shared-load instructions 8x vs scalar bf16 loads.
+// 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);
@@ -34,7 +30,7 @@ __device__ __forceinline__ void ld8(const bf16* p, float* o) {
     }
 }
 
-template <int HEAD_DIM, int G, int ROWS, int P_BC>
+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;
 
@@ -57,7 +53,7 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
                   + 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]) * p.scale;
+            qreg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
     }
 
     float m = -FLT_MAX, l = 0.0f;
@@ -68,13 +64,11 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
 
     // 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;
+    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;
 
-    // per-group shuffle mask: only the G lanes of this row's group participate,
-    // so causal masking (differing loop bounds across rows in a warp) is safe.
     int lane_in_warp = lid & 31;
     unsigned gmask = (G == 32) ? 0xFFFFFFFFu
                                : (((1u << G) - 1u) << (lane_in_warp & ~(G - 1)));
@@ -95,13 +89,15 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
         __syncthreads();
 
         int lim = tlen;
-        if (p.causal_offset >= 0 && q_row < p.q_len) {
+        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++) {
@@ -115,11 +111,13 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
                 for (int j = 0; j < 8; j++)
                     part = fmaf(qreg[i + j], k8[j], part);
             }
-            float dot = group_reduce_sum<G>(part, gmask);
+            float dot = group_reduce_sum<G>(part, gmask) * p.scale;
 
             int kv_idx = kv0 + s;
-            if (p.use_mask && p.mask && !p.mask[mask_row_base + kv_idx])
+            if constexpr (HasMask) {
+                if (!p.mask[mask_row_base + kv_idx])
                     dot = -FLT_MAX;
+            }
 
             float nm = fmaxf(m, dot);
             float al = __expf(m - nm);
@@ -141,10 +139,9 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
     }
 
     if (q_row < p.q_len) {
-        // O: stride-based write
         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-10f) ? (1.0f / l) : 0.0f;
+        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

@@ -4,121 +4,76 @@
 #include "attn_common.h"
 #include "attn_mma_utils.cuh"
 
-using bf16 = __nv_bfloat16;
-
 // 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). Q fragments are loaded once
-// straight from global into the mma A-operand layout (no smem staging) and
-// kept resident in registers across the tile loop. S, O, and the online-softmax
-// stats (m, l) also live in registers.
-// Shared memory is statically sized via template parameters — no dynamic
-// allocation. The mma fragment layout is used directly: the S accumulator
-// (f32) maps element-for-element onto the P matrix_a (bf16) operand, so
-// softmax needs no shuffle repack; row reductions fold across the 4-lane
-// thread group. Templated on <HEAD_DIM, WARPS, BC> with BC a multiple of 16.
+// cores via mma.sync.m16n8k16 (f32 accumulate).
 //
-// Software pipeline: K/V are double-buffered and loaded via cp.async one tile
-// ahead, so the next tile streams from global memory while the current tile's
-// tensor-core math runs — hiding load latency (long_scoreboard). A single
-// __syncthreads per tile both publishes the freshly loaded tile cross-warp and
-// (because it runs before the next prefetch) guards the buffer being refilled,
-// so no second barrier is needed. Predicated cp.async (cp_async_16_pred)
-// zero-fills rows past kv_len, unifying full and partial tiles on one path.
-// BC=32 (D<=128) amortizes the per-tile wait+barrier+loop overhead over more
-// tensor-core work — this kernel is latency-bound (low occupancy from high
-// register pressure), so fewer, larger tiles beat many tiny ones.
+// IsCausal and HasMask are compile-time bools — the compiler eliminates all
+// dead branches in the inner compute loop (FA2-style).
 //
-// Optimizations: load Q fragments directly from global in mma A-operand layout
-// (no sQ staging, no prologue barriers); post-multiply scale in float after
-// S=Q@K^T to avoid bf16 precision loss; packed bf16x2 output stores;
-// causal tile skipping (block-level prefetch bound + warp-level compute skip);
-// XOR swizzle (swiz_col) → eliminates ldmatrix bank conflicts without LD
-// padding (LD=HEAD_DIM).
-
-template <int HEAD_DIM, int WARPS, int BC>
+// 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) {
-    constexpr int BR = 16;
-    constexpr int KD = HEAD_DIM / 16;  // Q/K k-tiles
-    constexpr int NC8 = BC / 8;        // S n-tiles (N=8 each)
-    constexpr int KT2 = BC / 16;       // P k-tiles (K=16 each)
-    constexpr int DN8 = HEAD_DIM / 8;  // O n-tiles (N=8 each)
-    constexpr int LD = HEAD_DIM;   // XOR swizzle (swiz_col) handles bank conflicts
-    constexpr int SWIZ_MASK = (HEAD_DIM >= 64) ? 7 : (HEAD_DIM / 8 - 1);  // chunk bits, stay within LD
-
     const int warp = threadIdx.x / 32;
     const int lane = threadIdx.x % 32;
-    const int gid = lane >> 2;   // 0..7  → rows gid, gid+8
+    const int gid = lane >> 2;   // 0..7
     const int tid4 = lane & 3;   // 0..3
-    const int nthreads = WARPS * 32;
 
     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 * WARPS + warp) * BR;
+    const int qrow0 = (blockIdx.x * Traits::WARPS + warp) * Traits::BR;
 
-    // ---- Static shared memory: double-buffered K/V ----
-    // K/V are double-buffered (STAGES=2): the next tile's cp.async load runs
-    // while the current tile's tensor-core math executes, hiding global-load
-    // latency (FA2-style software pipeline). No dynamic smem / carveout opt-in.
-    constexpr int STAGES = 2;
-    __shared__ __align__(16) bf16 sK[STAGES * BC * LD];
-    __shared__ __align__(16) bf16 sV[STAGES * BC * LD];
+    // 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.
-    // stride_row = p.q_stride_l for prefill (multi-q rows across q_len).
-    // See attn_mma_utils.cuh for the shared template.
     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[KD][4];
-    load_q_mma_frags<KD>(p.q + q_base, p.q_stride_l, p.q_stride_d,
+    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[DN8][4];
+    float Oacc[Traits::DN8][4];
     #pragma unroll
-    for (int j = 0; j < DN8; j++)
+    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 + BC - 1) / BC;
-    const int qr0 = qrow0 + gid;      // row for c0/c1
-    const int qr1 = qrow0 + gid + 8;  // row for c2/c3
+    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 (no-op when causal_offset < 0)
-    const int use_skip = (p.causal_offset >= 0) ? 1 : 0;
-    const int max_kv = qrow0 + BR - 1 + p.causal_offset;
+    // 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 * WARPS * BR + WARPS * BR - 1 + p.causal_offset;
-    const int has_mask = p.use_mask && p.mask;
+        blockIdx.x * Traits::WARPS * Traits::BR + Traits::WARPS * Traits::BR - 1
+        + p.causal_offset;
 
-    // Last active tile: block-level causal bound (all warps in the block share
-    // the K/V load, so the prefetch range is the block max, not per-warp).
     int t_end = tiles - 1;
-    if (use_skip) {
-        int bt = block_max_kv / BC;
+    if constexpr (IsCausal) {
+        int bt = block_max_kv / Traits::BC;
         if (bt < t_end) t_end = bt;
     }
 
-    constexpr int VEC = 8;  // bf16 per cp.async unit (16 bytes)
-    constexpr int TOTAL = BC * HEAD_DIM;
-
     // ---- Load tile lambda: predicated cp.async ----
-    // Issue cp.async loads for tile `ti` into shared buffer `buf`. Predicated
-    // loads zero-fill rows past kv_len, so partial tiles need no scalar path.
     auto load_tile = [&](int ti, int buf) {
-        int kv0 = ti * BC;
-        bf16* dK = sK + buf * BC * LD;
-        bf16* dV = sV + buf * BC * LD;
+        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 * VEC; i < TOTAL; i += nthreads * VEC) {
-            int r = i / HEAD_DIM, d = i % HEAD_DIM;
+        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 * LD + swiz_col(d, r, SWIZ_MASK);
+            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);
@@ -132,65 +87,60 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
     for (int ti = 0; ti <= t_end; ti++) {
         int buf = ti & 1;
 
-        // Wait for the current tile's async copies, then a single barrier: it
-        // both publishes this tile's data cross-warp AND guarantees the prior
-        // compute on the buffer we are about to refill has finished. Issuing
-        // the next tile's load *after* this barrier lets one barrier cover both
-        // hazards (vs two), while the load still overlaps this tile's math.
+        // 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 * BC * LD;
-        const bf16* bV = sV + buf * BC * LD;
-        int kv0 = ti * BC;
+        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
-        if (!use_skip || kv0 <= max_kv) {
+        // Warp-level causal skip (dead branch eliminated when IsCausal == false)
+        if (!IsCausal || kv0 <= max_kv) {
 
-        // S = Q @ K^T + scale + online softmax + O += P @ V
-        float Sacc[NC8][4];
-        mma_compute_scores<KD, NC8>(Qa, bK, LD, SWIZ_MASK, lane, Sacc);
+            float Sacc[Traits::NC8][4];
+            mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
 
-        // post-multiply scale in float (no bf16 precision loss from pre-scaling Q)
+            // Post-multiply scale in float (no bf16 precision loss)
             #pragma unroll
-        for (int n8 = 0; n8 < NC8; n8++)
+            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 = (p.causal_offset >= 0) ? min(p.kv_len, qr0 + p.causal_offset + 1)
+            int maxc0 = IsCausal ? min(p.kv_len, qr0 + p.causal_offset + 1)
                                  : p.kv_len;
-        int maxc1 = (p.causal_offset >= 0) ? min(p.kv_len, qr1 + p.causal_offset + 1)
+            int maxc1 = IsCausal ? min(p.kv_len, qr1 + p.causal_offset + 1)
                                  : p.kv_len;
-        mma_softmax_tile<NC8, DN8>(kv0, maxc0, maxc1,
+            mma_softmax_tile<Traits, HasMask>(kv0, maxc0, maxc1,
                                                qr0, qr1,
-                                    p.mask_b_stride, p.mask_q_stride,
-                                    batch,
-                                    p.mask, has_mask,
+                                               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<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);
-        }  // if active (warp-level causal skip)
+            mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+        }
     }
 
-    // ---- write output ---- (packed bf16x2 stores: one 32-bit STG per pair,
-    // halves store count and removes the uncoalesced scalar-store penalty)
+    // ---- 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;
-    // O: stride-based write
     const int o_base = batch * p.q_stride_b + q_head * p.q_stride_h;
     #pragma unroll
-    for (int dn8 = 0; dn8 < DN8; dn8++) {
+    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;
+            *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;
+            *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

@@ -1,73 +1,33 @@
 /*
-Pure-C test:
+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_decode_split_kv.cuh"
-#ifndef ASTRAI_NO_MMA
-#include "../kernels/attn_decode_split_kv_mma.cuh"
-#endif
+#include "../kernels/attn_dispatchers.cuh"
 
-// Split-K scratch (torch-free): the production launcher allocates these from
-// torch; here we pass pre-allocated device buffers so the bench loop doesn't
-// pay a cudaMalloc per iteration. Size for the maximum split count (32).
+// Split-K scratch (torch-free)
 struct DecodeScratch {
     float* o_part = nullptr;
     float* ml_part = nullptr;
 };
 
-// Launch the production decode path (tensor-core head-packing MMA on sm_80+,
-// scalar fallback otherwise), mirroring dispatch_decode() in attn_decode.cu.
-#ifndef ASTRAI_NO_MMA
-static bool decode_use_mma(const AttentionParams<bf16>& p) {
-    int G = p.q_head / p.kv_head;
-    return !p.use_mask && G > 1 && G <= 16;
+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));
 }
 
-template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
-static void launch_mma_decode(AttentionParams<bf16>& p, DecodeScratch& sc) {
-    int tiles_total = (p.kv_len + BC - 1) / BC;
-    p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
-    p.o_part = sc.o_part;
-    p.ml_part = sc.ml_part;
-
-    attn_decode_split_kv_mma_kernel<HEAD_DIM, BC, STAGES>
-        <<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
-    attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
-}
-#endif
-
-static void launch_scalar_decode(AttentionParams<bf16>& p, DecodeScratch& sc) {
-    int gs = p.q_head / p.kv_head;
-    int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
-    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
-    p.o_part = sc.o_part;
-    p.ml_part = sc.ml_part;
-
-    size_t smem = DC_CHUNK * p.head_dim * sizeof(bf16);
-    attn_decode_split_kv_kernel<<<dim3(p.batch * p.kv_head, 1, p.num_splits), dim3(32, gs), smem>>>(p);
-    attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
-}
-
-template <int HEAD_DIM>
-static void dispatch_decode_t(AttentionParams<bf16>& p, DecodeScratch& sc) {
-#ifndef ASTRAI_NO_MMA
-    if (decode_use_mma(p)) { launch_mma_decode<HEAD_DIM, 32>(p, sc); return; }
-#endif
-    launch_scalar_decode(p, sc);
-}
-
-static void dispatch_decode(AttentionParams<bf16>& p, DecodeScratch& sc) {
-    dispatch_by_head_dim(p.head_dim, [&]<int D>() { dispatch_decode_t<D>(p, sc); });
+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},    // B, Hq, Hk, kv_len, D
+        {1, 32, 4, 512, 128},
         {1, 32, 4, 1024, 128},
         {1, 32, 4, 2048, 128},
         {1, 32, 4, 4096, 128},
@@ -104,10 +64,10 @@ static void bench() {
         p.q = dQ; p.k = dK; p.v = dV; p.mask = nullptr; p.o = dO;
 
         DecodeScratch sc;
-        cudaMalloc(&sc.o_part, (size_t)B*Hq*32*D*sizeof(float));
-        cudaMalloc(&sc.ml_part, (size_t)B*Hq*32*2*sizeof(float));
+        setup_scratch(p, sc);
+        p.o_part = sc.o_part; p.ml_part = sc.ml_part;
 
-        auto launch = [&]() { dispatch_decode(p, sc); };
+        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);
@@ -119,22 +79,14 @@ static void bench() {
         print_bench_row(cfg, r);
 
         cudaFree(dQ); cudaFree(dK); cudaFree(dV); cudaFree(dO);
-        cudaFree(sc.o_part); cudaFree(sc.ml_part);
+        free_scratch(sc);
     }
 }
 
-int main() {
-    const int configs[][5] = {
-        {1, 2, 1, 64, 32},    // B,Hq,Hk,seq_len,D
-        {1, 32, 4, 512, 128},
-        {1, 32, 4, 1024, 128},
-    };
-    int n_cfgs = sizeof(configs) / sizeof(configs[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], gs = Hq / Hk;
-        printf("=== B=%d Hq=%d Hk=%d seq=%d D=%d gs=%d ===\n", B,Hq,Hk,sl,D,gs);
+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];
@@ -161,18 +113,17 @@ int main() {
 
     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.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;
 
-        // Split-K scratch (max 32 splits), sized for the production MMA path.
     DecodeScratch sc;
-        cudaMalloc(&sc.o_part, (size_t)B*Hq*32*D*sizeof(float));
-        cudaMalloc(&sc.ml_part, (size_t)B*Hq*32*2*sizeof(float));
+    setup_scratch(p, sc);
+    p.o_part = sc.o_part; p.ml_part = sc.ml_part;
 
     double t0=now_ms();
-        dispatch_decode(p, sc);
+    dispatch_by_head_dim(D, [&]<int H>() { dispatch_decode<H>(p); });
     cudaDeviceSynchronize();
     double kms=now_ms()-t0;
     cudaError_t err=cudaGetLastError();
@@ -182,18 +133,51 @@ int main() {
     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, -1);
+    cpu_attention_ref(hQ, hK, hV, hMask, ref, B, Hq, Hk, 1, sl, D, causal ? 0 : -1);
 
-        float max_err=0;
+    float max_abs_err=0, max_rel_err=0;
     for (size_t i=0;i<nQ;i++){
-            float d=fabsf(bf2f(hOut[i])-ref[i]);
-            if(d>max_err) max_err=d;
+        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;
     }
-        printf("kernel: %.3f ms  max_err: %.6e\n\n",kms,max_err);
+    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);
-        cudaFree(sc.o_part);cudaFree(sc.ml_part);
+    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();

csrc/tests/attn_paged_decode_test.cu

@@ -5,12 +5,8 @@
 
 #include <cstring>
 #include "test_utils.cuh"
-#include "../kernels/attn_paged_decode_split_kv.cuh"
-#ifndef ASTRAI_NO_MMA
-#include "../kernels/attn_paged_decode_split_kv_mma.cuh"
-#endif
+#include "../kernels/attn_dispatchers.cuh"
 
