feat: add frequency penalty to inference sampling pipeline

- Add FrequencyPenaltyStrategy (logit -= penalty * count)
- Per-task rep_window for penalty history lookup
- Wire through engine, task, executor, API layer
- Add --frequency_penalty and --rep_window to stream_chat.py
- 9 unit tests for frequency penalty strategy

.github/workflows/release.yml

@@ -0,0 +1,71 @@
+name: Release
+
+on:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  build-pure:
+    name: Build pure-Python wheel
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Build wheel (no CUDA)
+        run: |
+          pip wheel . --no-deps -w dist/
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: pure-wheel
+          path: dist/*.whl
+
+  build-cuda-linux:
+    name: Build CUDA wheel (Linux)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install torch (CUDA 12.8)
+        run: |
+          pip install torch --index-url https://download.pytorch.org/whl/cu128
+
+      - name: Setup CUDA
+        uses: Jimver/cuda-toolkit@v0.2.35
+        with:
+          cuda: "12.8.0"
+
+      - name: Build wheel (with CUDA kernels)
+        run: |
+          CSRC_KERNELS=true pip wheel . --no-deps --no-build-isolation -w dist/
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: cuda-wheel-linux
+          path: dist/*.whl
+
+  release:
+    name: Attach wheels to release
+    needs: [build-pure, build-cuda-linux]
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: "*-wheel"
+          merge-multiple: true
+
+      - name: Create release & upload assets
+        uses: softprops/action-gh-release@v2
+        with:
+          files: ./*.whl
+          tag_name: ${{ github.ref_name }}
+          generate_release_notes: true

.gitignore

@@ -7,8 +7,14 @@
 # Allow specific file types and root files
 !astrai/**/*.py
 !scripts/**/*.py
-!scripts/**/*.sh
 !tests/**/*.py
+!csrc/**/*.py
+
+!csrc/**/*.cu
+!csrc/**/*.h
+!csrc/**/*.cuh
+
+!scripts/**/*.sh
 
 # Allow GitHub files
 !/.github/**
@@ -23,3 +29,8 @@
 !/LICENSE
 !/pyproject.toml
 !/README.md
+# Allow extension modules (only source .py)
+!/astrai/extension/**/*.py
+
+# Allow build files
+!/setup.py

README.md

@@ -9,7 +9,7 @@
 <div align="center">
   <img src="https://img.shields.io/badge/python-3.12+-blue.svg" alt="python">
   <img src="https://img.shields.io/badge/license-GPL--3.0-blue.svg" alt="license">
-  <img src="https://img.shields.io/github/v/release/ViperEkura/AstrAI?label=Release&color=76bad9" alt="release">
+  <img src="https://img.shields.io/github/v/tag/ViperEkura/AstrAI?label=Release&color=76bad9" alt="release">
   <img src="https://img.shields.io/github/stars/ViperEkura/AstrAI?style=flat&label=Stars&color=76bad9" alt="stars">
   <img src="https://img.shields.io/github/forks/ViperEkura/AstrAI?style=flat&label=Forks&color=76bad9" alt="forks">
 </div>
@@ -59,8 +59,9 @@ End-to-end walkthrough in 5 steps:
 ```bash
 git clone https://github.com/ViperEkura/AstrAI.git
 cd AstrAI
-pip install -e .
+pip install -e .                                          # pure PyTorch (no CUDA kernels)
-# pip install -e ".[dev]"    # optional: dev dependencies (pytest, ruff)
+# CSRC_KERNELS=true pip install -e . --no-build-isolation  # optional: fused CUDA kernels
+# pip install -e ".[dev]"                                  # dev dependencies (pytest, ruff)
 ```
 
 **2. Download model**
@@ -102,9 +103,7 @@ nohup python scripts/tools/train.py \
     --warmup_ratio=0.05 \
     --max_lr=1e-4 \
     --max_grad_norm=1.0 \
-    --adamw_beta1=0.9 \
+    --weight_decay=0.1 \
-    --adamw_beta2=0.95 \
-    --adamw_weight_decay=0.01 \
     --window_size=2048 \
     --ckpt_interval=10000 \
     --ckpt_dir=./checkpoint \

assets/docs/README-zh-CN.md

@@ -15,7 +15,7 @@
 <div align="center">
   <img src="https://img.shields.io/badge/python-3.12+-blue.svg" alt="python">
   <img src="https://img.shields.io/badge/license-GPL--3.0-blue.svg" alt="license">
-  <img src="https://img.shields.io/github/v/release/ViperEkura/AstrAI?label=Release&color=76bad9" alt="release">
+  <img src="https://img.shields.io/github/v/tag/ViperEkura/AstrAI?label=Release&color=76bad9" alt="release">
   <img src="https://img.shields.io/github/stars/ViperEkura/AstrAI?style=flat&label=Stars&color=76bad9" alt="stars">
   <img src="https://img.shields.io/github/forks/ViperEkura/AstrAI?style=flat&label=Forks&color=76bad9" alt="forks">
 </div>
@@ -65,7 +65,8 @@
 ```bash
 git clone https://github.com/ViperEkura/AstrAI.git
 cd AstrAI
-pip install -e .
+pip install -e .                                          # 纯 PyTorch（不含 CUDA 内核）
+# CSRC_KERNELS=true pip install -e . --no-build-isolation  # 可选：融合 CUDA 内核加速
 # pip install -e ".[dev]"                                  # 可选：开发依赖（pytest, ruff）
 ```
 
@@ -108,9 +109,7 @@ nohup python scripts/tools/train.py \
     --warmup_ratio=0.05 \
     --max_lr=1e-4 \
     --max_grad_norm=1.0 \
-    --adamw_beta1=0.9 \
+    --weight_decay=0.1 \
-    --adamw_beta2=0.95 \
-    --adamw_weight_decay=0.01 \
     --window_size=2048 \
     --ckpt_interval=10000 \
     --ckpt_dir=./checkpoint \

assets/docs/architecture.md

@@ -63,7 +63,6 @@ classDiagram
             +Optional[int] n_heads
             +Optional[int] n_kv_heads
             +Optional[bool] use_qk_norm
-            +Optional[bool] use_gated_attention
             +str ffn_type
             +Optional[dict] rope_scaling
             +Optional[str] pooling_type
@@ -125,7 +124,6 @@ classDiagram
             +str ckpt_dir
             +int ckpt_interval
             +str log_dir
-            +int log_interval
             +List[str] metrics
             +Optional[LoRAConfig] lora
             +int random_seed
@@ -501,7 +499,6 @@ classDiagram
             +float clip_eps
             +float kl_coef
             +int group_size
-            +str reduction
             +int sync_interval
             +compute_loss(batch) Tensor
             +sync_ref_model()
@@ -559,7 +556,7 @@ classDiagram
         }
 
         class GradientCheckpointingCallback {
-            +tuple modules
+            +Optional[List[type]] modules
             +on_train_begin(context)
             +on_train_end(context)
         }
@@ -573,31 +570,29 @@ classDiagram
             +on_batch_end(context)
             +on_train_end(context)
             +on_error(context)
-            +save_extra(context) dict$
+            +save_extra(context) dict
         }
 
         class ProgressBarCallback {
             +int num_epoch
             +int log_interval
             +IO file
+            +tqdm progress_bar
             +on_epoch_begin(context)
-            +on_batch_end(context)
+            +on_optimizer_step(context)
             +on_epoch_end(context)
         }
 
-        class MetricLoggerCallback {
+        class MetricCallback {
             +Path log_dir
             +int save_interval
-            +int log_interval
             +List[str] metrics
-            +on_batch_end(context)
+            +int val_step
+            +on_optimizer_step(context)
+            +on_epoch_end(context)
             +on_train_end(context)
             +on_error(context)
-        }
-
-        class ValidationCallback {
             -_run_validation(context)
-            +on_optimizer_step(context)
         }
 
         class CallbackFactory {
@@ -684,20 +679,44 @@ classDiagram
         }
 
         class KVCache {
-            -PagePool _pool
+            <<abstract>>
-            -Storage _storage
-            -TaskTable _table
-            +int page_size
             +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)
-            +make_table_tensor(task_ids, device) Tensor
+            +bind_tasks(task_ids, total_len, device) CacheView
-            +bind(page_table, total_len) KvcacheView
         }
 
-        class KvcacheView {
+        class PageCache {
+            +int page_size
+            -PagePool _pool
+            -Storage _storage
+            -TaskTable _table
+            +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
@@ -705,6 +724,14 @@ classDiagram
             +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]
@@ -727,7 +754,6 @@ classDiagram
         +int output_tokens
         +float arrival_time
         +Optional[float] finish_time
-            +Optional[Callable] stream_callback
         +int next_pos
         +is_finished(stop_ids) bool
     }
@@ -1035,14 +1061,14 @@ classDiagram
     TrainCallback <|-- GradientCheckpointingCallback
     TrainCallback <|-- CheckpointCallback
     TrainCallback <|-- ProgressBarCallback
-    TrainCallback <|-- MetricLoggerCallback
+    TrainCallback <|-- MetricCallback
-    TrainCallback <|-- ValidationCallback
     BaseDataset <|-- SEQDataset
     BaseDataset <|-- SFTDataset
     BaseDataset <|-- DPODataset
     BaseDataset <|-- GRPODataset
     Store <|-- H5Store
     Store <|-- MmapStore
+    Store <|-- JsonlStore
     BaseSamplingStrategy <|-- TemperatureStrategy
     BaseSamplingStrategy <|-- TopKStrategy
     BaseSamplingStrategy <|-- TopPStrategy
@@ -1075,11 +1101,15 @@ classDiagram
     ResponseBuilder <|-- OpenAIResponseBuilder
     ResponseBuilder <|-- AnthropicResponseBuilder
     BaseMaskBuilder <|-- SectionedMaskBuilder
+    KVCache <|-- PageCache
+    KVCache <|-- ContiguousCache
+    CacheView <|-- PageCacheView
+    CacheView <|-- ContiguousCacheView
 
     %% --- Composition (strong ownership, part destroyed with whole) ---
-    KVCache *-- PagePool
+    PageCache *-- PagePool
-    KVCache *-- Storage
+    PageCache *-- Storage
-    KVCache *-- TaskTable
+    PageCache *-- TaskTable
     InferenceEngine *-- InferenceScheduler
     InferenceScheduler *-- KVCache
     InferenceScheduler *-- Executor
@@ -1107,7 +1137,8 @@ classDiagram
     TrainContext o-- BaseScheduler
     TrainContext o-- Checkpoint
     TrainContext o-- BaseExecutor
-    KvcacheView o-- Storage
+    PageCacheView o-- Storage
+    ContiguousCacheView o-- ContiguousCache
     SamplingPipeline o-- BaseSamplingStrategy
     BaseDataset o-- Store
     Pipeline o-- PipelineConfig
@@ -1129,6 +1160,7 @@ classDiagram
     DecoderBlock ..> FFNFactory : uses
     StoreFactory ..> H5Store : creates
     StoreFactory ..> MmapStore : creates
+    StoreFactory ..> JsonlStore : creates
     ConfigFactory ..> AutoRegressiveLMConfig : creates
     ConfigFactory ..> EncoderConfig : creates
     ExecutorFactory ..> NoneExecutor : creates
@@ -1142,7 +1174,8 @@ classDiagram
     TrainContextBuilder ..> ResumableDistributedSampler : creates
     Checkpoint ..> Checkpoint : serializes
     CheckpointCallback ..> Checkpoint : creates
-    KVCache ..> KvcacheView : binds
+    PageCache ..> PageCacheView : binds
+    ContiguousCache ..> ContiguousCacheView : binds
     InferenceEngine ..> GenerationRequest : uses
     InferenceEngine ..> GenerateResult : creates
     OpenAIResponseBuilder ..> ChatCompletionRequest : receives
@@ -1171,12 +1204,12 @@ classDiagram
 |--------|------------|-------------|
 | **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, Pipeline, filter_by_length, PackingStrategy, PackingStrategyFactory, PositionIdStrategy, PositionIdStrategyFactory, StoreWriter, StoreWriterFactory | Declarative JSON-driven data preprocessing |
-| **astrai.dataset** | BaseDataset–GRPODataset, Store–MmapStore, StoreFactory, ResumableDistributedSampler, DatasetFactory | Dataset loading and management |
+| **astrai.dataset** | BaseDataset–GRPODataset, Store–JsonlStore/MmapStore/H5Store, StoreFactory, ResumableDistributedSampler, 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.tokenize** | AutoTokenizer, ChatTemplate | Tokenizer and chat template |
-| **astrai.trainer** | Trainer, TrainContext, TrainContextBuilder, BaseStrategy–GRPOStrategy, StrategyFactory, BaseScheduler–WSDScheduler, SchedulerFactory, TrainCallback(Protocol)–ValidationCallback, CallbackFactory | Training workflow |
+| **astrai.trainer** | Trainer, TrainContext, TrainContextBuilder, BaseStrategy–GRPOStrategy, StrategyFactory, BaseScheduler–WSDScheduler, SchedulerFactory, TrainCallback(Protocol)–MetricCallback, CallbackFactory | Training workflow |
-| **astrai.inference** | InferenceEngine, InferenceScheduler, Executor, KVCache–KvcacheView, Allocator–Storage, Task, TaskManager, TaskStatus, GenerationRequest, GenerateResult, BaseSamplingStrategy–SamplingPipeline, ProtocolHandler, ResponseBuilder, OpenAIResponseBuilder, AnthropicResponseBuilder, StopChecker, GenContext, ChatMessage–MessagesRequest, app | Inference service |
+| **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.factory** | BaseFactory | Component registration |
 | **astrai.protocols** | OptimizerProtocol, SchedulerProtocol | Structural subtyping for optimizer/scheduler wrappers |
@@ -1185,7 +1218,7 @@ classDiagram
 
 | Pattern | Classes | Purpose |
 |---------|---------|---------|
-| **Factory** | `AttnFactory`, `FFNFactory`, `StrategyFactory`, `DatasetFactory`, `SchedulerFactory`, `CallbackFactory`, `StoreFactory`, `ConfigFactory`, `ExecutorFactory` | Decorator-based component creation |
+| **Factory** | `AttnFactory`, `FFNFactory`, `StrategyFactory`, `DatasetFactory`, `SchedulerFactory`, `CallbackFactory`, `StoreFactory`, `ConfigFactory`, `ExecutorFactory`, `MaskBuilderFactory`, `StoreWriterFactory`, `PackingStrategyFactory`, `PositionIdStrategyFactory` | 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 |
@@ -1195,7 +1228,7 @@ classDiagram
 | **Context** | `TrainContext` | Unified training state bag |
 | **Object Pool** | `Allocator`, `PagePool` | Page-based KV cache with LRU eviction |
 | **Executor** | `BaseExecutor`, `NoneExecutor`, `DDPExecutor`, `FSDPExecutor` | Gradient accumulation & model distribution |
-| **Storage** | `Store`, `H5Store`, `MmapStore` | Format-agnostic data access with multi-segment support |
+| **Storage** | `Store`, `H5Store`, `MmapStore`, `JsonlStore` | Format-agnostic data access with multi-segment support |
 | **Producer-Consumer** | `InferenceScheduler`, `Task`, queues | Continuous batching |
 | **AutoModel Registry** | `AutoModel`, `AutoRegressiveLM`, `EmbeddingEncoder` | Model-type dynamic loading |
 
@@ -1207,10 +1240,10 @@ classDiagram
 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`
 6. **Distributed**: `spawn_parallel_fn` + `setup_parallel` for multi-process DDP
-7. **Dataset Loading**: `DatasetFactory` creates datasets, `Store` (H5Store/MmapStore) loads data with explicit `_length` and multi-segment `_data`
+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`
 9. **Scheduler**: `SchedulerFactory` creates `CosineScheduler`/`SGDRScheduler`/`WSDScheduler`
 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: 2026-05-30
+> Document Update Time: 2026-07-09

assets/docs/dataflow.md

@@ -48,8 +48,8 @@ The output `meta.json` records the storage format, key names, dtype, total token
 
 `detect_format(load_path)` inspects the path:
 
-- If `load_path` is a file: checks suffix — `.h5`/`.hdf5` → `"h5"`, unknown suffix raises `ValueError`
+- If `load_path` is a file: checks suffix — `.h5`/`.hdf5` → `"h5"`, `.jsonl` → `"jsonl"`, unknown suffix raises `ValueError`
-- If `load_path` is a directory: recursively globs for `*.h5`/`*.hdf5` files → `"h5"`, or `*.bin` + `**/meta.json` → `"bin"`
+- If `load_path` is a directory: recursively globs for `*.h5`/`*.hdf5` files → `"h5"`, `*.bin` + `**/meta.json` → `"bin"`, or `*.jsonl` + `dataset_config.json` → `"jsonl"`
 
 ### Store Backends
 
@@ -58,13 +58,16 @@ Storage format is auto-detected by `detect_format()`; backends are dispatched vi
 ```
 StoreFactory.create("h5")    → H5Store
 StoreFactory.create("bin")   → MmapStore
+StoreFactory.create("jsonl") → JsonlStore
 ```
 
-**H5Store**: Reads HDF5 files, supports `share_memory_()` for multi-process DataLoader workers (copies tensors to shared memory).
+**H5Store**: Reads HDF5 files. Tensors are loaded into host memory and normalized into segmented storage.
 
 **MmapStore**: Memory-maps `.bin` files. OS page cache sharing is native — no explicit `share_memory_()` needed. Uses `torch.from_numpy(np.memmap(...))`.
 
-Both backends normalise tensors into `Store._data[Dict[str, List[Tensor]]]` + `Store._cum[Dict[str, List[int]]]` (cumulative lengths for bisect-based indexing).
+**JsonlStore**: On-the-fly tokenization of raw JSONL files at load time. Requires a `dataset_config.json` alongside the `.jsonl` files following the same `PipelineConfig` schema with an additional `tokenizer_path` field.
+
+All backends normalise tensors into `Store._data[Dict[str, List[Tensor]]]` + `Store._cum[Dict[str, List[int]]]` (cumulative lengths for bisect-based indexing).
 
 ## Data Keys by Training Type
 
@@ -106,4 +109,4 @@ DatasetFactory.load(train_type, load_path, window_size, stride=None, storage_typ
 
 Standard PyTorch `DataLoader` with configurable `batch_size`, `num_workers`, `pin_memory`, `prefetch_factor`. Sampler produces indices; dataloader fetches tensor batches via `__getitem__`.
 
-> Document Update Time: 2026-06-19
+> Document Update Time: 2026-07-09

assets/docs/inference.md

@@ -23,29 +23,40 @@ RoPE is applied **before** KV cache write, not after — otherwise position enco
 
 ## KVCache System
 
-Seven classes working together:
+Seven classes working together, with two concrete cache implementations:
+
+### ContiguousCache (default)
 
 ```
-KVCache (facade)
+ContiguousCache (simple contiguous per-slot cache)
+  ├── ContiguousCacheView  bundles k/v tensors + slot indices for attention layers
+```
+
+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.
+
+### PageCache (paged with prefix sharing)
+
+```
+PageCache (paged KV cache with prefix sharing, alternative)
   ├── PagePool               orchestrates page allocation + prefix matching
-  │     ├── Allocator   bitmask-based page allocator + ref-count + LRU eviction (inside PagePool)
+  │     ├── Allocator         bitmask-based page allocator + ref-count + LRU
-  │     └── PrefixCache hash-based prefix matching (page_hash via polynomial hash) (inside PagePool)
+  │     └── 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)
-  └── KvcacheView      bundles Storage + page_table + total_len for attention layers (returned by bind())
+  └── PageCacheView          bundles Storage + page_table + total_len for attention layers
 ```
 
-`KVCache.bind(page_table, total_len)` returns a `KvcacheView` used by attention layers via `write()` / `gather()`.
+`isinstance(cache, KVCache)` checks dispatch to the correct view. Both implement the abstract `KVCache` interface used by `Executor` and `InferenceScheduler`.
 
 ## Continuous Batching
 
 `InferenceScheduler` runs a daemon thread with a 4-phase loop:
 
 ```
-1. Cleanup → Remove finished tasks, free KV pages
+1. Cleanup → Remove finished tasks, free KV cache slots/pages
-2. Refill  → Pop from waiting_queue, task_alloc pages, activate
+2. Refill  → Pop from waiting_queue, task_alloc resources, activate
 3. Prefill → Group by (prompt_len, start_pos), run full forward
-4. Decode  → Pick largest same-position group, single-token forward
+4. Decode  → Run single-token forward for each same-position group
 ```
 
 ## Sampling (Strategy Pattern)
@@ -238,4 +249,4 @@ async for token in engine.generate_async("Hello", ...):    # -> AsyncGenerator[s
     print(token)
 ```
 
-> Document Update Time: 2026-06-19
+> Document Update Time: 2026-07-09

assets/docs/params.md

@@ -28,13 +28,17 @@
 | `--max_lr` | Maximum learning rate (cosine decay after warmup) | 3e-4 |
 | `--max_grad_norm` | Maximum gradient norm for clipping | 1.0 |
 
-### Optimizer (AdamW)
+### Optimizer (MuonMix)
+
+Combined optimizer: matrix parameters via **Muon**, non-matrix via **AdamW** (`fused=True`).
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--adamw_beta1` | AdamW beta1 | 0.9 |
+| `--weight_decay` | Weight decay (applied to Muon matrix params; non-matrix use 0) | 0.1 |
-| `--adamw_beta2` | AdamW beta2 | 0.95 |
+| `--muon_momentum` | Muon momentum factor | 0.95 |
-| `--adamw_weight_decay` | AdamW weight decay | 0.01 |
+| `--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` |
 
 ### Data Loading
 
@@ -67,7 +71,6 @@
 | Parameter | Description | Default |
 |-----------|-------------|---------|
 | `--log_dir` | Directory for metric logs | checkpoint/logs |
-| `--log_interval` | Number of optimizer steps between metric logs | 1 |
 | `--metrics` | Metrics to log (e.g. --metrics loss lr val_loss) | ["loss", "lr", "grad_norm"] |
 
 ### Gradient Checkpointing
@@ -105,7 +108,7 @@
 | Parameter | Description | Default |
 |-----------|-------------|---------|
 | `--schedule_type` | LR scheduler type (`cosine`, `sgdr`, `wsd`) | cosine |
-| `--min_rate` | Minimum LR as fraction of base LR | None (scheduler default) |
+| `--min_rate` | Minimum LR as fraction of base LR | None (scheduler default: 0.01) |
 | `--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) |
@@ -127,9 +130,7 @@ nohup python scripts/tools/train.py \
     --warmup_ratio=0.05 \
     --max_lr=1e-4 \
     --max_grad_norm=1.0 \
-    --adamw_beta1=0.9 \
+    --weight_decay=0.1 \
-    --adamw_beta2=0.95 \
-    --adamw_weight_decay=0.01 \
     --window_size=2048 \
     --ckpt_interval=10000 \
     --ckpt_dir=./checkpoint \
@@ -200,4 +201,4 @@ See [Preprocessing Guide](preprocessing.md) for config file format and examples.
 
 ---
 
-> Document Update Time: 2026-06-19
+> Document Update Time: 2026-07-09

assets/docs/preprocessing.md

@@ -1,6 +1,6 @@
 # Preprocessing Pipeline
 
-Declarative JSON-driven data preprocessing. One `SectionedMaskBuilder` handles all formats via `input.sections` (single-output) or `input.sources` (multi-output).
+Declarative JSON-driven data preprocessing. `MaskBuilderFactory` supports three registered builders: `"single"` (single-output via `input.sections`), `"multi"` (multi-output via `input.sources`), and `"sectioned"` (façade dispatching to `single` or `multi` based on config).
 
 ## Contents
 
@@ -361,4 +361,4 @@ Pipeline(
 ).run()
 ```
 
-> Document Update Time: 2026-06-03
+> Document Update Time: 2026-07-09

assets/docs/training.md

@@ -80,13 +80,13 @@ on_train_end
 | `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`, `MetricLoggerCallback`, `ValidationCallback` |
+| `on_optimizer_step` | Every accumulation window | `GradientClippingCallback`, `MetricCallback`, `ProgressBarCallback` |
-| `on_batch_end` | Every batch | `CheckpointCallback`, `MetricLoggerCallback`, `ProgressBarCallback` |
+| `on_batch_end` | Every batch | `CheckpointCallback` |
-| `on_epoch_end` | End of each epoch | `ProgressBarCallback` |
+| `on_epoch_end` | End of each epoch | `MetricCallback`, `ProgressBarCallback` |
-| `on_error` | On exception during training | `CheckpointCallback`, `MetricLoggerCallback` |
+| `on_error` | On exception during training | `CheckpointCallback`, `MetricCallback` |
-| `on_train_end` | Training ends (always via finally) | `CheckpointCallback`, `MetricLoggerCallback`, `GradientCheckpointingCallback` |
+| `on_train_end` | Training ends (always via finally) | `CheckpointCallback`, `MetricCallback`, `GradientCheckpointingCallback` |
 
-Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `validation` (periodic validation on val_dataset), `metric_logger` (JSONL, rank-0), `progress_bar` (tqdm), `gradient_clipping`.
+Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `metric` (JSONL + validation, rank-0), `progress_bar` (tqdm), `gradient_clipping`.
 
 ## Strategies
 
@@ -108,7 +108,7 @@ $$
 L_{\text{SFT}} = -\sum_{t=P+1}^{P+L} \log P(s_t \mid s_{\lt t}; \theta)
 $$
 
-Keys: `input_ids`, `target_ids`, `loss_mask`. Optional: `label_smoothing`.
+Keys: `input_ids`, `target_ids`, `loss_mask`, `position_ids`. Optional: `label_smoothing`.
 
 ### DPO (Direct Preference Optimization)
 
@@ -122,17 +122,23 @@ Parameters: `beta=0.1`, `reduction="mean"`. Keys: `chosen`, `rejected`, `chosen_
 
 ### GRPO (Group Relative Policy Optimization)
 
-On-policy PPO with group-normalized advantages:
+Token-level PPO with group-normalized advantages. Advantages are derived from
+scalar per-response rewards, group-normalized, and broadcast across all response
+tokens. Only response tokens contribute to the loss (prompt tokens are masked
+out):
 
 $$
 \text{Advantage}_i = \frac{r_i - \mu}{\sigma + \epsilon}
 $$
 
 $$
-L_{\text{GRPO}} = -\mathbb{E}\left[\min\left(\frac{\pi_\theta}{\pi_{\text{ref}}}A,\; \text{clip}\left(\frac{\pi_\theta}{\pi_{\text{ref}}}, 1-\epsilon, 1+\epsilon\right)A\right)\right] + \lambda \cdot \mathbb{E}\left[(\log\pi_\theta - \log\pi_{\text{ref}})^2\right]
+L_{\text{GRPO}} = -\mathbb{E}_t\left[\min\left(\rho_t A,\; \text{clip}\left(\rho_t, 1-\epsilon, 1+\epsilon\right)A\right)\right] + \lambda \cdot \mathbb{E}_t\left[\frac{\pi_{\text{ref}}}{\pi_\theta} - \log\frac{\pi_{\text{ref}}}{\pi_\theta} - 1\right]
 $$
 
-Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`, `sync_interval=200`, `reduction="mean"`.
+where $\rho_t = \pi_\theta(a_t|s_t) / \pi_{\text{ref}}(a_t|s_t)$ is the
+per-token probability ratio and the expectations are over valid response tokens.
+
+Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`, `sync_interval=200`.
 
 Keys: `prompts`, `responses`, `masks`, `rewards`.
 