-// Copy contiguous K/V from page pool (reference gather)
 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,
@@ -28,7 +24,8 @@ static void gather_kv_cpu(
                 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;
+                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));
             }
@@ -37,54 +34,29 @@ static void gather_kv_cpu(
 }
 
 template <int HEAD_DIM>
-static void launch_paged_decode(PagedAttentionParams<bf16, float>& p) {
-#ifndef ASTRAI_NO_MMA
-    int G_check = p.q_head / p.kv_head;
-    bool use_mma = !p.use_mask && G_check >= 1 && G_check <= 16 && p.page_size >= 32;
-    if (use_mma) {
-        constexpr int STAGES = (HEAD_DIM <= 128) ? 2 : 1;
-        int tiles_total = (p.kv_len + 32 - 1) / 32;
-        p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
-        paged_attn_decode_split_kv_mma_kernel<HEAD_DIM, 32, STAGES>
-            <<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
-    } else
-#endif
-    {
-        int group_sz = p.q_head / p.kv_head;
-        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);
-        paged_attn_decode_split_kv_kernel<<<
-            dim3(p.batch * p.kv_head, 1, p.num_splits),
-            dim3(32, group_sz), smem>>>(p);
-    }
-    paged_attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
-}
-
-template <int HEAD_DIM>
-static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int seed) {
-    printf("B=%d Hq=%d Hkv=%d kv_len=%d page_sz=%d head_dim=%d ... ", B, Hq, Hkv, kv_len, page_size, 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);
-    int max_splits = 32;
     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, *d_o_ref;
+    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_o_ref, sz_o);
     cudaMalloc(&d_k_pool, sz_kv);
     cudaMalloc(&d_v_pool, sz_kv);
     cudaMalloc(&d_pt, sz_pt);
@@ -107,7 +79,8 @@ static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int seed)
             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;
+                    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);
                 }
@@ -138,22 +111,22 @@ static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int seed)
     }
 
     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, -1);
+    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);
 
-    float scale_val = 1.0f / sqrtf((float)HEAD_DIM);
-    PagedAttentionParams<bf16, float> p;
+    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 = -1;
+    p.use_mask = 0; p.causal_offset = causal ? 0 : -1;
     set_default_paged_strides(p);
-    p.num_splits = 1; p.scale = scale_val;
+    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;
 
-    launch_paged_decode<HEAD_DIM>(p);
+    dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(p); });
     cudaDeviceSynchronize();
 
     bf16* h_o_bf16 = (bf16*)malloc(sz_o);
@@ -162,23 +135,30 @@ static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int seed)
     for (int i = 0; i < B * Hq * HEAD_DIM; i++)
         h_o_paged[i] = __bfloat162float(h_o_bf16[i]);
 
-    float max_err = 0.0f;
+    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_err) { max_err = e; bad_idx = 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;
     }
 
-    bool pass = max_err < 0.02f;
+    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)\n", max_err);
+        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 at [%d,%d,%d]: ref=%.4f got=%.4f)\n",
-               max_err, b, h, d, h_o_ref[bad_idx], h_o_paged[bad_idx]);
+        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]);
@@ -192,7 +172,7 @@ static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int seed)
     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_o_ref);
+    cudaFree(d_q); cudaFree(d_o_paged);
     cudaFree(d_k_pool); cudaFree(d_v_pool); cudaFree(d_pt);
     cudaFree(d_op); cudaFree(d_ml);
 
@@ -201,48 +181,43 @@ static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int seed)
 
 struct TestCase {
     int head_dim;
-    int B, Hq, Hkv, kv_len, page_size, seed;
+    int B, Hq, Hkv, kv_len, page_size, causal, seed;
 };
 
 static const TestCase TESTS[] = {
-    {128, 1, 1, 1, 8, 128, 1},
-    {128, 1, 4, 4, 128, 128, 2},
-    {128, 2, 4, 4, 256, 128, 3},
-    {128, 1, 4, 1, 64, 64, 4},
-    {128, 1, 8, 2, 64, 128, 5},
-    {128, 2, 16, 4, 128, 128, 6},
-    {64, 1, 4, 2, 32, 128, 7},
-    {256, 1, 2, 1, 16, 128, 8},
-    {32, 1, 4, 2, 32, 64, 9},
-    {128, 3, 8, 2, 256, 128, 10},
-    {128, 2, 32, 8, 512, 128, 11},
-#ifndef ASTRAI_NO_MMA
-    {128, 1, 16, 2, 256, 128, 12},
-    {128, 2, 32, 4, 512, 128, 13},
-#endif
+    {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) {
-    bool matched = false;
     int r = 0;
     dispatch_by_head_dim(tc.head_dim, [&]<int D>() {
-        matched = true;
-        r = run_test<D>(tc.B, tc.Hq, tc.Hkv, tc.kv_len, tc.page_size, tc.seed);
+        r = run_test<D>(tc.B, tc.Hq, tc.Hkv, tc.kv_len, tc.page_size, tc.causal, tc.seed);
     });
-    return matched ? r : 1;
+    return r;
 }
 
-// Warmed-up, CUDA-event timed sweep over paged decode configs.
-// Bytes = K + V read through page table (B*Hk*kv*D each), bf16.
 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);
-    int max_splits = 32;
     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);
 
@@ -269,13 +244,12 @@ static void bench_config(int B, int Hq, int Hkv, int kv_len, int page_size) {
     cudaMemcpy(d_pt, h_pt, sz_pt, cudaMemcpyHostToDevice);
     free(h_pt);
 
-    float scale_val = 1.0f / sqrtf((float)HEAD_DIM);
-    PagedAttentionParams<bf16, float> pa;
+    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.num_splits = 1; pa.scale = scale_val;
+    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;
@@ -283,7 +257,9 @@ static void bench_config(int B, int Hq, int Hkv, int kv_len, int page_size) {
     pa.o_part = d_op; pa.ml_part = d_ml;
 
     const int WARMUP = 10, ITERS = 100;
-    auto launch = [&]() { launch_paged_decode<HEAD_DIM>(pa); };
+    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));

csrc/tests/attn_prefill_test.cu

@@ -1,45 +1,14 @@
 /*
-Pure-C test:
+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_prefill_test.cu -o test && ./test
 */
 
 #include "test_utils.cuh"
-#include "../kernels/attn_prefill_split_q.cuh"
-#ifndef ASTRAI_NO_MMA
-#include "../kernels/attn_prefill_split_q_mma.cuh"
-#endif
-
-// Launch the production prefill path (tensor-core MMA on sm_80+, else the
-// scalar fallback), mirroring dispatch_prefill() in attn_prefill.cu.
-template <int HEAD_DIM>
-static void launch_prefill(AttentionParams<bf16>& p) {
-#ifndef ASTRAI_NO_MMA
-    constexpr int WARPS = 4, BR = 16;
-    constexpr int BC = (HEAD_DIM <= 128) ? 32 : 16;
-    dim3 grid((p.q_len + BR * WARPS - 1) / (BR * WARPS), p.q_head, p.batch);
-    dim3 block(WARPS * 32, 1, 1);
-    attn_prefill_split_q_mma_kernel<HEAD_DIM, WARPS, BC><<<grid, block>>>(p);
-#else
-    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, 1);
-    attn_prefill_split_q_kernel_t<HEAD_DIM, G, ROWS, P_BC><<<grid, block>>>(p);
-#endif
-}
-
-static void dispatch_prefill(AttentionParams<bf16>& p) {
-    switch (p.head_dim) {
-        case 64:  launch_prefill<64>(p);  break;
-        case 128: launch_prefill<128>(p); break;
-        default:  printf("bench: unsupported D=%d\n", p.head_dim);
-    }
-}
+#include "../kernels/attn_dispatchers.cuh"
 
 // Warmed-up, CUDA-event timed throughput sweep over the production MMA path.
-// Reports per-call latency and effective tensor-core TFLOP/s (2 matmuls:
-// QK^T and P@V, each 2*B*Hq*ql*kl*D flops; halved for causal).
 static void bench() {
     const int cfgs[][7] = {
         {1,32,4,512,512,128,0},
@@ -80,21 +49,21 @@ static void bench() {
         p.scale=1.0f/sqrtf((float)D);
         p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
 
-        for (int i=0;i<WARMUP;i++) dispatch_prefill(p);
+        auto launch = [&]() { dispatch_by_head_dim(D, [&]<int H>() { dispatch_prefill<H>(p); }); };
+        for (int i=0;i<WARMUP;i++) launch();
         cudaDeviceSynchronize();
         cudaError_t err=cudaGetLastError();
         if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return;}
 
         cudaEvent_t s,e; cudaEventCreate(&s); cudaEventCreate(&e);
         cudaEventRecord(s);
-        for (int i=0;i<ITERS;i++) dispatch_prefill(p);
+        for (int i=0;i<ITERS;i++) launch();
         cudaEventRecord(e); cudaEventSynchronize(e);
         float ms=0; cudaEventElapsedTime(&ms,s,e); ms/=ITERS;
 
         double flops = 4.0*B*Hq*(double)ql*kl*D;
         if (causal) flops *= 0.5;
         double tflops = flops/(ms*1e-3)/1e12;
-        // HBM traffic: Q + O (B*Hq*ql*D each) + K + V (B*Hk*kl*D each), bf16.
         double bytes = 2.0 * (2.0*nQ + 2.0*nKV);
         double gbps = bytes/(ms*1e-3)/1e9;
 
@@ -110,19 +79,7 @@ static void bench() {
     }
 }
 
-int main() {
-    const int configs[][7] = {
-        {1,2,1,64,128,64,0},     // tiny: B,Hq,Hk,q,kv,D,causal
-        {1,32,4,512,512,128,0},  // standard
-        {1,32,4,128,256,128,0},  // medium
-        {1,4,2,256,256,128,1},   // causal
-    };
-    int n_configs = sizeof(configs) / sizeof(configs[0]);
-
-    for (int ci = 0; ci < n_configs; ci++) {
-        int B=configs[ci][0], Hq=configs[ci][1], Hk=configs[ci][2];
-        int ql=configs[ci][3], kl=configs[ci][4], D=configs[ci][5];
-        int causal=configs[ci][6];
+static int run_test(int B, int Hq, int Hk, int ql, int kl, int D, int causal) {
     printf("=== B=%d Hq=%d Hk=%d q=%d kv=%d D=%d causal=%d ===\n",
            B,Hq,Hk,ql,kl,D,causal);
 
@@ -150,7 +107,7 @@ int main() {
     p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
 
     double t0=now_ms();
-        dispatch_prefill(p);
+    dispatch_by_head_dim(D, [&]<int H>() { dispatch_prefill<H>(p); });
     cudaDeviceSynchronize();
     double kms=now_ms()-t0;
     cudaError_t err=cudaGetLastError();
@@ -162,15 +119,49 @@ int main() {
     float* ref=new float[nQ];
     cpu_attention_ref(hQ, hK, hV, nullptr, ref, B, Hq, Hk, ql, kl, D, causal ? 0 : -1);
 
-        float max_err=0;
+    float max_abs_err=0, max_rel_err=0;
     for (size_t i=0;i<nQ;i++) {
-            float d=fabsf(bf2f(hOut[i])-ref[i]);
-            if(d>max_err) max_err=d;
+        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;
     }
-        printf("kernel: %.3f ms  max_err: %.6e\n\n",kms,max_err);
+    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);
     delete[]hQ;delete[]hK;delete[]hV;delete[]hOut;delete[]ref;delete[]tmp;
+
+    return pass ? 0 : 1;
+}
+
+int main() {
+    const int configs[][7] = {
+        {1,2,1,64,128,64,0},     // tiny: B,Hq,Hk,q,kv,D,causal
+        {1,32,4,512,512,128,0},  // standard
+        {1,32,4,128,256,128,0},  // medium
+        {1,4,2,256,256,128,1},   // causal
+    };
+    int n_configs = sizeof(configs) / sizeof(configs[0]);
+    int fail = 0;
+
+    for (int ci = 0; ci < n_configs; ci++) {
+        int B=configs[ci][0], Hq=configs[ci][1], Hk=configs[ci][2];
+        int ql=configs[ci][3], kl=configs[ci][4], D=configs[ci][5];
+        int causal=configs[ci][6];
+        fail += run_test(B, Hq, Hk, ql, kl, D, causal);
+        if (fail) break;
+    }
+
+    if (fail) {
+        printf("FAILED\n");
+        return fail;
     }
     printf("All tests passed!\n");
     bench();

csrc/tests/test_utils.cuh

@@ -18,16 +18,6 @@ inline double now_ms() {
     return duration_cast<milliseconds>(steady_clock::now().time_since_epoch()).count();
 }
 
-inline int compute_num_splits(int base_blocks, int tiles_total) {
-    int sm_count = 0;
-    cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, 0);
-    int n = (2 * sm_count + base_blocks - 1) / base_blocks;
-    if (n > tiles_total) n = tiles_total;
-    if (n > 32) n = 32;
-    if (n < 1) n = 1;
-    return n;
-}
-
 #define CUDA_CHECK(call) \
     do { \
         cudaError_t _e = (call); \
@@ -113,6 +103,7 @@ inline void set_default_strides(P& p) {
     p.kv_stride_l = p.head_dim;
     p.kv_stride_d = 1;
     p.mask_b_stride = p.kv_len;
+    p.mask_h_stride = 0;
     p.mask_q_stride = 0;
 }
 
@@ -124,6 +115,7 @@ inline void set_default_paged_strides(P& p) {
     p.q_stride_l  = p.head_dim;
     p.q_stride_d  = 1;
     p.mask_b_stride = p.kv_len;
+    p.mask_h_stride = 0;
     p.mask_q_stride = 0;
 }
 

docker-compose.yml

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

docs/README-zh-CN.md

@@ -1,9 +1,9 @@
 <div align="center">
   
-  <img src="../images/logo.png" width="auto" alt="Logo">
+  <img src="./images/logo.png" width="auto" alt="Logo">
   
   <div>
-    <a href="../../README.md">English</a> • 
+    <a href="../README.md">English</a> • 
     <a href="#chinese">中文</a>
   </div>
   
@@ -23,7 +23,7 @@
 <br>
 
 <div align="center">
-  <a href="../../README.md">English</a> •
+  <a href="../README.md">English</a> •
   <a href="#chinese">中文</a> •
   <a href="https://github.com/ViperEkura/AstrAI/issues">问题追踪</a> •
   <a href="https://github.com/ViperEkura/AstrAI/discussions">讨论区</a> •
@@ -219,18 +219,23 @@ curl -X POST http://localhost:8000/v1/messages \
 curl http://localhost:8000/health
 ```
 
-SSE 流式格式、错误码和统计端点详见[推理文档](./inference.md)。
+SSE 流式格式、错误码和统计端点详见[推理文档](guides/inference.md)。
 
 ### 文档
 
 | 文档 | 说明 |
 |------|------|
-| [CLI 参考](./params.md) | 所有 CLI 工具参数（训练、服务、生成、预处理） |
-| [架构文档](./architecture.md) | 系统架构、类图与设计模式 |
-| [训练文档](./training.md) | 训练循环、策略与公式 |
-| [推理文档](./inference.md) | KVCache、连续批处理、采样与 HTTP API |
-| [数据流程](./dataflow.md) | 数据管道、存储后端与数据集架构 |
-| [数据预处理](./preprocessing.md) | 声明式 JSON 驱动数据预处理 |
+| [快速上手](./get-started.md) | 安装与快速入门 |
+| [CLI 参考](./guides/params.md) | 所有 CLI 工具参数（训练、服务、生成、预处理） |
+| [数据预处理](./guides/preprocessing.md) | 声明式 JSON 驱动数据预处理 |
+| [训练文档](./guides/training.md) | 训练循环、策略与公式 |
+| [推理文档](./guides/inference.md) | KVCache、连续批处理、采样与 HTTP API |
+| [评估文档](./guides/evaluation.md) | HumanEval、MMLU、PPL、ROUGE、IFD、IFEval |
+| [分布式训练](./guides/distributed.md) | 多卡 DDP / FSDP 训练 |
+| [架构文档](./developer/architecture.md) | 系统架构、类图与设计模式 |
+| [数据流程](./developer/dataflow.md) | 数据管道、存储后端与数据集架构 |
+| [内部实现](./developer/internals.md) | 训练原理：损失公式、回调生命周期、KV Cache |
+| [CUDA 内核](./developer/cuda_kernels.md) | 自定义 CUDA 注意力内核与基准测试 |
 
 ### 贡献
 

docs/developer/architecture.md

@@ -28,17 +28,17 @@ classDiagram
 
         class AutoRegressiveLMConfig {
             +Optional[int] vocab_size
-            +Optional[int] dim
-            +Optional[int] n_layers
-            +Optional[float] norm_eps
-            +Optional[int] dim_ffn
-            +Optional[bool] tie_weight
+            +Optional[int] hidden_size
+            +Optional[int] num_hidden_layers
+            +Optional[float] rms_norm_eps
+            +Optional[int] intermediate_size
+            +Optional[bool] tie_word_embeddings
             +Optional[dict] rope_scaling
-            +Optional[int] max_len
+            +Optional[int] max_position_embeddings
             +Optional[float] rope_theta
             +str attn_type
-            +Optional[int] n_heads
-            +Optional[int] n_kv_heads
+            +Optional[int] num_attention_heads
+            +Optional[int] num_key_value_heads
             +Optional[bool] use_qk_norm
             +Optional[bool] use_gated_attention
             +Optional[int] kv_lora_rank
@@ -53,15 +53,15 @@ classDiagram
 
         class EncoderConfig {
             +Optional[int] vocab_size
-            +Optional[int] dim
-            +Optional[int] n_layers
-            +Optional[float] norm_eps
-            +Optional[int] dim_ffn
-            +Optional[int] max_len
+            +Optional[int] hidden_size
+            +Optional[int] num_hidden_layers
+            +Optional[float] rms_norm_eps
+            +Optional[int] intermediate_size
+            +Optional[int] max_position_embeddings
             +Optional[float] rope_theta
             +str attn_type
-            +Optional[int] n_heads
-            +Optional[int] n_kv_heads
+            +Optional[int] num_attention_heads
+            +Optional[int] num_key_value_heads
             +Optional[bool] use_qk_norm
             +str ffn_type
             +Optional[dict] rope_scaling
@@ -141,6 +141,12 @@ classDiagram
             +int val_step
             +float neftune_alpha
             +str parallel_mode
+            +int rollout_interval
+            +float rollout_temperature
+            +int rollout_top_k
+            +float rollout_top_p
+            +int rollout_max_tokens
+            +Optional[Callable] reward_model_fn
             +dict executor_kwargs
             +dict extra_kwargs
             +validate()
@@ -166,13 +172,6 @@ classDiagram
             +__getitem__(index) Dict
         }
 
-        class RecordDataset {
-            +Optional[Callable] processor
-            +load(load_path, storage_type)
-            +__getitem__(index)
-            +__len__()
-        }
-
         class DPODataset {
             +__getitem__(index) Dict
         }
@@ -222,7 +221,12 @@ classDiagram
             +fetch_record(index, keys)
         }
 
-        class ResumableDistributedSampler {
+        class JsonlSource {
+            +Path path
+            +load() List[dict]
+        }
+
+        class RDSampler {
             +int epoch
             +int iter
         }
@@ -273,7 +277,7 @@ classDiagram
             +ModuleList layers
             +RMSNorm norm
             +Linear lm_head
-            +forward(input_ids, input_mask, paged_cache, position_ids) Dict[str, Tensor]
+            +forward(input_ids, input_mask, kv_cache, position_ids) Dict[str, Tensor]
             +load_state_dict(state_dict, strict, assign)
             +state_dict()
         }
@@ -295,7 +299,7 @@ classDiagram
             +RMSNorm input_norm
             +nn.Module mlp        # MLP or DeepSeekMoE via FFNFactory
             +RMSNorm post_attention_norm
-            +forward(x, rotary_emb, attention_mask, paged_cache) Tensor
+            +forward(x, rotary_emb, attention_mask, kv_cache) Tensor
         }
 
         class GQA {
@@ -310,7 +314,7 @@ classDiagram
             +Linear q_proj, k_proj, v_proj, o_proj
             +Linear gate  # only if use_gated_attention
             +RMSNorm q_norm, k_norm  # only if use_qk_norm
-            +forward(x, rotary_emb, attn_mask, paged_cache) Tensor
+            +forward(x, rotary_emb, attn_mask, kv_cache) Tensor
         }
 
         class MLA {
@@ -330,7 +334,7 @@ classDiagram
             +Linear gate  # only if use_gated_attention
             +RMSNorm kv_norm
             +RMSNorm q_norm, k_norm  # only if use_qk_norm
-            +forward(x, rotary_emb, attn_mask, paged_cache) Tensor
+            +forward(x, rotary_emb, attn_mask, kv_cache) Tensor
         }
 
         class MLP {
@@ -376,7 +380,9 @@ classDiagram
             +int max_len
             +float base
             +Optional[Dict] rope_scaling
-            +forward(x, position_ids=None) Tensor
+            +Tensor cos_table
+            +Tensor sin_table
+            +forward(x, position_ids=None) Tuple[Tensor, Tensor]
         }
 
         class Embedding {
@@ -385,19 +391,103 @@ classDiagram
             +forward(x) Tensor
             +set_neftune_alpha(alpha)
         }
+
+        class LoRAConfig {
+            +int r
+            +int alpha
+            +tuple target_modules
+        }
+
+        class LoRALinear {
+            +Linear weight
+            +Parameter lora_A, lora_B
+            +forward(x) Tensor
+            +merge()
+        }
     }
 
     namespace preprocessing {
+        class SectionRenderer {
+            +process_sections(item, sections, config, tokenizer) Tuple
+            +process_list_field(item, sections, config, tokenizer) Tuple
+        }
+
         class BaseMaskBuilder {
             <<abstract>>
             +build(item, config, tokenizer) Optional[dict]
         }
 
-        class SectionedMaskBuilder {
+        class SingleOutputMaskBuilder {
             +SectionRenderer renderer
             +build(item, config, tokenizer) Optional[dict]
-            +_build_single(item, config, tokenizer) Optional[dict]
-            +_build_multi(item, sources_spec, config, tokenizer) Optional[dict]
+        }
+
+        class MultiOutputMaskBuilder {
+            +SectionRenderer renderer
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class SectionedMaskBuilder {
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class PackingStrategy {
+            <<abstract>>
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class PackingStrategyFactory {
+            +create(name, *args, **kwargs) PackingStrategy
+        }
+
+        class SimplePacking {
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class BFDPacking {
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class BFDSplitPacking {
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class PositionIdStrategy {
+            <<abstract>>
+            +generate(sequences) List[int]
+        }
+
+        class PositionIdStrategyFactory {
+            +create(name, *args, **kwargs) PositionIdStrategy
+        }
+
+        class NoPositionId {
+            +generate(sequences) List[int]
+        }
+
+        class DocResetPositionId {
+            +generate(sequences) List[int]
+        }
+
+        class ContinuousPositionId {
+            +generate(sequences) List[int]
+        }
+
+        class StoreWriter {
+            <<abstract>>
+            +save(output_dir, domain, shard_idx, tensors)
+        }
+
+        class StoreWriterFactory {
+            +create(name, *args, **kwargs) StoreWriter
+        }
+
+        class BinWriter {
+            +save(output_dir, domain, shard_idx, tensors)
+        }
+
+        class H5Writer {
+            +save(output_dir, domain, shard_idx, tensors)
         }
 
         class Pipeline {
@@ -497,7 +587,7 @@ classDiagram
 
         class TrainContextBuilder {
             +TrainConfig config
-            +with_resume_dir(resume_dir) TrainContextBuilder
+            +with_param_path(param_path, resume) TrainContextBuilder
             +build() TrainContext
         }
 
@@ -544,6 +634,32 @@ classDiagram
             +sync_old_model()
         }
 
+        class RawRollout {
+            +Tensor prompts
+            +Tensor responses
+            +Tensor response_mask
+            +Tensor logprobs_old
+        }
+
+        class RolloutResult {
+            +Tensor rewards
+        }
+
+        class BaseRewardModel {
+            <<abstract>>
+            +score(prompts, responses) Tensor
+        }
+
+        class RolloutGenerator {
+            +generate(batch) RawRollout
+        }
+
+        class RolloutRunner {
+            +step()
+            +clear_cache()
+            +__call__(batch) Tuple[RolloutResult, bool]
+        }
+
         class BaseScheduler {
             +get_lr() List[float]
             +step()
@@ -710,75 +826,49 @@ classDiagram
             +record(page_idx, token_ids, logical_page_idx)
         }
 
-        class Storage {
-            +int page_size
-            +Tensor k_cache
-            +Tensor v_cache
-            +write(layer_id, page_table, start_pos, k, v)
-            +gather(layer_id, page_table, total_len) Tuple[Tensor, Tensor]
+        class KVStorage {
+            +int size
+            +Tensor k_buffer
+            +Tensor v_buffer
+            +get_key_buffer(layer_id) Tensor
+            +get_value_buffer(layer_id) Tensor
+            +set_kv_buffer(layer_id, loc, k, v)
+        }
+
+        class ReqToTokenPool {
+            +int size
+            +int max_context_len
+            +Tensor req_to_token
+            +alloc(num_reqs) List[int]
+            +free(req_indices)
+            +write(indices, values)
         }
 
         class KVCache {
-            <<abstract>>
-            +task_alloc(task_id, prompt_ids) bool
-            +task_free(task_id)
-            +task_extend(task_id, pos) bool
-            +task_cached(task_id) int
-            +task_record_hashes(task_id, prompt_ids, start_logical_page)
-            +bind_tasks(task_ids, total_len, device) CacheView
+            +Tensor k_buffer
+            +Tensor v_buffer
+            +Tensor req_to_token
+            +Tensor req_pool_indices
+            +Tensor seq_lens
+            +Tensor out_cache_loc
+            +int max_len
+            +Optional[Tensor] page_table
+            +Optional[Tensor] decode_mask
         }
 
-        class PageCache {
+        class PagePool {
             +int page_size
-            -PagePool _pool
-            -Storage _storage
-            -TaskTable _table
+            +bool contiguous
+            -KVStorage _storage
+            -ReqToTokenPool _req_pool
+            -Allocator _alloc
+            -PrefixCache _prefix
             +task_alloc(task_id, prompt_ids) bool
             +task_free(task_id)
             +task_extend(task_id, pos) bool
             +task_cached(task_id) int
             +task_record_hashes(task_id, prompt_ids, start_logical_page)
-            +bind_tasks(task_ids, total_len, device) PageCacheView
-        }
-
-        class ContiguousCache {
-            +int max_seq_len
-            +Tensor k, v
-            +task_alloc(task_id, prompt_ids) bool
-            +task_free(task_id)
-            +task_extend(task_id, pos) bool
-            +bind_tasks(task_ids, total_len, device) ContiguousCacheView
-        }
-
-        class CacheView {
-            <<abstract>>
-            +write(layer_id, k, v)
-            +gather(layer_id) Tuple[Tensor, Tensor]
-        }
-
-        class PageCacheView {
-            -Storage _storage
-            +Tensor _page_table
-            +int _total_len
-            +write(layer_id, k, v)
-            +gather(layer_id) Tuple[Tensor, Tensor]
-        }
-
-        class ContiguousCacheView {
-            -ContiguousCache _cache
-            +Tensor _batch_indices
-            +int _total_len
-            +write(layer_id, k, v)
-            +gather(layer_id) Tuple[Tensor, Tensor]
-        }
-
-        class TaskTable {
-            +set(task_id, page_table, cached)
-            +get(task_id) List[int]
-            +get_cached(task_id) int
-            +get_ref(task_id) List[int]
-            +pop(task_id) Tuple[List[int], int]
-            +table_tensor(task_ids, device) Tensor
+            +bind_tasks(task_ids, seq_lens, device, start_pos) KVCache
         }
 
     class Task {
@@ -810,7 +900,6 @@ classDiagram
             +AutoTokenizer tokenizer
             +int max_batch_size
             +int max_seq_len
-            +int max_prompt_len
             +Deque waiting_queue
             +List active_tasks
             +add_task(prompt, max_tokens, temperature, top_p, top_k, stream_callback) str
@@ -857,12 +946,21 @@ classDiagram
             +apply(logits, filter_value) Tensor
         }
 
+        class FrequencyPenaltyStrategy {
+            +float penalty
+            +apply(logits, filter_value, input_ids, input_mask) Tensor
+        }
+
         class SamplingPipeline {
             +List[BaseSamplingStrategy] strategies
             +apply(logits, filter_value) Tensor
             +sample(logits, filter_value) Tensor
         }
 
+        class StreamDecoder {
+            +push(token_id) str
+        }
+
         class GenerateResult {
             +List[Tuple[int, str]] tokens
             +List[str] results
@@ -881,6 +979,17 @@ classDiagram
             +Optional[str] tool_call_id
         }
 
+        class FunctionDef {
+            +str name
+            +Optional[str] description
+            +Optional[Dict] parameters
+        }
+
+        class ToolDef {
+            +str type
+            +FunctionDef function
+        }
+
         class ChatCompletionRequest {
             +str model
             +List[ChatMessage] messages
@@ -969,9 +1078,20 @@ classDiagram
             +str yielded
         }
 
-        class get_app {
-            <<module>>
-            +get_app() FastAPI
+        class BaseToolParser {
+            <<abstract>>
+            +feed(body, current_token_ids, delta_token_ids) List[Dict]
+            +parse_complete(body) Optional[Dict]
+            +has_tool_calls (property) bool
+        }
+
+        class ToolParserFactory {
+            +create(name, *args, **kwargs) BaseToolParser
+        }
+
+        class SimpleJsonToolParser {
+            +feed(body, current_token_ids, delta_token_ids) List[Dict]
+            +parse_complete(body) Optional[Dict]
         }
     }
 
@@ -994,14 +1114,17 @@ classDiagram
     }
 
     namespace parallel {
-        class setup {
-            <<module>>
-            +spawn_parallel_fn(func, world_size, backend, master_addr, master_port, device_type, start_method, **kwargs)
-            +setup_parallel(rank, world_size, backend, master_addr, master_port, device_type) contextmanager
-            +get_current_device() str
-            +get_world_size() int
-            +get_rank() int
-            +only_on_rank(rank, sync=False) decorator
+        class LaunchStrategy {
+            <<abstract>>
+            +launch(func, **kwargs)
+        }
+
+        class TorchrunStrategy {
+            +launch(func, **kwargs)
+        }
+
+        class LocalStrategy {
+            +launch(func, **kwargs)
         }
 
         class GradientState {
@@ -1030,7 +1153,7 @@ classDiagram
 
         class BaseExecutor {
             +GradientState gradient_state
-            +prepare(model, optimizer, dataloader, scheduler) tuple
+            +prepare(model_fn, optimizer_fn, scheduler_fn, before_wrap) tuple
             +accumulate(model) context manager
             +backward(loss)
             +unwrap_model(model) dict
@@ -1049,6 +1172,7 @@ classDiagram
 
         class FSDPExecutor {
             -_prepare_model(model) nn.Module
+            -_no_sync(model) context manager
             +unwrap_model(model) dict
         }
 
@@ -1104,9 +1228,8 @@ classDiagram
     TrainCallback <|-- MetricCallback
     BaseDataset <|-- SEQDataset
     BaseDataset <|-- SFTDataset
-    BaseDataset <|-- RecordDataset
-    RecordDataset <|-- DPODataset
-    RecordDataset <|-- GRPODataset
+    BaseDataset <|-- DPODataset
+    BaseDataset <|-- GRPODataset
     Store <|-- H5Store
     Store <|-- MmapStore
     Store <|-- JsonlStore
@@ -1119,6 +1242,7 @@ classDiagram
     BaseSamplingStrategy <|-- TemperatureStrategy
     BaseSamplingStrategy <|-- TopKStrategy
     BaseSamplingStrategy <|-- TopPStrategy
+    BaseSamplingStrategy <|-- FrequencyPenaltyStrategy
     ParallelModel <|-- RowParallelLinear
     ParallelModel <|-- ColumnParallelLinear
     AutoModel <|-- AutoRegressiveLM
@@ -1142,23 +1266,37 @@ classDiagram
     BaseFactory <|-- ExecutorFactory
     BaseFactory <|-- ConfigFactory
     BaseFactory <|-- MaskBuilderFactory
+    BaseFactory <|-- PackingStrategyFactory
+    BaseFactory <|-- PositionIdStrategyFactory
+    BaseFactory <|-- StoreWriterFactory
+    BaseFactory <|-- ToolParserFactory
     BaseExecutor <|-- NoneExecutor
     BaseExecutor <|-- DDPExecutor
     BaseExecutor <|-- FSDPExecutor
     ResponseBuilder <|-- OpenAIResponseBuilder
     ResponseBuilder <|-- AnthropicResponseBuilder
+    BaseToolParser <|-- SimpleJsonToolParser
     BaseMaskBuilder <|-- SectionedMaskBuilder
-    KVCache <|-- PageCache
-    KVCache <|-- ContiguousCache
-    CacheView <|-- PageCacheView
-    CacheView <|-- ContiguousCacheView
-
+    BaseMaskBuilder <|-- SingleOutputMaskBuilder
+    BaseMaskBuilder <|-- MultiOutputMaskBuilder
+    PackingStrategy <|-- SimplePacking
+    PackingStrategy <|-- BFDPacking
+    BFDPacking <|-- BFDSplitPacking
+    PositionIdStrategy <|-- NoPositionId
+    PositionIdStrategy <|-- DocResetPositionId
+    PositionIdStrategy <|-- ContinuousPositionId
+    StoreWriter <|-- BinWriter
+    StoreWriter <|-- H5Writer
+    RawRollout <|-- RolloutResult
+    LaunchStrategy <|-- TorchrunStrategy
+    LaunchStrategy <|-- LocalStrategy
     %% --- Composition (strong ownership, part destroyed with whole) ---
-    PageCache *-- PagePool
-    PageCache *-- Storage
-    PageCache *-- TaskTable
+    PagePool *-- KVStorage
+    PagePool *-- ReqToTokenPool
+    PagePool *-- Allocator
+    PagePool *-- PrefixCache
     InferenceEngine *-- InferenceScheduler
-    InferenceScheduler *-- KVCache
+    InferenceScheduler *-- PagePool
     InferenceScheduler *-- Executor
     InferenceScheduler *-- TaskManager
     AutoRegressiveLM *-- DecoderBlock
@@ -1169,6 +1307,8 @@ classDiagram
     EmbeddingEncoder *-- Embedding
     DecoderBlock *-- RMSNorm
     ChatCompletionRequest *-- ChatMessage