@@ -201,9 +207,7 @@ nohup python scripts/tools/train.py \
     --warmup_ratio=0.05 \
     --max_lr=1e-4 \
     --max_grad_norm=1.0 \
-    --adamw_beta1=0.9 \
+    --weight_decay=0.1 \
-    --adamw_beta2=0.95 \
-    --adamw_weight_decay=0.01 \
     --window_size=2048 \
     --ckpt_interval=10000 \
     --ckpt_dir=./checkpoint \
@@ -214,4 +218,4 @@ nohup python scripts/tools/train.py \
 
 Full parameter reference at [params.md](params.md).
 
-> Document Update Time: 2026-05-30
+> Document Update Time: 2026-07-09

astrai/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "1.3.8"
+__version__ = "1.3.9"
 __author__ = "ViperEkura"
 
 from astrai.config import (

astrai/dataset/__init__.py

@@ -1,6 +1,7 @@
 from astrai.dataset.dataset import (
     BaseDataset,
     DatasetFactory,
+    grpo_collate_fn,
 )
 from astrai.dataset.sampler import ResumableDistributedSampler
 from astrai.dataset.storage import (
@@ -21,6 +22,7 @@ from astrai.serialization import (
 __all__ = [
     "BaseDataset",
     "DatasetFactory",
+    "grpo_collate_fn",
     "Store",
     "StoreFactory",
     "H5Store",

astrai/dataset/dataset.py

@@ -15,6 +15,49 @@ from astrai.dataset.storage import (
 from astrai.factory import BaseFactory
 
 
+def grpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
+    """Collate variable-length GRPO samples into padded 3-D tensors.
+
+    Input: list of dicts, each with:
+      - prompts:  [P_i]
+      - responses: list of G tensors, each [R_ij]
+      - masks:     list of G tensors, each [R_ij]
+      - rewards:  [G]
+
+    Output:
+      - prompts:   [B, P_max]
+      - responses: [B, G, R_max]
+      - masks:     [B, G, R_max]
+      - rewards:   [B, G]
+    """
+    B = len(batch)
+    G = len(batch[0]["responses"])
+    P_max = max(b["prompts"].size(0) for b in batch)
+    R_max = max(r.size(0) for b in batch for r in b["responses"])
+
+    prompts = torch.zeros(B, P_max, dtype=torch.long)
+    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"]
+        rewards[i, : b["rewards"].size(0)] = b["rewards"]
+        for g in range(min(G, len(b["responses"]))):
+            r_len = b["responses"][g].size(0)
+            responses[i, g, :r_len] = b["responses"][g]
+            if g < len(b["masks"]):
+                masks[i, g, :r_len] = b["masks"][g]
+
+    return {
+        "prompts": prompts,
+        "responses": responses,
+        "masks": masks,
+        "rewards": rewards,
+    }
+
+
 class BaseDataset(Dataset, ABC):
     """Abstract base class for all dataset types.
 
@@ -250,28 +293,85 @@ class DPODataset(BaseDataset):
 
 @DatasetFactory.register("grpo")
 class GRPODataset(BaseDataset):
-    """Dataset for Group Relative Policy Optimization training."""
+    """Dataset for offline Group Relative Policy Optimization.
+
+    Unlike the window-based datasets (SEQ/SFT/DPO), GRPO data is
+    record-structured: each sample is one prompt with its group of
+    responses and scalar rewards.  There is no windowing or stride —
+    every record is an independent training unit.
+
+    Expected storage layout (produced by JsonlStore or pre-tokenized):
+
+    - ``prompts``:   List[Tensor]  — one 1-D token tensor per record
+    - ``responses``: List[List[Tensor]] — G response tensors per record
+    - ``masks``:     List[List[Tensor]] — G mask tensors per record
+    - ``rewards``:   List[Tensor]  — one 1-D float tensor (len G) per record
+    """
+
+    def __init__(self, window_size: int = 0, stride: int = 0, **kwargs):
+        super().__init__(window_size=window_size, stride=stride or window_size)
+        self._records: List[dict] = []
 
     @property
     def required_keys(self) -> List[str]:
         return ["prompts", "responses", "masks", "rewards"]
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
+    def load(self, load_path: str, storage_type: Optional[str] = None, **kwargs):
-        return self.storage.fetch(begin_idx, end_idx, key)
+        if storage_type is None:
+            storage_type = detect_format(load_path)
+        self.storage = StoreFactory.create(storage_type, **kwargs)
+        self._load_path = load_path
+        self.storage.load(load_path, **kwargs)
+        self._validate_keys()
+        self._build_records()
+
+    def _validate_keys(self):
+        actual_keys = set(self.storage.keys)
+        missing = [k for k in self.required_keys if k not in actual_keys]
+        if missing:
+            raise KeyError(
+                f"GRPODataset requires keys {self.required_keys}, "
+                f"but storage only has {sorted(actual_keys)}. Missing: {missing}"
+            )
+
+    def _build_records(self):
+        """Unfold segmented storage into per-record lists.
+
+        ``prompts`` is a flat list of 1-D tensors (one per record).
+        ``responses`` / ``masks`` are nested lists (G tensors per record).
+        ``rewards`` is a flat list of 1-D tensors (len G per record).
+        """
+        prompt_segs = self.storage._data.get("prompts", [])
+        response_segs = self.storage._data.get("responses", [])
+        mask_segs = self.storage._data.get("masks", [])
+        reward_segs = self.storage._data.get("rewards", [])
+
+        n_records = len(prompt_segs)
+        self._records = []
+        for i in range(n_records):
+            self._records.append(
+                {
+                    "prompts": prompt_segs[i],
+                    "responses": response_segs[i] if i < len(response_segs) else [],
+                    "masks": mask_segs[i] if i < len(mask_segs) else [],
+                    "rewards": reward_segs[i]
+                    if i < len(reward_segs)
+                    else torch.tensor([]),
+                }
+            )
+
+    @property
+    def count(self) -> int:
+        return len(self._records)
+
+    def __len__(self) -> int:
+        return len(self._records)
 
     def __getitem__(self, index: int) -> Dict[str, Tensor]:
-        begin_idx, end_idx = self.get_index(index)
+        rec = self._records[index]
-
-        prompts = self._fetch_data(begin_idx, end_idx, "prompts").to(dtype=torch.long)
-        responses = self._fetch_data(begin_idx, end_idx, "responses").to(
-            dtype=torch.long
-        )
-        masks = self._fetch_data(begin_idx, end_idx, "masks").to(dtype=torch.bool)
-        rewards = self._fetch_data(begin_idx, end_idx, "rewards")
-
         return {
-            "prompts": prompts,
+            "prompts": rec["prompts"].to(dtype=torch.long),
-            "responses": responses,
+            "responses": [r.to(dtype=torch.long) for r in rec["responses"]],
-            "masks": masks,
+            "masks": [m.to(dtype=torch.bool) for m in rec["masks"]],
-            "rewards": rewards,
+            "rewards": rec["rewards"].to(dtype=torch.float32),
         }

astrai/dataset/storage.py

@@ -81,6 +81,11 @@ def detect_format(load_path: str) -> str:
     ]
     if jsonl_files:
         return "jsonl"
+    json_files = [
+        Path(p) for p in glob.glob(str(root / "**" / "*.json"), recursive=True)
+    ]
+    if json_files:
+        return "jsonl"
     raise FileNotFoundError(f"No supported data files found at {load_path}")
 
 
@@ -143,26 +148,37 @@ class Store(ABC):
 
         return results[0] if len(results) == 1 else torch.cat(results, dim=0)
 
-    def _normalize(self, raw: Dict[str, List[Tensor]]):
+    def _normalize(self, raw: Dict[str, list]):
         """Register segments and pre-compute cumulative lengths.
 
         Does NOT concatenate — segments are kept as-is to avoid OOM on
         large datasets.  Sets ``self._length`` to the minimum total
-        element count across all keys.
+        element count across all flat-tensor keys.
+
+        For GRPO multi-response keys, values may be ``List[List[Tensor]]``
+        (one list of G tensors per record).  These are stored as-is and
+        excluded from the cumulative-length bookkeeping since they are
+        accessed record-by-record via ``_data`` rather than via ``fetch``.
         """
+        flat_lengths = []
         for key, tensors in raw.items():
             self._data[key] = tensors
+            if not tensors:
+                self._cum[key] = []
+                flat_lengths.append(0)
+                continue
+            # Skip nested lists (GRPO responses/masks) — record-level access
+            if isinstance(tensors[0], list):
+                self._cum[key] = []
+                continue
             cum = []
             total = 0
             for t in tensors:
                 total += t.shape[0]
                 cum.append(total)
             self._cum[key] = cum
-        self._length = (
+            flat_lengths.append(cum[-1] if cum else 0)
-            min((cum[-1] if cum else 0) for cum in self._cum.values())
+        self._length = min(flat_lengths) if flat_lengths else 0
-            if self._cum
-            else 0
-        )
 
 
 class StoreFactory(BaseFactory["Store"]):
@@ -245,12 +261,7 @@ class JsonlStore(Store):
         with open(config_path, "r", encoding="utf-8") as f:
             raw_config = json.load(f)
 
-        tokenizer_path = raw_config.pop("tokenizer_path", None)
+        tokenizer_path = raw_config.pop("tokenizer_path", None) or str(root)
-        if tokenizer_path is None:
-            raise ValueError(
-                f"JSONL dataset config must specify 'tokenizer_path': {config_path}"
-            )
-
         self.config = PipelineConfig.from_dict(raw_config)
         tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
         mask_builder = MaskBuilderFactory.create("sectioned")
@@ -261,6 +272,27 @@ class JsonlStore(Store):
         raw: Dict[str, List[Tensor]] = {}
         doc_sequences: List[List[int]] = []
 
+        def _process_item(item: dict) -> None:
+            nonlocal raw, doc_sequences
+            result = mask_builder.build(item, self.config, tokenizer)
+            if result is None:
+                return
+            result.pop("domain", None)
+            primary_ids = self._primary_ids(result)
+            if not primary_ids:
+                return
+            doc_sequences.append(primary_ids)
+            for key, ids in result.items():
+                if key not in raw:
+                    raw[key] = []
+                if ids and isinstance(ids[0], list):
+                    # GRPO multi-response: List[List[int]] → List[Tensor]
+                    raw[key].append(
+                        [torch.tensor(sub, dtype=self._infer_dtype(sub)) for sub in ids]
+                    )
+                else:
+                    raw[key].append(torch.tensor(ids, dtype=self._infer_dtype(ids)))
+
         for jsonl_path in sorted(root.glob("*.jsonl")):
             with open(jsonl_path, "r", encoding="utf-8") as f:
                 for line in f:
@@ -274,21 +306,22 @@ class JsonlStore(Store):
                             "Failed to parse JSON line in %s, skipping", jsonl_path
                         )
                         continue
+                    _process_item(item)
 
-                    result = mask_builder.build(item, self.config, tokenizer)
+        for json_path in sorted(root.glob("*.json")):
-                    if result is None:
+            if json_path.name == self.CONFIG_NAME:
                 continue
-
+            with open(json_path, "r", encoding="utf-8") as f:
-                    result.pop("domain", None)
+                try:
-                    primary_ids = self._primary_ids(result)
+                    data = json.load(f)
-                    if not primary_ids:
+                except json.JSONDecodeError:
+                    logger.warning("Failed to parse JSON file %s, skipping", json_path)
                     continue
-
+            if isinstance(data, list):
-                    doc_sequences.append(primary_ids)
+                for item in data:
-                    for key, ids in result.items():
+                    _process_item(item)
-                        if key not in raw:
+            elif isinstance(data, dict):
-                            raw[key] = []
+                _process_item(data)
-                        raw[key].append(torch.tensor(ids, dtype=self._infer_dtype(ids)))
 
         pos_ids = position_strategy.generate(doc_sequences)
         if pos_ids:
@@ -298,7 +331,7 @@ class JsonlStore(Store):
 
     @staticmethod
     def _primary_ids(result: dict) -> List[int]:
-        """Return the first integer list in *result* as the primary id sequence."""
+        """Return the first flat integer list in *result* as the primary id sequence."""
         for val in result.values():
             if isinstance(val, list) and val and isinstance(val[0], int):
                 return val

astrai/extension/__init__.py

@@ -0,0 +1,29 @@
+"""CUDA attention kernel wrappers with torch fallback.
+
+Public API:
+    - ``attn_decode`` — single-query decode attention
+    - ``attn_prefill`` — multi-query prefill attention
+    - ``attn_paged_decode`` — paged decode attention (direct page-table access)
+
+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"
+
+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``.
+"""
+
+from astrai.extension.loader import KERNEL_NAMES, is_available
+from astrai.extension.ops import attn_decode, attn_paged_decode, attn_prefill
+
+__all__ = [
+    "attn_decode",
+    "attn_paged_decode",
+    "attn_prefill",
+    "is_available",
+    "KERNEL_NAMES",
+]

astrai/extension/loader.py

@@ -0,0 +1,36 @@
+"""Dynamic discovery and loading of compiled CUDA kernel modules.
+
+Each kernel is registered in ``csrc/build.py`` and built into a ``.so`` placed
+in this package directory. On import we try to load each one; kernels that
+failed to build (or are running on a CPU-only machine) are marked unavailable
+so the wrapper functions can fall back to ``torch`` SDPA.
+"""
+
+import importlib
+import logging
+
+logger = logging.getLogger(__name__)
+
+KERNEL_NAMES = ["attn_decode", "attn_prefill", "attn_paged_decode"]
+
+_available: dict[str, bool] = {}
+_modules: dict[str, object] = {}
+
+for _name in KERNEL_NAMES:
+    try:
+        _mod = importlib.import_module(f".{_name}", package=__package__)
+        _available[_name] = True
+        _modules[_name] = _mod
+    except ImportError:
+        _available[_name] = False
+        _modules[_name] = None
+
+
+def is_available(name: str) -> bool:
+    """Return ``True`` if the compiled kernel ``name`` was loaded."""
+    return _available.get(name, False)
+
+
+def get_module(name: str) -> object:
+    """Return the loaded kernel module for ``name``, or ``None`` if unavailable."""
+    return _modules.get(name)

astrai/extension/ops.py

@@ -0,0 +1,246 @@
+"""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/inference/__init__.py

@@ -6,7 +6,7 @@ Layers:
   - protocols/:   Response builders (OpenAI, Anthropic)
   - transport/:   SSE transport utilities
   - engine.py:    Facade (InferenceEngine), Value Object (GenerationRequest)
-  - sample.py:    Strategy pattern (TemperatureStrategy, TopKStrategy, TopPStrategy)
+  - sample.py:    Strategy pattern (TemperatureStrategy, TopKStrategy, TopPStrategy, FrequencyPenaltyStrategy)
 """
 
 from astrai.inference.api import (
@@ -50,6 +50,7 @@ from astrai.inference.core import (
 from astrai.inference.engine import GenerationRequest, InferenceEngine
 from astrai.inference.sample import (
     BaseSamplingStrategy,
+    FrequencyPenaltyStrategy,
     SamplingPipeline,
     TemperatureStrategy,
     TopKStrategy,
@@ -83,6 +84,7 @@ __all__ = [
     "TemperatureStrategy",
     "TopKStrategy",
     "TopPStrategy",
+    "FrequencyPenaltyStrategy",
     "SamplingPipeline",
     "ProtocolHandler",
     "StopChecker",

astrai/inference/api/openai.py

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

astrai/inference/api/protocol.py

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

astrai/inference/core/executor.py

@@ -75,6 +75,33 @@ class Executor:
         temperatures = torch.tensor([t.temperature for t in tasks], device=self.device)
         top_ks = torch.tensor([t.top_k for t in tasks], device=self.device)
         top_ps = torch.tensor([t.top_p for t in tasks], device=self.device)
+        freq_penalties = torch.tensor(
+            [t.frequency_penalty for t in tasks], device=self.device
+        )
+
+        history_lists = []
+        mask_lists = []
+        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))
+
+        max_len = max(len(h) for h in history_lists)
+        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(
+                h, dtype=torch.long, device=self.device
+            )
+            padded_mask[i, : len(m)] = torch.tensor(
+                m, dtype=torch.bool, device=self.device
+            )
 
         with torch.inference_mode():
             outputs = self.model(
@@ -89,4 +116,7 @@ class Executor:
             temperature=temperatures,
             top_k=top_ks,
             top_p=top_ps,
+            frequency_penalty=freq_penalties,
+            input_ids=padded_ids,
+            input_mask=padded_mask,
         ).tolist()

astrai/inference/core/task.py

@@ -33,6 +33,8 @@ class Task:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
     ):
         self.task_id = task_id
         self.prompt_ids = prompt_ids
@@ -40,6 +42,8 @@ class Task:
         self.temperature = temperature
         self.top_p = top_p
         self.top_k = top_k
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
 
         self.status = TaskStatus.PENDING
         self.output_ids: List[int] = []
@@ -92,6 +96,8 @@ class TaskManager:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
         stream_callback: Optional[Callable[[str], None]] = None,
     ) -> str:
         task_id = f"task_{int(time.time())}_{uuid.uuid4().hex[:8]}"
@@ -116,6 +122,8 @@ class TaskManager:
             temperature=temperature,
             top_p=top_p,
             top_k=top_k,
+            frequency_penalty=frequency_penalty,
+            rep_window=rep_window,
         )
 
         with self._lock:

astrai/inference/engine.py

@@ -74,6 +74,8 @@ class GenerationRequest:
         top_p: float = 1.0,
         temperature: float = 1.0,
         max_tokens: Optional[int] = None,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
         stream: bool = False,
     ):
         if not (isinstance(top_k, int) and top_k >= 0):
@@ -82,12 +84,21 @@ class GenerationRequest:
             raise ValueError("top_p must be a float between 0.0 and 1.0")
         if not (isinstance(temperature, (int, float)) and temperature > 0):
             raise ValueError("temperature must be a positive number")
+        if not (
+            isinstance(frequency_penalty, (int, float))
+            and -2.0 <= frequency_penalty <= 2.0
+        ):
+            raise ValueError("frequency_penalty must be between -2.0 and 2.0")
+        if not (isinstance(rep_window, int) and rep_window > 0):
+            raise ValueError("rep_window must be a positive integer")
 
         self.messages = messages
         self.top_k = top_k
         self.top_p = top_p
         self.temperature = temperature
         self.max_tokens = max_tokens
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
         self.stream = stream
 
 
@@ -132,17 +143,33 @@ class InferenceEngine:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
     ) -> Union[Generator, str, List[str]]:
         is_batch = isinstance(prompt, list)
         prompts = prompt if is_batch else [prompt]
 
         if stream:
             return self._generate_streaming(
-                prompts, is_batch, max_tokens, temperature, top_p, top_k
+                prompts,
+                is_batch,
+                max_tokens,
+                temperature,
+                top_p,
+                top_k,
+                frequency_penalty,
+                rep_window,
             )
         else:
             return self._generate_non_streaming(
-                prompts, is_batch, max_tokens, temperature, top_p, top_k
+                prompts,
+                is_batch,
+                max_tokens,
+                temperature,
+                top_p,
+                top_k,
+                frequency_penalty,
+                rep_window,
             )
 
     def generate_async(
@@ -152,9 +179,18 @@ class InferenceEngine:
         temperature: float = 1.0,
         top_p: float = 1.0,
         top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
     ) -> AsyncGenerator[str, None]:
         sync_gen = self._generate_streaming(
-            [prompt], False, max_tokens, temperature, top_p, top_k
+            [prompt],
+            False,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
         )
 
         async def _agen():
@@ -185,6 +221,8 @@ class InferenceEngine:
             temperature=request.temperature,
             top_p=request.top_p,
             top_k=request.top_k,
+            frequency_penalty=request.frequency_penalty,
+            rep_window=request.rep_window,
         )
 
     def _submit_tasks(
@@ -194,6 +232,8 @@ class InferenceEngine:
         temperature: float,
         top_p: float,
         top_k: int,
+        frequency_penalty: float,
+        rep_window: int,
     ) -> Tuple[GenerateResult, List[str]]:
         n = len(prompts)
         result = GenerateResult(count=n)
@@ -206,6 +246,8 @@ class InferenceEngine:
                 temperature=temperature,
                 top_p=top_p,
                 top_k=top_k,
+                frequency_penalty=frequency_penalty,
+                rep_window=rep_window,
                 stream_callback=cb,
             )
             task_ids.append(task_id)
@@ -226,9 +268,17 @@ class InferenceEngine:
         temperature: float,
         top_p: float,
         top_k: int,
+        frequency_penalty: float,
+        rep_window: int,
     ) -> Generator:
         result, task_ids = self._submit_tasks(
-            prompts, max_tokens, temperature, top_p, top_k
+            prompts,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
         )
         n = len(prompts)
         remaining = n
@@ -262,9 +312,17 @@ class InferenceEngine:
         temperature: float,
         top_p: float,
         top_k: int,
+        frequency_penalty: float,
+        rep_window: int,
     ) -> Union[str, List[str]]:
         result, task_ids = self._submit_tasks(
-            prompts, max_tokens, temperature, top_p, top_k
+            prompts,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            frequency_penalty,
+            rep_window,
         )
 
         try:

astrai/inference/sample.py

@@ -1,15 +1,15 @@
 """Composable sampling strategies for logit transformation.
 
 Implements the Strategy pattern: each sampling technique
-(temperature, top-k, top-p) is a pluggable strategy that
+(temperature, top-k, top-p, frequency penalty) is a pluggable
-can be composed into a pipeline.
+strategy that can be composed into a pipeline.
 
 All strategies accept both scalar and per-sample tensor
 parameters, so a single pipeline works for any batch size.
 """
 
 from abc import ABC, abstractmethod
-from typing import List, Union
+from typing import List, Optional, Union
 
 import torch
 from torch import Tensor
@@ -19,12 +19,23 @@ class BaseSamplingStrategy(ABC):
     """Abstract base for a logit transformation strategy."""
 
     @abstractmethod
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         """Applies the strategy to logits.
 
         Args:
             logits: Raw logits tensor (batch, vocab_size).
             filter_value: Value assigned to filtered-out positions.
+            input_ids: Previously generated token IDs ``[batch, seq_len]``,
+                padded with 0. Used by frequency penalty.
+            input_mask: Boolean mask ``[batch, seq_len]``, True for real
+                tokens, False for padding. Used to exclude padding from
+                penalty computation.
 
         Returns:
             Transformed logits tensor.
@@ -42,7 +53,13 @@ class TemperatureStrategy(BaseSamplingStrategy):
     def __init__(self, temperature: Union[float, Tensor] = 1.0):
         self.temperature = temperature
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         t = self.temperature
         if isinstance(t, Tensor):
             t = t.to(logits.device, non_blocking=True).view(-1, 1)
@@ -64,7 +81,13 @@ class TopKStrategy(BaseSamplingStrategy):
     def __init__(self, top_k: Union[int, Tensor] = 0):
         self.top_k = top_k
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         tk = self.top_k
         if isinstance(tk, Tensor):
             tk = tk.to(logits.device, non_blocking=True).long().clamp(min=0)
@@ -114,7 +137,13 @@ class TopPStrategy(BaseSamplingStrategy):
         logits[mask] = filter_value
         return logits
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         tp = self.top_p
         if isinstance(tp, Tensor):
             tp = tp.to(logits.device, non_blocking=True)
@@ -125,6 +154,84 @@ class TopPStrategy(BaseSamplingStrategy):
         return logits
 
 
+class FrequencyPenaltyStrategy(BaseSamplingStrategy):
+    """Penalizes tokens based on how many times they appeared in history.
+
+    Subtracts ``penalty * count(token)`` from each token's logit, where
+    ``count(token)`` is the number of occurrences in the generation history
+    (prompt + output). A penalty of ``0.0`` disables the strategy.
+
+    Unlike repetition penalty (which only checks *presence*), frequency
+    penalty scales linearly with occurrence count: the first use is
+    penalized once, the third use three times. This allows natural
+    repetition of common words while suppressing degenerate loops.
+
+    Reference: OpenAI API ``frequency_penalty`` parameter.
+
+    Args:
+        penalty: Scalar or ``[batch]`` tensor (0.0 disables, range -2.0~2.0).
+    """
+
+    def __init__(self, penalty: Union[float, Tensor] = 0.0):
+        self.penalty = penalty
+
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
+        if input_ids is None:
+            return logits
+
+        p = self.penalty
+        if isinstance(p, Tensor):
+            p = p.to(logits.device, non_blocking=True).view(-1, 1)
+            if (p == 0.0).all():
+                return logits
+        elif p == 0.0:
+            return logits
+
+        input_ids = input_ids.to(logits.device, non_blocking=True)
+
+        if input_mask is not None:
+            input_mask = input_mask.to(logits.device, non_blocking=True)
+            masked_ids = input_ids.clone()
+            masked_ids[~input_mask] = -1
+        else:
+            masked_ids = input_ids
+
+        batch_sz, seq_len = masked_ids.shape
+        vocab_size = logits.size(-1)
+
+        if isinstance(p, Tensor):
+            penalty_per_row = p.expand(batch_sz, 1)
+        else:
+            penalty_per_row = torch.full(
+                (batch_sz, 1), float(p), device=logits.device, dtype=logits.dtype
+            )
+
+        counts = torch.zeros(
+            batch_sz, vocab_size, device=logits.device, dtype=logits.dtype
+        )
+        valid_mask = masked_ids >= 0
+        if valid_mask.any():
+            valid_ids = masked_ids[valid_mask]
+            row_indices = (
+                torch.arange(batch_sz, device=logits.device)
+                .unsqueeze(1)
+                .expand_as(masked_ids)[valid_mask]
+            )
+            counts.index_put_(
+                (row_indices, valid_ids),
+                torch.ones_like(valid_ids, dtype=logits.dtype),
+                accumulate=True,
+            )
+
+        return logits - penalty_per_row * counts
+
+
 class SamplingPipeline(BaseSamplingStrategy):
     """Composes multiple sampling strategies into a single transformation.
 
@@ -145,23 +252,39 @@ class SamplingPipeline(BaseSamplingStrategy):
     def __init__(self, strategies: List[BaseSamplingStrategy]):
         self.strategies = strategies
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         for strategy in self.strategies:
-            logits = strategy.apply(logits, filter_value)
+            logits = strategy.apply(logits, filter_value, input_ids, input_mask)
         return logits
 
-    @torch.no_grad()
+    @torch.inference_mode()
-    def sample(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def sample(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         """Apply strategies then sample (softmax + multinomial).
 
         Args:
             logits: Raw logits ``[batch, vocab_size]``.
+            input_ids: Previously generated token IDs ``[batch, seq_len]``.
+            input_mask: Boolean mask for ``input_ids`` padding.
 
         Returns:
             Sampled token IDs ``[batch]``.
         """
         return torch.multinomial(
-            torch.softmax(self.apply(logits, filter_value), dim=-1),
+            torch.softmax(
+                self.apply(logits, filter_value, input_ids, input_mask), dim=-1
+            ),
             num_samples=1,
         ).squeeze(-1)
 
@@ -172,6 +295,9 @@ def sample(
     temperature: Union[float, Tensor] = 1.0,
     top_k: Union[int, Tensor] = 0,
     top_p: Union[float, Tensor] = 1.0,
+    frequency_penalty: Union[float, Tensor] = 0.0,
+    input_ids: Optional[Tensor] = None,
+    input_mask: Optional[Tensor] = None,
     filter_value: float = -float("inf"),
 ) -> Tensor:
     """Apply sampling strategies then sample (softmax + multinomial).
@@ -180,6 +306,10 @@ def sample(
 
     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.
 