+    ChatCompletionRequest *-- ToolDef
+    ToolDef *-- FunctionDef
     MessagesRequest *-- AnthropicMessage
     BaseExecutor *-- GradientState
     AccumOptimizer o-- GradientState
@@ -1184,13 +1324,14 @@ classDiagram
     TrainContext o-- BaseScheduler
     TrainContext o-- Checkpoint
     TrainContext o-- BaseExecutor
-    PageCacheView o-- Storage
-    ContiguousCacheView o-- ContiguousCache
     SamplingPipeline o-- BaseSamplingStrategy
     BaseDataset o-- Store
     Pipeline o-- PipelineConfig
     Pipeline o-- BaseMaskBuilder
     Pipeline o-- AutoTokenizer
+    Pipeline o-- PackingStrategy
+    Pipeline o-- PositionIdStrategy
+    Pipeline o-- StoreWriter
     TokenizeTransform o-- AutoTokenizer
     TokenizeTransform o-- BaseMaskBuilder
 
@@ -1198,6 +1339,9 @@ classDiagram
     TrainConfig ..> BaseStrategy : selects
     PipelineConfig ..> MaskBuilderFactory : selects
     MaskBuilderFactory ..> BaseMaskBuilder : creates
+    PackingStrategyFactory ..> PackingStrategy : creates
+    PositionIdStrategyFactory ..> PositionIdStrategy : creates
+    StoreWriterFactory ..> StoreWriter : creates
     StrategyFactory ..> BaseStrategy : creates
     SchedulerFactory ..> BaseScheduler : creates
     DatasetFactory ..> BaseDataset : creates
@@ -1216,22 +1360,24 @@ classDiagram
     ExecutorFactory ..> NoneExecutor : creates
     ExecutorFactory ..> DDPExecutor : creates
     ExecutorFactory ..> FSDPExecutor : creates
+    ToolParserFactory ..> BaseToolParser : creates
     TrainContextBuilder ..> ExecutorFactory : creates
     Trainer ..> TrainContextBuilder : uses
     TrainContextBuilder ..> TrainContext : creates
-    Trainer ..> Functions : spawns
     TrainContextBuilder ..> StrategyFactory : uses
-    TrainContextBuilder ..> ResumableDistributedSampler : creates
+    TrainContextBuilder ..> RDSampler : creates
     Checkpoint ..> Checkpoint : serializes
     CheckpointCallback ..> Checkpoint : creates
-    PageCache ..> PageCacheView : binds
-    ContiguousCache ..> ContiguousCacheView : binds
+    PagePool ..> KVCache : binds
     InferenceEngine ..> GenerationRequest : uses
     InferenceEngine ..> GenerateResult : creates
     OpenAIResponseBuilder ..> ChatCompletionRequest : receives
     AnthropicResponseBuilder ..> MessagesRequest : receives
     ProtocolHandler ..> StopChecker : creates
     ProtocolHandler ..> GenContext : creates
+    RolloutGenerator ..> InferenceScheduler : uses
+    RolloutRunner ..> RolloutGenerator : uses
+    RolloutRunner ..> BaseRewardModel : uses
 
     %% --- Association (general usage) ---
     Trainer --> TrainConfig
@@ -1253,14 +1399,15 @@ classDiagram
 | Module | Components | Description |
 |--------|------------|-------------|
 | **astrai.config** | BaseConfig, BaseModelConfig, AutoRegressiveLMConfig, EncoderConfig, ConfigFactory, TrainConfig, PipelineConfig, InputConfig, ProcessingConfig, OutputConfig | Configuration management (to_dict/from_dict, to_file/from_file) |
-| **astrai.preprocessing** | BaseMaskBuilder, MaskBuilderFactory, SectionedMaskBuilder, SingleOutputMaskBuilder, MultiOutputMaskBuilder, Pipeline, TokenizeTransform, filter_by_length, PackingStrategy, PackingStrategyFactory, plan_bfd, PositionIdStrategy, PositionIdStrategyFactory, StoreWriter, StoreWriterFactory, core (shared helpers) | Declarative JSON-driven data preprocessing |
-| **astrai.dataset** | BaseDataset–RecordDataset–DPO/GRPODataset, SEQDataset, SFTDataset, Store, Streamable, Recordable, H5Store, MmapStore, JsonlStore, StoreFactory, ResumableDistributedSampler, DatasetFactory | Dataset loading and management |
+| **astrai.preprocessing** | SectionRenderer, BaseMaskBuilder, MaskBuilderFactory, SectionedMaskBuilder, SingleOutputMaskBuilder, MultiOutputMaskBuilder, Pipeline, TokenizeTransform, PackingStrategy, PackingStrategyFactory, SimplePacking, BFDPacking, BFDSplitPacking, PositionIdStrategy, PositionIdStrategyFactory, NoPositionId, DocResetPositionId, ContinuousPositionId, StoreWriter, StoreWriterFactory, BinWriter, H5Writer | Declarative JSON-driven data preprocessing |
+| **astrai.dataset** | BaseDataset, SEQDataset, SFTDataset, DPODataset, GRPODataset, Store, Streamable, Recordable, H5Store, MmapStore, JsonlSource, JsonlStore, StoreFactory, RDSampler, DatasetFactory | Dataset loading and management |
 | **astrai.serialization** | Checkpoint | Model serialization |
-| **astrai.model** | AutoModel, AutoRegressiveLM, EmbeddingEncoder, DecoderBlock, GQA, MLA, MLP, DeepSeekMoE, AttnFactory, FFNFactory, RMSNorm, Linear, RotaryEmbedding, Embedding | Neural network model |
+| **astrai.model** | AutoModel, AutoRegressiveLM, EmbeddingEncoder, DecoderBlock, GQA, MLA, MLP, DeepSeekMoE, AttnFactory, FFNFactory, RMSNorm, Linear, LoRAConfig, LoRALinear, RotaryEmbedding, Embedding | Neural network model |
 | **astrai.tokenize** | AutoTokenizer, ChatTemplate | Tokenizer and chat template |
-| **astrai.trainer** | Trainer, TrainContext, TrainContextBuilder, BaseStrategy–GRPOStrategy, StrategyFactory, BaseScheduler–WSDScheduler, SchedulerFactory, TrainCallback(Protocol)–MetricCallback, CallbackFactory | Training workflow |
-| **astrai.inference** | InferenceEngine, InferenceScheduler, Executor, KVCache–ContiguousCache/PageCache, CacheView–ContiguousCacheView/PageCacheView, Allocator–Storage, Task, TaskManager, TaskStatus, GenerationRequest, GenerateResult, BaseSamplingStrategy–SamplingPipeline, ProtocolHandler, ResponseBuilder, OpenAIResponseBuilder, AnthropicResponseBuilder, StopChecker, GenContext, ChatMessage–MessagesRequest, app | Inference service |
-| **astrai.parallel** | spawn_parallel_fn, setup_parallel, get_rank/get_world_size/get_current_device, only_on_rank, BaseExecutor, ExecutorFactory, NoneExecutor, DDPExecutor, FSDPExecutor, GradientState, AccumOptimizer, AccumScheduler, ParallelModel, RowParallelLinear, ColumnParallelLinear | Distributed parallel & gradient accumulation |
+| **astrai.trainer** | Trainer, TrainContext, TrainContextBuilder, BaseStrategy–GRPOStrategy, StrategyFactory, BaseScheduler–WSDScheduler, SchedulerFactory, TrainCallback(Protocol)–MetricCallback, CallbackFactory, RawRollout, RolloutResult, BaseRewardModel, RolloutGenerator, RolloutRunner | Training workflow |
+| **astrai.inference** | InferenceEngine, InferenceScheduler, Executor, PagePool, KVStorage, ReqToTokenPool, KVCache, Allocator, PrefixCache, Task, TaskManager, TaskStatus, StreamDecoder, GenerationRequest, GenerateResult, BaseSamplingStrategy–SamplingPipeline, FrequencyPenaltyStrategy, ProtocolHandler, ResponseBuilder, OpenAIResponseBuilder, AnthropicResponseBuilder, StopChecker, GenContext, StopInfo, ChatMessage, FunctionDef, ToolDef, ChatCompletionRequest, AnthropicMessage, MessagesRequest, BaseToolParser, ToolParserFactory, SimpleJsonToolParser | Inference service |
+| **astrai.extension** | AttentionBackend, TorchNativeBackend, CudaBackend, attn_backend, ATTN_BACKEND, attn_decode, attn_prefill, attn_paged_decode, rotary_emb, apply_rotary_emb, rotary_backend, is_available | CUDA attention + rotary kernels, backend abstraction, auto-dispatch |
+| **astrai.parallel** | spawn_parallel_fn, setup_parallel, get_rank/get_world_size/get_current_device, only_on_rank, LaunchStrategy, TorchrunStrategy, LocalStrategy, BaseExecutor, ExecutorFactory, NoneExecutor, DDPExecutor, FSDPExecutor, GradientState, AccumOptimizer, AccumScheduler | Distributed parallel & gradient accumulation |
 | **astrai.factory** | BaseFactory | Component registration |
 | **astrai.protocols** | OptimizerProtocol, SchedulerProtocol | Structural subtyping for optimizer/scheduler wrappers |
 
@@ -1268,7 +1415,7 @@ classDiagram
 
 | Pattern | Classes | Purpose |
 |---------|---------|---------|
-| **Factory** | `AttnFactory`, `FFNFactory`, `StrategyFactory`, `DatasetFactory`, `SchedulerFactory`, `CallbackFactory`, `StoreFactory`, `ConfigFactory`, `ExecutorFactory`, `MaskBuilderFactory`, `StoreWriterFactory`, `PackingStrategyFactory`, `PositionIdStrategyFactory` | Decorator-based component creation |
+| **Factory** | `AttnFactory`, `FFNFactory`, `StrategyFactory`, `DatasetFactory`, `SchedulerFactory`, `CallbackFactory`, `StoreFactory`, `ConfigFactory`, `ExecutorFactory`, `MaskBuilderFactory`, `StoreWriterFactory`, `PackingStrategyFactory`, `PositionIdStrategyFactory`, `ToolParserFactory` | Decorator-based component creation |
 | **Registry** | `BaseFactory` | Component registration |
 | **Strategy** | `SEQStrategy`, `SFTStrategy`, `DPOStrategy`, `GRPOStrategy` | Training strategy switching |
 | **Strategy (Sampling)** | `TemperatureStrategy`, `TopKStrategy`, `TopPStrategy`, `SamplingPipeline` | Composable logit transformations |
@@ -1277,6 +1424,8 @@ classDiagram
 | **Observer** | `TrainCallback`, callback implementations | Training process monitoring |
 | **Context** | `TrainContext` | Unified training state bag |
 | **Object Pool** | `Allocator`, `PagePool` | Page-based KV cache with LRU eviction |
+| **Strategy (Attention)** | `AttentionBackend`, `TorchNativeBackend`, `CudaBackend` | Attention computation backend switching via context manager |
+| **Auto-dispatch (Rotary)** | `apply_rotary_emb`, `rotary_backend.py`, `rotary_ops.py` | Rotary embedding CUDA kernel auto-dispatch with torch fallback |
 | **Executor** | `BaseExecutor`, `NoneExecutor`, `DDPExecutor`, `FSDPExecutor` | Gradient accumulation & model distribution |
 | **Storage** | `Store`, `H5Store`, `MmapStore`, `JsonlStore` | Format-agnostic data access with multi-segment support |
 | **Producer-Consumer** | `InferenceScheduler`, `Task`, queues | Continuous batching |
@@ -1288,7 +1437,7 @@ classDiagram
 2. **Training Flow**: `Trainer` → `TrainContextBuilder` → `TrainContext`, uses `BaseStrategy` for loss, `BaseExecutor` for gradient accumulation + model distribution
 3. **Strategy Selection**: `StrategyFactory` creates strategy by `train_type`
 4. **Executor Selection**: `ExecutorFactory.create(cfg.parallel_mode, grad_accum_steps=cfg.grad_accum_steps, **cfg.executor_kwargs)` → `NoneExecutor` / `DDPExecutor` / `FSDPExecutor`
-5. **Inference Flow**: `InferenceEngine` → `InferenceScheduler` → `AutoRegressiveLM`, backed by `KVCache` + `SamplingPipeline`
+5. **Inference Flow**: `InferenceEngine` → `InferenceScheduler` → `AutoRegressiveLM`, backed by `PagePool` + `KVCache` + `SamplingPipeline`. Attention backend selected via `attn_backend()` context manager (`TorchNativeBackend` default, `CudaBackend` for CUDA kernels). Rotary embedding auto-dispatches to CUDA kernel when available (inference mode), else torch complex multiply (training).
 6. **Distributed**: `spawn_parallel_fn` + `setup_parallel` for multi-process DDP
 7. **Dataset Loading**: `DatasetFactory` creates datasets, `Store` (H5Store/MmapStore/JsonlStore) loads data with explicit `_length` and multi-segment `_data`
 8. **Checkpoint**: `Checkpoint` saves/loads safetensors + metadata (rank-0 only), extra state saved as `{key}.pt`
@@ -1296,4 +1445,4 @@ classDiagram
 10. **AutoModel**: `from_pretrained()` loads `config.json` + `model.safetensors`, `_disable_random_init` replaces `nn.init.*` with no-ops
 11. **Protocols**: `OptimizerProtocol` / `SchedulerProtocol` — structural subtyping for `AccumOptimizer` / `AccumScheduler` wrappers
 
-> Document Update Time: 2026-07-19
+> Document Update Time: 2026-07-31

docs/developer/cuda_kernels.md

@@ -0,0 +1,175 @@
+# CUDA Kernels
+
+AstrAI includes optional custom CUDA kernels for attention and rotary embedding. These are built when `nvcc` is available and CUDA is detected, and are dispatched via the `CudaBackend` attention backend or auto-dispatched for rotary.
+
+## Overview
+
+| Kernel | File | Description |
+|--------|------|-------------|
+| `attn_decode` | `attn_decode.cu` | GQA decode attention (split-KV) |
+| `attn_prefill` | `attn_prefill.cu` | GQA prefill attention (split-Q) |
+| `attn_paged_decode` | `attn_paged_decode.cu` | Paged KV cache decode attention |
+| `rotary_emb` | `rotary_emb.cu` | Fused rotary embedding (cos/sin lookup + rotation) |
+
+Additionally, optimized `.cuh` variants with tensor-core MMA (Matrix Multiply-Accumulate) exist:
+
+| Variant | File | Optimization |
+|---------|------|--------------|
+| Split-KV MMA decode | `attn_decode_split_kv_mma.cuh` | Split KV across warps + MMA (sm_80+) |
+| Split-Q MMA prefill | `attn_prefill_split_q_mma.cuh` | Split Q across warps + MMA (sm_80+) |
+| Paged split-KV MMA decode | `attn_paged_decode_split_kv_mma.cuh` | Paged cache + split-KV + MMA |
+
+### Rotary Embedding Kernel
+
+The `rotary_emb` kernel (`csrc/kernels/rotary_emb.cu`) fuses cos/sin lookup and rotation into a single kernel:
+
+- One thread per (head, dim-pair), vectorized `__nv_bfloat162` load/store
+- f32 cos/sin input, bf16 compute and output
+- 256-thread blocks, grid-stride loop
+- Auto-dispatched via `apply_rotary_emb` in `astrai/extension/rotary_backend.py` (CUDA when available + inference mode, else torch complex-multiply fallback)
+- No context-manager backend needed — rotary is backend-agnostic, both attention backends benefit
+
+Standalone benchmark vs torch complex-multiply (48 calls = 24 layers × q+k): 6-9x faster, max diff 0 (decode) to 3e-2 (large prefill, bf16).
+
+## Build System
+
+### Auto-detection
+
+Kernels are built when **both** of these conditions are met:
+1. `nvcc` is available on `PATH`
+2. `torch.cuda.is_available()` returns `True`
+
+Unless `CSRC_KERNELS=false` is set explicitly.
+
+### Manual build
+
+```bash
+# During install
+CSRC_KERNELS=true pip install -e . --no-build-isolation
+
+# Rebuild after editing .cu/.cuh files
+CSRC_KERNELS=true python setup.py build_ext --inplace
+# Output: astrai/extension/lib/*.so
+```
+
+### Architecture flags
+
+`csrc/build.py` auto-detects the GPU compute capability and generates the appropriate `nvcc` gencode flag:
+
+- **sm_80+** (Ampere and later): enables tensor-core MMA path (`mma.sync.m16n8k16.bf16`)
+- **Below sm_80**: adds `-DASTRAI_NO_MMA` to disable the MMA path at compile time
+
+### Build configuration
+
+```
+NVCC_FLAGS = -O3 --expt-relaxed-constexpr --use_fast_math
+             --ptxas-options=-O3,-v --extra-device-vectorization --threads=8
+```
+
+The `REGISTRY` in `csrc/build.py` lists all registered kernels (currently 4). Each entry maps a kernel name to its source files and build flags.
+
+## Attention Backend
+
+`astrai/extension/attention_backend.py` provides the backend abstraction:
+
+- **`AttentionBackend`** (ABC): `fwd_decode` / `fwd_prefill` abstract methods, `forward` dispatches by q_len
+- **`TorchNativeBackend`**: SDPA with indirect KV cache gather (default)
+- **`CudaBackend`**: CUDA kernel dispatch — decode via `attn_paged_decode` (page_size=1), prefill via `attn_prefill`
+
+Select a backend via context manager (mirrors `torch.nn.attention.sdpa_kernel`):
+
+```python
+from astrai.extension import attn_backend, ATTN_BACKEND
+
+with attn_backend(ATTN_BACKEND.CUDA):
+    engine.generate("hello")
+```
+
+`CudaBackend` falls back to `TorchNativeBackend` when a kernel is not available.
+
+### Rotary Backend
+
+`astrai/extension/rotary_backend.py` provides `apply_rotary_emb(x, (cos, sin))` with auto-dispatch:
+
+- **CUDA path**: calls `rotary_emb` kernel directly when available, input is bf16 on CUDA, and `torch.is_grad_enabled()` is `False` (inference)
+- **Torch fallback**: complex multiply (`torch.view_as_complex` → `torch.complex` multiply → `torch.view_as_real`), used during training (supports autograd) or when kernel unavailable
+
+No context-manager switching needed — the dispatch is automatic per call.
+
+## Python Wrappers
+
+`astrai/extension/attention_ops.py` provides Python wrappers for each compiled attention kernel. Each wrapper calls its CUDA kernel directly and raises `RuntimeError` if the `.so` is not available. Fallback to torch SDPA is handled by the attention backend, not the wrapper functions.
+
+`astrai/extension/rotary_ops.py` provides the wrapper for the rotary embedding kernel. Fallback to torch complex multiply is handled by `rotary_backend.py`.
+
+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)
+```
+
+Layout convention: all q/k/v are `[batch, seq_len, n_heads, head_dim]` (blhd). Scale is always `1/sqrt(head_dim)`.
+
+## Standalone Testing
+
+Each `csrc/tests/*.cu` file has the `nvcc` compile command in its header comment. Example:
+
+```bash
+nvcc -I csrc -arch=sm_89 -O3 --use_fast_math \
+     --ptxas-options=-O3,-v --extra-device-vectorization \
+     csrc/tests/attn_decode_test.cu -o /tmp/test && /tmp/test
+```
+
+Test files:
+- `attn_decode_test.cu` — basic decode kernel
+- `attn_paged_decode_test.cu` — paged decode kernel
+- `attn_prefill_test.cu` — prefill kernel
+
+## Benchmarks
+
+Hardware: NVIDIA L20 (sm_89, 46 GB), CUDA 12.8, driver 570.86.
+
+Reproduce:
+```bash
+nvcc -I csrc -arch=sm_89 -O3 --use_fast_math \
+     --ptxas-options=-O3,-v --extra-device-vectorization \
+     csrc/tests/attn_<name>_test.cu -o /tmp/test && /tmp/test
+```
+
+## Known Optimization Targets
+
+- **Decode D=256**: spill eliminated (BC=16 + STAGES=2), but still 248 regs — further tiling could help.
+- **Prefill single-batch**: bandwidth low (22 GB/s at q=kv=2048) — compute-bound at ~94 TFLOP/s (near L20 bf16 ceiling ~193 TFLOP/s for non-causal).
+- **Decode single-batch**: bandwidth low (113 GB/s at kv=512, 13% of 864 GB/s theoretical) — small kv underutilizes SMs despite split-KV; scales to 757 GB/s (88%) at B=16+.
+
+## File Layout
+
+```
+csrc/
+├── build.py              # Build system: REGISTRY, _arch_flags, nvcc flags
+├── kernels/
+│   ├── attn_common.h     # Shared attention params (AttentionParams, PagedAttentionParams)
+│   ├── attn_decode.cu    # Basic decode kernel (registered)
+│   ├── attn_prefill.cu   # Basic prefill kernel (registered)
+│   ├── attn_paged_decode.cu             # Paged decode kernel (registered)
+│   ├── rotary_emb.cu                    # Fused rotary embedding kernel (registered)
+│   ├── attn_decode_split_kv.cuh         # Split-KV variant
+│   ├── attn_decode_split_kv_mma.cuh     # Split-KV + MMA variant
+│   ├── attn_prefill_split_q.cuh         # Split-Q variant
+│   ├── attn_prefill_split_q_mma.cuh     # Split-Q + MMA variant
+│   ├── attn_paged_decode_split_kv.cuh   # Paged + split-KV variant
+│   ├── attn_paged_decode_split_kv_mma.cuh  # Paged + split-KV + MMA variant
+│   ├── attn_dispatchers.cuh             # Kernel dispatch macros
+│   ├── attn_entry_utils.cuh             # Entry point helpers
+│   ├── attn_mma_utils.cuh               # MMA utilities
+│   └── attn_warp_utils.cuh              # Warp-level utilities
+└── tests/
+    ├── test_utils.cuh           # Shared test utilities
+    ├── attn_decode_test.cu      # Decode kernel test
+    ├── attn_paged_decode_test.cu # Paged decode test
+    └── attn_prefill_test.cu     # Prefill kernel test
+```
+
+Compiled `.so` files are placed in `astrai/extension/lib/`, separate from Python source files.
+
+> Document Update Time: 2026-07-31

docs/developer/dataflow.md

@@ -1,6 +1,6 @@
 # Data Flow
 
-This document describes the data pipeline: from raw text to model input tensors. For creating preprocessing configs, see [Preprocessing Guide](preprocessing.md).
+This document describes the data pipeline: from raw text to model input tensors. For creating preprocessing configs, see [Preprocessing Guide](../guides/preprocessing.md).
 
 ## Contents
 
@@ -33,7 +33,7 @@ Raw text is tokenized via `AutoTokenizer.encode()` and saved as HDF5 (`.h5`) or
 
 ### Tokenization
 
-The `Pipeline` reads JSONL lines, applies the mask builder (see [Preprocessing](preprocessing.md)), and produces flat token sequences:
+The `Pipeline` reads JSONL lines, applies the mask builder (see [Preprocessing](../guides/preprocessing.md)), and produces flat token sequences:
 