     Returns:
         Sampled token IDs ``[batch]``.
@@ -189,5 +319,6 @@ def sample(
             TemperatureStrategy(temperature),
             TopKStrategy(top_k),
             TopPStrategy(top_p),
+            FrequencyPenaltyStrategy(frequency_penalty),
         ]
-    ).sample(logits, filter_value)
+    ).sample(logits, filter_value, input_ids, input_mask)

astrai/parallel/executor.py

@@ -7,6 +7,7 @@ from contextlib import contextmanager
 from typing import 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
@@ -120,6 +121,21 @@ class BaseExecutor:
     def unwrap_model(self, model: nn.Module):
         return model.state_dict()
 
+    @contextmanager
+    def checkpoint_context(self, model: nn.Module):
+        if self.use_distributed:
+            dist.barrier()
+        state_dict = self._gather_state_dict(model)
+        yield state_dict
+        if self.use_distributed:
+            dist.barrier()
+
+    def _gather_state_dict(self, model: nn.Module):
+        state_dict = self.unwrap_model(model)
+        if self.use_distributed and get_rank() != 0:
+            return None
+        return state_dict
+
     @property
     def use_distributed(self) -> bool:
         return get_world_size() > 1
@@ -208,6 +224,13 @@ class DDPExecutor(BaseExecutor):
             return model.module.state_dict()
         return model.state_dict()
 
+    def _gather_state_dict(self, model: nn.Module):
+        if not self.use_distributed:
+            return self.unwrap_model(model)
+        if get_rank() != 0:
+            return None
+        return self.unwrap_model(model)
+
 
 @ExecutorFactory.register("fsdp")
 class FSDPExecutor(BaseExecutor):
@@ -279,7 +302,7 @@ class FSDPExecutor(BaseExecutor):
             with FSDP.state_dict_type(
                 model,
                 StateDictType.FULL_STATE_DICT,
-                FullStateDictConfig(offload_to_cpu=True, rank0_only=False),
+                FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
             ):
                 return model.state_dict()
 

astrai/parallel/setup.py

@@ -1,14 +1,21 @@
 import os
+import socket
 from abc import ABC, abstractmethod
 from contextlib import contextmanager
 from functools import wraps
-from typing import Callable
+from typing import Callable, Optional
 
 import torch
 import torch.distributed as dist
 import torch.multiprocessing as mp
 
 
+def find_free_port() -> str:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+        s.bind(("", 0))
+        return str(s.getsockname()[1])
+
+
 def get_current_device():
     return os.environ["LOCAL_DEVICE"]
 
@@ -217,11 +224,13 @@ def spawn_parallel_fn(
     world_size: int,
     backend: str = "nccl",
     master_addr: str = "localhost",
-    master_port: str = "29500",
+    master_port: Optional[str] = None,
     device_type: str = "cuda",
     start_method: str = "spawn",
     **kwargs,
 ):
+    if master_port is None:
+        master_port = find_free_port()
     launcher = _detect_launcher()
     if launcher in ("torchelastic", "torchrun", "external"):
         strategy = TorchrunStrategy(

astrai/preprocessing/__init__.py

@@ -1,7 +1,9 @@
 from astrai.preprocessing.builder import (
     BaseMaskBuilder,
     MaskBuilderFactory,
+    MultiOutputMaskBuilder,
     SectionedMaskBuilder,
+    SingleOutputMaskBuilder,
 )
 from astrai.preprocessing.packing import (
     PackingStrategy,
@@ -20,12 +22,14 @@ from astrai.preprocessing.writer import (
 __all__ = [
     "BaseMaskBuilder",
     "MaskBuilderFactory",
+    "MultiOutputMaskBuilder",
     "PackingStrategy",
     "PackingStrategyFactory",
     "Pipeline",
     "PositionIdStrategy",
     "PositionIdStrategyFactory",
     "SectionedMaskBuilder",
+    "SingleOutputMaskBuilder",
     "StoreWriter",
     "StoreWriterFactory",
     "filter_by_length",

astrai/preprocessing/builder.py

@@ -1,8 +1,10 @@
 """Mask building for preprocessing pipeline.
 
 :class:`SectionRenderer` converts section specs into token ids and loss
-masks (template / text / value extraction).  :class:`SectionedMaskBuilder`
+masks (template / text / value extraction).  :class:`SingleOutputMaskBuilder`
-orchestrates single-output / multi-output (DPO / GRPO) assembly.
+handles single-output (SFT / pretrain), :class:`MultiOutputMaskBuilder`
+handles multi-output (DPO / GRPO), and :class:`SectionedMaskBuilder`
+orchestrates both modes as a façade.
 """
 
 from abc import ABC, abstractmethod
@@ -93,8 +95,15 @@ class SectionRenderer:
         return all_ids, loss_mask
 
     def process_list_field(self, item: dict, sections: list, config, tokenizer):
-        all_ids: list[int] = []
+        """Tokenize a list-valued field, preserving per-element boundaries.
-        loss_mask: list[int] = []
+
+        Returns ``(list_of_id_lists, list_of_mask_lists)`` where each
+        inner list corresponds to one element of the source list.  This
+        is critical for GRPO where each response must stay a separate
+        sequence so the strategy can form a ``[G, R]`` tensor.
+        """
+        per_item_ids: list[list[int]] = []
+        per_item_masks: list[list[int]] = []
 
         for sec in sections:
             field = sec["field"]
@@ -106,17 +115,13 @@ class SectionRenderer:
                 continue
 
             for val in values:
+                ids: list[int] = []
+                mask: list[int] = []
                 if use_template:
                     if isinstance(val, list):
                         wrapper = {field: val}
                         self._append_template(
-                            wrapper,
+                            wrapper, field, action, tokenizer, config, ids, mask
-                            field,
-                            action,
-                            tokenizer,
-                            config,
-                            all_ids,
-                            loss_mask,
                         )
                 else:
                     wrapper = {field: str(val)}
@@ -128,17 +133,19 @@ class SectionRenderer:
                         False,
                         False,
                         config,
-                        all_ids,
+                        ids,
-                        loss_mask,
+                        mask,
                     )
-
+                if ids:
                     max_len = config.preprocessing.max_seq_len
-        all_ids = all_ids[:max_len]
+                    ids = ids[:max_len]
-        loss_mask = loss_mask[: len(all_ids)]
+                    mask = mask[: len(ids)]
+                    per_item_ids.append(ids)
+                    per_item_masks.append(mask)
 
-        if not all_ids:
+        if not per_item_ids:
             return None, None
-        return all_ids, loss_mask
+        return per_item_ids, per_item_masks
 
     @staticmethod
     def is_value_section(sections: list) -> bool:
@@ -212,42 +219,17 @@ class MaskBuilderFactory(BaseFactory["BaseMaskBuilder"]):
     pass
 
 
-@MaskBuilderFactory.register("sectioned")
+@MaskBuilderFactory.register("single")
-class SectionedMaskBuilder(BaseMaskBuilder):
+class SingleOutputMaskBuilder(BaseMaskBuilder):
-    """Config-driven builder supporting single and multi-output modes.
+    """Build a single output sequence with optional loss mask.
 
-    Single-output::
+    Expects ``config.input.sections`` (list of section specs).
-
-        {"input": {"sections": [
-            {"field": "messages", "action": "$role", "template": true}
-        ]}}
-        → {"sequence": [...], "loss_mask": [...], "domain": "..."}
-
-    Multi-output (DPO / GRPO)::
-
-        {"input": {"sources": {
-            "chosen": {"sections": [{"field": "chosen", "action": "$role", "template": true}]},
-            "rejected": {"sections": [{"field": "rejected", "action": "$role", "template": true}]},
-        }}}
-        → {"chosen": [...], "chosen_mask": [...], "rejected": [...], "rejected_mask": [...], "domain": "..."}
-
-    Output spec fields::
-
-        sections      – list of section specs (same format as single-output)
-        list_field    – True when JSONL field holds a list (GRPO responses)
-        mask_key      – explicit loss-mask output key (default: ``"{output_key}_mask"``)
     """
 
-    def __init__(self):
+    def __init__(self, renderer: Optional[SectionRenderer] = None):
-        self.renderer = SectionRenderer()
+        self.renderer = renderer or SectionRenderer()
 
     def build(self, item: dict, config, tokenizer) -> Optional[dict]:
-        sources_spec = getattr(config.input, "sources", None)
-        if sources_spec:
-            return self._build_multi(item, sources_spec, config, tokenizer)
-        return self._build_single(item, config, tokenizer)
-
-    def _build_single(self, item: dict, config, tokenizer) -> Optional[dict]:
         sections = config.input.sections
         if not sections:
             return None
@@ -266,9 +248,22 @@ class SectionedMaskBuilder(BaseMaskBuilder):
             result["loss_mask"] = mask
         return result
 
-    def _build_multi(
+
-        self, item: dict, sources_spec: dict, config, tokenizer
+@MaskBuilderFactory.register("multi")
-    ) -> Optional[dict]:
+class MultiOutputMaskBuilder(BaseMaskBuilder):
+    """Build multiple output sequences (DPO / GRPO).
+
+    Expects ``config.input.sources`` (dict of output_key → spec).
+    """
+
+    def __init__(self, renderer: Optional[SectionRenderer] = None):
+        self.renderer = renderer or SectionRenderer()
+
+    def build(self, item: dict, config, tokenizer) -> Optional[dict]:
+        sources_spec = getattr(config.input, "sources", None)
+        if not sources_spec:
+            return None
+
         result: dict = {}
         any_output = False
 
@@ -292,7 +287,15 @@ class SectionedMaskBuilder(BaseMaskBuilder):
                 ids, mask = self.renderer.process_list_field(
                     item, sections, config, tokenizer
                 )
-            else:
+                if ids is None:
+                    continue
+                # ids is List[List[int]] — preserve per-response structure
+                result[output_key] = ids
+                if mask is not None:
+                    result[mask_key] = mask
+                any_output = True
+                continue
+
             ids, mask = self.renderer.process_sections(
                 item, sections, config, tokenizer, is_top_level=True
             )
@@ -313,3 +316,22 @@ class SectionedMaskBuilder(BaseMaskBuilder):
 
         result["domain"] = _extract_domain(item, config.output.domain_key)
         return result
+
+
+@MaskBuilderFactory.register("sectioned")
+class SectionedMaskBuilder(BaseMaskBuilder):
+    """Façade that dispatches to SingleOutputMaskBuilder or MultiOutputMaskBuilder.
+
+    Preserves backward compatibility for existing configs and code that rely
+    on the ``"sectioned"`` factory name.
+    """
+
+    def __init__(self):
+        self._single = SingleOutputMaskBuilder()
+        self._multi = MultiOutputMaskBuilder()
+
+    def build(self, item: dict, config, tokenizer) -> Optional[dict]:
+        sources_spec = getattr(config.input, "sources", None)
+        if sources_spec:
+            return self._multi.build(item, config, tokenizer)
+        return self._single.build(item, config, tokenizer)

astrai/preprocessing/packing.py

@@ -119,3 +119,50 @@ class BFDPacking(PackingStrategy):
                 bin_lengths.append(seq_len)
 
         return bins
+
+
+@PackingStrategyFactory.register("bfd_split")
+class BFDSplitPacking(BFDPacking):
+    """BFD packing with over-length sequences split into chunks.
+
+    Sequences longer than *max_packed_len* are split into consecutive
+    chunks of at most *max_packed_len* tokens instead of being
+    truncated.  Each chunk becomes an independent sequence that enters
+    BFD planning.  All keys (``loss_mask``, ``position_ids``, …) are
+    split in lockstep so per-token alignment is preserved.
+
+    Note: because each chunk is treated as a separate document, the
+    second chunk of a split sequence loses the preceding context.
+    """
+
+    def apply(
+        self,
+        keys: Dict[str, List[List[int]]],
+        max_packed_len: int,
+        truncation_mode: str,
+    ) -> Dict[str, List[List[int]]]:
+        sequences = keys.get("sequence", [])
+        if not sequences:
+            return keys
+        if max_packed_len <= 0:
+            return super().apply(keys, max_packed_len, truncation_mode)
+
+        split_keys = self._split_all(keys, max_packed_len)
+        return super().apply(split_keys, max_packed_len, truncation_mode)
+
+    @staticmethod
+    def _split_all(
+        keys: Dict[str, List[List[int]]], max_packed_len: int
+    ) -> Dict[str, List[List[int]]]:
+        """Split every sequence exceeding *max_packed_len* into chunks,
+        applying the same chunk boundaries to all keys."""
+        sequences = keys["sequence"]
+        chunk_bounds = [list(range(0, len(s), max_packed_len)) for s in sequences]
+        result: Dict[str, List[List[int]]] = {}
+        for key, vals in keys.items():
+            split_vals: List[List[int]] = []
+            for val, starts in zip(vals, chunk_bounds):
+                for start in starts:
+                    split_vals.append(val[start : start + max_packed_len])
+            result[key] = split_vals
+        return result

astrai/preprocessing/pipeline.py

@@ -149,6 +149,13 @@ class Pipeline:
     def _iter_items(self):
         for path in self.paths:
             with open(path, "r", encoding="utf-8") as f:
+                if path.endswith(".json"):
+                    data = json.load(f)
+                    if isinstance(data, dict):
+                        yield data
+                    elif isinstance(data, list):
+                        yield from data
+                else:
                     for line in f:
                         line = line.strip()
                         if not line:
@@ -173,6 +180,21 @@ class Pipeline:
                 dt = _STR_TO_DTYPE.get(
                     self.config.output.dtype.get(key, "int32"), torch.int32
                 )
+                # GRPO multi-response keys store List[List[int]] per record
+                # (responses/masks).  Rewards store List[float] per record.
+                # Both produce List[Tensor] (one tensor per record), but
+                # responses need inner flattening while rewards do not.
+                if ids_list and isinstance(ids_list[0], list):
+                    tensors[key] = [
+                        torch.tensor(
+                            list(chain.from_iterable(ids))
+                            if ids and isinstance(ids[0], list)
+                            else ids,
+                            dtype=dt,
+                        )
+                        for ids in ids_list
+                    ]
+                else:
                     tensors[key] = [
                         torch.tensor(list(chain.from_iterable(ids_list)), dtype=dt)
                     ]

astrai/serialization/checkpoint.py

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

astrai/serialization/dataset.py

@@ -26,6 +26,9 @@ 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:

astrai/trainer/strategy.py

@@ -223,14 +223,13 @@ class DPOStrategy(BaseStrategy):
         self,
         model: nn.Module,
         device: str,
+        ref_model: nn.Module,
         beta: float = 0.1,
         reduction: str = "mean",
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(
+        self.ref_model = ref_model
-            self.model_fn, self.executor.unwrap_model(model)
-        ).to(device=self.device)
         self.beta = beta
         self.reduction = reduction
 
@@ -267,42 +266,45 @@ class DPOStrategy(BaseStrategy):
 class GRPOStrategy(BaseStrategy):
     """Group Relative Policy Optimization strategy.
 
-    On-policy GRPO following DeepSeek-R1: the policy model is updated while
+    Implements GRPO following DeepSeek-R1 with token-level PPO clipping.
-    a frozen ref_model stores the old-policy log-probs.  ratio = exp(logπ_θ - logπ_ref),
+    Advantages are group-normalized from scalar per-response rewards and
-    clipped PPO objective.  Call ``sync_ref_model()`` after each data-generation round.
+    broadcast across all response tokens.  The loss is computed **only on
+    response tokens** — prompt tokens are masked out.
+
+    Three model roles are distinguished:
+
+    * **Policy** ``self.model`` — the model being trained.
+    * **Old policy** ``self.old_model`` — the behaviour policy that generated
+      the responses.  Used for the importance sampling ratio
+      ``ρ = π_θ / π_old``.  Synced externally after each data-generation round.
+    * **Reference model** ``self.ref_model`` — a frozen copy of the initial
+      policy (typically the SFT checkpoint) used **only** for the KL
+      regularisation term.  It is never updated during training.
     """
 
     def __init__(
         self,
         model: nn.Module,
         device: str,
+        old_model: nn.Module,
+        ref_model: nn.Module,
         clip_eps: float = 0.2,
         kl_coef: float = 0.01,
         group_size: int = 4,
-        reduction: str = "mean",
-        sync_interval: int = 200,
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(
+        self.old_model = old_model
-            self.model_fn, self.executor.unwrap_model(model)
+        self.ref_model = ref_model
-        ).to(device=self.device)
         self.clip_eps = clip_eps
         self.kl_coef = kl_coef
         self.group_size = group_size
-        self.reduction = reduction
-        self.sync_interval = sync_interval
-        self._step = 0
 
-    def sync_ref_model(self):
+    def sync_old_model(self):
-        """Copy current model weights to ref model."""
+        """Copy current policy weights to old model."""
-        self.ref_model.load_state_dict(self.executor.unwrap_model(self.model))
+        self.old_model.load_state_dict(self.executor.unwrap_model(self.model))
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
-        self._step += 1
-        if self._step % self.sync_interval == 0:
-            self.sync_ref_model()
-
         batch = move_to_device(batch, self.device)
         prompts = batch["prompts"]
         responses = batch["responses"]
@@ -313,33 +315,60 @@ 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_len = prompt_expanded.size(1)
 
         full_sequences = torch.cat([prompt_expanded, responses_flat], dim=-1)
-        full_masks = torch.cat([torch.ones_like(prompt_expanded), masks_flat], dim=-1)
+        # Prompt tokens are masked out (0) so logprobs are computed only for
-
+        # response tokens.  get_logprobs shifts the mask by one position, so
-        log_probs_policy = get_logprobs(
+        # the first response token's logprob (predicted from the last prompt
-            self.model, full_sequences, full_masks, self.reduction
+        # token) is correctly included.
-        )
+        full_masks = torch.cat([torch.zeros_like(prompt_expanded), masks_flat], dim=-1)
-        log_probs_policy = log_probs_policy.view(batch_size, group_size)
 
+        # 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"
+        )[:, prompt_len - 1 :]
         with torch.no_grad():
-            log_probs_ref = get_logprobs(
+            token_log_probs_old = get_logprobs(
-                self.ref_model, full_sequences, full_masks, self.reduction
+                self.old_model, full_sequences, full_masks, "none"
-            )
+            )[:, prompt_len - 1 :]
-            log_probs_ref = log_probs_ref.view(batch_size, group_size)
+            token_log_probs_ref = get_logprobs(
+                self.ref_model, full_sequences, full_masks, "none"
+            )[:, prompt_len - 1 :]
 
-        eps = torch.finfo(log_probs_policy.dtype).eps
+        # Reshape to [B, G, response_len]
+        token_log_probs_policy = token_log_probs_policy.view(batch_size, group_size, -1)
+        token_log_probs_old = token_log_probs_old.view(batch_size, group_size, -1)
+        token_log_probs_ref = token_log_probs_ref.view(batch_size, group_size, -1)
+        token_masks = masks_flat.view(batch_size, group_size, -1).float()
+
+        # Group-normalized advantages from scalar per-response rewards.
+        eps = 1e-8
         mean = rewards.mean(dim=-1, keepdim=True)
-        std = rewards.std(dim=-1, keepdim=True)
+        std = rewards.std(dim=-1, keepdim=True, unbiased=False)
         advantages = (rewards - mean) / (std + eps)
+        # Broadcast scalar advantage to every response token: [B, G, 1]
+        advantages = advantages.unsqueeze(-1)
 
-        ratio = torch.exp(log_probs_policy - log_probs_ref)
+        # Token-level ratio (π_θ / π_old) and PPO clipping.
+        log_ratio = token_log_probs_policy - token_log_probs_old
+        ratio = torch.exp(log_ratio)
 
         surr1 = ratio * advantages
         surr2 = torch.clamp(ratio, 1 - self.clip_eps, 1 + self.clip_eps) * advantages
+        per_token_policy_loss = -torch.min(surr1, surr2)
+        token_count = token_masks.sum().clamp(min=1.0)
+        policy_loss = (per_token_policy_loss * token_masks).sum() / token_count
+
+        # KL penalty to frozen reference model with k1 estimator (non-negative):
+        # k1 = π_ref / π_θ - log(π_ref / π_θ) - 1, where π_ref / π_θ = exp(log_ref - log_policy).
+        log_ref_ratio = token_log_probs_ref - token_log_probs_policy
+        r = torch.exp(log_ref_ratio)
+        kl_per_token = r - torch.log(r + eps) - 1.0
+        kl_penalty = self.kl_coef * (kl_per_token * token_masks).sum() / token_count
 
-        policy_loss = -torch.min(surr1, surr2).mean()
-        kl_penalty = self.kl_coef * (log_probs_policy - log_probs_ref).square().mean()
         total_loss = policy_loss + kl_penalty
 
         return total_loss

astrai/trainer/train_callback.py

@@ -14,7 +14,7 @@ from tqdm import tqdm
 
 from astrai.factory import BaseFactory
 from astrai.parallel import only_on_rank
-from astrai.parallel.setup import get_current_device, get_rank
+from astrai.parallel.setup import get_current_device
 from astrai.serialization import Checkpoint
 from astrai.trainer.metric_util import (
     ctx_get_grad_norm,
@@ -139,13 +139,16 @@ class CheckpointCallback(TrainCallback):
         self.interval = interval
         self.weight_only = weight_only
         self.save_extra_fn = save_extra_fn or CheckpointCallback.save_extra
-        self.last_ckpt_step = 0
+        self.last_ckpt_step = None
 
-    def _save_checkpoint(self, context: TrainContext):
+    def on_train_begin(self, context: TrainContext):
-        state_dict = context.executor.unwrap_model(context.model)
         self.last_ckpt_step = context.optimizer_step
 
-        if get_rank() == 0:
+    def _save_checkpoint(self, context: TrainContext):
+        self.last_ckpt_step = context.optimizer_step
+
+        with context.executor.checkpoint_context(context.model) as state_dict:
+            if state_dict is not None:
                 save_path = os.path.join(
                     self.save_dir,
                     f"epoch_{context.epoch}_step_{context.optimizer_step}",
@@ -209,9 +212,8 @@ class ProgressBarCallback(TrainCallback):
 
     @only_on_rank(0)
     def on_optimizer_step(self, context: TrainContext):
-        self.progress_bar.update(1)
         postfix = {
-            "step": context.optimizer_step,
+            "step": f"{context.optimizer_step:d}",
             "loss": f"{context.loss:.4f}",
             "lr": f"{context.optimizer.param_groups[-1]['lr']:.2e}",
         }
@@ -220,6 +222,7 @@ class ProgressBarCallback(TrainCallback):
         if context.val_loss is not None:
             postfix["val_loss"] = f"{context.val_loss:.4f}"
         self.progress_bar.set_postfix(postfix)
+        self.progress_bar.update(1)
 
     @only_on_rank(0)
     def on_epoch_end(self, context: TrainContext):
@@ -237,7 +240,7 @@ class MetricCallback(TrainCallback):
         metrics: List[str] = None,
         val_step: int = 0,
     ):
-        self.last_log_flush_step = 0
+        self.last_log_flush_step = None
         self.save_interval = save_interval
         self.metrics = metrics or ["loss", "lr"]
         self.val_step = val_step
@@ -298,6 +301,9 @@ class MetricCallback(TrainCallback):
         context.model.train()
         return avg_loss
 
+    def on_train_begin(self, context: TrainContext):
+        self.last_log_flush_step = context.optimizer_step
+
     @only_on_rank(0)
     def _flush(self, epoch, step):
         log_file = self.log_dir / f"epoch_{epoch}_step_{step}_metric.jsonl"
@@ -327,8 +333,12 @@ class MetricCallback(TrainCallback):
         self._append("epoch", context)
 
     def on_train_end(self, context):
-        if context.optimizer_step != self.last_log_flush_step:
+        if (
+            self.last_log_flush_step is None
+            or context.optimizer_step != self.last_log_flush_step
+        ):
             self._flush(context.epoch, context.optimizer_step)
+            self.last_log_flush_step = context.optimizer_step
 
     def on_error(self, context):
         self._flush(context.epoch, context.optimizer_step)

astrai/trainer/train_context.py

@@ -13,7 +13,7 @@ from astrai.parallel.executor import BaseExecutor, ExecutorFactory
 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
+from astrai.trainer.strategy import BaseStrategy, StrategyFactory, create_ref_model
 
 
 @dataclass
@@ -54,10 +54,12 @@ class TrainContextBuilder:
         config: TrainConfig,
     ):
         self.config = config
-        self._resume_dir: Optional[str] = None
+        self._param_path: Optional[str] = None
+        self._resume: bool = False
 
-    def with_resume_dir(self, resume_dir: Optional[str]) -> Self:
+    def with_param_path(self, param_path: Optional[str], resume: bool = False) -> Self:
-        self._resume_dir = resume_dir
+        self._param_path = param_path
+        self._resume = resume
         return self
 
     def build(self) -> TrainContext:
@@ -74,8 +76,8 @@ class TrainContextBuilder:
         model = model.to(device=device)
 
         model_config = {}
-        if self._resume_dir:
+        if self._param_path:
-            config_path = Path(self._resume_dir) / "config.json"
+            config_path = Path(self._param_path) / "config.json"
             if config_path.exists():
                 model_config = load_json(config_path)
 
@@ -91,17 +93,28 @@ class TrainContextBuilder:
             executor=executor,
         )
 
-        if self._resume_dir:
+        if self._param_path:
-            checkpoint = Checkpoint.load_any(self._resume_dir)
+            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:
-                    context.consumed_samples = checkpoint.consumed_samples
+                        per_step = (
+                            cfg.batch_per_device
+                            * context.world_size
+                            * cfg.grad_accum_steps
+                        )
+                        context.consumed_samples = (
+                            checkpoint.consumed_samples // per_step
+                        ) * per_step
                     else:
-                    context.consumed_samples = cfg.start_samples * context.world_size
+                        context.consumed_samples = (
+                            cfg.start_samples * context.world_size
+                        )
                     context.checkpoint = checkpoint
 
         if cfg.lora is not None:
@@ -177,13 +190,27 @@ class TrainContextBuilder:
                     if obj is not None:
                         obj.load_state_dict(extra[name])
 
+        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
+
+        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
+
         context.strategy = StrategyFactory.create(
             cfg.strategy,
             model=context.model,
             device=device,
             executor=executor,
             model_fn=cfg.model_fn,
-            **cfg.extra_kwargs,
+            **strategy_kwargs,
         )
 
         return context

astrai/trainer/trainer.py

@@ -52,9 +52,11 @@ class Trainer:
             if method:
                 method(context)
 
-    def _trainer_loop(self, resume_dir: Optional[str] = None):
+    def _trainer_loop(self, param_path: Optional[str] = None, resume: bool = False):
         context = (
-            TrainContextBuilder(self.train_config).with_resume_dir(resume_dir).build()
+            TrainContextBuilder(self.train_config)
+            .with_param_path(param_path, resume=resume)
+            .build()
         )
         executor = context.executor
         self._call_callbacks("on_train_begin", context)
@@ -95,7 +97,7 @@ class Trainer:
         finally:
             self._call_callbacks("on_train_end", context)
 
-    def train(self, resume_dir: Optional[str] = None):
+    def train(self, param_path: Optional[str] = None, resume: bool = False):
         cfg = self.train_config
         spawn_parallel_fn(
             self._trainer_loop,
@@ -105,5 +107,6 @@ class Trainer:
             master_port=cfg.master_port,
             device_type=cfg.device_type,
             start_method=cfg.start_method,
-            resume_dir=resume_dir,
+            param_path=param_path,
+            resume=resume,
         )

csrc/__init__.py

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

csrc/build.py

@@ -0,0 +1,48 @@
+from pathlib import Path
+
+
+def _arch_flags() -> list[str]:
+    import torch
+
+    if torch.cuda.is_available():
+        cap = torch.cuda.get_device_capability()
+    else:
+        cap = (8, 0)
+    ver = f"{cap[0]}{cap[1]}"
+    flags = [f"-gencode=arch=compute_{ver},code=sm_{ver}"]
+    # tensor-core mma path (mma.sync.m16n8k16.bf16) requires sm_80+; decide the
+    # kernel dispatch at build time via this define rather than at runtime.
+    if cap[0] < 8:
+        flags.append("-DASTRAI_NO_MMA")
+    return flags
+
+
+_kernels_dir = Path("csrc/kernels")
+REGISTRY: dict[str, dict] = {}
+
+CXX_FLAGS = ["-O3", "-funroll-loops"]
+NVCC_FLAGS = [
+    "-O3",
+    "--expt-relaxed-constexpr",
+    "--use_fast_math",
+    "--ptxas-options=-O3,-v",
+    "--extra-device-vectorization",
+    "--threads=8",
+]
+
+
+def register(name: str, sources: list[str] | None = None, **kwargs):
+    if sources is None:
+        sources = [str(_kernels_dir / f"{name}.cu")]
+    REGISTRY[name] = {
+        "sources": sources,
+        "cxx_flags": [*CXX_FLAGS],
+        "nvcc_flags": [*NVCC_FLAGS, *_arch_flags()],
+        "extra_link_args": kwargs.pop("extra_link_args", []),
+        **kwargs,
+    }
+
+
+register("attn_decode")
+register("attn_prefill")
+register("attn_paged_decode")

csrc/kernels/attn_common.h

@@ -0,0 +1,68 @@
+#pragma once
+
+
+template<typename T, typename AT = float>
+struct AttentionParams {
+    int batch;
+    int q_head;
+    int kv_head;
+    int q_len;
+    int kv_len;
+    int head_dim;
+    int use_mask;
+    int causal_offset;   // -1 = non-causal; >=0 = absolute position of first Q token
+    int num_splits;
+    float scale;
+
+    // Q strides (element offsets for each dim — layout-agnostic)
+    int q_stride_b, q_stride_h, q_stride_l, q_stride_d;
+    // KV strides (K and V share the same layout — only base pointers differ)
+    int kv_stride_b, kv_stride_h, kv_stride_l, kv_stride_d;
+
+    // Mask: 2D [batch, kv_len] (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
+
+    const T* __restrict__ q;
+    const T* __restrict__ k;
+    const T* __restrict__ v;
+    const bool* __restrict__ mask;
+
+    T* __restrict__ o;
+    AT* __restrict__ o_part;
+    AT* __restrict__ ml_part;
+};
+
+template<typename T, typename AT = float>
+struct PagedAttentionParams {
+    int batch;
+    int q_head;
+    int kv_head;
+    int q_len;
+    int kv_len;
+    int head_dim;
+    int use_mask;
+    int causal_offset;
+    float scale;
+
+    int num_splits;
+    int page_size;
+    int max_pages;
+
+    // Q strides (layout-agnostic)
+    int q_stride_b, q_stride_h, q_stride_l, q_stride_d;
+
+    // Mask strides (2D or 3D)
+    int mask_b_stride;
+    int mask_q_stride;
+
+    const T* __restrict__ q;
+    const T* __restrict__ k_cache;
+    const T* __restrict__ v_cache;
+    const bool* __restrict__ mask;
+    const int64_t* __restrict__ page_table;
+
+    T* __restrict__ o;
+    AT* __restrict__ o_part;
+    AT* __restrict__ ml_part;
+};

csrc/kernels/attn_decode.cu

@@ -0,0 +1,82 @@
+#include "attn_decode_split_kv.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,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout
+) {
+    AttentionParams<bf16> p;
+    attn_pack_params(q, k, v, mask, causal_offset, scale, layout, p);
+    TORCH_CHECK(p.q_len == 1, "Q seq_len must be 1");
+    TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
+
+    // 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();
+
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_decode, p);
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_decode", &attn_decode,
+        py::arg("q"),
+        py::arg("k"),
+        py::arg("v"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = 0,
+        "GQA decode (tensor-core head-packing on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_decode_split_kv.cuh

@@ -0,0 +1,132 @@
+#pragma once
+#include <cuda_bf16.h>
+#include <float.h>
+#include "attn_common.h"
+
+using bf16 = __nv_bfloat16;
+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;
+}
+
+__global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
+    int batch = blockIdx.x / p.kv_head;
+    int kv_head = blockIdx.x % p.kv_head;
+    int split = blockIdx.z;
+    int group_size = blockDim.y;
+    int q_head = kv_head * group_size + threadIdx.y;
+    int lane = threadIdx.x;
+    int hd_per_thread = p.head_dim / 32;
+
+    // Q: [batch, q_head, q_len=1, head_dim] — stride-based
+    float q_reg[8];
+    int q_off = batch * p.q_stride_b + q_head * p.q_stride_h
+              + lane * hd_per_thread * p.q_stride_d;
+    for (int i = 0; i < hd_per_thread; i++)
+        q_reg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
+
+    // KV: [batch, kv_head, kv_len, head_dim] — stride-based base
+    int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    int mask_base = batch * p.mask_b_stride;
+
+    float m = -FLT_MAX, d = 0.0f, acc_reg[8] = {0.0f};
+
+    extern __shared__ __align__(16) bf16 k_smem[];
+
+    // Split-KV: each split processes a contiguous subset of chunks
+    int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
+    int chunks_per_split = (chunks_total + p.num_splits - 1) / p.num_splits;
+    int ch_begin = split * chunks_per_split;
+    int ch_end = min(chunks_total, ch_begin + chunks_per_split);
+
+    for (int ci = ch_begin; ci < ch_end; ci++) {
+        int chunk_start = ci * DC_CHUNK;
+        int this_chunk = min(DC_CHUNK, p.kv_len - chunk_start);
+
+        // Load K into shared memory (gather from strided global)
+        int total = this_chunk * p.head_dim;
+        for (int i = threadIdx.y * 32 + lane; i < total; i += blockDim.x * blockDim.y) {
+            int s = i / p.head_dim;
+            int d_dim = i % p.head_dim;
+            int kv_idx = chunk_start + s;
+            int g_off = kv_base + kv_idx * p.kv_stride_l + d_dim * p.kv_stride_d;
+            k_smem[i] = p.k[g_off];
+        }
+        __syncthreads();
+
+        for (int s = 0; s < this_chunk; s++) {
+            float partial = 0.0f;
+            for (int i = 0; i < hd_per_thread; i++)
+                partial += q_reg[i] * __bfloat162float(k_smem[s * p.head_dim + lane * hd_per_thread + i]);
+            partial = warp_reduce_sum(partial) * p.scale;
+
+            int kv_idx = chunk_start + s;
+            if (p.use_mask && p.mask && !p.mask[mask_base + kv_idx])
+                partial = -FLT_MAX;
+            if (p.causal_offset >= 0 && kv_idx > p.causal_offset)
+                partial = -FLT_MAX;
+
+            float new_m = fmaxf(m, partial);
+            float alpha = expf(m - new_m);
+            float beta  = expf(partial - new_m);
+            d = d * alpha + beta;
+
+            // V: stride-based read
+            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;
+            m = new_m;
+        }
+        __syncthreads();
+    }
+
+    // ---- write UN-normalised partials for this split ----
+    size_t bh = (size_t)batch * p.q_head + q_head;
+    size_t slot = bh * p.num_splits + split;
+    int d0 = lane * hd_per_thread;
+    for (int i = 0; i < hd_per_thread; i++) {
+        int dd = d0 + i;
+        p.o_part[slot * p.head_dim + dd] = acc_reg[i];
+    }
+    if (lane == 0) {
+        p.ml_part[slot * 2] = m;
+        p.ml_part[slot * 2 + 1] = d;
+    }
+}
+
+// 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;
+    if (d >= p.head_dim) return;
+
+    int batch = bh / p.q_head;
+    int q_head = bh % p.q_head;
+
+    size_t split_base = (size_t)bh * p.num_splits;
+    const float* mlp = p.ml_part + split_base * 2;
+    const float* op = p.o_part + split_base * p.head_dim;
+
+    float m = -FLT_MAX, l = 0.0f, acc = 0.0f;
+    for (int s = 0; s < p.num_splits; s++) {
+        float mi = mlp[s * 2];
+        if (mi <= -FLT_MAX) continue;
+        float li = mlp[s * 2 + 1];
+        float nm = fmaxf(m, mi);
+        float corr = __expf(m - nm);
+        float e = __expf(mi - nm);
+        acc = acc * corr + op[s * p.head_dim + d] * e;
+        l = 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

@@ -0,0 +1,176 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+
+using bf16 = __nv_bfloat16;
+
+// Split-K (FlashDecoding) tensor-core decode via GQA head-packing.
+//
+// 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>
+__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 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];
+
+    // ---- 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,
+                          qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[DN8][4];
+#pragma unroll
+    for (int j = 0; j < 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_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 ----
+    auto load_tile = [&](int ti, int buf) {
+        int kv0 = ti * BC;
+        bf16* dK = sK + buf * BC * LD;
+        bf16* dV = sV + buf * BC * LD;
+#pragma unroll
+        for (int i = lane * VEC; i < TOTAL; i += 32 * VEC) {
+            int r = i / HEAD_DIM, d = i % 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 g_off = kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
+            cp_async_16_pred(&dK[off], &p.k[g_off], valid);
+            cp_async_16_pred(&dV[off], &p.v[g_off], valid);
+        }
+        cp_async_commit();
+    };
+
+    // ---- Prologue: issue first tile load ----
+    if (ti_begin < ti_end) {
+        load_tile(ti_begin, 0);
+    }
+
+    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;
+
+        // 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);
+
+        #pragma unroll
+        for (int n8 = 0; n8 < 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,
+                                    0, 0,
+                                    p.mask_b_stride, 0,
+                                    batch,
+                                    p.mask, has_mask,
+                                    Sacc, Oacc, m0, m1, l0, l1, lane);
+
+        mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);
+        __syncwarp();
+
+        if constexpr (STAGES == 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, 0);
+        }
+    }
+
+    // ---- write UN-normalised partials for this split ----
+    auto split_slot = [&](int h) -> size_t {
+        size_t bh = (size_t)batch * p.q_head + h;
+        return bh * p.num_splits + split;
+    };
+#pragma unroll
+    for (int dn8 = 0; dn8 < 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;
+            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;
+            op[d] = Oacc[dn8][2];
+            op[d + 1] = Oacc[dn8][3];
+        }
+    }
+    if (tid4 == 0) {
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m0; mp[1] = l0;
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m1; mp[1] = l1;
+        }
+    }
+}

csrc/kernels/attn_entry_utils.cuh

@@ -0,0 +1,177 @@
+#pragma once
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "attn_common.h"
+
+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.
+#define DISPATCH_HEAD_DIM(hd, fn, arg) \
+    switch (hd) { \
+        case 32:  fn<32>(arg); break; \
+        case 64:  fn<64>(arg); break; \
+        case 128: fn<128>(arg); break; \
+        case 256: fn<256>(arg); break; \
+        default: \
+            TORCH_CHECK(false, "unsupported head_dim ", hd, \
+                         " (supported: 32, 64, 128, 256)"); \
+    }
+
+template<typename P>
+inline void alloc_split_partials(P& p) {
+    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
+    auto o_part = torch::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);
+    p.o_part = (float*)o_part.data_ptr();
+    p.ml_part = (float*)ml_part.data_ptr();
+}
+
+// ---- Shared Q-dims + strides extraction ----
+template <typename P>
+inline void extract_q_dims_and_strides(torch::Tensor& q, int64_t layout, P& p) {
+    if (layout == 1) q = q.transpose(1, 2);
+    p.batch = (int)q.size(0);
+    p.q_head = (int)q.size(1);
+    p.q_len = (int)q.size(2);
+    p.head_dim = (int)q.size(3);
+    p.q_stride_b = (int)q.stride(0);
+    p.q_stride_h = (int)q.stride(1);
+    p.q_stride_l = (int)q.stride(2);
+    p.q_stride_d = (int)q.stride(3);
+}
+
+// ---- Shared mask packing ----
+template <typename P>
+inline void pack_mask(const c10::optional<torch::Tensor>& mask, P& p) {
+    if (p.use_mask) {
+        auto m = mask.value();
+        TORCH_CHECK(m.is_cuda(), "mask must be on CUDA");
+        TORCH_CHECK(m.dtype() == torch::kBool, "mask must be bool");
+        TORCH_CHECK(m.size(0) == p.batch, "mask batch mismatch");
+        TORCH_CHECK(m.size(m.dim() - 1) == p.kv_len, "mask kv_len mismatch");
+        if (m.dim() == 2) {
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_q_stride = 0;
+        } else if (m.dim() == 3) {
+            TORCH_CHECK(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);
+        } else {
+            TORCH_CHECK(false, "mask must be 2D [batch, kv_len] or 3D [batch, q_len, kv_len]");
+        }
+        p.mask = m.data_ptr<bool>();
+    } else {
+        p.mask = nullptr;
+        p.mask_b_stride = 0;
+        p.mask_q_stride = 0;
+    }
+}
+
+// ---- attn_pack_params (contiguous KV) ----
+template<typename T>
+inline void attn_pack_params(
+    torch::Tensor q,
+    torch::Tensor k,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout,
+    AttentionParams<T>& p
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+
+    TORCH_CHECK(q.is_cuda() && k.is_cuda() && v.is_cuda());
+    TORCH_CHECK(q.dtype() == torch::kBFloat16);
+    TORCH_CHECK(k.dtype() == torch::kBFloat16);
+    TORCH_CHECK(v.dtype() == torch::kBFloat16);
+    TORCH_CHECK(k.sizes() == v.sizes(), "K and V must have identical shapes");
+    TORCH_CHECK(q.dim() == 4 && k.dim() == 4, "Q/K/V must be 4D");
+
+    extract_q_dims_and_strides(q, layout, p);
+
+    if (layout == 1) k = k.transpose(1, 2), v = v.transpose(1, 2);
+
+    p.kv_head = (int)k.size(1);
+    p.kv_len = (int)k.size(2);
+    TORCH_CHECK(k.size(3) == p.head_dim, "K/V head_dim must match Q");
+
+    p.kv_stride_b = (int)k.stride(0);
+    p.kv_stride_h = (int)k.stride(1);
+    p.kv_stride_l = (int)k.stride(2);
+    p.kv_stride_d = (int)k.stride(3);
+
+    p.causal_offset = (int)causal_offset;
+    p.use_mask = mask.has_value() ? 1 : 0;
+    p.scale = (scale > 0.0) ? (float)scale : 1.0f / sqrtf((float)p.head_dim);
+
+    p.q = (const T*)q.data_ptr();
+    p.k = (const T*)k.data_ptr();
+    p.v = (const T*)v.data_ptr();
+    p.o = nullptr;
+    p.o_part = nullptr;
+    p.ml_part = nullptr;
+
+    pack_mask(mask, p);
+}
+
+// ---- attn_pack_paged_params ----
+template<typename T>
+inline void attn_pack_paged_params(
+    torch::Tensor q,
+    torch::Tensor page_table,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    int64_t page_size,
+    int64_t kv_len,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout,
+    PagedAttentionParams<T>& p
+) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+
+    TORCH_CHECK(q.is_cuda() && page_table.is_cuda() && k_cache.is_cuda() && v_cache.is_cuda());
+    TORCH_CHECK(q.dtype() == torch::kBFloat16, "q must be bf16");
+    TORCH_CHECK(k_cache.dtype() == torch::kBFloat16, "k_cache must be bf16");
+    TORCH_CHECK(v_cache.dtype() == torch::kBFloat16, "v_cache must be bf16");
+    TORCH_CHECK(page_table.dtype() == torch::kLong, "page_table must be int64");
+    TORCH_CHECK(k_cache.sizes() == v_cache.sizes(), "k_cache and v_cache must have identical shapes");
+
+    extract_q_dims_and_strides(q, layout, p);
+
+    p.kv_head = (int)k_cache.size(2);
+    p.kv_len = (int)kv_len;
+    p.page_size = (int)page_size;
+    p.max_pages = (int)page_table.size(1);
+
+    TORCH_CHECK(q.size(2) == 1, "Q seq_len must be 1 (decode)");
+    TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
+    TORCH_CHECK(k_cache.size(1) == page_size,
+                "k_cache dim 1 must equal page_size, got ",
+                k_cache.size(1), " vs ", page_size);
+
+    p.causal_offset = (int)causal_offset;
+    p.use_mask = (mask.has_value() && mask.value().defined()) ? 1 : 0;
+    p.scale = (scale > 0.0) ? (float)scale : 1.0f / sqrtf((float)p.head_dim);
+
+    p.page_table = page_table.data_ptr<int64_t>();
+    p.k_cache = (const T*)k_cache.data_ptr();
+    p.v_cache = (const T*)v_cache.data_ptr();
+    p.q = (const T*)q.data_ptr();
+    p.o = nullptr;
+    p.o_part = nullptr;
+    p.ml_part = nullptr;
+
+    pack_mask(mask, p);
+}

csrc/kernels/attn_mma_utils.cuh

@@ -0,0 +1,293 @@
+#pragma once
+#include <cfloat>
+#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.
+
+// 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(
+        "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+        "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};"
+        : "=f"(d[0]), "=f"(d[1]), "=f"(d[2]), "=f"(d[3])
+        : "r"(a[0]), "r"(a[1]), "r"(a[2]), "r"(a[3]), "r"(b[0]), "r"(b[1]),
+          "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]));
+}
+
+// read two adjacent bf16 from smem as one packed .b32 (elem0 low, elem1 high)
+__device__ __forceinline__ unsigned ld2(const bf16* p) {
+    return *reinterpret_cast<const unsigned*>(p);
+}
+
+// pack two floats into one bf16x2 as .b32
+__device__ __forceinline__ unsigned pk2(float a, float b) {
+    __nv_bfloat162 v = __floats2bfloat162_rn(a, b);
+    return *reinterpret_cast<unsigned*>(&v);
+}
+
+// pack two (non-contiguous) bf16 into one .b32
+__device__ __forceinline__ unsigned pkb(bf16 a, bf16 b) {
+    __nv_bfloat162 v;
+    v.x = a;
+    v.y = b;
+    return *reinterpret_cast<unsigned*>(&v);
+}
+
+// ldmatrix: cooperatively load mma fragments from smem (one instruction per
+// 16x16 / 16x8 tile) with the exact register layout mma expects — 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.
+__device__ __forceinline__ void ldmatrix_x4(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x4.shared.b16 {%0,%1,%2,%3}, [%4];"
+                 : "=r"(r[0]), "=r"(r[1]), "=r"(r[2]), "=r"(r[3])
+                 : "r"(a));
+}
+__device__ __forceinline__ void ldmatrix_x2(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x2.shared.b16 {%0,%1}, [%2];"
+                 : "=r"(r[0]), "=r"(r[1])
+                 : "r"(a));
+}
+__device__ __forceinline__ void ldmatrix_x2_trans(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x2.trans.shared.b16 {%0,%1}, [%2];"
+                 : "=r"(r[0]), "=r"(r[1])
+                 : "r"(a));
+}
+
+// XOR swizzle for shared-memory column at 8-bf16 chunk granularity.
+// 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+.
+__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.
+__device__ __forceinline__ void cp_async_16_pred(bf16* smem_ptr,
+                                                  const void* gmem_ptr,
+                                                  bool pred) {
+    unsigned smem_addr = __cvta_generic_to_shared(smem_ptr);
+    int src_size = pred ? 16 : 0;
+    asm volatile("cp.async.ca.shared.global [%0], [%1], 16, %2;"
+                 :: "r"(smem_addr), "l"(gmem_ptr), "r"(src_size));
+}
+
+__device__ __forceinline__ void cp_async_commit() {
+    asm volatile("cp.async.commit_group;");
+}
+
+__device__ __forceinline__ void cp_async_wait_all() {
+    asm volatile("cp.async.wait_all;");
+}
+
+// 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));
+}
+
+// ---------------------------------------------------------------------------
+// Q-load: load query rows directly from global memory into mma A-operand
+// register layout. One call replaces ~15 duplicated lines in each MMA kernel.
+// stride_row is p.q_stride_h for decode (q_len=1, G heads) or
+//              p.q_stride_l for prefill (multi-q rows).
+// ---------------------------------------------------------------------------
+template <int KD>
+__device__ inline void load_q_mma_frags(
+    const bf16* __restrict__ q,
+    int stride_row,
+    int stride_d,
+    int qra, int qrb,
+    bool va, bool vb,
+    int tid4,
+    unsigned Qa[KD][4])
+{
+    #pragma unroll
+    for (int kt = 0; kt < KD; kt++) {
+        int c = kt * 16 + tid4 * 2;
+        const unsigned* pau = reinterpret_cast<const unsigned*>(
+            &q[qra * stride_row + c * stride_d]);
+        const unsigned* pbu = reinterpret_cast<const unsigned*>(
+            &q[qrb * stride_row + c * stride_d]);
+        Qa[kt][0] = va ? pau[0] : 0u;
+        Qa[kt][1] = vb ? pbu[0] : 0u;
+        Qa[kt][2] = va ? pau[4] : 0u;
+        Qa[kt][3] = vb ? pbu[4] : 0u;
+    }
+}
+
+// ---------------------------------------------------------------------------
+// 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>
+__device__ inline void mma_compute_scores(
+    const unsigned Qa[KD][4],
+    const bf16* __restrict__ sK,
+    int LD, 
+    int SWIZ_MASK, 
+    int lane,
+    float Sacc[NC8][4])
+{
+    #pragma unroll
+    for (int n8 = 0; n8 < 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++) {
+            unsigned b[2];
+            ldmatrix_x2(b, &sK[krow_l * LD + swiz_col(kt * 16 + kcol_h, krow_l, 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>
+__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,
+    const bool* __restrict__ mask,
+    bool has_mask,
+    float Sacc[NC8][4],
+    float Oacc[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;
+    #pragma unroll
+    for (int n8 = 0; n8 < 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]);
+        float s0 = b0 ? -FLT_MAX : Sacc[n8][0];
+        float s1 = b1 ? -FLT_MAX : Sacc[n8][1];
+        float s2 = b2 ? -FLT_MAX : Sacc[n8][2];
+        float s3 = b3 ? -FLT_MAX : Sacc[n8][3];
+        Sacc[n8][0] = s0; Sacc[n8][1] = s1;
+        Sacc[n8][2] = s2; Sacc[n8][3] = s3;
+        rmax0 = fmaxf(rmax0, fmaxf(s0, s1));
+        rmax1 = fmaxf(rmax1, fmaxf(s2, s3));
+    }
+    // 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++) {
+        float p0 = pn0 * __expf(Sacc[n8][0] - nm0);
+        float p1 = pn0 * __expf(Sacc[n8][1] - nm0);
+        float p2 = pn1 * __expf(Sacc[n8][2] - nm1);
+        float p3 = pn1 * __expf(Sacc[n8][3] - nm1);
+        Sacc[n8][0] = p0; Sacc[n8][1] = p1;
+        Sacc[n8][2] = p2; Sacc[n8][3] = p3;
+        rsum0 += p0 + p1;
+        rsum1 += p2 + p3;
+    }
+    rsum0 += __shfl_xor_sync(0xFFFFFFFF, rsum0, 1);
+    rsum0 += __shfl_xor_sync(0xFFFFFFFF, rsum0, 2);
+    rsum1 += __shfl_xor_sync(0xFFFFFFFF, rsum1, 1);
+    rsum1 += __shfl_xor_sync(0xFFFFFFFF, rsum1, 2);
+    l0 = l0 * corr0 + rsum0;
+    l1 = l1 * corr1 + rsum1;
+    m0 = nm0; m1 = nm1;
+
+    #pragma unroll
+    for (int j = 0; j < 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>
+__device__ inline void mma_pv_accumulate(
+    float Sacc[][4],
+    const bf16* __restrict__ sV,
+    int LD, int SWIZ_MASK, int lane,
+    float Oacc[DN8][4])
+{
+    #pragma unroll
+    for (int kt2 = 0; kt2 < KT2; kt2++) {
+        unsigned Pa[4];
+        Pa[0] = pk2(Sacc[kt2 * 2][0], Sacc[kt2 * 2][1]);
+        Pa[1] = pk2(Sacc[kt2 * 2][2], Sacc[kt2 * 2][3]);
+        Pa[2] = pk2(Sacc[kt2 * 2 + 1][0], Sacc[kt2 * 2 + 1][1]);
+        Pa[3] = pk2(Sacc[kt2 * 2 + 1][2], Sacc[kt2 * 2 + 1][3]);
+        int vrow_l = kt2 * 16 + (lane & 15);
+        #pragma unroll
+        for (int dn8 = 0; dn8 < DN8; dn8++) {
+            unsigned b[2];
+            ldmatrix_x2_trans(b, &sV[vrow_l * LD + swiz_col(dn8 * 8, vrow_l, SWIZ_MASK)]);
+            mma16816(Oacc[dn8], Pa, b, Oacc[dn8]);
+        }
+    }
+}

csrc/kernels/attn_paged_decode.cu

@@ -0,0 +1,82 @@
+#include "attn_paged_decode_split_kv.cuh"
+#ifndef ASTRAI_NO_MMA
+#include "attn_paged_decode_split_kv_mma.cuh"
+#endif
+
+#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,
+    torch::Tensor k_cache,
+    torch::Tensor v_cache,
+    int64_t page_size,
+    int64_t kv_len,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout
+) {
+    PagedAttentionParams<bf16> p;
+    attn_pack_paged_params(q, page_table, k_cache, v_cache,
+                           page_size, kv_len, mask, causal_offset, scale, layout, p);
+
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
+    auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_decode, p);
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_paged_decode", &attn_paged_decode,
+        py::arg("q"),
+        py::arg("page_table"),
+        py::arg("k_cache"),
+        py::arg("v_cache"),
+        py::arg("page_size"),
+        py::arg("kv_len"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = 0,
+        "Paged GQA decode — split-KV with direct page-table access.");
+}