 ```python
 # Per JSONL line: messages → chat template → token IDs + loss mask
@@ -83,9 +83,11 @@ All backends normalise tensors into `Store._data[Dict[str, List[Tensor]]]` + `St
 ## Dataset Architecture
 
 ```
-DatasetFactory.load(train_type, load_path, window_size, stride=None,
+DatasetFactory.load(
+    train_type, load_path=None, window_size=0, stride=None,
     storage_type=None, tokenizer_path=None,
-                    max_len=2048, store=None)
+    max_len=2048, store=None
+)
   → BaseDataset.load(load_path, storage_type=None)
     → detect_format(load_path)
     → StoreFactory.create(storage_type)

docs/developer/internals.md

@@ -0,0 +1,233 @@
+# Internals
+
+Mathematical foundations and internal algorithms for AstrAI's training, inference, and preprocessing pipelines. For practical usage guides, see [Training](../guides/training.md), [Inference](../guides/inference.md), and [Preprocessing](../guides/preprocessing.md).
+
+## Contents
+
+- [Autoregression & Causal Masking](#autoregression--causal-masking)
+- [Rotary Position Embedding (RoPE)](#rotary-position-embedding-rope)
+- [Training Loss Formulas](#training-loss-formulas)
+- [Training Loop Internals](#training-loop-internals)
+- [Callback Lifecycle](#callback-lifecycle)
+- [KV Cache Mathematics](#kv-cache-mathematics)
+- [Mask Algorithm Internals](#mask-algorithm-internals)
+- [Gradient Accumulation Mechanics](#gradient-accumulation-mechanics)
+
+## Autoregression & Causal Masking
+
+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.
+
+```
+sequence : [[1, 2, 3, 4, 5, 6]]
+input_ids: [[1, 2, 3, 4, 5]]
+target_ids: [[2, 3, 4, 5, 6]]
+```
+
+A lower-triangular causal 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]]
+```
+
+This ensures position $i$ can only attend to positions $\leq i$, which is essential for autoregressive generation.
+
+## Rotary Position Embedding (RoPE)
+
+RoPE embeds position into Q/K vectors via complex rotation:
+
+$$ q_i = R_i W_q x_i, \quad k_j = R_j W_k x_j, \quad q_i^T k_j = x_i^T W_q^T R_{i-j} W_k x_j $$
+
+`RotaryEmbedding` pre-computes `cos_table` and `sin_table` (f32, `[max_len, dim/2]`). `forward()` returns a `(cos, sin)` tuple indexed by `position_ids`. `apply_rotary_emb` applies the rotation: during training it uses torch complex multiply (autograd-compatible); during inference it auto-dispatches to a fused CUDA kernel when available. The key property is that the dot product $q_i^T k_j$ depends only on the relative position $i - j$, not the absolute positions.
+
+**Critical for inference**: RoPE is applied **before** KV cache write, not after. If applied after caching, position encoding drift occurs because cached K/V would have stale rotation factors.
+
+## Training Loss Formulas
+
+### SEQ (Pre-training)
+
+Next-token cross-entropy with optional label smoothing:
+
+$$ L_{\text{PT}} = -\sum_{t=1}^{T} \log P(x_t \mid x_{\lt t}; \theta) $$
+
+### SFT (Supervised Fine-Tuning)
+
+Masked cross-entropy (`ignore_index=-100`) over response tokens only:
+
+$$ L_{\text{SFT}} = -\sum_{t=P+1}^{P+L} \log P(s_t \mid s_{\lt t}; \theta) $$
+
+Prompt tokens are masked out via `loss_mask`; only response tokens contribute to the loss.
+
+### 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`, `reduction="sum"`.
+
+### GRPO (Group Relative Policy Optimization)
+
+Token-level PPO with group-normalized advantages:
+
+$$ \text{Advantage}_i = \frac{r_i - \mu}{\sigma + \epsilon} $$
+
+$$ L_{\text{GRPO}} = -\mathbb{E}_t\left[\min\left(\rho_t A,\; \text{clip}\left(\rho_t, 1-\epsilon, 1+\epsilon\right)A\right)\right] + \lambda \cdot \mathbb{E}_t\left[\frac{\pi_{\text{ref}}}{\pi_\theta} - \log\frac{\pi_{\text{ref}}}{\pi_\theta} - 1\right] $$
+
+Where $\rho_t = \pi_\theta(a_t|s_t) / \pi_{\text{old}}(a_t|s_t)$ is the per-token importance sampling ratio. Advantages are derived from scalar per-response rewards, group-normalized, and broadcast across all response tokens. Only response tokens contribute to the loss.
+
+Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`.
+
+## Training Loop Internals
+
+Two-level loop: **epoch** → **batch**. Optimizer step fires every `grad_accum_steps` batches.
+
+```
+on_train_begin
+  model.train()
+  on_epoch_begin
+    for batch in dataloader:
+      on_batch_begin
+      with executor.accumulate(model):
+        loss = strategy.compute_loss(batch)
+        context.loss = loss.item()
+        stand_loss = loss / executor.grad_accum_steps
+        executor.backward(stand_loss)
+        context.consumed_samples += (
+            context.config.batch_per_device * context.world_size
+        )
+        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
+```
+
+The loss is divided by `grad_accum_steps` before `backward()`, so accumulated gradients sum to the correct mean.
+
+## Callback Lifecycle
+
+| Hook | Fires | Default callback |
+|------|-------|-----------------|
+| `on_train_begin` | Before training starts | `GradientCheckpointingCallback` |
+| `on_epoch_begin` | Start of each epoch | `ProgressBarCallback` |
+| `on_batch_begin` | Every batch | — |
+| `on_optimizer_step` | Every accumulation window | `GradientClippingCallback`, `MetricCallback`, `ProgressBarCallback` |
+| `on_batch_end` | Every batch | `CheckpointCallback` |
+| `on_epoch_end` | End of each epoch | `MetricCallback`, `ProgressBarCallback` |
+| `on_error` | On exception during training | `CheckpointCallback`, `MetricCallback` |
+| `on_train_end` | Training ends (always via finally) | `CheckpointCallback`, `MetricCallback`, `GradientCheckpointingCallback` |
+
+Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `metric` (JSONL + validation, rank-0), `progress_bar` (tqdm), `gradient_clipping` (always registered; computes grad norm, clips only when `max_grad_norm` is not `None`).
+
+## KV Cache Mathematics
+
+At decode time, only the last query token matters. All previous K/V are cached to avoid recomputation:
+
+$$ o_n = \sum_j \text{softmax}\left(\frac{q_n k_j}{\sqrt{d_k}}\right) v_j $$
+
+The cache stores $k_j$ and $v_j$ for all previous positions. At each decode step, only $q_n$ (the current query) is computed fresh, and attention is computed against the cached K/V.
+
+**RoPE ordering**: RoPE is applied to Q/K **before** writing to the KV cache. This is essential because:
+1. The cached K values already contain the rotation for their original positions.
+2. The new Q is rotated for its current position.
+3. The dot product $q_n^T k_j$ then correctly depends on $n - j$ (relative position).
+
+If RoPE were applied after caching, the rotation factors would be inconsistent between cached and new tokens.
+
+### Cache Architecture
+
+Three-layer separation (SGLang-inspired):
+
+- **KVStorage**: Flat token-level buffers `[n_layers, size, n_kv_heads, head_dim]`.
+- **ReqToTokenPool**: Index table `[req_idx, pos] → physical token slot`, shared across all layers.
+- **Allocator + PrefixCache**: Paged-mode slot allocation with ref-counting, LRU eviction, and hash-based prefix sharing.
+
+`PagePool` orchestrates all three. In contiguous mode (default), `req_to_token` is a trivial linear mapping. In paged mode, slots are allocated on demand with prefix caching support. `bind_tasks()` returns a `KVCache` dataclass with precomputed `page_table` and `decode_mask` fields (computed once per decode step, shared across all layers). Attention layers access buffers directly — no methods, no abstraction.
+
+### Attention Backend
+
+Attention computation is decoupled from the model via `AttentionBackend` ABC (`astrai/extension/attention_backend.py`):
+
+- **`TorchNativeBackend`** (default): writes K/V to cache, gathers via `req_to_token` indirect indexing, calls `F.scaled_dot_product_attention`.
+- **`CudaBackend`**: decode path uses `attn_paged_decode` with `page_size=1` (the `req_to_token` table serves as the page table, each token slot is a single-token "page"); prefill path gathers K/V then calls `attn_prefill`. Falls back to `TorchNativeBackend` when kernel unavailable.
+
+Rotary embedding is applied via `apply_rotary_emb` in `astrai/extension/rotary_backend.py`, which auto-dispatches to the fused CUDA kernel (`rotary_emb.cu`) during inference or torch complex multiply during training (for autograd compatibility). Both attention backends share the same rotary dispatch.
+
+Backend selection is thread-safe via `contextvars`, mirroring `torch.nn.attention.sdpa_kernel`:
+
+```python
+from astrai.extension import attn_backend, ATTN_BACKEND
+
+with attn_backend(ATTN_BACKEND.CUDA):
+    engine.generate("hello")
+```
+
+Layout convention: all q/k/v are `[batch, seq_len, n_heads, head_dim]` (blhd). Scale is always `1/sqrt(head_dim)`.
+
+## Mask Algorithm Internals
+
+### Template mode (`template: true`)
+
+1. Prepend BOS token (masked)
+2. For each message in the field's array:
+   1. Render through `chat_template` for that single message
+   2. Encode rendered text
+   3. Apply mask rule for the message's role
+
+### Non-template mode
+
+Encode the field value as text. Mask value is 1 (train) or 0 (mask) per the section's `action`.
+
+### Text config detection
+
+When no section uses `template` and all sections have `action: "train"`, the builder omits `loss_mask` from the output — all tokens are trained.
+
+### Position ID strategies
+
+| Mode | Behavior |
+|------|----------|
+| `none` | No position IDs generated |
+| `doc_reset` | Reset position to 0 at each document boundary in packed sequences |
+| `continuous` | Continuous position IDs across packed documents |
+
+Default is `doc_reset`, which ensures each document in a packed bin starts from position 0, preventing position encoding drift between unrelated documents.
+
+## Gradient Accumulation Mechanics
+
+Three cooperating layers enable gradient accumulation:
+
+1. **`GradientState`** — tracks the micro-step counter. Fires `sync_gradients=True` every `grad_accum_steps` micro-batches. The counter is incremented at the **start** of `accumulate()`, before the forward pass.
+
+2. **`executor._no_sync(model)`** — suppresses gradient synchronization on non-sync micro-steps:
+   - `NoneExecutor`: `nullcontext` (nothing to skip)
+   - `DDPExecutor`: `model.no_sync()` (PyTorch's built-in — skips all-reduce of gradient buckets)
+   - `FSDPExecutor`: `set_requires_gradient_sync(False, recurse=True)` on each `FSDPModule` (FSDP2's native mechanism)
+
+3. **`AccumOptimizer` / `AccumScheduler`** — wrap the real optimizer/scheduler. `step()` and `zero_grad()` are gated on `sync_gradients` — they only forward to the inner optimizer when the sync flag is True.
+
+The loss is divided by `grad_accum_steps` before `backward()`, so gradients sum to the correct mean across micro-steps. `consumed_samples` increments by `batch_per_device * world_size` every micro-batch.
+
+### Effective batch size
+
+$$ \text{Effective batch} = \text{nprocs} \times \text{batch\_per\_device} \times \text{grad\_accum\_steps} $$
+
+### Total optimizer steps
+
+```
+samples_per_replica = ceil(dataset_len / nprocs)
+batches_per_replica  = ceil(samples_per_replica / batch_per_device)
+total_steps          = (batches_per_replica // grad_accum_steps) * n_epoch
+```
+
+This accounts for data-parallel sharding — each rank processes `1/nprocs` of the dataset.
+
+> Document Update Time: 2026-07-31

docs/get-started.md

@@ -0,0 +1,235 @@
+# Getting Started
+
+This guide walks you through installing AstrAI, downloading a model, running inference, preprocessing data, and launching your first training job.
+
+## Prerequisites
+
+- **Python 3.12+**
+- **PyTorch 2.11+** (CUDA 12.8 recommended for GPU support)
+- NVIDIA GPU with CUDA (optional but recommended; CPU works for inference)
+
+## 1. Install
+
+```bash
+git clone https://github.com/ViperEkura/AstrAI.git
+cd AstrAI
+
+# Basic install (pure PyTorch, no custom CUDA kernels)
+pip install -e .
+
+# With CUDA kernels (optional, for fused attention and rotary embedding)
+# CSRC_KERNELS=true pip install -e . --no-build-isolation
+
+# With dev dependencies (pytest, ruff)
+# pip install -e ".[dev]"
+```
+
+> **CUDA kernels** are opt-in. They are not built by default. When built, they can be activated via `with attn_backend(ATTN_BACKEND.CUDA):` for accelerated decode/prefill, and the fused rotary embedding kernel is auto-dispatched when available. You can skip them for normal usage.
+
+## 2. Download Model Weights
+
+AstrAI uses HuggingFace-style model directories. Download the default 1B instruction-tuned model:
+
+```bash
+python scripts/demo/download.py
+# → Downloads to params/
+```
+
+To use a different model:
+
+```bash
+python scripts/demo/download.py --repo-id <HF_REPO_ID> --local-dir ./my_model
+```
+
+The model directory contains:
+- `config.json` — model architecture configuration
+- `model.safetensors` — model weights
+- `tokenizer.json` + `tokenizer_config.json` — tokenizer files (including chat template)
+
+## 3. Run Inference
+
+### Interactive Chat (Simplest)
+
+```bash
+python scripts/demo/stream_chat.py
+# Type your message after >>, type !exit to quit
+```
+
+This starts a multi-turn interactive chat session with streaming output.
+
+### Start an HTTP Server
+
+```bash
+# Terminal 1: start server
+python scripts/tools/server.py --param_path ./params --device cuda
+
+# Terminal 2: query (OpenAI-compatible API)
+curl -X POST http://localhost:8000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{"messages":[{"role":"user","content":"Hello"}],"max_tokens":512}'
+```
+
+The server also supports the Anthropic API at `/v1/messages`. See [Inference Guide](guides/inference.md) for full API documentation.
+
+### Batch Generation from a File
+
+Create an input JSONL file (one JSON object per line):
+
+```json
+{"question": "What is machine learning?"}
+{"question": "Explain gradient descent."}
+```
+
+```bash
+python scripts/tools/generate.py \
+    --param_path ./params \
+    --input_json_file input.jsonl \
+    --output_json_file output.jsonl
+```
+
+## 4. Preprocess Data
+
+AstrAI uses a declarative JSON config to define the preprocessing pipeline. Create a config file for your training type:
+
+### Pretraining (seq)
+
+Input JSONL:
+```json
+{"text": "Artificial intelligence is..."}
+```
+
+Config (`pretrain.json`):
+```json
+{
+    "input": {
+        "sections": [{"field": "text", "action": "train"}]
+    },
+    "preprocessing": {"max_seq_len": 2048},
+    "output": {"storage_format": "bin"}
+}
+```
+
+### SFT (Supervised Fine-Tuning)
+
+Input JSONL:
+```json
+{"messages": [{"role": "user", "content": "Hi"}, {"role": "assistant", "content": "Hello!"}]}
+```
+
+Config (`sft.json`):
+```json
+{
+    "input": {
+        "sections": [{"field": "messages", "action": "$role", "template": true}]
+    },
+    "mask": {
+        "system": "mask",
+        "user": "mask",
+        "assistant": "train"
+    },
+    "mask_default": "mask",
+    "preprocessing": {"max_seq_len": 2048},
+    "output": {"storage_format": "bin", "dtype": {"loss_mask": "bool"}}
+}
+```
+
+### Run Preprocessing
+
+```bash
+python scripts/tools/preprocess.py data/*.jsonl -o output/ -c pretrain.json
+```
+
+See [Preprocessing Guide](guides/preprocessing.md) for DPO/GRPO configs and all options.
+
+## 5. Train
+
+### Single GPU
+
+```bash
+python scripts/tools/train.py \
+    --train_type=seq \
+    --data_root_path=/path/to/dataset \
+    --param_path=./params \
+    --batch_per_device=4 \
+    --grad_accum_steps=8 \
+    --max_lr=1e-4 \
+    --window_size=2048 \
+    --ckpt_dir=./checkpoint \
+    --nprocs=1 \
+    --parallel_mode=none
+```
+
+### Multi-GPU (DDP)
+
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+export NCCL_P2P_DISABLE=1
+export NCCL_NET_GDR_LEVEL=0
+
+python scripts/tools/train.py \
+    --train_type=seq \
+    --data_root_path=/path/to/dataset \
+    --param_path=./params \
+    --parallel_mode=ddp \
+    --nprocs=4 \
+    --batch_per_device=4 \
+    --grad_accum_steps=8 \
+    --max_lr=1e-4 \
+    --window_size=2048 \
+    --ckpt_dir=./checkpoint
+```
+
+### Training Types
+
+| `--train_type` | Description | Data Keys |
+|----------------|-------------|-----------|
+| `seq` | Pre-training (next-token prediction) | `sequence` |
+| `sft` | Supervised fine-tuning (masked loss) | `sequence`, `loss_mask` |
+| `dpo` | Direct Preference Optimization | `chosen`, `rejected`, `*_mask` |
+| `grpo` | Group Relative Policy Optimization | `prompts`, `responses`, `masks`, `rewards` |
+
+See [Training Guide](guides/training.md) for loss formulas and strategies. See [Distributed Guide](guides/distributed.md) for DDP/FSDP details.
+
+## 6. Evaluate
+
+```bash
+# HumanEval (code generation, auto-downloads dataset)
+python scripts/eval/evaluate_humaneval.py --param_path ./params --num_samples 20
+
+# MMLU (knowledge, auto-downloads dataset)
+python scripts/eval/evaluate_mmlu.py --param_path ./params --n_shot 5
+
+# Perplexity on custom data
+python scripts/eval/evaluate_ppl.py --param_path ./params --input_path data.jsonl --output_dir ppl_results/
+```
+
+See [Evaluation Guide](guides/evaluation.md) for all benchmarks.
+
+## 7. Docker
+
+```bash
+# Build
+docker build -t astrai:latest .
+
+# Run inference server with GPU
+docker run --gpus all -p 8000:8000 astrai:latest \
+  python -m scripts.tools.server --port 8000 --device cuda
+
+# Docker Compose (GPU)
+docker compose up -d
+```
+
+## Next Steps
+
+| Topic | Document |
+|-------|----------|
+| CLI parameters (train, server, generate, preprocess) | [CLI Reference](guides/params.md) |
+| Preprocessing pipeline details | [Preprocessing Guide](guides/preprocessing.md) |
+| Training loop, strategies, schedulers | [Training Guide](guides/training.md) |
+| KV cache, continuous batching, HTTP API | [Inference Guide](guides/inference.md) |
+| Evaluation benchmarks | [Evaluation Guide](guides/evaluation.md) |
+| Multi-GPU DDP / FSDP | [Distributed Guide](guides/distributed.md) |
+| System architecture | [Architecture](developer/architecture.md) |
+| Data pipeline internals | [Data Flow](developer/dataflow.md) |
+
+> Document Update Time: 2026-07-31

docs/guides/distributed.md

@@ -0,0 +1,255 @@
+# Distributed Training
+
+AstrAI supports three parallel modes: **single GPU** (`none`), **Data Parallel** (`ddp`), and **Fully Sharded Data Parallel** (`fsdp`). This guide covers when to use each, how to launch multi-GPU training, and how gradient accumulation works.
+
+## Quick Start
+
+### Single GPU
+
+```bash
+python scripts/tools/train.py \
+    --train_type=sft \
+    --param_path ./params \
+    --data_root_path ./dataset \
+    --parallel_mode=none \
+    --nprocs=1 \
+    --batch_per_device=4 \
+    --grad_accum_steps=8
+```
+
+### Multi-GPU DDP (4 GPUs)
+
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+export NCCL_P2P_DISABLE=1
+export NCCL_NET_GDR_LEVEL=0
+
+python scripts/tools/train.py \
+    --train_type=sft \
+    --param_path ./params \
+    --data_root_path ./dataset \
+    --parallel_mode=ddp \
+    --nprocs=4 \
+    --batch_per_device=4 \
+    --grad_accum_steps=8
+```
+
+### Multi-GPU FSDP (4 GPUs)
+
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+export NCCL_P2P_DISABLE=1
+export NCCL_NET_GDR_LEVEL=0
+
+python scripts/tools/train.py \
+    --train_type=sft \
+    --param_path ./params \
+    --data_root_path ./dataset \
+    --parallel_mode=fsdp \
+    --nprocs=4 \
+    --batch_per_device=4 \
+    --grad_accum_steps=8
+```
+
+> `--parallel_mode` defaults to `fsdp`. You can omit it for FSDP.
+
+## Parallel Modes
+
+| Mode | `--parallel_mode` | Param Layout | Memory | When to Use |
+|------|-------------------|--------------|--------|-------------|
+| Single GPU | `none` | Full, replicated | Highest | Small models, DPO/GRPO, debugging |
+| DDP | `ddp` | Full, replicated | High | Most multi-GPU training |
+| FSDP | `fsdp` | Sharded (DTensor) | Lowest | Large models that don't fit in single GPU |
+
+### NoneExecutor
+
+No wrapping. The model runs as-is on a single device. Gradient accumulation still works via `AccumOptimizer`/`AccumScheduler` (they gate `step()` on the sync counter). Checkpoint saving is a plain `state_dict()` call.
+
+### DDPExecutor
+
+Wraps the model with `torch.nn.parallel.DistributedDataParallel`. Each rank has a full copy of the model; gradients are all-reduced across ranks. Uses `gradient_as_bucket_view=True` and `broadcast_buffers=False` by default (hardcoded in `train.py`).
+
+During gradient accumulation, non-sync micro-steps use `model.no_sync()` to skip gradient all-reduce. Only the final micro-step triggers the all-reduce.
+
+### FSDPExecutor (FSDP2 / `fully_shard`)
+
+Uses PyTorch's FSDP2 per-module API (`torch.distributed.fsdp.fully_shard`). Each model child (e.g., each `DecoderBlock`) is individually sharded — parameters become `DTensor`s distributed across ranks. No `FlatParameter`, original parameter names are preserved.
+
+Key differences from DDP:
+- **Lower memory**: parameters are sharded, not replicated.
+- **Custom grad norm**: FSDP gradients are `DTensor`s, so `clip_grad_norm` computes the local norm, then all-reduces to get the global norm.
+- **Collective checkpoint ops**: `unshard()` and `full_tensor()` are collective — all ranks must call them even though only rank-0 saves. The executor handles this via `dist.barrier()` in `checkpoint_context`.
+- **Root skipped**: `fully_shard` is applied to direct children only (not the root model) due to an `ABC + Generic[T]` MRO incompatibility.
+
+## Gradient Accumulation
+
+Gradient accumulation lets you simulate a larger effective batch size by accumulating gradients over multiple micro-batches before calling `optimizer.step()`.
+
+```
+Effective batch = nprocs × batch_per_device × grad_accum_steps
+```
+
+Example: 4 GPUs × batch 4 × accum 8 = effective batch 256.
+
+### How it works
+
+Three cooperating layers:
+
+1. **`GradientState`** — tracks the micro-step counter. Fires `sync_gradients=True` every `grad_accum_steps` micro-batches.
+2. **`executor._no_sync(model)`** — suppresses gradient synchronization on non-sync micro-steps:
+   - `none`: `nullcontext` (nothing to skip)
+   - `ddp`: `model.no_sync()` (skips all-reduce)
+   - `fsdp`: `set_requires_gradient_sync(False)` on each `FSDPModule`
+3. **`AccumOptimizer` / `AccumScheduler`** — gate `step()` and `zero_grad()` on `sync_gradients`, so the optimizer only fires on the last micro-step.
+
+The loss is divided by `grad_accum_steps` before `backward()`, so gradients sum to the correct mean.
+
+## Process Launching
+
+AstrAI auto-detects the launch method:
+
+| Detection | Strategy | Use Case |
+|-----------|----------|----------|
+| `torchelastic` / `torchrun` env vars | `TorchrunStrategy` | External orchestrator (torchrun, SLURM, K8s) |
+| `RANK` + `WORLD_SIZE` env vars | `TorchrunStrategy` | External launch |
+| Neither | `LocalStrategy` | `python scripts/tools/train.py` (in-process spawn) |
+
+### Local (default)
+
+When you run `python scripts/tools/train.py --nprocs=4`, AstrAI uses `torch.multiprocessing.start_processes` to spawn 4 child processes. The parent process manages signal forwarding (SIGTERM/SIGINT) and waits for all children to finish.
+
+### Torchrun
+
+For multi-node or SLURM environments:
+
+```bash
+torchrun --nproc_per_node=4 scripts/tools/train.py \
+    --train_type=sft \
+    --parallel_mode=ddp \
+    --param_path ./params \
+    --data_root_path ./dataset \
+    --batch_per_device=4
+```
+
+When launched via torchrun, AstrAI reads `RANK`, `WORLD_SIZE`, `LOCAL_RANK` from the environment and uses `TorchrunStrategy`. The `--nprocs` flag is ignored (the orchestrator controls process count).
+
+## NCCL Environment Variables
+
+For multi-GPU training, you **must** set these environment variables:
+
+```bash
+export NCCL_P2P_DISABLE=1
+export NCCL_NET_GDR_LEVEL=0
+```
+
+These are required on certain hardware configurations (see `AGENTS.md`). Without them, NCCL may hang or crash during collective operations. These are set in the training shell scripts (`train-seq.sh`, `train-sft.sh`, `train-dpo.sh`) but not in Python code — you must export them before launching.
+
+## Checkpoint Saving
+
+Checkpoints are saved by **rank-0 only**. The flow:
+
+1. `executor.checkpoint_context(model)` — wraps with `dist.barrier()` before and after (distributed only).
+2. `executor.unwrap_model(model)` — gathers the full state dict:
+   - `none`: `model.state_dict()`
+   - `ddp`: `model.module.state_dict()`
+   - `fsdp`: `unshard()` → `full_tensor()` → `reshard()` (collective on all ranks, result kept only on rank-0)
+3. Non-rank-0 ranks get `None` — the save is skipped.
+4. Rank-0 writes `meta.json`, `config.json`, `model.safetensors`, and optional `{key}.pt` (optimizer/scheduler state).
+
+> **FSDP note**: Even though only rank-0 saves, all ranks must participate in `unwrap_model` because `unshard()` and `full_tensor()` are collective operations. The barriers in `checkpoint_context` keep all ranks in lockstep.
+
+## Total Steps Calculation
+
+The scheduler's total step count accounts for data-parallel sharding:
+
+```
+samples_per_replica = ceil(dataset_len / nprocs)
+batches_per_replica  = ceil(samples_per_replica / batch_per_device)
+total_steps          = (batches_per_replica // grad_accum_steps) * n_epoch
+```
+
+This ensures the LR schedule is correctly scaled regardless of the number of GPUs.
+
+## Real Examples
+
+### Pretraining (seq, DDP, 4 GPUs)
+
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+export NCCL_P2P_DISABLE=1
+export NCCL_NET_GDR_LEVEL=0
+
+python scripts/tools/train.py \
+    --train_type=seq \
+    --param_path ./params \
+    --data_root_path ./dataset/cached \
+    --parallel_mode=ddp \
+    --nprocs=4 \
+    --n_epoch=1 \
+    --max_lr=2e-4 \
+    --schedule_type=wsd \
+    --warmup_ratio=0.02 \
+    --window_size=2048 \
+    --batch_per_device=4 \
+    --grad_accum_steps=32 \
+    --ckpt_interval=2000
+# Effective batch = 4 × 4 × 32 = 512
+```
+
+### SFT (DDP, 4 GPUs)
+
+```bash
+python scripts/tools/train.py \
+    --train_type=sft \
+    --param_path ./AstrAI-V1-base \
+    --data_root_path ./dataset/cached_sft \
+    --parallel_mode=ddp \
+    --nprocs=4 \
+    --n_epoch=2 \
+    --max_lr=2e-5 \
+    --schedule_type=cosine \
+    --warmup_ratio=0.02 \
+    --min_rate=0.05 \
+    --window_size=2048 \
+    --batch_per_device=4 \
+    --grad_accum_steps=8
+# Effective batch = 4 × 4 × 8 = 128
+```
+
+### DPO (Single GPU)
+
+```bash
+python scripts/tools/train.py \
+    --train_type=dpo \
+    --param_path ./checkpoint/epoch_1_step_6000 \
+    --data_root_path ./alpaca_dpo.jsonl \
+    --parallel_mode=none \
+    --nprocs=1 \
+    --max_lr=5e-6 \
+    --schedule_type=cosine \
+    --warmup_ratio=0.1 \
+    --min_rate=0.1 \
+    --window_size=1024 \
+    --batch_per_device=4 \
+    --grad_accum_steps=8 \
+    --dpo_beta=0.1 \
+    --max_grad_norm=50
+```
+
+## CLI Parameters
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--nprocs` | 1 | Number of GPUs / processes |
+| `--parallel_mode` | `fsdp` | `none`, `ddp`, or `fsdp` |
+| `--start_method` | `spawn` | Multiprocessing start method (`spawn`, `fork`, `forkserver`) |
+| `--backend` | `nccl` | Distributed backend (`nccl`, `gloo`) |
+| `--master_addr` | `localhost` | Master node address |
+| `--master_port` | `29500` | Master node port |
+| `--device_type` | `cuda` | Device type |
+
+> `--tp_size` is parsed but **not yet wired** — tensor parallelism is future work. `ColumnParallelLinear` / `RowParallelLinear` exist in `astrai/parallel/module.py` but are not used by the model.
+
+Full parameter reference: [CLI Reference](params.md). Training loop and strategies: [Training Guide](training.md).
+
+> Document Update Time: 2026-07-30

docs/guides/evaluation.md

@@ -0,0 +1,252 @@
+# Evaluation
+
+AstrAI provides 7 evaluation scripts in `scripts/eval/` covering code generation, knowledge QA, perplexity, summarization, data quality, instruction following, and weight analysis.
+
+## Overview
+
+| Script | Metric | Model Invocation | External Dataset |
+|--------|--------|-------------------|-------------------|
+| `evaluate_humaneval.py` | Code-gen pass@1/10/100 | `InferenceEngine.generate` | HF `openai/openai_humaneval` (auto-download) |
+| `evaluate_mmlu.py` | MCQ accuracy (log-likelihood) | Direct `model()` forward | HF `cais/mmlu` (auto-download) |
+| `evaluate_ppl.py` | Perplexity / token loss | Direct `model()` forward | User JSONL |
+| `evaluate_rouge.py` | ROUGE-1/2/L | None (pure metric) | User JSONL |
+| `evaluate_ifd.py` | Instruction-Following Difficulty | Direct `model()` forward | User JSONL |
+| `evaluate_ifeval.py` | Instruction-following constraints | `InferenceEngine.generate` | HF `google/IFEval` (auto-download) |
+| `analyze_weights.py` | SVD effective rank / weight stats | None (loads safetensors) | Checkpoint dir |
+
+Two invocation patterns exist:
+- **Generation benchmarks** (HumanEval, IFEval): use `InferenceEngine` to generate responses, then score them.
+- **Scoring benchmarks** (MMLU, PPL, IFD): call `model()` directly under `torch.inference_mode()` for log-likelihood computation.
+
+Common defaults: `--param_path` defaults to `./params`; dtype defaults to `bfloat16` on CUDA, `float32` on CPU.
+
+---
+
+## HumanEval (Code Generation)
+
+Generates completions for 164 programming problems, executes them against hidden tests, and reports pass@k.
+
+```bash
+python scripts/eval/evaluate_humaneval.py \
+    --param_path ./params \
+    --num_samples 20 \
+    --batch_size 32 \
+    --max_tokens 512 \
+    --output results/humaneval.json
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--param_path` | `./params` | Model directory |
+| `--data_path` | `./humaneval/HumanEval.jsonl` | HumanEval JSONL (auto-downloaded if missing) |
+| `--output` | None | Save results JSON (also writes `_completions.json`) |
+| `--test_only` | None | Test an existing completions JSON (skip generation) |
+| `--generate_only` | False | Only generate, skip execution/testing |
+| `--num_samples` | 200 | Completions per problem (pass@k needs >= k) |
+| `--max_tokens` | 512 | Max generation length |
+| `--temperature` | 0.8 | Sampling temperature |
+| `--top_p` | 0.95 | Nucleus sampling threshold |
+| `--top_k` | 50 | Top-k sampling |
+| `--batch_size` | 32 | Generation batch size |
+| `--test_workers` | 8 | ProcessPoolExecutor workers for test execution |
+| `--test_timeout` | 3.0 | Per-subprocess timeout (seconds) |
+| `--problems` | None | Restrict to specific problem indices |
+
+**Output**: stdout prints `pass@1`, `pass@10`, `pass@100`. With `--output`, writes per-problem results + `_summary` aggregate and a `_completions.json` file.
+
+**Data**: Auto-downloads `openai/openai_humaneval` from HuggingFace on first run. Each problem has `task_id`, `entry_point`, `prompt`, `test`.
+
+---
+
+## MMLU (Knowledge QA)
+
+57-subject multiple-choice accuracy via log-likelihood comparison. Supports n-shot few-shot prompting and option permutation.
+
+```bash
+python scripts/eval/evaluate_mmlu.py \
+    --param_path ./params \
+    --n_shot 5 \
+    --subjects math_algebra history_us \
+    --output results/mmlu.json
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--param_path` | `./params` | Model directory |
+| `--data_dir` | `./mmlu_data` | MMLU data directory (per-subject CSVs) |
+| `--download` | False | Force re-download |
+| `--n_shot` | 5 | Few-shot examples (0 = zero-shot) |
+| `--subjects` | all 57 | Specific subjects to evaluate |
+| `--output` | None | Output JSON path |
+| `--split` | `test` | `test` or `val` |
+| `--device` | auto | Device (`cuda` / `cpu`) |
+| `--dtype` | auto | `bfloat16` on CUDA, `float32` on CPU |
+| `--seed` | 0 | Seed for option permutation (0 = enabled, -1 = disabled) |
+
+**How it works**: For each question, builds a prompt with n-shot examples, then scores each choice (A/B/C/D) by computing the summed log-likelihood of the choice token given the context. The choice with the highest log-prob is the prediction.
+
+**Output**: stdout prints per-subject accuracy and overall. With `--output`, writes per-subject `{accuracy, correct, total}` + `_overall` aggregate.
+
+**Data**: Auto-downloads `cais/mmlu` from HuggingFace. Stored as per-subject CSVs in `<data_dir>/<split>/` and `<data_dir>/dev/` (for few-shot).
+
+---
+
+## Perplexity (PPL)
+
+Token-level negative-log-likelihood and perplexity on arbitrary text data. Supports streaming mode (memory-efficient) and non-streaming mode (exact per-token stats).
+
+```bash
+python scripts/eval/evaluate_ppl.py \
+    --param_path ./params \
+    --input_path data.jsonl \
+    --output_dir ppl_results/ \
+    --batch_size 4 \
+    --max_length 2048
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--param_path` | required | Model directory |
+| `--input_path` | required | Input file, glob, or directory |
+| `--output_dir` | required | Output directory for `summary.json` + token JSONL |
+| `--text_key` | `text` | Key for the text field in input data |
+| `--batch_size` | 4 | Batch size |
+| `--max_length` | 2048 | Max sequence length (tokens) |
+| `--token_level` | False | Store per-token log_probs + token-type analysis |
+| `--max_samples` | None | Random subsample per file |
+| `--device` | auto | Device |
+| `--dtype` | auto | Torch dtype |
+
+**Input**: JSONL or JSON files. Each item must have a field named by `--text_key` (default `text`). If `--input_path` is a directory, recursively collects `*.jsonl` and `*.json`.
+
+**Output**: `summary.json` with per-file stats (tokens, mean/median loss, perplexity, p50/p90/p95/p99). With `--token_level`, also writes per-token JSONL with token IDs and log-probs.
+
+---
+
+## ROUGE
+
+ROUGE-1/2/L (precision, recall, F1) for summarization. Self-contained implementation with no external dependencies.
+
+```bash
+python scripts/eval/evaluate_rouge.py \
+    --data_path predictions.jsonl \
+    --output results/rouge.json
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--data_path` | required | JSONL with `reference`/`candidate` per line |
+| `--output` | None | Output JSON path |
+
+**Input**: JSONL, one object per line:
+```json
+{"reference": "Ground truth text", "candidate": "Model output text"}
+```
+
+**Output**: stdout prints `rouge-1`, `rouge-2`, `rouge-l` each as P/R/F1. With `--output`, writes JSON with `aggregate` and `per_item` scores.
+
+Can also be imported as a library:
+```python
+from scripts.eval.evaluate_rouge import compute_rouge
+scores = compute_rouge(reference, candidate)
+```
+
+---
+
+## IFD (Instruction-Following Difficulty)
+
+Data quality metric: `IFD = L_conditional / L_unconditional`. Measures how much harder it is to predict a response given its instruction vs. without it. Useful for filtering instruction-tuning data.
+
+```bash
+python scripts/eval/evaluate_ifd.py \
+    --param_path ./params \
+    --input_path sft_data.jsonl \
+    --output_dir ifd_results/ \
+    --format messages \
+    --batch_size 8
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--param_path` | required | Model directory |
+| `--input_path` | required | Input file, glob, or directory |
+| `--output_dir` | required | Output directory |
+| `--max_len` | 2048 | Max token length |
+| `--format` | `plain` | `plain` (instruction/response fields) or `messages` (chat format) |
+| `--instr_key` | `instruction` | Instruction field key (plain format) |
+| `--resp_key` | `response` | Response field key (plain format) |
+| `--batch_size` | 8 | Items per model-forward flush |
+| `--device` | auto | Device |
+| `--dtype` | auto | Torch dtype |
+| `--sentinel_text` | `\n` | Prefix for unconditional pass (`""` → bos/pad fallback) |
+| `--per_token` | False | Include per-token IFD breakdown |
+| `--max_samples` | None | Random subsample per file |
+
+**How it works**: Two forward passes per batch — (1) conditional: packed BFD sequence with context + response, (2) unconditional: response prefixed with a sentinel. IFD = mean_conditional_loss / mean_unconditional_loss. IFD > 1 means the instruction makes the response harder to predict (higher quality data).
+
+**Output**: Per-file `<label>_ifd.jsonl` with IFD scores per item. `summary.json` aggregates per-file stats.
+
+---
+
+## IFEval (Instruction Following)
+
+Google's IFEval benchmark: generates responses and verifies 27 types of constraints (keywords, format, length, case, punctuation, etc.).
+
+```bash
+python scripts/eval/evaluate_ifeval.py \
+    --param_path ./params \
+    --num_samples 1 \
+    --max_tokens 512 \
+    --output results/ifeval.json
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--param_path` | `./params` | Model directory |
+| `--data_path` | `./ifeval/input_data.jsonl` | IFEval JSONL (auto-downloaded if missing) |
+| `--output` | None | Output JSON path |
+| `--max_tokens` | 512 | Max generation tokens |
+| `--temperature` | 0.1 | Sampling temperature (low for instruction-following) |
+| `--top_p` | 0.95 | Top-p sampling |
+| `--top_k` | 50 | Top-k sampling |
+| `--num_samples` | 1 | Samples per problem (best-of-n scoring) |
+| `--batch_size` | 1 | Inference batch size |
+| `--limit` | None | Limit to first N problems (quick testing) |
+| `--dump_responses` | None | Path to dump raw responses as JSONL |
+
+**Output**: stdout prints overall accuracy + per-constraint-type accuracy table. With `--output`, writes per-problem results + `_summary`.
+
+**Data**: Auto-downloads `google/IFEval` from HuggingFace. Each problem has `key`, `prompt`, `instruction_id_list`, `kwargs`.
+
+---
+
+## Weight Analysis
+
+SVD-based effective rank and weight statistics for checkpoint diagnostics. Does not load the model graph or run any forward pass.
+
+```bash
+python scripts/eval/analyze_weights.py \
+    --ckpt_dir ./checkpoint/epoch_1_step_6000 \
+    --output results/weights.json
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `--ckpt_dir` | required | Checkpoint dir with `model.safetensors` + `config.json` |
+| `--compare` | None | Additional checkpoint dirs to compare |
+| `--no_svd` | False | Skip SVD; show only weight stats (faster) |
+| `--output` | None | Save results as JSON |
+| `--device` | `cuda` | Device for SVD |
+
+**Output**: SVD effective rank by component (ER@90/95/99%, entropic rank, condition number), per-layer effective rank grid, and weight value statistics (mean/std/min/max). Provides a utilization verdict (HIGH >0.85 / MODERATE >0.5 / LOW).
+
+---
+
+## Tips
+
+- **Quick test**: Use `--limit` (IFEval) or `--problems` (HumanEval) to run on a small subset first.
+- **Auto-download**: HumanEval, MMLU, and IFEval auto-download their datasets on first run. The other scripts expect user-provided data.
+- **Output formats**: `--output` writes a single JSON for most scripts. PPL and IFD write an `--output_dir` containing `summary.json` plus per-file artifacts.
+- **CPU mode**: All scripts auto-detect CUDA. To force CPU, use `--device cpu --dtype float32`.
+
+> Document Update Time: 2026-07-30

docs/guides/inference.md

@@ -4,6 +4,7 @@
 
 - [KV Cache](#kv-cache)
 - [KVCache System](#kvcache-system)
+- [Attention Backend](#attention-backend)
 - [Continuous Batching](#continuous-batching)
 - [Sampling](#sampling-strategy-pattern)
 - [Protocol Handlers](#protocol-handlers-strategy-pattern)
@@ -23,30 +24,70 @@ RoPE is applied **before** KV cache write, not after — otherwise position enco
 
 ## KVCache System
 
-Seven classes working together, with two concrete cache implementations:
-
-### ContiguousCache (default)
+Three-layer separation (SGLang-inspired): storage, index table, allocator.
 