csrc/kernels/attn_paged_decode_split_kv.cuh

@@ -0,0 +1,147 @@
+#pragma once
+#include <cuda_bf16.h>
+#include <float.h>
+#include "attn_common.h"
+
+using bf16 = __nv_bfloat16;
+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.
+__global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p) {
+    int batch = blockIdx.x / p.kv_head;
+    int kv_head = blockIdx.x % p.kv_head;
+    int split = blockIdx.z;
+    int group_size = blockDim.y;
+    int q_head = kv_head * group_size + threadIdx.y;
+    int lane = threadIdx.x;
+    int hd_per_thread = p.head_dim / 32;
+
+    // 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;
+    #pragma unroll
+    for (int i = 0; i < hd_per_thread; i++)
+        q_reg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
+
+    float m = -FLT_MAX, d = 0.0f, acc_reg[8] = {0.0f};
+
+    extern __shared__ __align__(16) bf16 k_smem[];
+
+    int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_CHUNK;
+    int chunks_per_split = (chunks_total + p.num_splits - 1) / p.num_splits;
+    int ch_begin = split * chunks_per_split;
+    int ch_end = min(chunks_total, ch_begin + chunks_per_split);
+
+    const int mask_base = batch * p.mask_b_stride;
+
+    for (int ci = ch_begin; ci < ch_end; ci++) {
+        int chunk_start = ci * PDC_CHUNK;
+        int this_chunk = min(PDC_CHUNK, p.kv_len - chunk_start);
+
+        int total = this_chunk * p.head_dim;
+        for (int i = threadIdx.y * 32 + lane; i < total; i += blockDim.x * blockDim.y) {
+            int s = i / p.head_dim;
+            int d_dim = i % p.head_dim;
+            int pos = chunk_start + s;
+            int logical_page = pos / p.page_size;
+            int page_offset = pos % p.page_size;
+            int phys_page = p.page_table[batch * p.max_pages + logical_page];
+            if (phys_page >= 0) {
+                int64_t off = (int64_t)phys_page * p.page_size * p.kv_head * p.head_dim
+                            + (int64_t)page_offset * p.kv_head * p.head_dim
+                            + (int64_t)kv_head * p.head_dim
+                            + d_dim;
+                k_smem[i] = p.k_cache[off];
+            } else {
+                k_smem[i] = __float2bfloat16(0.0f);
+            }
+        }
+        __syncthreads();
+
+        for (int s = 0; s < this_chunk; s++) {
+            float partial = 0.0f;
+            #pragma unroll
+            for (int i = 0; i < hd_per_thread; i++)
+                partial += q_reg[i] * __bfloat162float(k_smem[s * p.head_dim + lane * hd_per_thread + i]);
+            partial = paged_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)
+                partial = -FLT_MAX;
+
+            float new_m = fmaxf(m, partial);
+            float alpha = expf(m - new_m);
+            float beta  = expf(partial - new_m);
+            d = d * alpha + beta;
+
+            int pos = chunk_start + s;
+            int logical_page = pos / p.page_size;
+            int page_offset = pos % p.page_size;
+            int phys_page = p.page_table[batch * p.max_pages + logical_page];
+            if (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;
+            } else {
+                #pragma unroll
+                for (int i = 0; i < hd_per_thread; i++)
+                    acc_reg[i] = acc_reg[i] * alpha + 0.0f * beta;
+            }
+            m = new_m;
+        }
+        __syncthreads();
+    }
+
+    size_t bh = (size_t)batch * p.q_head + q_head;
+    size_t slot = bh * p.num_splits + split;
+    int d0 = lane * hd_per_thread;
+    #pragma unroll
+    for (int i = 0; i < hd_per_thread; i++)
+        p.o_part[slot * p.head_dim + (d0 + i)] = acc_reg[i];
+    if (lane == 0) {
+        p.ml_part[slot * 2] = m;
+        p.ml_part[slot * 2 + 1] = d;
+    }
+}
+
+__global__ void paged_attn_decode_combine_kernel(PagedAttentionParams<bf16> p) {
+    int bh = blockIdx.x;
+    int d = threadIdx.x;
+    if (d >= p.head_dim) return;
+
+    int batch = bh / p.q_head;
+    int q_head = bh % p.q_head;
+
+    size_t split_base = (size_t)bh * p.num_splits;
+    const float* mlp = p.ml_part + split_base * 2;
+    const float* op = p.o_part + split_base * p.head_dim;
+
+    float m = -FLT_MAX, l = 0.0f, acc = 0.0f;
+    for (int s = 0; s < p.num_splits; s++) {
+        float mi = mlp[s * 2];
+        if (mi <= -FLT_MAX) continue;
+        float li = mlp[s * 2 + 1];
+        float nm = fmaxf(m, mi);
+        float corr = __expf(m - nm);
+        float e = __expf(mi - nm);
+        acc = acc * corr + op[s * p.head_dim + d] * e;
+        l = l * corr + li * e;
+        m = nm;
+    }
+
+    float inv = (l > 1e-20f) ? (1.0f / l) : 0.0f;
+    int o_off = batch * p.q_stride_b + q_head * p.q_stride_h + d * p.q_stride_d;
+    p.o[o_off] = __float2bfloat16(acc * inv);
+}

csrc/kernels/attn_paged_decode_split_kv_mma.cuh

@@ -0,0 +1,170 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+
+using bf16 = __nv_bfloat16;
+
+// 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>
+__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 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];
+
+    // ---- 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,
+                          qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[DN8][4];
+#pragma unroll
+    for (int j = 0; j < 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_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;
+
+    // ---- Load tile lambda: predicated cp.async, paged addressing ----
+    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);
+#pragma unroll
+        for (int i = lane * VEC; i < TOTAL; i += 32 * VEC) {
+            int r = i / HEAD_DIM, d = i % HEAD_DIM;
+            int kc = kv0 + r;
+            bool valid = (kc < p.kv_len) && page_valid;
+            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);
+            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);
+    }
+
+    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;
+
+        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);
+
+#pragma unroll
+        for (int n8 = 0; n8 < 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,
+                                    0, 0,
+                                    p.mask_b_stride, 0,
+                                    batch,
+                                    p.mask, has_mask,
+                                    Sacc, Oacc, m0, m1, l0, l1, lane);
+
+        mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);
+        __syncwarp();
+
+        if constexpr (STAGES == 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, 0);
+        }
+    }
+
+    // ---- write UN-normalised partials for this split ----
+    auto split_slot = [&](int h) -> size_t {
+        size_t bh = (size_t)batch * p.q_head + h;
+        return bh * p.num_splits + split;
+    };
+#pragma unroll
+    for (int dn8 = 0; dn8 < 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;
+            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;
+            op[d] = Oacc[dn8][2];
+            op[d + 1] = Oacc[dn8][3];
+        }
+    }
+    if (tid4 == 0) {
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m0; mp[1] = l0;
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m1; mp[1] = l1;
+        }
+    }
+}

csrc/kernels/attn_prefill.cu

@@ -0,0 +1,64 @@
+#include "attn_prefill_split_q.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,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout
+) {
+    AttentionParams<bf16> p;
+    attn_pack_params(q, k, v, mask, causal_offset, scale, layout, p);
+    TORCH_CHECK(p.head_dim % 16 == 0, "head_dim must be multiple of 16");
+
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
+    auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_prefill, p);
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_prefill", &attn_prefill,
+        py::arg("q"),
+        py::arg("k"),
+        py::arg("v"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = 0,
+        "GQA prefill (tensor-core mma on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_prefill_split_q.cuh

@@ -0,0 +1,152 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+
+using bf16 = __nv_bfloat16;
+
+// v9: group-split register blocking. G threads cooperate on one query row,
+// each owning HEAD_DIM/G dims of qreg[]/acc[]. 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.
+
+template <int G>
+__device__ __forceinline__ float group_reduce_sum(float v, unsigned mask) {
+#pragma unroll
+    for (int o = G / 2; o > 0; o >>= 1)
+        v += __shfl_xor_sync(mask, v, o);
+    return v;
+}
+
+// load 8 contiguous bf16 from (16-byte aligned) smem as one float4, unpack to
+// 8 floats — cuts shared-load instructions 8x vs scalar bf16 loads.
+__device__ __forceinline__ void ld8(const bf16* p, float* o) {
+    float4 raw = *reinterpret_cast<const float4*>(p);
+    const __nv_bfloat162* h = reinterpret_cast<const __nv_bfloat162*>(&raw);
+#pragma unroll
+    for (int j = 0; j < 4; j++) {
+        float2 f = __bfloat1622float2(h[j]);
+        o[2 * j] = f.x;
+        o[2 * j + 1] = f.y;
+    }
+}
+
+template <int HEAD_DIM, int G, int ROWS, int P_BC>
+__global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
+    constexpr int DPT = HEAD_DIM / G;
+
+    int q_tile = blockIdx.x;
+    int q_head = blockIdx.y;
+    int batch  = blockIdx.z;
+    int gpos   = threadIdx.x;  // 0..G-1  (which d-chunk)
+    int row    = threadIdx.y;  // 0..ROWS-1
+    int q_row  = q_tile * ROWS + row;
+
+    int kv_head = q_head / (p.q_head / p.kv_head);
+
+    __shared__ __align__(16) bf16 sK[P_BC * HEAD_DIM];
+    __shared__ __align__(16) bf16 sV[P_BC * HEAD_DIM];
+
+    // Q: stride-based load [batch, q_head, q_len, head_dim]
+    float qreg[DPT];
+    if (q_row < p.q_len) {
+        int q_off = batch * p.q_stride_b + q_head * p.q_stride_h
+                  + q_row * p.q_stride_l + gpos * DPT * p.q_stride_d;
+#pragma unroll
+        for (int i = 0; i < DPT; i++)
+            qreg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]) * p.scale;
+    }
+
+    float m = -FLT_MAX, l = 0.0f;
+    float acc[DPT];
+#pragma unroll
+    for (int i = 0; i < DPT; i++)
+        acc[i] = 0.0f;
+
+    // KV: stride-based base
+    int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    int mask_batch_base = batch * p.mask_b_stride;
+    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)));
+
+    for (int ti = 0; ti < tiles; ti++) {
+        int kv0  = ti * P_BC;
+        int tlen = min(P_BC, p.kv_len - kv0);
+
+        // Load K/V into shared memory from strided global
+        for (int i = lid; i < tlen * HEAD_DIM; i += tt) {
+            int s = i / HEAD_DIM;
+            int d_dim = i % HEAD_DIM;
+            int kv_idx = kv0 + s;
+            int g_off = kv_base + kv_idx * p.kv_stride_l + d_dim * p.kv_stride_d;
+            sK[i] = p.k[g_off];
+            sV[i] = p.v[g_off];
+        }
+        __syncthreads();
+
+        int lim = tlen;
+        if (p.causal_offset >= 0 && q_row < p.q_len) {
+            int ep = q_row + p.causal_offset + 1;
+            if (kv0 >= ep)
+                lim = 0;
+            else if (kv0 + tlen > ep)
+                lim = ep - kv0;
+        }
+
+        int mask_row_base = mask_batch_base + q_row * p.mask_q_stride;
+        for (int s = 0; s < lim; s++) {
+            const bf16* kr = sK + s * HEAD_DIM + gpos * DPT;
+            float part = 0.0f;
+#pragma unroll
+            for (int i = 0; i < DPT; i += 8) {
+                float k8[8];
+                ld8(kr + i, k8);
+#pragma unroll
+                for (int j = 0; j < 8; j++)
+                    part = fmaf(qreg[i + j], k8[j], part);
+            }
+            float dot = group_reduce_sum<G>(part, gmask);
+
+            int kv_idx = kv0 + s;
+            if (p.use_mask && p.mask && !p.mask[mask_row_base + kv_idx])
+                dot = -FLT_MAX;
+
+            float nm = fmaxf(m, dot);
+            float al = __expf(m - nm);
+            float be = __expf(dot - nm);
+            l = l * al + be;
+
+            const bf16* vr = sV + s * HEAD_DIM + gpos * DPT;
+#pragma unroll
+            for (int i = 0; i < DPT; i += 8) {
+                float v8[8];
+                ld8(vr + i, v8);
+#pragma unroll
+                for (int j = 0; j < 8; j++)
+                    acc[i + j] = fmaf(v8[j], be, acc[i + j] * al);
+            }
+            m = nm;
+        }
+        __syncthreads();
+    }
+
+    if (q_row < p.q_len) {
+        // 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;
+#pragma unroll
+        for (int i = 0; i < DPT; i++)
+            p.o[o_off + i * p.q_stride_d] = __float2bfloat16(acc[i] * rl);
+    }
+}

csrc/kernels/attn_prefill_split_q_mma.cuh

@@ -0,0 +1,196 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#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.
+//
+// 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.
+//
+// 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>
+__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 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;
+
+    // ---- 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];
+
+    // 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,
+                          qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[DN8][4];
+#pragma unroll
+    for (int j = 0; j < 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
+
+    // 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;
+    const int block_max_kv =
+        blockIdx.x * WARPS * BR + WARPS * BR - 1 + p.causal_offset;
+    const int has_mask = p.use_mask && p.mask;
+
+    // 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 (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;
+#pragma unroll
+        for (int i = threadIdx.x * VEC; i < TOTAL; i += nthreads * VEC) {
+            int r = i / HEAD_DIM, d = i % HEAD_DIM;
+            int kc = kv0 + r;
+            bool valid = kc < p.kv_len;
+            int off = r * LD + swiz_col(d, r, SWIZ_MASK);
+            int g_off = kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
+            cp_async_16_pred(&dK[off], &p.k[g_off], valid);
+            cp_async_16_pred(&dV[off], &p.v[g_off], valid);
+        }
+        cp_async_commit();
+    };
+
+    // ---- Prologue: issue first tile load ----
+    load_tile(0, 0);
+
+    for (int ti = 0; ti <= t_end; ti++) {
+        int buf = ti & 1;
+
+        // Wait for 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.
+        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;
+
+        // Warp-level causal skip
+        if (!use_skip || 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);
+
+        // post-multiply scale in float (no bf16 precision loss from pre-scaling Q)
+        #pragma unroll
+        for (int n8 = 0; n8 < 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)
+                                        : p.kv_len;
+        int maxc1 = (p.causal_offset >= 0) ? min(p.kv_len, qr1 + p.causal_offset + 1)
+                                        : p.kv_len;
+        mma_softmax_tile<NC8, DN8>(kv0, maxc0, maxc1,
+                                    qr0, qr1,
+                                    p.mask_b_stride, p.mask_q_stride,
+                                    batch,
+                                    p.mask, has_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)
+    }
+
+    // ---- write output ---- (packed bf16x2 stores: one 32-bit STG per pair,
+    // halves store count and removes the uncoalesced scalar-store penalty)
+    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++) {
+        int d = dn8 * 8 + 2 * tid4;
+        if (qr0 < p.q_len) {
+            __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][0] * rl0,
+                                                     Oacc[dn8][1] * rl0);
+            *reinterpret_cast<__nv_bfloat162*>(&p.o[o_base + qr0 * p.q_stride_l + d * p.q_stride_d]) = v;
+        }
+        if (qr1 < p.q_len) {
+            __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][2] * rl1,
+                                                     Oacc[dn8][3] * rl1);
+            *reinterpret_cast<__nv_bfloat162*>(&p.o[o_base + qr1 * p.q_stride_l + d * p.q_stride_d]) = v;
+        }
+    }
+}

csrc/tests/attn_decode_test.cu

@@ -0,0 +1,201 @@
+/*
+Pure-C test:
+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
+
+// 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).
+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;
+}
+
+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); });
+}
+
+// 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, 1024, 128},
+        {1, 32, 4, 2048, 128},
+        {1, 32, 4, 4096, 128},
+        {16, 32, 4, 2048, 128},
+        {32, 32, 4, 1024, 128},
+    };
+    const int WARMUP = 10, ITERS = 100;
+    printf("\n===== DECODE BENCH (warmup=%d iters=%d) =====\n", WARMUP, ITERS);
+    print_bench_header();
+
+    for (int ci = 0; ci < 6; ci++) {
+        int B = cfgs[ci][0], Hq = cfgs[ci][1], Hk = cfgs[ci][2];
+        int sl = cfgs[ci][3], D = cfgs[ci][4];
+        size_t nQ = (size_t)B * Hq * D;
+        size_t nKV = (size_t)B * Hk * sl * D;
+
+        bf16 *dQ, *dK, *dV, *dO;
+        cudaMalloc(&dQ, nQ*2); cudaMalloc(&dK, nKV*2);
+        cudaMalloc(&dV, nKV*2); cudaMalloc(&dO, nQ*2);
+        size_t big = nQ > nKV ? nQ : nKV; bf16* tmp = new bf16[big];
+        for (size_t i = 0; i < nQ; i++)  tmp[i] = f2bf(randf());
+        cudaMemcpy(dQ, tmp, nQ*2, cudaMemcpyHostToDevice);
+        for (size_t i = 0; i < nKV; i++) tmp[i] = f2bf(randf());
+        cudaMemcpy(dK, tmp, nKV*2, cudaMemcpyHostToDevice);
+        for (size_t i = 0; i < nKV; i++) tmp[i] = f2bf(randf());
+        cudaMemcpy(dV, tmp, nKV*2, cudaMemcpyHostToDevice);
+        delete[] tmp;
+
+        AttentionParams<bf16> p;
+        p.batch = B; p.q_head = Hq; p.kv_head = Hk; p.q_len = 1; p.kv_len = sl;
+        p.head_dim = D; p.use_mask = 0; p.causal_offset = -1;
+        p.scale = 1.0f / sqrtf((float)D);
+        set_default_strides(p);
+        p.q = dQ; p.k = dK; p.v = dV; p.mask = nullptr; p.o = dO;
+
+        DecodeScratch sc;
+        cudaMalloc(&sc.o_part, (size_t)B*Hq*32*D*sizeof(float));
+        cudaMalloc(&sc.ml_part, (size_t)B*Hq*32*2*sizeof(float));
+
+        auto launch = [&]() { dispatch_decode(p, sc); };
+        double flops = 4.0 * B * Hq * (double)sl * D;
+        double bytes = 2.0 * (2.0 * nKV * sizeof(bf16));
+        BenchResult r = bench_kernel(launch, WARMUP, ITERS, flops, bytes);
+
+        char cfg[64];
+        snprintf(cfg, sizeof(cfg),
+                 "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d causal=%d",
+                 B, Hq, Hk, 1, sl, D, 0);
+        print_bench_row(cfg, r);
+
+        cudaFree(dQ); cudaFree(dK); cudaFree(dV); cudaFree(dO);
+        cudaFree(sc.o_part); cudaFree(sc.ml_part);
+    }
+}
+
+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);
+
+        size_t nQ = B*Hq*1*D, nKV = B*Hk*sl*D;
+        float *hQ=new float[nQ], *hK=new float[nKV], *hV=new float[nKV];
+        for (size_t i=0;i<nQ;i++) hQ[i]=randf();
+        for (size_t i=0;i<nKV;i++){hK[i]=randf();hV[i]=randf();}
+
+        bool* hMask=new bool[B*sl];
+        for (int i=0;i<B*sl;i++) hMask[i]=true;
+
+        bf16 *dQ,*dK,*dV,*dO,*tmp;
+        bool* dMask;
+        cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+        cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+        cudaMalloc(&dMask,B*sl);
+
+        tmp=new bf16[max(nQ,nKV)];
+        for (size_t i=0;i<nQ;i++) tmp[i]=f2bf(hQ[i]);
+        cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hK[i]);
+        cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hV[i]);
+        cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+        cudaMemcpy(dMask,hMask,B*sl,cudaMemcpyHostToDevice);
+
+        AttentionParams<bf16> p;
+        p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=1; p.kv_len=sl; p.head_dim=D;
+        p.use_mask=0; p.causal_offset=-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));
+
+        double t0=now_ms();
+        dispatch_decode(p, sc);
+        cudaDeviceSynchronize();
+        double kms=now_ms()-t0;
+        cudaError_t err=cudaGetLastError();
+        if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return 1;}
+
+        bf16* hOut=new bf16[nQ];
+        cudaMemcpy(hOut,dO,nQ*2,cudaMemcpyDeviceToHost);
+
+        float* ref=new float[nQ];
+        cpu_attention_ref(hQ, hK, hV, hMask, ref, B, Hq, Hk, 1, sl, D, -1);
+
+        float max_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;
+        }
+        printf("kernel: %.3f ms  max_err: %.6e\n\n",kms,max_err);
+
+        cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);cudaFree(dMask);
+        cudaFree(sc.o_part);cudaFree(sc.ml_part);
+        delete[]hQ;delete[]hK;delete[]hV;delete[]hMask;delete[]hOut;delete[]ref;delete[]tmp;
+    }
+    printf("All tests passed!\n");
+    bench();
+    return 0;
+}

csrc/tests/attn_paged_decode_test.cu

@@ -0,0 +1,332 @@
+// Compile:
+//   nvcc -I csrc -arch=sm_89 -O3 --use_fast_math --ptxas-options=-O3 \
+//        --extra-device-vectorization csrc/tests/attn_paged_decode_test.cu \
+//        -o /tmp/test_paged && /tmp/test_paged
+
+#include <cstring>
+#include "test_utils.cuh"
+#include "../kernels/attn_paged_decode_split_kv.cuh"
+#ifndef ASTRAI_NO_MMA
+#include "../kernels/attn_paged_decode_split_kv_mma.cuh"
+#endif
+
+// 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,
+    int page_size, int head_dim,
+    bf16* h_k, bf16* h_v)
+{
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    size_t page_stride = (size_t)page_size * Hkv * head_dim;
+    for (int b = 0; b < B; b++) {
+        for (int pos = 0; pos < kv_len; pos++) {
+            int log_pg = pos / page_size;
+            int pg_off = pos % page_size;
+            int phys = (int)h_pt[b * max_pages + log_pg];
+            for (int h = 0; h < Hkv; h++) {
+                size_t src_base = (size_t)phys * page_stride
+                                + (size_t)pg_off * Hkv * head_dim
+                                + h * head_dim;
+                size_t dst_base = ((size_t)b * Hkv + h) * kv_len * head_dim + (size_t)pos * head_dim;
+                memcpy(h_k + dst_base, h_k_pool + src_base, head_dim * sizeof(bf16));
+                memcpy(h_v + dst_base, h_v_pool + src_base, head_dim * sizeof(bf16));
+            }
+        }
+    }
+}
+
+template <int HEAD_DIM>
+static 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);
+    fflush(stdout);
+
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    int n_phys_pages = B * max_pages;
+
+    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_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);
+    cudaMalloc(&d_op, sz_op);
+    cudaMalloc(&d_ml, sz_ml);
+
+    srand(seed);
+    auto rnd = [&]() { return (rand() / (float)RAND_MAX) * 2.0f - 1.0f; };
+
+    bf16* h_q = (bf16*)malloc(sz_q);
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++)
+        h_q[i] = __float2bfloat16(rnd());
+    cudaMemcpy(d_q, h_q, sz_q, cudaMemcpyHostToDevice);
+
+    bf16* h_k_pool = (bf16*)malloc(sz_kv);
+    bf16* h_v_pool = (bf16*)malloc(sz_kv);
+    size_t ps = (size_t)page_size * Hkv * HEAD_DIM;
+    for (int pg = 0; pg < n_phys_pages; pg++) {
+        for (int off = 0; off < page_size; off++) {
+            for (int h = 0; h < Hkv; h++) {
+                for (int d = 0; d < HEAD_DIM; d++) {
+                    float v = sinf((float)(pg * 7919 + off * 1049 + h * 331 + d));
+                    size_t idx = (size_t)pg * ps + (size_t)off * Hkv * HEAD_DIM + h * HEAD_DIM + d;
+                    h_k_pool[idx] = __float2bfloat16(v);
+                    h_v_pool[idx] = __float2bfloat16(v * 0.3f);
+                }
+            }
+        }
+    }
+    cudaMemcpy(d_k_pool, h_k_pool, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, h_v_pool, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_pt = (int64_t*)malloc(sz_pt);
+    int next_pg = 0;
+    for (int b = 0; b < B; b++)
+        for (int p = 0; p < max_pages; p++)
+            h_pt[b * max_pages + p] = next_pg++;
+    cudaMemcpy(d_pt, h_pt, sz_pt, cudaMemcpyHostToDevice);
+
+    bf16* h_k_cont = (bf16*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(bf16));
+    bf16* h_v_cont = (bf16*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(bf16));
+    gather_kv_cpu(h_k_pool, h_v_pool, h_pt, B, Hkv, kv_len, page_size, HEAD_DIM, h_k_cont, h_v_cont);
+
+    float* h_q_f = (float*)malloc((size_t)B * Hq * HEAD_DIM * sizeof(float));
+    float* h_k_f = (float*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(float));
+    float* h_v_f = (float*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) h_q_f[i] = bf2f(h_q[i]);
+    for (int i = 0; i < B * kv_len * Hkv * HEAD_DIM; i++) {
+        h_k_f[i] = bf2f(h_k_cont[i]);
+        h_v_f[i] = bf2f(h_v_cont[i]);
+    }
+
+    float* h_o_ref = (float*)calloc(B * Hq * HEAD_DIM, sizeof(float));
+    cpu_attention_ref(h_q_f, h_k_f, h_v_f, nullptr, h_o_ref, B, Hq, Hkv, 1, kv_len, HEAD_DIM, -1);
+
+    float scale_val = 1.0f / sqrtf((float)HEAD_DIM);
+    PagedAttentionParams<bf16, float> 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;
+    set_default_paged_strides(p);
+    p.num_splits = 1; p.scale = scale_val;
+    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);
+    cudaDeviceSynchronize();
+
+    bf16* h_o_bf16 = (bf16*)malloc(sz_o);
+    cudaMemcpy(h_o_bf16, d_o_paged, sz_o, cudaMemcpyDeviceToHost);
+    float* h_o_paged = (float*)malloc(B * Hq * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++)
+        h_o_paged[i] = __bfloat162float(h_o_bf16[i]);
+
+    float max_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; }
+    }
+
+    bool pass = max_err < 0.02f;
+
+    if (pass) {
+        printf("PASS (max_abs_err=%.4e)\n", max_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("  ref[0..7]:");
+        for (int i = 0; i < 8 && i < HEAD_DIM; i++)
+            printf(" %.4f", h_o_ref[i]);
+        printf("\n  got[0..7]:");
+        for (int i = 0; i < 8 && i < HEAD_DIM; i++)
+            printf(" %.4f", h_o_paged[i]);
+        printf("\n");
+    }
+
+    free(h_q); free(h_k_pool); free(h_v_pool); free(h_pt);
+    free(h_k_cont); free(h_v_cont);
+    free(h_q_f); free(h_k_f); free(h_v_f);
+    free(h_o_ref); free(h_o_bf16); free(h_o_paged);
+    cudaFree(d_q); cudaFree(d_o_paged); cudaFree(d_o_ref);
+    cudaFree(d_k_pool); cudaFree(d_v_pool); cudaFree(d_pt);
+    cudaFree(d_op); cudaFree(d_ml);
+
+    return pass ? 0 : 1;
+}
+
+struct TestCase {
+    int head_dim;
+    int B, Hq, Hkv, kv_len, page_size, 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
+};
+
+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);
+    });
+    return matched ? r : 1;
+}
+
+// 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;
+
+    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);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t* d_pt;
+    float *d_op, *d_ml;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_pt, sz_pt);
+    cudaMalloc(&d_op, sz_op); cudaMalloc(&d_ml, sz_ml);
+
+    bf16* tmp = (bf16*)malloc(sz_kv > sz_q ? sz_kv : sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_q, tmp, sz_q, cudaMemcpyHostToDevice);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_k_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_pt = (int64_t*)malloc(sz_pt);
+    int next_pg = 0;
+    for (int b = 0; b < B; b++)
+        for (int p = 0; p < max_pages; p++)
+            h_pt[b * max_pages + p] = next_pg++;
+    cudaMemcpy(d_pt, h_pt, sz_pt, cudaMemcpyHostToDevice);
+    free(h_pt);
+
+    float scale_val = 1.0f / sqrtf((float)HEAD_DIM);
+    PagedAttentionParams<bf16, float> 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.page_size = page_size; pa.max_pages = max_pages;
+    pa.page_table = d_pt;
+    pa.k_cache = d_k_pool; pa.v_cache = d_v_pool;
+    pa.q = d_q; pa.mask = nullptr; pa.o = d_o;
+    pa.o_part = d_op; pa.ml_part = d_ml;
+
+    const int WARMUP = 10, ITERS = 100;
+    auto launch = [&]() { launch_paged_decode<HEAD_DIM>(pa); };
+    double flops = 4.0 * B * Hq * (double)kv_len * HEAD_DIM;
+    size_t nKV = (size_t)B * Hkv * kv_len * HEAD_DIM;
+    double bytes = 2.0 * (2.0 * nKV * sizeof(bf16));
+    BenchResult r = bench_kernel(launch, WARMUP, ITERS, flops, bytes);
+
+    char cfg[64];
+    snprintf(cfg, sizeof(cfg),
+             "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d page=%3d",
+             B, Hq, Hkv, 1, kv_len, HEAD_DIM, page_size);
+    print_bench_row(cfg, r);
+
+    free(tmp);
+    cudaFree(d_q); cudaFree(d_o);
+    cudaFree(d_k_pool); cudaFree(d_v_pool); cudaFree(d_pt);
+    cudaFree(d_op); cudaFree(d_ml);
+}
+
+static void bench() {
+    printf("\n===== PAGED DECODE BENCH =====\n");
+    print_bench_header();
+    bench_config<128>(1, 32, 4, 512, 128);
+    bench_config<128>(1, 32, 4, 1024, 128);
+    bench_config<128>(1, 32, 4, 2048, 128);
+    bench_config<128>(1, 32, 4, 4096, 128);
+    bench_config<128>(16, 32, 4, 2048, 128);
+    bench_config<128>(32, 32, 4, 1024, 128);
+}
+
+int main() {
+    int n = sizeof(TESTS) / sizeof(TESTS[0]);
+    int fail = 0;
+    printf("=== Paged Decode vs CPU reference (%d cases) ===\n\n", n);
+
+    for (int i = 0; i < n; i++) {
+        fail += dispatch_test(TESTS[i]);
+        if (fail) break;
+    }
+
+    if (fail) {
+        printf("\nFAILED (%d/%d tests failed)\n", fail, n);
+        return fail;
+    }
+    printf("\nAll %d tests passed!\n", n);
+    bench();
+    return 0;
+}

csrc/tests/attn_prefill_test.cu

@@ -0,0 +1,178 @@
+/*
+Pure-C test:
+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);
+    }
+}
+
+// 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},
+        {1,32,4,1024,1024,128,0},
+        {1,32,4,2048,2048,128,0},
+        {1,32,4,2048,2048,128,1},
+        {4,32,4,2048,2048,128,1},
+        {1,32,4,4096,4096,128,1},
+    };
+    int n = sizeof(cfgs)/sizeof(cfgs[0]);
+    const int WARMUP = 10, ITERS = 50;
+    printf("\n===== PREFILL BENCH (warmup=%d iters=%d) =====\n", WARMUP, ITERS);
+    printf("%-46s | %10s | %10s | %10s\n",
+           "config", "latency", "bandwidth", "throughput");
+    printf("---------------------------------------------------------------"
+           "----------------------------\n");
+
+    for (int ci = 0; ci < n; ci++) {
+        int B=cfgs[ci][0], Hq=cfgs[ci][1], Hk=cfgs[ci][2];
+        int ql=cfgs[ci][3], kl=cfgs[ci][4], D=cfgs[ci][5], causal=cfgs[ci][6];
+        size_t nQ=(size_t)B*Hq*ql*D, nKV=(size_t)B*Hk*kl*D;
+
+        bf16 *dQ,*dK,*dV,*dO,*tmp;
+        cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+        cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+        size_t big = nQ>nKV?nQ:nKV; tmp=new bf16[big];
+        for (size_t i=0;i<nQ;i++)  tmp[i]=f2bf(randf());
+        cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(randf());
+        cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(randf());
+        cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+
+        AttentionParams<bf16> p;
+        p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=ql; p.kv_len=kl; p.head_dim=D;
+        p.use_mask=0; p.causal_offset=causal?0:-1;
+        set_default_strides(p);
+        p.scale=1.0f/sqrtf((float)D);
+        p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
+
+        for (int i=0;i<WARMUP;i++) dispatch_prefill(p);
+        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);
+        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;
+
+        char cfg[64];
+        snprintf(cfg, sizeof(cfg),
+                 "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d causal=%d",
+                 B,Hq,Hk,ql,kl,D,causal);
+        printf("%-46s | %7.4f ms | %7.1f GB/s | %6.2f TFLOP/s\n",
+               cfg, ms, gbps, tflops);
+
+        cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);
+        delete[]tmp; cudaEventDestroy(s); cudaEventDestroy(e);
+    }
+}
+
+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];
+        printf("=== B=%d Hq=%d Hk=%d q=%d kv=%d D=%d causal=%d ===\n",
+               B,Hq,Hk,ql,kl,D,causal);
+
+        size_t nQ = B*Hq*ql*D, nKV = B*Hk*kl*D;
+        float *hQ=new float[nQ], *hK=new float[nKV], *hV=new float[nKV];
+        for (size_t i=0;i<nQ;i++) hQ[i]=randf();
+        for (size_t i=0;i<nKV;i++){hK[i]=randf();hV[i]=randf();}
+
+        bf16 *dQ,*dK,*dV,*dO,*tmp;
+        cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+        cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+        tmp=new bf16[max(nQ,nKV)];
+        for (size_t i=0;i<nQ;i++) tmp[i]=f2bf(hQ[i]);
+        cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hK[i]);
+        cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+        for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hV[i]);
+        cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+
+        AttentionParams<bf16> p;
+        p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=ql; p.kv_len=kl; p.head_dim=D;
+        p.use_mask=0; p.causal_offset=causal?0:-1;
+        set_default_strides(p);
+        p.scale=1.0f/sqrtf((float)D);
+        p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
+
+        double t0=now_ms();
+        dispatch_prefill(p);
+        cudaDeviceSynchronize();
+        double kms=now_ms()-t0;
+        cudaError_t err=cudaGetLastError();
+        if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return 1;}
+
+        bf16* hOut=new bf16[nQ];
+        cudaMemcpy(hOut,dO,nQ*2,cudaMemcpyDeviceToHost);
+
+        float* ref=new float[nQ];
+        cpu_attention_ref(hQ, hK, hV, nullptr, ref, B, Hq, Hk, ql, kl, D, causal ? 0 : -1);
+
+        float max_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;
+        }
+        printf("kernel: %.3f ms  max_err: %.6e\n\n",kms,max_err);
+
+        cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);
+        delete[]hQ;delete[]hK;delete[]hV;delete[]hOut;delete[]ref;delete[]tmp;
+    }
+    printf("All tests passed!\n");
+    bench();
+    return 0;
+}

csrc/tests/test_utils.cuh

@@ -0,0 +1,181 @@
+#pragma once
+
+#include <cstdio>
+#include <cstdlib>
+#include <cmath>
+#include <chrono>
+#include <cuda_bf16.h>
+
+using bf16 = __nv_bfloat16;
+
+inline bf16 f2bf(float x) { return __float2bfloat16(x); }
+inline float bf2f(bf16 x) { return __bfloat162float(x); }
+
+inline float randf() { return (float)rand() / (float)RAND_MAX - 0.5f; }
+
+inline double now_ms() {
+    using namespace std::chrono;
+    return duration_cast<milliseconds>(steady_clock::now().time_since_epoch()).count();
+}
+
+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); \
+        if (_e != cudaSuccess) { \
+            printf("CUDA error %s at %s:%d\n", cudaGetErrorString(_e), __FILE__, __LINE__); \
+            exit(1); \
+        } \
+    } while (0)
+
+struct BenchResult {
+    float ms;
+    double gbps;
+    double tflops;
+};
+
+template <typename Fn>
+BenchResult bench_kernel(Fn launch, int warmup, int iters,
+                         double flops, double bytes) {
+    for (int i = 0; i < warmup; i++) launch();
+    cudaDeviceSynchronize();
+    cudaError_t err = cudaGetLastError();
+    if (err != cudaSuccess) {
+        printf("CUDA error before bench: %s\n", cudaGetErrorString(err));
+        return {0, 0, 0};
+    }
+
+    cudaEvent_t s, e;
+    cudaEventCreate(&s); cudaEventCreate(&e);
+    cudaEventRecord(s);
+    for (int i = 0; i < iters; i++) launch();
+    cudaEventRecord(e); cudaEventSynchronize(e);
+    float ms = 0; cudaEventElapsedTime(&ms, s, e); ms /= iters;
+    cudaEventDestroy(s); cudaEventDestroy(e);
+
+    return {ms, bytes / (ms * 1e-3) / 1e9, flops / (ms * 1e-3) / 1e12};
+}
+
+inline void print_bench_header() {
+    printf("%-46s | %10s | %10s | %10s\n",
+           "config", "latency", "bandwidth", "throughput");
+    printf("---------------------------------------------------------------"
+           "----------------------------\n");
+}
+
+inline void print_bench_row(const char* cfg, const BenchResult& r) {
+    printf("%-46s | %7.4f ms | %7.1f GB/s | %6.2f TFLOP/s\n",
+           cfg, r.ms, r.gbps, r.tflops);
+}
+
+template <int... Ds>
+struct _HeadSwitch;
+
+template <int D>
+struct _HeadSwitch<D> {
+    template <typename Fn>
+    static void call(int hd, Fn&& fn) { if (hd == D) fn.template operator()<D>(); }
+};
+
+template <int D, int... Rest>
+struct _HeadSwitch<D, Rest...> {
+    template <typename Fn>
+    static void call(int hd, Fn&& fn) {
+        if (hd == D) fn.template operator()<D>();
+        else _HeadSwitch<Rest...>::call(hd, fn);
+    }
+};
+
+// Default set: 32, 64, 128, 256
+template <typename Fn>
+void dispatch_by_head_dim(int head_dim, Fn&& fn) {
+    _HeadSwitch<32, 64, 128, 256>::call(head_dim, fn);
+}
+
+// Set default strides for contiguous b h l d layout on AttentionParams.
+template<typename P>
+inline void set_default_strides(P& p) {
+    p.q_stride_b  = p.q_head * p.q_len * p.head_dim;
+    p.q_stride_h  = p.q_len * p.head_dim;
+    p.q_stride_l  = p.head_dim;
+    p.q_stride_d  = 1;
+    p.kv_stride_b = p.kv_head * p.kv_len * p.head_dim;
+    p.kv_stride_h = p.kv_len * p.head_dim;
+    p.kv_stride_l = p.head_dim;
+    p.kv_stride_d = 1;
+    p.mask_b_stride = p.kv_len;
+    p.mask_q_stride = 0;
+}
+
+// Set default Q strides for contiguous b h l d layout on PagedAttentionParams.
+template<typename P>
+inline void set_default_paged_strides(P& p) {
+    p.q_stride_b  = p.q_head * p.q_len * p.head_dim;
+    p.q_stride_h  = p.q_len * p.head_dim;
+    p.q_stride_l  = p.head_dim;
+    p.q_stride_d  = 1;
+    p.mask_b_stride = p.kv_len;
+    p.mask_q_stride = 0;
+}
+
+// Generic CPU reference for multi-query / grouped-query attention.
+// Tensor shapes (all float*):
+//   Q : [B, Hq, q_len, D]
+//   K : [B, Hk, kv_len, D]
+//   V : [B, Hk, kv_len, D]
+//   O : [B, Hq, q_len, D]
+// mask: if q_len == 1, shape is [B, kv_len]; otherwise mask is not supported.
+// causal_offset: -1 = non-causal; >=0 = absolute position of first Q token.
+static void cpu_attention_ref(
+    const float* Q, const float* K, const float* V, const bool* mask,
+    float* O, int B, int Hq, int Hk, int q_len, int kv_len, int D,
+    int causal_offset
+) {
+    float scale = 1.0f / sqrtf((float)D);
+    int n_rep = Hq / Hk;
+    for (int b = 0; b < B; b++) {
+        for (int h = 0; h < Hq; h++) {
+            int kv_h = h / n_rep;
+            for (int qi = 0; qi < q_len; qi++) {
+                float mv = -INFINITY, sv = 0.0f;
+                float accum[256] = {0.0f};
+                int lim = kv_len;
+                if (causal_offset >= 0) {
+                    int c = qi + causal_offset + 1;
+                    lim = (c < kv_len) ? c : kv_len;
+                }
+                for (int kj = 0; kj < lim; kj++) {
+                    if (mask != nullptr && q_len == 1) {
+                        if (!mask[b * kv_len + kj]) continue;
+                    }
+                    float dot = 0.0f;
+                    size_t q_idx = ((size_t)b * Hq + h) * q_len + qi;
+                    size_t kv_idx = ((size_t)b * Hk + kv_h) * kv_len + kj;
+                    for (int d = 0; d < D; d++)
+                        dot += Q[q_idx * D + d] * K[kv_idx * D + d];
+                    dot *= scale;
+                    float nm = fmaxf(mv, dot);
+                    float a = expf(mv - nm);
+                    float b_exp = expf(dot - nm);
+                    sv = sv * a + b_exp;
+                    for (int d = 0; d < D; d++)
+                        accum[d] = accum[d] * a + V[kv_idx * D + d] * b_exp;
+                    mv = nm;
+                }
+                float inv = 1.0f / sv;
+                size_t o_idx = ((size_t)b * Hq + h) * q_len + qi;
+                for (int d = 0; d < D; d++)
+                    O[o_idx * D + d] = accum[d] * inv;
+            }
+        }
+    }
+}

scripts/demo/stream_chat.py

@@ -42,6 +42,19 @@ def parse_args():
         default=2048,
         help="Maximum tokens to generate",
     )
+    parser.add_argument(
+        "--frequency_penalty",
+        type=float,
+        default=0.5,
+        help="Penalty per occurrence for repeated tokens (0.0 disables, "
+        "range -2.0~2.0, typical 0.3-1.0)",
+    )
+    parser.add_argument(
+        "--rep_window",
+        type=int,
+        default=64,
+        help="Number of recent prompt tokens to include in penalty history",
+    )
     parser.add_argument(
         "--system_prompt",
         type=str,
@@ -79,6 +92,8 @@ def chat():
             temperature=args.temperature,
             top_p=args.top_p,
             top_k=args.top_k,
+            frequency_penalty=args.frequency_penalty,
+            rep_window=args.rep_window,
         ):
             print(token, end="", flush=True)
             full_response += token

scripts/eval/analyze_weights.py

@@ -117,7 +117,7 @@ def print_component_summary(results: dict[str, dict], title: str):
         r["er_99_norm"]
         for vs in matrix_groups.values()
         for r in vs
-        if "_norm" not in r or not r.get("is_1d")
+        if not r.get("is_1d")
     ]
     if all_er:
         m = sum(all_er) / len(all_er)
@@ -232,8 +232,16 @@ def main():
         action="store_true",
         help="Skip SVD analysis, only show weight statistics (mean/std/min/max).",
     )
+    parser.add_argument(
+        "--output",
+        type=str,
+        default=None,
+        help="Save results as JSON to this path.",
+    )
     args = parser.parse_args()
 
+    all_results = {}
+
     def analyze_one(ckpt_dir: str, label: str):
         ckpt_dir = Path(ckpt_dir)
         weights_path = ckpt_dir / "model.safetensors"
@@ -294,13 +302,19 @@ def main():
             )
             print_layer_grid(results)
         print_weight_stats(results)
+        all_results[label] = results
         return results
 
     analyze_one(args.ckpt_dir, "Primary")
 
     if args.compare:
         for cdir in args.compare:
-            analyze_one(cdir, "Compare")
+            analyze_one(cdir, f"Compare_{cdir}")
+
+    if args.output:
+        with open(args.output, "w", encoding="utf-8") as f:
+            json.dump(all_results, f, indent=2)
+        print(f"\nResults saved to {args.output}")
 
 
 if __name__ == "__main__":

scripts/eval/evaluate_humaneval.py

@@ -8,7 +8,6 @@ Config is a single dataclass; side effects are isolated at pipeline boundaries.
 """
 
 import argparse
-import itertools
 import json
 import os
 import re
@@ -233,7 +232,7 @@ def execute_one(args: tuple) -> bool:
         return False
 
 
-def test_one(item: dict, cfg: EvalConfig) -> Tuple[str, int, int]:
+def test_one(item: dict, cfg: EvalConfig, pool=None) -> Tuple[str, int, int]:
     from concurrent.futures import ProcessPoolExecutor
 
     task_id = item["task_id"]
@@ -247,11 +246,16 @@ def test_one(item: dict, cfg: EvalConfig) -> Tuple[str, int, int]:
         for c in completions
     ]
     n = len(codes)
-    passed = 0
+
-    with ProcessPoolExecutor(max_workers=cfg.test_workers) as pool:
+    def _run(p):
-        for ok in pool.map(execute_one, codes):
+        return sum(1 for ok in p.map(execute_one, codes) if ok)
-            if ok:
+
-                passed += 1
+    if pool is not None:
+        passed = _run(pool)
+    else:
+        with ProcessPoolExecutor(max_workers=cfg.test_workers) as p:
+            passed = _run(p)
+
     return task_id, n, passed
 
 
@@ -259,8 +263,14 @@ def test_all(
     items: Sequence[dict],
     cfg: EvalConfig,
 ) -> Iterator[Tuple[str, int, int]]:
+    from concurrent.futures import ProcessPoolExecutor
+
+    pool = ProcessPoolExecutor(max_workers=cfg.test_workers)
+    try:
         for item in tqdm.tqdm(items, desc="Testing", unit="problem"):
-        yield test_one(item, cfg)
+            yield test_one(item, cfg, pool)
+    finally:
+        pool.shutdown(wait=True)
 
 
 def pass_at_k(n: int, c: int, k: int) -> float:
@@ -273,26 +283,32 @@ def score_results(
     results: Iterator[Tuple[str, int, int]],
     k_values: Tuple[int, ...],
 ) -> Dict:
-    # filter to k <= n (peek first result to get n)
+    """Score pass@k for each problem.
-    first = next(results)
-    results = itertools.chain([first], results)
-    n = first[1]
-    k_values = tuple(k for k in k_values if k <= n)
 
+    k values are filtered per-problem: if a problem has n < k samples
+    (e.g. after deduplication), pass@k is not computed for that problem.
+    The summary averages only over problems where the k was computed.
+    """
     scores = {k: [] for k in k_values}
     output = {}
     for task_id, n, passed in results:
         entry = {"task_id": task_id, "n": n, "passed": passed}
         for k in k_values:
+            if k <= n:
                 pk = round(pass_at_k(n, passed, k), 4)
                 entry[f"pass@{k}"] = pk
                 scores[k].append(pk)
+            else:
+                entry[f"pass@{k}"] = None
         output[task_id] = entry
 
     summary = {}
     for k in k_values:
         vals = scores[k]
+        if vals:
             summary[f"pass@{k}"] = round(float(np.mean(vals)), 4)
+        else:
+            summary[f"pass@{k}"] = None
     output["_summary"] = summary
     return output
 
@@ -375,7 +391,10 @@ def report(scored: Dict):
     summary = scored.pop("_summary", {})
     print(f"\n{'=' * 60}")
     for k, v in summary.items():
+        if v is not None:
             print(f"  {k}: {v:.2%}")
+        else:
+            print(f"  {k}: N/A")
     print(f"{'=' * 60}")
     scored["_summary"] = summary
 

scripts/eval/evaluate_ifd.py

@@ -16,7 +16,9 @@ v2 changelog:
 """
 
 import argparse
+import glob
 import json
+import os
 import statistics
 
 import torch
@@ -65,6 +67,29 @@ def _resolve_sentinel_ids(tokenizer, sentinel_text):
     return [0]
 
 
+def _collect_input_files(input_path: str) -> list:
+    """Resolve *input_path* to a list of JSONL/JSON files."""
+    if os.path.isdir(input_path):
+        files = []
+        for ext in ("*.jsonl", "*.json"):
+            files.extend(
+                sorted(glob.glob(os.path.join(input_path, "**", ext), recursive=True))
+            )
+        return files
+    return sorted(glob.glob(input_path))
+
+
+def _load_items(filepath: str) -> list:
+    """Load JSONL or JSON (array / single dict) into a list of dicts."""
+    with open(filepath, "r", encoding="utf-8") as f:
+        if filepath.lower().endswith(".json"):
+            data = json.load(f)
+            if isinstance(data, dict):
+                return [data]
+            return data
+        return [json.loads(line) for line in f if line.strip()]
+
+
 @torch.inference_mode()
 def _score_batch(
     pairs, model, device, max_len=2048, sentinel_ids=None, per_token=False
@@ -204,69 +229,9 @@ def _trim(context_ids, resp_ids, max_len):
     return context_ids[overflow:], resp_ids
 
 
-def score_plain(
+def process_file(
     model,
     tokenizer,
-    instruction,
-    response,
-    device,
-    max_len=2048,
-    sentinel_ids=None,
-    per_token=False,
-):
-    """Compute IFD for a single instruction-response pair (plain format)."""
-    ctx_ids = tokenizer.encode(instruction, add_special_tokens=False)
-    resp_ids = tokenizer.encode(response, add_special_tokens=False)
-    ctx_ids, resp_ids = _trim(ctx_ids, resp_ids, max_len)
-    if not ctx_ids or not resp_ids:
-        return {
-            "L_cond": None,
-            "L_uncond": None,
-            "ifd": None,
-            "skip_reason": "empty ctx or resp",
-        }
-    return _score_batch(
-        [(ctx_ids, resp_ids)],
-        model,
-        device,
-        max_len,
-        sentinel_ids=sentinel_ids,
-        per_token=per_token,
-    )[0]
-
-
-def score_messages(
-    model, tokenizer, messages, device, max_len=2048, sentinel_ids=None, per_token=False
-):
-    """Compute IFD for each assistant turn in a messages array."""
-    turns = []
-    for i, msg in enumerate(messages):
-        if msg.get("role") != "assistant":
-            continue
-        ctx_text = "\n\n".join(m["content"] for m in messages[:i])
-        ctx_ids = tokenizer.encode(ctx_text)
-        resp_ids = tokenizer.encode(msg["content"], add_special_tokens=False)
-        ctx_ids, resp_ids = _trim(ctx_ids, resp_ids, max_len)
-        if ctx_ids and resp_ids:
-            turns.append((ctx_ids, resp_ids))
-    if not turns:
-        return None
-    raw_scores = _score_batch(
-        turns, model, device, max_len, sentinel_ids=sentinel_ids, per_token=per_token
-    )
-    valid = [s for s in raw_scores if s is not None and s.get("ifd") is not None]
-    if not valid:
-        return {"ifd": None, "ifd_turns": raw_scores}
-    avg = sum(s["ifd"] for s in valid) / len(valid)
-    return {
-        "ifd": avg,
-        "ifd_detail": valid[0] if len(valid) == 1 else None,
-        "ifd_turns": raw_scores,
-    }
-
-
-def process_file(
-    param_path,
     input_file,
     output_file,
     instr_key,
@@ -275,28 +240,31 @@ def process_file(
     data_format="plain",
     batch_size=1,
     device=None,
-    sentinel_text="\n",
+    sentinel_ids=None,
     per_token=False,
+    max_samples=None,
 ):
+    """Score a single file, write per-sample JSONL, return summary stats."""
     if device is None:
         device = "cuda" if torch.cuda.is_available() else "cpu"
-    dtype = torch.bfloat16 if "cuda" in device else torch.float32
 
-    model = AutoModel.from_pretrained(param_path)
+    if sentinel_ids is None:
-    tokenizer = AutoTokenizer.from_pretrained(param_path)
+        sentinel_ids = _resolve_sentinel_ids(tokenizer, "\n")
-    model.to(device=device, dtype=dtype)
-    model.eval()
 
-    sentinel_ids = _resolve_sentinel_ids(tokenizer, sentinel_text)
+    data = _load_items(input_file)
 
-    with open(input_file, encoding="utf-8") as f:
+    if max_samples and len(data) > max_samples:
-        data = [json.loads(line) for line in f if line.strip()]
+        import random
+
+        data = random.sample(data, max_samples)
 
     results = []
     all_ifds = []
     buffer = []
 
-    for item in tqdm.tqdm(data, desc="Computing IFD", unit="sample"):
+    label = os.path.splitext(os.path.basename(input_file))[0]
+
+    for item in tqdm.tqdm(data, desc=f"  {label}", unit="sample", leave=False):
         if data_format == "messages":
             turns = []
             for i, msg in enumerate(item.get("messages", [])):
@@ -356,8 +324,22 @@ def process_file(
             f.write(json.dumps(item, ensure_ascii=False) + "\n")
 
     valid_ifd = [v for v in all_ifds if v is not None]
+    stats = {
+        "samples": len(data),
+        "valid_ifd": len(valid_ifd),
+        "skipped": len(data) - len(valid_ifd),
+    }
     if valid_ifd:
+        stats["mean_ifd"] = statistics.mean(valid_ifd)
+        stats["median_ifd"] = statistics.median(valid_ifd)
+        if len(valid_ifd) > 1:
+            stats["stdev_ifd"] = statistics.stdev(valid_ifd)
+        stats["min_ifd"] = min(valid_ifd)
+        stats["max_ifd"] = max(valid_ifd)
+
         print(f"\n{'=' * 50}")
+        print(f"  [{label}]")
+        print(f"{'=' * 50}")
         print(f"  Samples:       {len(data)}")
         print(f"  Valid IFD:     {len(valid_ifd)}")
         print(f"  Skipped:       {len(data) - len(valid_ifd)}")
@@ -369,6 +351,7 @@ def process_file(
         print(f"  Max IFD:       {max(valid_ifd):.4f}")
         print(f"{'=' * 50}")
     print(f"  Results saved to {output_file}")
+    return stats
 
 
 def _flush_buffer(
@@ -422,8 +405,18 @@ def main():
         description="Compute IFD scores for instruction-response data"
     )
     parser.add_argument("--param_path", type=str, required=True, help="Model directory")
-    parser.add_argument("--input", type=str, required=True, help="Input JSONL file")
+    parser.add_argument(
-    parser.add_argument("--output", type=str, required=True, help="Output JSONL file")
+        "--input_path",
+        type=str,
+        required=True,
+        help="Input file, glob pattern, or directory.",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        required=True,
+        help="Directory for output files (summary.json + per-file JSONL).",
+    )
     parser.add_argument("--max_len", type=int, default=2048, help="Max token length")
     parser.add_argument(
         "--format",
@@ -442,6 +435,12 @@ def main():
         "--batch_size", type=int, default=8, help="Batch size for model forward passes"
     )
     parser.add_argument("--device", type=str, default=None, help="Device (e.g. cuda:0)")
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        default="bfloat16" if torch.cuda.is_available() else "float32",
+        help="Torch dtype",
+    )
     parser.add_argument(
         "--sentinel_text",
         type=str,
@@ -453,21 +452,60 @@ def main():
         action="store_true",
         help="Include per-token IFD breakdown in output",
     )
+    parser.add_argument(
+        "--max_samples",
+        type=int,
+        default=None,
+        help="Maximum number of samples per file (random subsample). Default: all.",
+    )
     args = parser.parse_args()
 
-    process_file(
+    if args.device is None:
-        args.param_path,
+        args.device = "cuda" if torch.cuda.is_available() else "cpu"
-        args.input,
+    dtype = getattr(torch, args.dtype)
-        args.output,
+
-        args.instr_key,
+    print(f"Loading model from {args.param_path} ...")
-        args.resp_key,
+    model = AutoModel.from_pretrained(args.param_path)
-        args.max_len,
+    tokenizer = AutoTokenizer.from_pretrained(args.param_path)
+    model.to(device=args.device, dtype=dtype)
+    model.eval()
+
+    sentinel_ids = _resolve_sentinel_ids(tokenizer, args.sentinel_text)
+
+    input_files = _collect_input_files(args.input_path)
+    if not input_files:
+        print(f"No input files found at {args.input_path}")
+        return
+
+    print(f"Found {len(input_files)} file(s) to evaluate")
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    all_stats = {}
+    for filepath in input_files:
+        label = os.path.splitext(os.path.basename(filepath))[0]
+        output_file = os.path.join(args.output_dir, f"{label}_ifd.jsonl")
+
+        stats = process_file(
+            model=model,
+            tokenizer=tokenizer,
+            input_file=filepath,
+            output_file=output_file,
+            instr_key=args.instr_key,
+            resp_key=args.resp_key,
+            max_len=args.max_len,
             data_format=args.format,
             batch_size=args.batch_size,
             device=args.device,
-        sentinel_text=args.sentinel_text,
+            sentinel_ids=sentinel_ids,
             per_token=args.per_token,
+            max_samples=args.max_samples,
         )
+        all_stats[label] = stats
+
+    summary_path = os.path.join(args.output_dir, "summary.json")
+    with open(summary_path, "w", encoding="utf-8") as f:
+        json.dump(all_stats, f, ensure_ascii=False, indent=2)
+    print(f"\nSummary saved to {summary_path}")
 
 
 if __name__ == "__main__":

scripts/eval/evaluate_ifeval.py

@@ -5,7 +5,7 @@ Supports all IFEval constraint types except language detection.
 