 ```
-ContiguousCache (simple contiguous per-slot cache)
-  ├── ContiguousCacheView  bundles k/v tensors + slot indices for attention layers
+PagePool (top-level manager, orchestrates all layers)
+  ├── KVStorage              k_buffer / v_buffer [n_layers, size, n_kv_heads, head_dim]
+  ├── ReqToTokenPool         req_to_token [num_reqs, max_ctx_len] → physical token slot
+  ├── Allocator              bitmask-based page allocator + ref-count + LRU (paged mode only)
+  └── PrefixCache            hash-based prefix matching (paged mode only)
 ```
 
-Created by default when no cache is passed to `InferenceScheduler`. Each task occupies a fixed slot of `[max_seq_len, n_kv_heads, head_dim]`. Simple and efficient for small-to-medium batch sizes.
+`PagePool` supports two modes:
 
-### PageCache (paged with prefix sharing)
+- **Contiguous (default)**: pre-allocates `max_batch_size * max_seq_len` token slots. `req_to_token` is a trivial linear mapping (`slot = req_idx * max_seq_len + pos`). No dynamic allocation.
+- **Paged** (`page_size=1` or `>1` with `n_tokens` set): shared token pool with on-demand allocation. Allocator + PrefixCache enable prefix sharing and LRU eviction.
+
+`bind_tasks()` returns a `KVCache` dataclass — pure data, no methods:
 
 ```
-PageCache (paged KV cache with prefix sharing, alternative)
-  ├── PagePool               orchestrates page allocation + prefix matching
-  │     ├── Allocator         bitmask-based page allocator + ref-count + LRU
-  │     └── PrefixCache       hash-based prefix matching (page_hash via polynomial hash)
-  ├── 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)
-  └── PageCacheView          bundles Storage + page_table + total_len for attention layers
+KVCache
+  ├── k_buffer, v_buffer     [n_layers, size, n_kv_heads, head_dim]
+  ├── req_to_token           [num_reqs, max_ctx_len]
+  ├── req_pool_indices       [batch_size]
+  ├── seq_lens               [batch_size]
+  ├── out_cache_loc          [batch, seq_len] — write indices for this forward
+  ├── max_len                int — max(seq_lens), avoids GPU sync in decode
+  ├── page_table             [batch, max_len] — precomputed gather indices for decode (None for prefill)
+  └── decode_mask            [batch, max_len] bool — precomputed position validity mask (None for single-batch decode)
 ```
 
-`isinstance(cache, KVCache)` checks dispatch to the correct view. Both implement the abstract `KVCache` interface used by `Executor` and `InferenceScheduler`.
+Attention layers do raw buffer indexing: `k_buffer[layer_id, out_cache_loc] = k` to write, `k_buffer[layer_id, indices]` to gather.
+
+## Attention Backend
+
+Attention computation (cache I/O + SDPA/kernel dispatch) is decoupled from the model via `AttentionBackend` ABC:
+
+```
+AttentionBackend (ABC)
+  ├── TorchNativeBackend   SDPA + indirect KV cache gather (default)
+  └── CudaBackend          CUDA kernel dispatch (attn_paged_decode, attn_prefill)
+```
+
+Select via context manager (mirrors `torch.nn.attention.sdpa_kernel`):
+
+```python
+from astrai.extension import attn_backend, ATTN_BACKEND
+
+with attn_backend(ATTN_BACKEND.CUDA):
+    engine.generate("hello")
+```
+
+`CudaBackend` decode path: writes K/V to cache, then calls `attn_paged_decode` with `page_size=1` — the `req_to_token` table serves directly as the page table, each token slot is a single-token "page". No explicit K/V gather needed.
+
+`CudaBackend` prefill path: writes K/V, gathers full-sequence K/V via indirect indexing (same as `TorchNativeBackend`), then calls `attn_prefill`.
+
+Fallback: `CudaBackend` delegates to `TorchNativeBackend` when a CUDA kernel is not available.
+
+### Rotary Embedding Backend
+
+Rotary embedding is applied via `apply_rotary_emb` in `astrai/extension/rotary_backend.py`, which auto-dispatches:
+
+- **CUDA kernel** (`rotary_emb.cu`): fused cos/sin lookup + rotation in a single kernel, used when the kernel is available, input is on CUDA, and `torch.is_grad_enabled()` is `False` (inference mode)
+- **Torch fallback**: complex multiply path (`torch.view_as_complex` → `torch.complex` multiply → `torch.view_as_real`), used during training (supports autograd backward) or when the CUDA kernel is not available
+
+`RotaryEmbedding` stores `cos_table`/`sin_table` as f32 buffers and returns a `(cos, sin)` tuple from `forward()`. Both attention backends share the same rotary dispatch — it is backend-agnostic.
 
 ## Continuous Batching
 
@@ -154,6 +195,10 @@ Supports `stop_sequences` and streaming via `event: content_block_delta`.
 | `temperature` | float | 1.0 | Sampling temperature (> 0.0) |
 | `max_tokens` | Optional[int] | None | Max generation length |
 | `stream` | bool | False | Stream output |
+| `stop` | Optional[Union[str, List[str]]] | None | Stop sequences |
+| `frequency_penalty` | float | 0.0 | Frequency penalty |
+| `tools` | Optional[List[dict]] | None | Tool definitions for function calling |
+| `tool_choice` | Optional[str] | None | Tool selection mode |
 
 ### SSE Streaming Format
 
@@ -249,4 +294,4 @@ async for token in engine.generate_async("Hello", ...):    # -> AsyncGenerator[s
     print(token)
 ```
 