 Usage::
 
-    python scripts/tools/evaluate_ifeval.py --param_path ./params \
+    python scripts/eval/evaluate_ifeval.py --param_path ./params \
         --data_path ifeval.jsonl --output results.json \
         --temperature 0.1 --max_tokens 512
 """

scripts/eval/evaluate_mmlu.py

@@ -139,17 +139,12 @@ def load_csv(path: str) -> list[dict]:
     return data
 
 
-def build_prompt(
+def build_prompt(question: str, choices: dict, subject: str) -> str:
-    question: str, choices: dict, subject: str, n_shot: int, dev_data: list[dict]
+    """Build the raw question prompt (without few-shot examples).
-) -> str:
+
-    prompt = ""
+    Few-shot examples are handled by ``apply_chat`` to avoid duplication.
-    if n_shot > 0 and dev_data:
+    """
     prompt = f"The following are multiple choice questions (with answers) about {subject}.\n\n"
-        for item in dev_data[:n_shot]:
-            prompt += f"Question: {item['question']}\n"
-            for k in ("A", "B", "C", "D"):
-                prompt += f"{k}. {item[k]}\n"
-            prompt += f"Answer: {item['answer']}\n\n"
     prompt += f"Question: {question}\n"
     for k in ("A", "B", "C", "D"):
         prompt += f"{k}. {choices[k]}\n"
@@ -218,9 +213,7 @@ def evaluate_subject(
     correct = 0
     total = 0
     for item in tqdm.tqdm(test_data, desc=f"{subject:40s}", leave=False):
-        raw_prompt = build_prompt(
+        raw_prompt = build_prompt(item["question"], item, subject)
-            item["question"], item, subject, n_shot, dev_data or []
-        )
         context = apply_chat(tokenizer, raw_prompt, n_shot, dev_data or [])
         context_ids = tokenizer.encode(context)
         scores = {

scripts/eval/evaluate_ppl.py

@@ -1,5 +1,9 @@
 import argparse
+import glob
 import json
+import os
+import statistics
+from typing import Dict, List, Optional, Tuple
 
 import torch
 import torch.nn.functional as F
@@ -9,95 +13,400 @@ from astrai.model import AutoModel
 from astrai.tokenize import AutoTokenizer
 
 
-def process_file(
+def _collect_input_files(input_path: str) -> List[str]:
-    param_path: str, input_file: str, output_file: str, batch_size: int, text_key: str
+    """Resolve *input_path* to a list of JSONL/JSON files."""
-):
+    if os.path.isdir(input_path):
-    # Load model and tokenizer
+        files = []
-    model = AutoModel.from_pretrained(param_path)
+        for ext in ("*.jsonl", "*.json"):
-    tokenizer = AutoTokenizer.from_pretrained(param_path)
+            files.extend(
-    model.to(device="cuda", dtype=torch.bfloat16)
+                sorted(glob.glob(os.path.join(input_path, "**", ext), recursive=True))
+            )
+        return files
+    return sorted(glob.glob(input_path))
 
-    with open(input_file, "r", encoding="utf-8") as f:
-        input_data = [json.loads(line) for line in f]
 
-    texts = [item[text_key] for item in input_data]
+def _load_items(filepath: str) -> List[dict]:
+    """Load JSONL or JSON (array / single dict) into a list of dicts."""
+    with open(filepath, "r", encoding="utf-8") as f:
+        if filepath.lower().endswith(".json"):
+            data = json.load(f)
+            if isinstance(data, dict):
+                return [data]
+            return data
+        return [json.loads(line) for line in f if line.strip()]
 
-    # Encode all texts
-    print(f"Encoding {len(texts)} texts...")
-    encoded_texts = [tokenizer.encode(text) for text in texts]
 
-    output_data = []
+def _encode_batch(
-    total_batches = (len(encoded_texts) + batch_size - 1) // batch_size
+    tokenizer: AutoTokenizer, texts: List[str], max_length: int
+) -> Tuple[List[List[int]], List[List[int]]]:
+    """Encode *texts* and return (token_ids, attention_masks).
 
-    for i in tqdm.tqdm(
+    Each sequence is left-aligned and padded to the batch max length.
-        range(0, len(encoded_texts), batch_size),
+    """
-        total=total_batches,
+    encoded = [tokenizer.encode(t)[:max_length] for t in texts]
-        desc="Computing perplexity",
+    if not encoded:
-    ):
+        return [], []
-        batch_encoded = encoded_texts[i : i + batch_size]
+    max_len = max(len(seq) for seq in encoded)
-        batch_texts = texts[i : i + batch_size]
-
-        # Find max length in batch and pad
-        max_len = max(len(seq) for seq in batch_encoded)
     padded_ids = []
     masks = []
-
+    for seq in encoded:
-        for seq in batch_encoded:
         pad_len = max_len - len(seq)
-            padded_seq = seq + [tokenizer.pad_id] * pad_len
+        padded_ids.append(seq + [tokenizer.pad_id] * pad_len)
-            mask = [True] * len(seq) + [False] * pad_len
+        masks.append([1] * len(seq) + [0] * pad_len)
-            padded_ids.append(padded_seq)
+    return padded_ids, masks
-            masks.append(mask)
 
-        # Convert to tensors
-        input_ids = torch.tensor(padded_ids, device="cuda", dtype=torch.long)
-        input_mask = torch.tensor(masks, device="cuda", dtype=torch.bool)
 
-        # Compute perplexity
+def _compute_batch(
-        output = model(input_ids, input_mask=input_mask)
+    model,
-        logits = output["logits"]
+    input_ids: torch.Tensor,
+    attention_mask: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Forward pass and return (log_probs, valid_mask) of shape [B, S-1].
 
-        # Shift for causal language modeling
+    log_probs[i, j] = log P(token j+1 | tokens 0..j)
-        shifted_logits = logits[:, :-1, :]  # [batch_size, seq_len-1, vocab_size]
+    """
-        shifted_input_ids = input_ids[:, 1:]  # [batch_size, seq_len-1]
+    output = model(input_ids, input_mask=attention_mask)
-        shifted_mask = input_mask[:, 1:]  # [batch_size, seq_len-1]
+    logits = output["logits"][:, :-1, :]  # [B, S-1, V]
+    targets = input_ids[:, 1:]  # [B, S-1]
+    valid = attention_mask[:, 1:].float()  # [B, S-1]
 
-        # Compute cross entropy loss
+    log_probs = F.log_softmax(logits.float(), dim=-1)  # [B, S-1, V]
-        loss = F.cross_entropy(
+    token_log_probs = log_probs.gather(2, targets.unsqueeze(-1)).squeeze(-1)  # [B, S-1]
-            shifted_logits.flatten(0, 1),
+
-            shifted_input_ids.flatten(0, 1),
+    return token_log_probs, valid
-            reduction="none",
+
+
+def _token_type(token_id: int, stop_ids: frozenset, decode_fn) -> str:
+    """Classify a token into a coarse type for analysis.
+
+    *stop_ids* is a pre-built set of special token IDs.
+    *decode_fn* is ``tokenizer.decode`` (or a wrapper) for single-token
+    decoding.
+    """
+    if token_id in stop_ids:
+        return "special"
+    decoded = decode_fn([token_id], skip_special_tokens=True)
+    if any("\u4e00" <= ch <= "\u9fff" for ch in decoded):
+        return "cjk"
+    if any(ord(ch) > 127 for ch in decoded):
+        return "non_ascii"
+    return "ascii"
+
+
+def _percentiles(values: List[float]) -> Dict[str, float]:
+    """Compute common percentiles from a list of floats.
+
+    Uses linear interpolation between closest ranks (same convention
+    as NumPy's default).
+    """
+    if not values:
+        return {}
+    sorted_vals = sorted(values)
+    n = len(sorted_vals)
+
+    def _pct(p: float) -> float:
+        if n == 1:
+            return sorted_vals[0]
+        k = p * (n - 1)
+        f = int(k)
+        c = min(f + 1, n - 1)
+        return sorted_vals[f] + (sorted_vals[c] - sorted_vals[f]) * (k - f)
+
+    return {
+        "p50": _pct(0.50),
+        "p90": _pct(0.90),
+        "p95": _pct(0.95),
+        "p99": _pct(0.99),
+    }
+
+
+class LossAccumulator:
+    """Accumulate per-token losses with optional streaming mode.
+
+    When *stream* is True (token_level=False), losses are not kept
+    in memory individually — only a running sum/count and a histogram
+    (for approximate percentiles) are maintained.  When *stream* is
+    False, all losses are retained for exact statistics and per-record
+    output.
+    """
+
+    _HIST_BINS = 1000
+    _HIST_MAX = 20.0  # clamp losses above this for histogram
+
+    def __init__(self, stream: bool):
+        self.stream = stream
+        self.losses: List[float] = [] if not stream else []
+        self.total: float = 0.0
+        self.count: int = 0
+        self.hist = torch.zeros(self._HIST_BINS, dtype=torch.long)
+        # per-type losses (only populated when not streaming)
+        self.by_type: Dict[str, List[float]] = {}
+        self.type_total: Dict[str, float] = {}
+        self.type_count: Dict[str, int] = {}
+
+    def add(self, losses: List[float]):
+        self.total += sum(losses)
+        self.count += len(losses)
+        if self.stream:
+            clamped = [min(max(l, 0.0), self._HIST_MAX) for l in losses]
+            idx = torch.tensor(clamped) / self._HIST_MAX * (self._HIST_BINS - 1)
+            self.hist += torch.bincount(
+                idx.long().clamp(0, self._HIST_BINS - 1),
+                minlength=self._HIST_BINS,
             )
+        else:
+            self.losses.extend(losses)
 
-        loss = loss.view(shifted_input_ids.shape)  # [batch_size, seq_len-1]
+    def add_typed(self, ttype: str, losses: List[float]):
-        loss = loss * shifted_mask
+        if not self.stream:
-        sentence_loss = loss.sum(dim=1) / shifted_mask.sum(dim=1).clamp(min=1)
+            self.by_type.setdefault(ttype, []).extend(losses)
-        perplexity = torch.exp(sentence_loss)  # [batch_size]
+        self.type_total[ttype] = self.type_total.get(ttype, 0.0) + sum(losses)
+        self.type_count[ttype] = self.type_count.get(ttype, 0) + len(losses)
 
-        for text, ppl in zip(batch_texts, perplexity):
+    def stats(self) -> Dict:
-            output_data.append({text_key: text, "ppl": float(ppl.item())})
+        result: Dict = {}
+        if self.count == 0:
+            return result
+        mean_loss = self.total / self.count
+        result["overall"] = {
+            "num_tokens": self.count,
+            "mean_loss": mean_loss,
+            "ppl": float(torch.exp(torch.tensor(mean_loss))),
+        }
+        if self.stream:
+            result["overall"].update(self._hist_percentiles())
+        else:
+            result["overall"]["median_loss"] = statistics.median(self.losses)
+            result["overall"].update(_percentiles(self.losses))
 
-    # Write results
+        if self.type_count:
+            result["by_token_type"] = {}
+            for ttype in sorted(self.type_count.keys()):
+                cnt = self.type_count[ttype]
+                tmean = self.type_total[ttype] / cnt
+                entry: Dict = {
+                    "num_tokens": cnt,
+                    "mean_loss": tmean,
+                    "ppl": float(torch.exp(torch.tensor(tmean))),
+                }
+                if not self.stream and ttype in self.by_type:
+                    entry["median_loss"] = statistics.median(self.by_type[ttype])
+                    entry.update(_percentiles(self.by_type[ttype]))
+                result["by_token_type"][ttype] = entry
+        return result
+
+    def _hist_percentiles(self) -> Dict[str, float]:
+        """Approximate percentiles from the histogram."""
+        total = self.hist.sum().item()
+        if total == 0:
+            return {}
+        cum = torch.cumsum(self.hist.float(), dim=0)
+        result = {}
+        for label, p in [("p50", 0.5), ("p90", 0.9), ("p95", 0.95), ("p99", 0.99)]:
+            target = p * total
+            idx = int(torch.searchsorted(cum, target).item())
+            idx = min(idx, self._HIST_BINS - 1)
+            result[label] = (idx + 0.5) / self._HIST_BINS * self._HIST_MAX
+        return result
+
+
+def process_file(
+    model,
+    tokenizer: AutoTokenizer,
+    items: List[dict],
+    text_key: str,
+    batch_size: int,
+    max_length: int,
+    token_level: bool,
+    max_samples: Optional[int],
+    output_file: Optional[str],
+    label: str,
+    device: str = "cuda",
+) -> Dict:
+    """Evaluate a single dataset (list of items), return summary stats.
+
+    If *token_level* is True and *output_file* is set, per-record token_ids
+    and log_probs are written as JSONL alongside the summary.
+    """
+    if max_samples and len(items) > max_samples:
+        import random
+
+        items = random.sample(items, max_samples)
+
+    texts = [item[text_key] for item in items if text_key in item]
+    print(f"  [{label}] {len(texts)} samples, text_key='{text_key}'")
+
+    acc = LossAccumulator(stream=not token_level)
+    per_sample: List[dict] = []
+
+    if token_level:
+        stop_ids = frozenset(tokenizer.stop_ids)
+        decode_fn = tokenizer.decode
+
+    num_batches = (len(texts) + batch_size - 1) // batch_size
+    for i in tqdm.tqdm(
+        range(0, len(texts), batch_size),
+        total=num_batches,
+        desc=f"  {label}",
+        leave=False,
+    ):
+        batch_texts = texts[i : i + batch_size]
+        padded_ids, masks = _encode_batch(tokenizer, batch_texts, max_length)
+
+        input_ids = torch.tensor(padded_ids, device=device, dtype=torch.long)
+        attention_mask = torch.tensor(masks, device=device, dtype=torch.bool)
+
+        token_log_probs, valid = _compute_batch(model, input_ids, attention_mask)
+
+        for b in range(len(batch_texts)):
+            seq_len = int(valid[b].sum().item())
+            lps = token_log_probs[b, :seq_len].tolist()
+            losses = [-lp for lp in lps]
+            acc.add(losses)
+
+            if token_level:
+                # log_probs correspond to positions 1..seq_len (predicted
+                # from position 0..seq_len-1), so token_ids must skip BOS
+                # at position 0 to stay aligned with log_probs.
+                ids = padded_ids[b][1 : seq_len + 1]
+                per_sample.append(
+                    {
+                        "text": batch_texts[b][:200],
+                        "token_ids": ids,
+                        "log_probs": [round(lp, 4) for lp in lps],
+                        "ppl": float(torch.exp(torch.tensor(statistics.mean(losses))))
+                        if losses
+                        else None,
+                    }
+                )
+                typed_losses: Dict[str, List[float]] = {}
+                for tid, loss in zip(ids, losses):
+                    ttype = _token_type(tid, stop_ids, decode_fn)
+                    typed_losses.setdefault(ttype, []).append(loss)
+                for ttype, tl in typed_losses.items():
+                    acc.add_typed(ttype, tl)
+
+    stats = acc.stats()
+
+    if token_level and output_file:
         with open(output_file, "w", encoding="utf-8") as f:
-        for item in output_data:
+            for item in per_sample:
                 f.write(json.dumps(item, ensure_ascii=False) + "\n")
 
-    print(f"Perplexity computation complete. Results saved to {output_file}")
+    return stats
+
+
+def print_stats(label: str, stats: Dict):
+    """Pretty-print summary statistics."""
+    print(f"\n{'=' * 60}")
+    print(f"  {label}")
+    print(f"{'=' * 60}")
+    ov = stats.get("overall", {})
+    if ov:
+        print(f"  tokens:       {ov['num_tokens']:,}")
+        print(f"  mean loss:    {ov['mean_loss']:.4f}")
+        if "median_loss" in ov:
+            print(f"  median loss:  {ov['median_loss']:.4f}")
+        print(f"  ppl:          {ov['ppl']:.2f}")
+        if "p50" in ov:
+            print(
+                f"  p50/p90/p95/p99: "
+                f"{ov['p50']:.2f} / {ov['p90']:.2f} / {ov['p95']:.2f} / {ov['p99']:.2f}"
+            )
+    by_type = stats.get("by_token_type", {})
+    if by_type:
+        print(f"\n  by token type:")
+        print(f"  {'type':<12} {'count':>8} {'mean_loss':>10} {'ppl':>8}")
+        print(f"  {'-' * 12} {'-' * 8} {'-' * 10} {'-' * 8}")
+        for ttype, s in by_type.items():
+            print(
+                f"  {ttype:<12} {s['num_tokens']:>8,} "
+                f"{s['mean_loss']:>10.4f} {s['ppl']:>8.2f}"
+            )
+
+
+def main(
+    param_path: str,
+    input_path: str,
+    output_dir: str,
+    text_key: str,
+    batch_size: int,
+    max_length: int,
+    token_level: bool,
+    max_samples: Optional[int],
+    device: str = "cuda",
+    dtype: str = "bfloat16",
+):
+    print(f"Loading model from {param_path} ...")
+    model = AutoModel.from_pretrained(param_path)
+    tokenizer = AutoTokenizer.from_pretrained(param_path)
+    torch_dtype = getattr(torch, dtype)
+    model.to(device=device, dtype=torch_dtype)
+    model.eval()
+
+    input_files = _collect_input_files(input_path)
+    if not input_files:
+        print(f"No input files found at {input_path}")
+        return
+
+    print(f"Found {len(input_files)} file(s) to evaluate")
+    os.makedirs(output_dir, exist_ok=True)
+
+    all_stats = {}
+    for filepath in input_files:
+        label = os.path.splitext(os.path.basename(filepath))[0]
+        items = _load_items(filepath)
+        if not items:
+            print(f"  [{label}] empty, skipping")
+            continue
+
+        token_output = (
+            os.path.join(output_dir, f"{label}_tokens.jsonl") if token_level else None
+        )
+
+        stats = process_file(
+            model=model,
+            tokenizer=tokenizer,
+            items=items,
+            text_key=text_key,
+            batch_size=batch_size,
+            max_length=max_length,
+            token_level=token_level,
+            max_samples=max_samples,
+            output_file=token_output,
+            label=label,
+            device=device,
+        )
+        all_stats[label] = stats
+        print_stats(label, stats)
+
+        if token_output:
+            print(f"  token-level output: {token_output}")
+
+    summary_path = os.path.join(output_dir, "summary.json")
+    with open(summary_path, "w", encoding="utf-8") as f:
+        json.dump(all_stats, f, ensure_ascii=False, indent=2)
+    print(f"\nSummary saved to {summary_path}")
 
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Perplexity evaluation on JSONL text.")
+    parser = argparse.ArgumentParser(
+        description="Perplexity and token-level loss evaluation on JSONL/JSON data."
+    )
     parser.add_argument(
         "--param_path", type=str, required=True, help="Path to the model directory."
     )
     parser.add_argument(
-        "--input_file", type=str, required=True, help="Path to the input file."
+        "--input_path",
+        type=str,
+        required=True,
+        help="Path to input file, glob pattern, or directory.",
     )
     parser.add_argument(
-        "--output_file", type=str, required=True, help="Path to the output file."
+        "--output_dir",
-    )
+        type=str,
-    parser.add_argument(
+        required=True,
-        "--batch_size", type=int, default=4, help="Batch size for evaluation."
+        help="Directory for output files (summary.json + per-file token JSONL).",
     )
     parser.add_argument(
         "--text_key",
@@ -105,7 +414,51 @@ if __name__ == "__main__":
         default="text",
         help="Key for the text field in the input data.",
     )
+    parser.add_argument(
+        "--batch_size", type=int, default=4, help="Batch size for evaluation."
+    )
+    parser.add_argument(
+        "--max_length",
+        type=int,
+        default=2048,
+        help="Maximum sequence length (tokens). Longer sequences are truncated.",
+    )
+    parser.add_argument(
+        "--token_level",
+        action="store_true",
+        help="Store per-token log_probs and token type analysis. "
+        "Default: off (only aggregate stats).",
+    )
+    parser.add_argument(
+        "--max_samples",
+        type=int,
+        default=None,
+        help="Maximum number of samples per file (random subsample). Default: all.",
+    )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda" if torch.cuda.is_available() else "cpu",
+        help="Device for model inference.",
+    )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        default="bfloat16" if torch.cuda.is_available() else "float32",
+        help="Torch dtype for model weights.",
+    )
     args = parser.parse_args()
 
     with torch.inference_mode():
-        process_file(**vars(args))
+        main(
+            param_path=args.param_path,
+            input_path=args.input_path,
+            output_dir=args.output_dir,
+            text_key=args.text_key,
+            batch_size=args.batch_size,
+            max_length=args.max_length,
+            token_level=args.token_level,
+            max_samples=args.max_samples,
+            device=args.device,
+            dtype=args.dtype,
+        )

scripts/tools/benchmark.py

@@ -1,12 +1,13 @@
 """Benchmark AutoRegressiveLM with KVCache"""
 
+import argparse
 from dataclasses import dataclass
 from typing import Any, Dict
 
 import torch
 
 from astrai.config import AutoRegressiveLMConfig
-from astrai.inference import KVCache
+from astrai.inference import ContiguousCache, PageCache
 from astrai.model.transformer import AutoRegressiveLM
 
 
@@ -24,41 +25,14 @@ class GenerationBenchmark:
         config: AutoRegressiveLMConfig,
         device: str = "cuda",
         dtype: torch.dtype = torch.bfloat16,
-        page_size: int = 128,
+        cache_type: str = "contiguous",
     ):
         self.config = config
         self.device = device
         self.dtype = dtype
+        self.cache_type = cache_type
         self.model = AutoRegressiveLM(config).to(device=device, dtype=dtype)
         self.model.eval()
-        head_dim = config.dim // config.n_heads
-        n_pages = (config.max_len * 4 + page_size - 1) // page_size
-        self._page_cache = KVCache(
-            config.n_layers,
-            n_pages,
-            page_size,
-            config.n_kv_heads,
-            head_dim,
-            device,
-            dtype,
-        )
-
-    def _prepare_inputs(self, batch_size: int, prompt_length: int, total_length: int):
-        prompt_ids = torch.randint(
-            low=0,
-            high=self.config.vocab_size,
-            size=(batch_size, prompt_length),
-            device=self.device,
-            dtype=torch.long,
-        )
-        gen_ids = torch.randint(
-            low=0,
-            high=self.config.vocab_size,
-            size=(batch_size, total_length - prompt_length),
-            device=self.device,
-            dtype=torch.long,
-        )
-        return prompt_ids, gen_ids
 
     @torch.inference_mode()
     def run_prefill_benchmark(
@@ -68,8 +42,12 @@ class GenerationBenchmark:
         num_trials: int = 10,
     ) -> BenchmarkResult:
         for _ in range(3):
-            prompt_ids, _ = self._prepare_inputs(
+            prompt_ids = torch.randint(
-                batch_size, prompt_length, prompt_length
+                0,
+                self.config.vocab_size,
+                (batch_size, prompt_length),
+                device=self.device,
+                dtype=torch.long,
             )
             _ = self.model(prompt_ids)
         torch.cuda.synchronize()
@@ -78,12 +56,15 @@ class GenerationBenchmark:
         total_tokens = batch_size * prompt_length * num_trials
 
         for trial in range(num_trials):
-            prompt_ids, _ = self._prepare_inputs(
+            prompt_ids = torch.randint(
-                batch_size, prompt_length, prompt_length
+                0,
+                self.config.vocab_size,
+                (batch_size, prompt_length),
+                device=self.device,
+                dtype=torch.long,
             )
             start = torch.cuda.Event(enable_timing=True)
             end = torch.cuda.Event(enable_timing=True)
-
             start.record()
             _ = self.model(prompt_ids)
             end.record()
@@ -107,6 +88,7 @@ class GenerationBenchmark:
                 "prompt_length": prompt_length,
                 "dtype": str(self.dtype),
                 "device": self.device,
+                "cache": "none",
             },
         )
 
@@ -120,29 +102,56 @@ class GenerationBenchmark:
     ) -> BenchmarkResult:
         total_time = 0.0
         total_tokens = batch_size * gen_length * num_trials
-        page_size = self._page_cache.page_size
 
         for trial in range(num_trials):
-            prompt_ids, gen_ids = self._prepare_inputs(
+            prompt_ids = torch.randint(
-                batch_size,
+                0,
-                prompt_length,
+                self.config.vocab_size,
-                prompt_length + gen_length,
+                (batch_size, prompt_length),
-            )
-
-            n_pages = (prompt_length + gen_length + page_size - 1) // page_size
-            total = n_pages * batch_size
-            pages = []
-            for _ in range(total):
-                p = self._page_cache._pool.alloc()
-                assert p >= 0, "OOM"
-                pages.append(p)
-            page_table = torch.tensor(
-                [pages[i * n_pages : (i + 1) * n_pages] for i in range(batch_size)],
-                dtype=torch.long,
                 device=self.device,
+                dtype=torch.long,
+            )
+            gen_ids = torch.randint(
+                0,
+                self.config.vocab_size,
+                (batch_size, gen_length),
+                device=self.device,
+                dtype=torch.long,
             )
 
-            cv = self._page_cache.bind(page_table, total_len=prompt_length)
+            head_dim = self.config.dim // self.config.n_heads
+            max_seq = prompt_length + gen_length
+
+            if self.cache_type == "contiguous":
+                cache = ContiguousCache(
+                    self.config.n_layers,
+                    batch_size,
+                    max_seq,
+                    self.config.n_kv_heads,
+                    head_dim,
+                    self.device,
+                    self.dtype,
+                )
+            else:
+                page_size = 128
+                n_pages = (max_seq + page_size - 1) // page_size * batch_size
+                cache = PageCache(
+                    self.config.n_layers,
+                    n_pages,
+                    page_size,
+                    self.config.n_kv_heads,
+                    head_dim,
+                    self.device,
+                    self.dtype,
+                )
+
+            task_ids = [f"b{i}" for i in range(batch_size)]
+            for tid in task_ids:
+                cache.task_alloc(tid, [0] * max_seq)
+                for p in range(max_seq):
+                    cache.task_extend(tid, p)
+
+            cv = cache.bind_tasks(task_ids, prompt_length, self.device)
             _ = self.model(
                 prompt_ids,
                 paged_cache=cv,
@@ -152,37 +161,35 @@ class GenerationBenchmark:
                 .unsqueeze(0)
                 .expand(batch_size, -1),
             )
-
             torch.cuda.synchronize()
 
             start = torch.cuda.Event(enable_timing=True)
             end = torch.cuda.Event(enable_timing=True)
-
             start.record()
-            current_pos = prompt_length
+
             for i in range(gen_length):
-                input_token = gen_ids[:, i : i + 1]
+                pos = prompt_length + i
-                cv = self._page_cache.bind(page_table, total_len=current_pos + 1)
+                cv = cache.bind_tasks(task_ids, pos + 1, self.device)
                 _ = self.model(
-                    input_token,
+                    gen_ids[:, i : i + 1],
                     paged_cache=cv,
                     position_ids=torch.full(
                         (batch_size, 1),
-                        current_pos,
+                        pos,
                         dtype=torch.long,
                         device=self.device,
                     ),
                 )
-                current_pos += 1
+
             end.record()
             torch.cuda.synchronize()
 
+            for tid in task_ids:
+                cache.task_free(tid)
+
             trial_time = start.elapsed_time(end) / 1000
             total_time += trial_time
 
-            for idx in pages:
-                self._page_cache._pool.free(idx)
-
             print(
                 f"  Trial {trial + 1}/{num_trials}: {gen_length} tokens in {trial_time:.3f}s "
                 f"({gen_length / trial_time:.1f} tok/s)"
@@ -199,6 +206,7 @@ class GenerationBenchmark:
                 "gen_length": gen_length,
                 "dtype": str(self.dtype),
                 "device": self.device,
+                "cache": self.cache_type,
             },
         )
 