-> Document Update Time: 2026-07-09
+> Document Update Time: 2026-07-31

docs/guides/params.md

@@ -13,7 +13,7 @@
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--train_type` | Training type (`seq`, `sft`, `dpo`, `grpo`) | required |
+| `--train_type` | Training type (`seq`, `sft`, `dpo`, `grpo`, `online_grpo`, `online_dpo`) | required |
 | `--data_root_path` | Dataset root directory | required |
 | `--param_path` | Model parameters or checkpoint path | required |
 | `--n_epoch` | Total training epochs | 1 |
@@ -26,25 +26,55 @@
 |-----------|-------------|---------|
 | `--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 (None disables) | None |
+| `--max_grad_norm` | Maximum gradient norm for clipping (None disables) | 1.0 |
 
-### Optimizer (MuonMix)
+### Optimizer
 
-Combined optimizer: matrix parameters via **Muon**, non-matrix via **AdamW** (`fused=True`).
+The default `muon_adamw` optimizer sends matrix parameters through **Muon** and
+non-matrix parameters through **AdamW** (`fused=True`).
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
+| `--optimizer` | Built-in optimizer (`muon_adamw`, `nora_nadamw`, `mano_adamw`) | `muon_adamw` |
 | `--weight_decay` | Weight decay (applied to Muon matrix params; non-matrix use 0) | 0.1 |
 | `--muon_momentum` | Muon momentum factor | 0.95 |
 | `--muon_nesterov` | Enable Nesterov momentum for Muon | True |
 | `--muon_ns_steps` | Newton-Schulz iteration steps for Muon | 5 |
 | `--muon_adjust_lr` | Muon LR adjustment strategy (`original`, `match_rms_adamw`) | `match_rms_adamw` |
 
+`nora_nadamw` routes internal `Linear.weight` matrices to **Nora** and
+embeddings, the LM head, norms, biases, LoRA factors, and fallback parameters to
+**NAdamW**. Parameters are classified by module role and identity, so tied
+embedding/head weights occur in exactly one group. Nora requires complete rows
+under DTensor sharding and rejects layouts sharded along the last dimension.
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--nora_lr` | Nora learning rate | 5e-3 |
+| `--nora_beta` | Nora momentum-buffer EMA factor | 0.95 |
+| `--nora_momentum` | Nora Nesterov interpolation factor | 0.95 |
+| `--nora_weight_decay` | Nora matrix weight decay | 0.0 |
+
+`mano_adamw` routes internal `Linear.weight` matrices to **Mano** (manifold
+normalized optimizer) and the remaining parameters to **NAdamW**. Mano projects
+the momentum onto the tangent space of the Oblique manifold and normalizes it,
+alternating the projection axis (row/column) each step — replacing Muon's
+Newton-Schulz iteration with a cheaper normalization.
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--mano_momentum` | Mano momentum factor | 0.95 |
+| `--mano_nesterov` | Enable Nesterov momentum for Mano | True |
+
+Optimizer identity and hyperparameters are saved in checkpoint metadata. Optimizer
+states are intentionally not interchangeable: resume older MuonAdamW checkpoints
+with `--optimizer=muon_adamw`.
+
 ### Data Loading
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--window_size` | Max input sequence length | model config `max_len` |
+| `--window_size` | Max input sequence length | model config `max_position_embeddings` |
 | `--stride` | Stride for sliding window over sequences | None |
 | `--random_seed` | Random seed for reproducibility | 3407 |
 | `--num_workers` | DataLoader worker processes | 4 |
@@ -84,7 +114,7 @@ Combined optimizer: matrix parameters via **Muon**, non-matrix via **AdamW** (`f
 | Parameter | Description | Default |
 |-----------|-------------|---------|
 | `--nprocs` | Number of GPUs / processes | 1 |
-| `--parallel_mode` | Parallel strategy (`none`, `ddp`, or `fsdp`) | none |
+| `--parallel_mode` | Parallel strategy (`none`, `ddp`, `fsdp`) | fsdp |
 | `--device_type` | Device type | cuda |
 | `--start_method` | Multiprocessing start method (`spawn`, `fork`, `forkserver`) | spawn |
 | `--backend` | Distributed training backend | nccl |
@@ -100,18 +130,31 @@ Combined optimizer: matrix parameters via **Muon**, non-matrix via **AdamW** (`f
 | `--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` |
 | `--neftune_alpha` | NEFTune noise alpha (0=disabled, typical: 5.0) | 0.0 | `sft` |
 
+### Online Rollout
+
+These options apply to `online_grpo` and `online_dpo`. Online strategies require
+a `BaseRewardModel` factory in `TrainConfig`; `train.py` does not currently
+provide a command-line option for configuring one.
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--rollout_interval` | Optimizer steps between rollout refreshes | 512 |
+| `--rollout_temperature` | Rollout sampling temperature | 0.7 |
+| `--rollout_top_k` | Rollout top-k filtering (`0` disables) | 0 |
+| `--rollout_top_p` | Rollout nucleus sampling threshold | 0.9 |
+| `--rollout_max_tokens` | Maximum generated tokens per response | 1024 |
+
 ### Scheduler
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
 | `--schedule_type` | LR scheduler type (`cosine`, `sgdr`, `wsd`) | cosine |
-| `--min_rate` | Minimum LR as fraction of base LR | None (scheduler default: 0.01) |
+| `--min_rate` | Minimum LR as fraction of base LR | None (scheduler default: 0.05 for cosine/SGDR, 0.0 for WSD) |
 | `--cycle_length` | SGDR first cycle length in steps | None (total_steps - warmup_steps) |
 | `--t_mult` | SGDR cycle length multiplier per restart | 2 |
-| `--stable_steps` | WSD stable plateau steps | None (required for wsd) |
+| `--stable_steps` | WSD stable plateau steps | None (80% of post-warmup steps) |
 | `--decay_steps` | WSD decay steps | None (total_steps - warmup_steps - stable_steps) |
 
 ### Usage Example
@@ -151,6 +194,7 @@ nohup python scripts/tools/train.py \
 | `--device` | str | `cuda` | Device to load model on |
 | `--dtype` | str | `bfloat16` | Model weights dtype (`bfloat16`, `float16`, `float32`) |
 | `--max_batch_size` | int | `16` | Maximum batch size for continuous batching |
+| `--max_seq_len` | int | model config `max_position_embeddings` | Maximum sequence length (KV cache size + prompt truncation) |
 | `--reload` | flag | `False` | Enable auto-reload for development |
 
 Usage:
@@ -160,6 +204,14 @@ python scripts/tools/server.py --param_path ./params --device cuda --dtype bfloa
 
 See [Inference Guide](inference.md) for HTTP API documentation.
 
+# Preprocess
+
+```bash
+python scripts/tools/preprocess.py data/*.jsonl -o output/ -c config.json
+```
+
+See [Preprocessing Guide](preprocessing.md) for config file format and examples.
+
 ## Generate (`generate.py`)
 
 | Parameter | Type | Default | Description |
@@ -173,7 +225,11 @@ See [Inference Guide](inference.md) for HTTP API documentation.
 | `--top_k` | int | `30` | Top-k filtering |
 | `--top_p` | float | `0.95` | Nucleus sampling threshold |
 | `--batch_size` | int | `1` | Batch size for generation |
-| `--max_tokens` | int | model config `max_len` | Maximum tokens to generate |
+| `--num_samples` | int | `1` | Responses per prompt |
+| `--max_tokens` | int | model config `max_position_embeddings` | Maximum tokens to generate |
+| `--cache_len` | int | `2048` | KV cache length |
+| `--frequency_penalty` | float | `0.0` | Frequency penalty |
+| `--rep_window` | int | `64` | Window size for frequency penalty |
 
 Usage:
 ```bash
@@ -201,4 +257,4 @@ See [Preprocessing Guide](preprocessing.md) for config file format and examples.
 
 ---
 
-> Document Update Time: 2026-07-19
+> Document Update Time: 2026-07-20

docs/guides/preprocessing.md

@@ -256,6 +256,7 @@ When `sources` is set, `sections` is ignored.
 | `min_chars` | int | `50` | Skip text-mode items shorter than this |
 | `max_chars` | int | `2000000` | Skip text-mode items longer than this |
 | `max_items` | int or null | `null` | Stop after N documents |
+| `batch_size` | int | `256` | Records per tokenization batch |
 | `packing_strategy` | str | `"simple"` | Packing strategy: `"simple"`, `"bfd"`, `"bfd_split"` |
 | `max_packed_len` | int | `8192` | Maximum length of a packed bin |
 | `truncation_mode` | str | `"keep_start"` | How to truncate sequences: `"keep_start"` or `"keep_end"` |
@@ -265,7 +266,7 @@ When `sources` is set, `sections` is ignored.
 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
 | `domain_key` | str or null | `null` | JSONL key for domain grouping |
-| `storage_format` | str | `"bin"` | `"bin"` (mmap) or `"h5"` |
+| `storage_format` | str | `"bin"` | `"bin"` (mmap). Reading also supports `"jsonl"` for on-the-fly tokenization |
 | `max_tokens_per_shard` | int | `100000000` | Flush threshold in cumulative tokens |
 | `dtype` | dict[str, str] | `{}` | Per-key tensor dtype override (e.g. `{"loss_mask": "bool"}`) |
 | `position_ids_mode` | str | `"doc_reset"` | How to compute position_ids: `"none"`, `"doc_reset"`, `"continuous"` |

docs/guides/training.md

@@ -6,7 +6,7 @@
 - [Causal Mask](#causal-mask)
 - [Rotary Position Embedding (RoPE)](#rotary-position-embedding-rope)
 - [Training Loop](#training-loop)
-- [Strategies](#strategies) — SEQ, SFT, DPO, GRPO
+- [Strategies](#strategies) — SEQ, SFT, DPO, GRPO, online rollout
 - [LR Schedulers](#lr-schedulers)
 - [Gradient Checkpointing](#gradient-checkpointing)
 - [Checkpoint](#checkpoint)
@@ -41,7 +41,7 @@ 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.
+`RotaryEmbedding` pre-computes `cos_table` and `sin_table` (f32, `[max_len, dim/2]`). `forward()` returns a `(cos, sin)` tuple indexed by `position_ids`. `apply_rotary_emb` applies the rotation: during training it uses torch complex multiply (autograd-compatible); during inference it auto-dispatches to a fused CUDA kernel when available.
 
 ## Training Loop
 
@@ -146,6 +146,19 @@ Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`. External sync of `ol
 
 Keys: `prompts`, `responses`, `masks`, `rewards`.
 
+### Online Rollout
+
+`online_grpo` and `online_dpo` use the respective GRPO and DPO strategies with
+a `RolloutRunner`. The runner renders prompts through the tokenizer chat
+template, generates grouped responses through `InferenceScheduler`, then scores
+them with a `BaseRewardModel`. It refreshes cached rollouts every
+`rollout_interval` optimizer steps. `online_grpo` synchronizes `old_model` when
+a fresh rollout is produced.
+
+Online strategies require `TrainConfig.reward_model_fn`. `train.py` exposes the
+rollout sampling parameters but does not yet offer a CLI argument for the reward
+model factory.
+
 ## LR Schedulers
 
 | Type | Class | Description |
@@ -162,6 +175,7 @@ Trades compute for memory by recomputing activations during backward pass. Speci
 
 ```python
 from astrai.model.components.decoder_block import DecoderBlock
+
 config = TrainConfig(..., gradient_checkpointing_modules=[DecoderBlock])
 ```
 
@@ -181,18 +195,14 @@ Model config (`context.model_config`) saved into `config.json` during training v
 ## TrainContextBuilder (Builder Pattern)
 
 ```python
-context = (
-    TrainContextBuilder(config)
-        .with_resume_dir(resume_dir)
-        .build()
-)
+context = TrainContextBuilder(config).with_param_path(param_path, resume=True).build()
 # Returns TrainContext with model, strategy, optimizer, scheduler, dataloader, checkpoint
 ```
 
-- Loads checkpoint weights if provided
+- Loads checkpoint weights before the model is wrapped
 - 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
+- Calls `executor.prepare(model_fn, optimizer_fn, scheduler_fn, before_wrap=...)`; the executor creates, wraps, then builds the optimizer and scheduler for the wrapped model
+- Creates `RDSampler` for shuffle+resume
 - Builds strategy via `StrategyFactory.create(train_type, model, device, **kwargs)`
 
 ## Training CLI
@@ -222,4 +232,4 @@ nohup python scripts/tools/train.py \
 
 Full parameter reference at [params.md](params.md).
 
-> Document Update Time: 2026-07-19
+> Document Update Time: 2026-07-31

pyproject.toml

@@ -8,7 +8,6 @@ name = "astrai"
 readme = "README.md"
 requires-python = ">=3.12"
 dependencies = [
-    "h5py==3.15.1",
     "numpy==2.4.4",
     "torch==2.11.0",
     "tokenizers==0.21.4",
@@ -16,10 +15,11 @@ dependencies = [
     "safetensors==0.5.3",
     "huggingface-hub==0.34.3",
     "jinja2>=3.0.0",
+    "pydantic>=2.0",
     "fastapi",
     "uvicorn[standard]",
-    "httpx",
-    "requests",
+    "click>=8.0",
+    "pyyaml>=6.0",
 ]
 keywords = ["nlp", "datasets", "language-models", "machine-learning"]
 license = { text = "GPL-3.0" }
@@ -31,14 +31,11 @@ classifiers = [
 urls = { Homepage = "https://github.com/ViperEkura/AstrAI" }
 
 [project.optional-dependencies]
-dev = ["pytest==9.0.2", "ruff"]
+dev = ["pytest==9.0.2", "ruff", "httpx2"]
 
 [tool.setuptools.packages.find]
 where = ["."]
 
-[tool.pip]
-extra-index-url = "https://download.pytorch.org/whl/cu128"
-
 [tool.setuptools.dynamic]
 version = { attr = "astrai.__version__" }
 
@@ -50,3 +47,4 @@ quote-style = "double"
 indent-style = "space"
 skip-magic-trailing-comma = false
 line-ending = "auto"
+exclude = ["*.md", "*.json", "*.yml", "*.yaml"]

scripts/eval/analyze_weights.py

@@ -8,11 +8,11 @@ import safetensors.torch
 import torch
 
 
-def effective_rank_metrics(w: torch.Tensor) -> dict:
+def effective_rank_metrics(w: torch.Tensor, device: str = "cpu") -> dict:
     if w.ndim == 1:
         return {"shape": tuple(w.shape), "is_1d": True}
 
-    w = w.float()
+    w = w.float().to(device)
     s = torch.linalg.svdvals(w)
     s_sq = s**2
     total = s_sq.sum()
@@ -143,7 +143,7 @@ def print_layer_grid(results: dict[str, dict]):
     widths = [6] + [10] * len(comps)
     metric = "er_99_norm"
 
-    print(f"\n--- Per-Layer Effective Rank (99% energy) ---")
+    print("\n--- Per-Layer Effective Rank (99% energy) ---")
     print(format_header(["Layer"] + comps, widths))
     print("-" * sum(widths))
 
@@ -173,7 +173,7 @@ def print_layer_grid(results: dict[str, dict]):
 def print_weight_stats(results: dict[str, dict]):
     groups = group_by_component(results)
     widths = [20, 12, 12, 12, 12]
-    print(f"\n--- Weight Value Statistics ---")
+    print("\n--- Weight Value Statistics ---")
     print(format_header(["Component", "Mean", "Std", "Min", "Max"], widths))
     print("-" * sum(widths))
 
@@ -238,6 +238,12 @@ def main():
         default=None,
         help="Save results as JSON to this path.",
     )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        help="Device for SVD computation (e.g., 'cuda:0', 'cpu').",
+    )
     args = parser.parse_args()
 
     all_results = {}
@@ -265,7 +271,7 @@ def main():
             )
         print(f"{'=' * 70}")
 
-        print(f"Loading weights...")
+        print("Loading weights...")
         sd = safetensors.torch.load_file(str(weights_path))
         print(f"  {len(sd)} keys loaded")
 
@@ -277,10 +283,12 @@ def main():
 
         results = {}
         if not args.no_svd:
-            print(f"Computing SVD on {len(weight_keys)} tensors...")
+            print(
+                f"Computing SVD on {len(weight_keys)} tensors (device={args.device})..."
+            )
             for i, k in enumerate(sorted(weight_keys)):
                 print(f"  [{i + 1}/{len(weight_keys)}] {k:<60s}", end="\r")
-                results[k] = effective_rank_metrics(sd[k])
+                results[k] = effective_rank_metrics(sd[k], device=args.device)
             print()
         else:
             print(f"Computing stats on {len(weight_keys)} tensors (no SVD)...")