@@ -216,6 +224,42 @@ def print_benchmark_result(result: BenchmarkResult):
 
 
 if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="AutoRegressiveLM benchmark")
+    parser.add_argument(
+        "--device", type=str, default="cuda", help="Device (default: cuda)"
+    )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        default="bfloat16",
+        choices=["bfloat16", "float16", "float32"],
+        help="Dtype",
+    )
+    parser.add_argument(
+        "--cache",
+        type=str,
+        default="contiguous",
+        choices=["contiguous", "paged"],
+        help="KV cache type",
+    )
+    parser.add_argument("--batch_size", type=int, default=4, help="Batch size")
+    parser.add_argument("--prompt_length", type=int, default=512, help="Prompt length")
+    parser.add_argument("--gen_length", type=int, default=128, help="Generation length")
+    parser.add_argument("--num_trials", type=int, default=5, help="Number of trials")
+    parser.add_argument(
+        "--prefill_only", action="store_true", help="Run prefill benchmark only"
+    )
+    parser.add_argument(
+        "--decode_only", action="store_true", help="Run decoding benchmark only"
+    )
+    args = parser.parse_args()
+
+    dtype_map = {
+        "bfloat16": torch.bfloat16,
+        "float16": torch.float16,
+        "float32": torch.float32,
+    }
+
     config = AutoRegressiveLMConfig(
         vocab_size=10000,
         dim=1536,
@@ -227,23 +271,29 @@ if __name__ == "__main__":
         norm_eps=1e-5,
     )
 
-    benchmark = GenerationBenchmark(config)
+    benchmark = GenerationBenchmark(
+        config, device=args.device, dtype=dtype_map[args.dtype], cache_type=args.cache
+    )
 
     print("=" * 80)
-    print("Running AutoRegressiveLM Generation Benchmark (KVCache)")
+    print(
+        f"Running AutoRegressiveLM Benchmark (device={args.device}, dtype={args.dtype})"
+    )
     print("=" * 80)
 
+    if not args.decode_only:
         prefill_result = benchmark.run_prefill_benchmark(
-        batch_size=4,
+            batch_size=args.batch_size,
-        prompt_length=512,
+            prompt_length=args.prompt_length,
-        num_trials=5,
+            num_trials=args.num_trials,
         )
         print_benchmark_result(prefill_result)
 
+    if not args.prefill_only:
         gen_result = benchmark.run_decoding_benchmark(
-        batch_size=4,
+            batch_size=args.batch_size,
-        prompt_length=512,
+            prompt_length=args.prompt_length,
-        gen_length=128,
+            gen_length=args.gen_length,
-        num_trials=5,
+            num_trials=args.num_trials,
         )
         print_benchmark_result(gen_result)

scripts/tools/train.py

@@ -1,9 +1,11 @@
 import argparse
 import os
 from functools import partial
+from typing import Any, Dict
 
 import torch
 import torch.optim as optim
+from torch import Tensor, nn
 
 from astrai.config import AutoRegressiveLMConfig, TrainConfig
 from astrai.dataset import DatasetFactory
@@ -12,6 +14,85 @@ from astrai.model.components.decoder_block import DecoderBlock
 from astrai.trainer import SchedulerFactory, Trainer
 
 
+class MuonMix(optim.Optimizer):
+    """Combined Muon (matrix) + AdamW (non-matrix) optimizer."""
+
+    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 = dict(
+            lr=lr,
+            weight_decay=weight_decay,
+            momentum=momentum,
+            nesterov=nesterov,
+            ns_steps=ns_steps,
+            adjust_lr_fn=adjust_lr_fn,
+        )
+        params = [p for p in model.parameters() if p.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 = [*self.muon.param_groups, *self.adamw.param_groups]
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        self.muon.step(closure)
+        self.adamw.step(closure)
+
+    def zero_grad(self, set_to_none: bool = True):
+        self.muon.zero_grad(set_to_none)
+        self.adamw.zero_grad(set_to_none)
+
+    def state_dict(self) -> Dict[str, Any]:
+        return {
+            "muon": self.muon.state_dict(),
+            "adamw": self.adamw.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict: Dict[str, Any]):
+        self.muon.load_state_dict(state_dict["muon"])
+        self.adamw.load_state_dict(state_dict["adamw"])
+        self.param_groups = [*self.muon.param_groups, *self.adamw.param_groups]
+
+
 def parse_args() -> argparse.Namespace:
 
     parser = argparse.ArgumentParser(description="Train the AutoRegressiveLM model.")
@@ -35,6 +116,13 @@ def parse_args() -> argparse.Namespace:
         required=True,
         help="Path to the model parameters or resume checkpoint.",
     )
+    parser.add_argument(
+        "--resume",
+        action="store_true",
+        default=False,
+        help="Resume training from checkpoint at --param_path "
+        "(restore epoch, consumed_samples, optimizer & scheduler state).",
+    )
 
     parser.add_argument(
         "--n_epoch", type=int, default=1, help="Number of epochs to train."
@@ -64,22 +152,35 @@ def parse_args() -> argparse.Namespace:
         help="Max gradient norm for clipping.",
     )
     parser.add_argument(
-        "--adamw_beta1",
+        "--weight_decay",
         type=float,
-        default=0.9,
+        default=0.1,
-        help="Beta1 for AdamW optimizer.",
+        help="Weight decay (applied to Muon matrix params; non-matrix use 0).",
     )
     parser.add_argument(
-        "--adamw_beta2",
+        "--muon_momentum",
         type=float,
         default=0.95,
-        help="Beta2 for AdamW optimizer.",
+        help="Momentum factor for Muon optimizer.",
     )
     parser.add_argument(
-        "--adamw_weight_decay",
+        "--muon_nesterov",
-        type=float,
+        action=argparse.BooleanOptionalAction,
-        default=0.01,
+        default=True,
-        help="Weight decay for AdamW optimizer.",
+        help="Enable Nesterov momentum for Muon.",
+    )
+    parser.add_argument(
+        "--muon_ns_steps",
+        type=int,
+        default=5,
+        help="Newton-Schulz iteration steps for Muon.",
+    )
+    parser.add_argument(
+        "--muon_adjust_lr",
+        type=str,
+        default="match_rms_adamw",
+        choices=["original", "match_rms_adamw"],
+        help="Muon learning rate adjustment strategy.",
     )
     parser.add_argument(
         "--random_seed", type=int, default=3407, help="Random seed for reproducibility."
@@ -159,12 +260,6 @@ def parse_args() -> argparse.Namespace:
         default="checkpoint/logs",
         help="Directory for metric logs.",
     )
-    parser.add_argument(
-        "--grpo_sync_interval",
-        type=int,
-        default=200,
-        help="GRPO ref model sync interval (steps).",
-    )
     parser.add_argument(
         "--start_epoch", type=int, default=0, help="Start epoch for training."
     )
@@ -265,21 +360,8 @@ def create_model(config):
     return AutoRegressiveLM(config).to(dtype=torch.bfloat16)
 
 
-def create_optimizer(model, **kwargs) -> optim.Optimizer:
+def create_optimizer(model, **kwargs) -> MuonMix:
-    decay_params = []
+    return MuonMix(model, **kwargs)
-    no_decay_params = []
-    for name, param in model.named_parameters():
-        if not param.requires_grad:
-            continue
-        if param.dim() < 2 or "norm" in name or "bias" in name:
-            no_decay_params.append(param)
-        else:
-            decay_params.append(param)
-    param_groups = [
-        {"params": decay_params, "weight_decay": kwargs.pop("weight_decay", 0.01)},
-        {"params": no_decay_params, "weight_decay": 0.0},
-    ]
-    return optim.AdamW(param_groups, fused=True, **kwargs)
 
 
 def create_scheduler(
@@ -310,7 +392,7 @@ def train(
     train_type: str,
     param_path: str,
     data_root_path: str,
-    max_lr: float,
+    resume: bool,
     n_epoch: int,
     batch_per_device: int,
     start_epoch: int,
@@ -323,16 +405,7 @@ def train(
     val_step: int,
     metrics: list[str],
     log_dir: str,
-    dpo_beta: float,
-    grpo_clip_eps: float,
-    grpo_kl_coef: float,
-    group_size: int,
-    grpo_sync_interval: int,
-    adamw_beta1: float,
-    adamw_beta2: float,
-    adamw_weight_decay: float,
     max_grad_norm: float,
-    label_smoothing: float,
     random_seed: int,
     num_workers: int,
     pin_memory: bool,
@@ -353,6 +426,7 @@ def train(
     t_mult: int,
     stable_steps: int,
     decay_steps: int,
+    **kwargs,
 ):
     assert train_type in ["seq", "sft", "dpo", "grpo"]
     assert os.path.exists(param_path)
@@ -368,12 +442,11 @@ def train(
         window_size = config.max_len
 
     strategy_kwargs = {
-        "beta": dpo_beta,
+        "beta": kwargs.pop("dpo_beta"),
-        "label_smoothing": label_smoothing,
+        "label_smoothing": kwargs.pop("label_smoothing"),
-        "clip_eps": grpo_clip_eps,
+        "clip_eps": kwargs.pop("grpo_clip_eps"),
-        "kl_coef": grpo_kl_coef,
+        "kl_coef": kwargs.pop("grpo_kl_coef"),
-        "group_size": group_size,
+        "group_size": kwargs.pop("group_size"),
-        "sync_interval": grpo_sync_interval,
     }
 
     executor_kwargs = {
@@ -391,11 +464,12 @@ def train(
 
     optimizer_fn = partial(
         create_optimizer,
-        **{
+        lr=kwargs.pop("max_lr"),
-            "lr": max_lr,
+        weight_decay=kwargs.pop("weight_decay"),
-            "betas": (adamw_beta1, adamw_beta2),
+        momentum=kwargs.pop("muon_momentum"),
-            "weight_decay": adamw_weight_decay,
+        nesterov=kwargs.pop("muon_nesterov"),
-        },
+        ns_steps=kwargs.pop("muon_ns_steps"),
+        adjust_lr_fn=kwargs.pop("muon_adjust_lr"),
     )
 
     total_steps = compute_total_steps(
@@ -465,7 +539,7 @@ def train(
     )
 
     trainer = Trainer(train_config)
-    trainer.train(resume_dir=param_path)
+    trainer.train(param_path=param_path, resume=resume)
 
 
 if __name__ == "__main__":

setup.py

@@ -0,0 +1,61 @@
+import os
+import sys
+from pathlib import Path
+
+from setuptools import setup
+from setuptools.command.build_ext import build_ext as _build_ext
+
+sys.path.insert(0, str(Path(__file__).parent))
+os.makedirs("astrai/extension", exist_ok=True)
+
+
+def _should_build():
+    force = os.environ.get("CSRC_KERNELS", "").strip().lower()
+    if force == "true":
+        return True
+    if force == "false":
+        return False
+    try:
+        import shutil
+
+        import torch
+
+        return shutil.which("nvcc") is not None and torch.cuda.is_available()
+    except Exception:
+        return False
+
+
+ext_modules = []
+cmdclass = {}
+
+if _should_build():
+    import torch
+    from torch.utils.cpp_extension import BuildExtension, CUDAExtension
+
+    from csrc.build import REGISTRY
+
+    _torch_lib = torch.utils.cpp_extension.library_paths()[0]
+
+    for name, info in REGISTRY.items():
+        ext_modules.append(
+            CUDAExtension(
+                f"astrai.extension.{name}",
+                info["sources"],
+                extra_compile_args={
+                    "cxx": info["cxx_flags"],
+                    "nvcc": info["nvcc_flags"],
+                },
+                extra_link_args=[f"-Wl,-rpath,{_torch_lib}"],
+            )
+        )
+    cmdclass["build_ext"] = BuildExtension
+
+if not cmdclass:
+
+    class _NullBuildExt(_build_ext):
+        def build_extensions(self):
+            pass
+
+    cmdclass["build_ext"] = _NullBuildExt
+
+setup(ext_modules=ext_modules, cmdclass=cmdclass)

tests/data/conftest.py

@@ -10,7 +10,11 @@ from astrai.config.preprocess_config import (
     PipelineConfig,
     ProcessingConfig,
 )
-from astrai.preprocessing.builder import SectionedMaskBuilder
+from astrai.preprocessing.builder import (
+    MultiOutputMaskBuilder,
+    SectionedMaskBuilder,
+    SingleOutputMaskBuilder,
+)
 from astrai.tokenize import AutoTokenizer
 
 _SPECIAL_TOKENS_CONFIG = {
@@ -210,6 +214,16 @@ def builder():
     return SectionedMaskBuilder()
 
 
+@pytest.fixture
+def single_builder():
+    return SingleOutputMaskBuilder()
+
+
+@pytest.fixture
+def multi_builder():
+    return MultiOutputMaskBuilder()
+
+
 @pytest.fixture
 def tokenizer_dir(temp_dir, test_tokenizer):
     d = os.path.join(temp_dir, "tok")

tests/data/test_dataset.py

@@ -327,43 +327,82 @@ def test_normalize_mixed_empty_key():
 
 
 def test_grpo_dataset_dtype(base_test_env):
+    """GRPO dataset returns correct dtypes for per-record structured data."""
+    from astrai.dataset.dataset import GRPODataset
+
     test_dir = base_test_env["test_dir"]
-    dummy_data = {
+    G = 4
-        "prompts": [torch.randint(0, 100, (100,), dtype=torch.int32)],
+    dataset = GRPODataset()
-        "responses": [torch.randint(0, 100, (100,), dtype=torch.int32)],
+    dataset.storage = type(
-        "masks": [torch.ones(100, dtype=torch.int32)],
+        "FakeStore",
-        "rewards": [torch.ones(100, dtype=torch.float32)],
+        (),
-    }
+        {
-    dataset = _make_seq_dataset(
+            "keys": ["prompts", "responses", "masks", "rewards"],
-        test_dir, "grpo_dtype", train_type="grpo", data=dummy_data, window_size=32
+            "_data": {
-    )
+                "prompts": [torch.randint(0, 100, (10,), dtype=torch.int32)],
+                "responses": [
+                    [torch.randint(0, 100, (5,), dtype=torch.int32) for _ in range(G)]
+                ],
+                "masks": [[torch.ones(5, dtype=torch.int32) for _ in range(G)]],
+                "rewards": [torch.rand(G, dtype=torch.float32)],
+            },
+        },
+    )()
+    dataset._build_records()
     item = dataset[0]
 
     assert item["prompts"].dtype == torch.long
-    assert item["responses"].dtype == torch.long
+    assert all(r.dtype == torch.long for r in item["responses"])
-    assert item["masks"].dtype == torch.bool
+    assert all(m.dtype == torch.bool for m in item["masks"])
     assert item["rewards"].dtype == torch.float32
 
 
 def test_grpo_dataset_load(base_test_env):
+    """GRPO dataset loads record-structured data with per-response boundaries."""
+    from astrai.dataset.dataset import GRPODataset
+
     test_dir = base_test_env["test_dir"]
-    dummy_data = {
+    G = 3
-        "prompts": [_rand_seq(200)],
+    prompt_len = 8
-        "responses": [_rand_seq(200)],
+    resp_lens = [5, 7, 4]
-        "masks": [torch.ones(200, dtype=torch.int64)],
+    dataset = GRPODataset()
-        "rewards": [torch.rand(200, dtype=torch.float32)],
+    dataset.storage = type(
-    }
+        "FakeStore",
-    dataset = _make_seq_dataset(
+        (),
-        test_dir, "grpo_test", train_type="grpo", data=dummy_data
+        {
-    )
+            "keys": ["prompts", "responses", "masks", "rewards"],
-    assert len(dataset) > 0
+            "_data": {
+                "prompts": [torch.randint(0, 100, (prompt_len,))],
+                "responses": [[torch.randint(0, 100, (rl,)) for rl in resp_lens]],
+                "masks": [[torch.ones(rl, dtype=torch.int64) for rl in resp_lens]],
+                "rewards": [torch.tensor([0.9, 0.3, 0.7], dtype=torch.float32)],
+            },
+        },
+    )()
+    dataset._build_records()
+
+    assert len(dataset) == 1
     item = dataset[0]
     assert "prompts" in item
     assert "responses" in item
     assert "masks" in item
     assert "rewards" in item
-    assert item["prompts"].shape[0] == 64
+
-    assert item["responses"].shape[0] == 64
+    # Prompts is 1-D
+    assert item["prompts"].shape == (prompt_len,)
+
+    # Responses is a list of G tensors with correct lengths
+    assert len(item["responses"]) == G
+    for i, r in enumerate(item["responses"]):
+        assert r.shape == (resp_lens[i],)
+
+    # Masks align with responses
+    assert len(item["masks"]) == G
+    for i, m in enumerate(item["masks"]):
+        assert m.shape == (resp_lens[i],)
+
+    # Rewards has G elements
+    assert item["rewards"].shape == (G,)
 
 
 def test_detect_format_bin_dir(base_test_env):
@@ -411,6 +450,32 @@ def test_dataset_load_explicit_storage_type(base_test_env):
     assert dataset.count == 200
 
 
+def _write_json_dataset(test_dir, tokenizer_path, records, config_overrides=None):
+    """Write JSON (not JSONL) dataset — array of objects."""
+    data_dir = os.path.join(test_dir, "json_data")
+    os.makedirs(data_dir, exist_ok=True)
+
+    with open(os.path.join(data_dir, "data.json"), "w", encoding="utf-8") as f:
+        json.dump(records, f, ensure_ascii=False)
+
+    config = {
+        "tokenizer_path": tokenizer_path,
+        "version": 1,
+        "input": {"sections": [{"field": "text", "action": "train"}]},
+        "preprocessing": {"max_seq_len": 128, "min_chars": 0},
+        "output": {"position_ids_mode": "continuous"},
+    }
+    if config_overrides:
+        config.update(config_overrides)
+
+    with open(
+        os.path.join(data_dir, "dataset_config.json"), "w", encoding="utf-8"
+    ) as f:
+        json.dump(config, f, ensure_ascii=False, indent=2)
+
+    return data_dir
+
+
 def test_detect_format_jsonl_dir(base_test_env):
     test_dir = base_test_env["test_dir"]
     tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
@@ -422,6 +487,89 @@ def test_detect_format_jsonl_dir(base_test_env):
     assert detect_format(data_dir) == "jsonl"
 
 
+def test_detect_format_json_dir(base_test_env):
+    """detect_format returns 'jsonl' for directory with .json files."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
+    data_dir = _write_json_dataset(
+        test_dir,
+        tokenizer_path,
+        [{"text": "hello world"}, {"text": "foo bar baz qux"}],
+    )
+    assert detect_format(data_dir) == "jsonl"
+
+
+def test_json_store_seq(base_test_env):
+    """JsonlStore loads .json array correctly."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
+    data_dir = _write_json_dataset(
+        test_dir,
+        tokenizer_path,
+        [{"text": "hello world"}, {"text": "foo bar baz qux"}],
+    )
+
+    store = StoreFactory.create("jsonl")
+    store.load(data_dir)
+    assert len(store) > 0
+    assert "sequence" in store.keys
+
+    dataset = DatasetFactory.load("seq", data_dir, window_size=8)
+    assert len(dataset) > 0
+    item = dataset[0]
+    assert "input_ids" in item
+    assert "target_ids" in item
+
+
+def test_json_store_no_tokenizer_path(base_test_env):
+    """JsonlStore uses dataset dir as tokenizer_path when omitted."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer = base_test_env["tokenizer"]
+    tokenizer.set_chat_template(
+        "{% for message in messages %}{{ message['role'] }}:{{ message['content'] }}\n{% endfor %}"
+    )
+
+    data_dir = os.path.join(test_dir, "self_contained")
+    os.makedirs(data_dir, exist_ok=True)
+
+    # Save tokenizer files directly in the dataset directory
+    tokenizer.save_pretrained(data_dir)
+
+    # Write .json data
+    records = [
+        {
+            "messages": [
+                {"role": "user", "content": "hi"},
+                {"role": "assistant", "content": "hello"},
+            ]
+        }
+    ]
+    with open(os.path.join(data_dir, "data.json"), "w", encoding="utf-8") as f:
+        json.dump(records, f, ensure_ascii=False)
+
+    # dataset_config.json WITHOUT tokenizer_path
+    config = {
+        "version": 1,
+        "input": {
+            "sections": [{"field": "messages", "action": "$role", "template": True}]
+        },
+        "mask": {"user": "mask", "assistant": "train"},
+        "mask_default": "mask",
+        "preprocessing": {"max_seq_len": 128, "min_chars": 0},
+        "output": {"position_ids_mode": "continuous"},
+    }
+    with open(
+        os.path.join(data_dir, "dataset_config.json"), "w", encoding="utf-8"
+    ) as f:
+        json.dump(config, f, ensure_ascii=False, indent=2)
+
+    store = StoreFactory.create("jsonl")
+    store.load(data_dir)
+    assert len(store) > 0
+    assert "sequence" in store.keys
+    assert "loss_mask" in store.keys
+
+
 def test_jsonl_store_seq(base_test_env):
     test_dir = base_test_env["test_dir"]
     tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
@@ -512,3 +660,231 @@ def test_jsonl_store_pipeline_config_roundtrip(base_test_env):
     config = PipelineConfig.from_dict(raw)
     assert config.output.position_ids_mode == "doc_reset"
     assert config.preprocessing.max_seq_len == 64
+
+
+# ---------------------------------------------------------------------------
+# GRPO end-to-end: builder → JsonlStore → GRPODataset → collate_fn
+# ---------------------------------------------------------------------------
+
+
+def _write_grpo_jsonl(test_dir, tokenizer_path, records):
+    """Write a GRPO JSONL dataset directory with config."""
+    data_dir = os.path.join(test_dir, "grpo_jsonl")
+    os.makedirs(data_dir, exist_ok=True)
+
+    with open(os.path.join(data_dir, "data.jsonl"), "w", encoding="utf-8") as f:
+        for rec in records:
+            f.write(json.dumps(rec, ensure_ascii=False) + "\n")
+
+    config = {
+        "tokenizer_path": tokenizer_path,
+        "version": 1,
+        "input": {
+            "sources": {
+                "prompts": {
+                    "sections": [
+                        {
+                            "field": "prompt",
+                            "action": "mask",
+                            "add_special_tokens": True,
+                        }
+                    ]
+                },
+                "responses": {
+                    "sections": [{"field": "responses", "action": "train"}],
+                    "list_field": True,
+                    "mask_key": "masks",
+                },
+                "rewards": {
+                    "sections": [{"field": "rewards", "action": "value"}],
+                },
+            }
+        },
+        "mask": {"user": "mask", "assistant": "train"},
+        "mask_default": "mask",
+        "preprocessing": {"max_seq_len": 128},
+        "output": {"position_ids_mode": "none"},
+    }
+
+    with open(
+        os.path.join(data_dir, "dataset_config.json"), "w", encoding="utf-8"
+    ) as f:
+        json.dump(config, f, ensure_ascii=False, indent=2)
+
+    return data_dir
+
+
+def test_grpo_builder_preserves_response_boundaries(base_test_env):
+    """MultiOutputMaskBuilder with list_field returns List[List[int]] for responses."""
+    from astrai.preprocessing.builder import SectionedMaskBuilder
+    from tests.data.conftest import make_grpo_no_template_config
+
+    tokenizer = base_test_env["tokenizer"]
+    tokenizer_path = _save_test_tokenizer(base_test_env["test_dir"], tokenizer)
+
+    builder = SectionedMaskBuilder()
+    config = make_grpo_no_template_config()
+    config.preprocessing.max_seq_len = 128
+
+    item = {
+        "prompt": "What is 2+2?",
+        "responses": ["4", "four", "2+2=4"],
+        "rewards": [0.9, 0.1, 0.5],
+    }
+
+    result = builder.build(item, config, tokenizer)
+    assert result is not None
+
+    # prompts should be flat list of ints
+    assert isinstance(result["prompts"], list)
+    assert isinstance(result["prompts"][0], int)
+
+    # responses should be list of lists (one per response)
+    assert isinstance(result["responses"], list)
+    assert isinstance(result["responses"][0], list)
+    assert isinstance(result["responses"][0][0], int)
+    assert len(result["responses"]) == 3
+
+    # masks should match responses structure
+    assert isinstance(result["masks"], list)
+    assert len(result["masks"]) == 3
+    for i in range(3):
+        assert len(result["masks"][i]) == len(result["responses"][i])
+
+    # rewards should be flat list of floats
+    assert isinstance(result["rewards"], list)
+    assert all(isinstance(r, float) for r in result["rewards"])
+    assert len(result["rewards"]) == 3
+
+
+def test_grpo_end_to_end_jsonl(base_test_env):
+    """Full GRPO pipeline: JSONL → JsonlStore → GRPODataset → collate_fn."""
+    from astrai.dataset.dataset import grpo_collate_fn
+
+    test_dir = base_test_env["test_dir"]
+    tokenizer = base_test_env["tokenizer"]
+    tokenizer_path = _save_test_tokenizer(test_dir, tokenizer)
+
+    records = [
+        {
+            "prompt": "What is 2+2?",
+            "responses": ["4", "four", "The answer is 4"],
+            "rewards": [0.9, 0.1, 0.5],
+        },
+        {
+            "prompt": "Write a haiku",
+            "responses": ["Leaves fall", "Cherry blossoms bloom in spring"],
+            "rewards": [0.3, 0.8],
+        },
+    ]
+
+    data_dir = _write_grpo_jsonl(test_dir, tokenizer_path, records)
+
+    dataset = DatasetFactory.load("grpo", data_dir, window_size=0)
+    assert len(dataset) == 2
+
+    # Item 0: 3 responses
+    item0 = dataset[0]
+    assert item0["prompts"].ndim == 1
+    assert len(item0["responses"]) == 3
+    assert len(item0["masks"]) == 3
+    assert item0["rewards"].shape == (3,)
+    for r, m in zip(item0["responses"], item0["masks"]):
+        assert r.shape == m.shape
+
+    # Item 1: 2 responses (different group size)
+    item1 = dataset[1]
+    assert len(item1["responses"]) == 2
+    assert item1["rewards"].shape == (2,)
+
+    # Collate: batch records with same G (item0 has G=3)
+    batch = grpo_collate_fn([item0, item0])
+    assert batch["prompts"].shape[0] == 2
+    assert batch["responses"].ndim == 3
+    assert batch["responses"].shape[0] == 2
+    assert batch["responses"].shape[1] == 3  # G=3
+    assert batch["masks"].shape == batch["responses"].shape
+    assert batch["rewards"].shape == (2, 3)
+
+
+def test_grpo_collate_variable_lengths():
+    """collate_fn pads variable-length responses to [B, G, R_max]."""
+    from astrai.dataset.dataset import grpo_collate_fn
+
+    batch = [
+        {
+            "prompts": torch.tensor([1, 2, 3]),
+            "responses": [torch.tensor([4, 5]), torch.tensor([6, 7, 8, 9])],
+            "masks": [torch.tensor([1, 1]), torch.tensor([1, 1, 1, 1])],
+            "rewards": torch.tensor([0.9, 0.1]),
+        },
+        {
+            "prompts": torch.tensor([10, 11]),
+            "responses": [torch.tensor([12]), torch.tensor([13, 14, 15])],
+            "masks": [torch.tensor([1]), torch.tensor([1, 1, 1])],
+            "rewards": torch.tensor([0.5, 0.5]),
+        },
+    ]
+
+    result = grpo_collate_fn(batch)
+
+    assert result["prompts"].shape == (2, 3)  # B=2, P_max=3
+    assert result["responses"].shape == (2, 2, 4)  # B=2, G=2, R_max=4
+    assert result["masks"].shape == (2, 2, 4)
+    assert result["rewards"].shape == (2, 2)
+
+    # Check padding: item 1 prompt is length 2, padded to 3
+    assert result["prompts"][1, 2] == 0
+
+    # Check response content: item 0, response 0 is [4,5] padded to 4
+    assert result["responses"][0, 0, 0] == 4
+    assert result["responses"][0, 0, 1] == 5
+    assert result["responses"][0, 0, 2] == 0  # padded
+    assert result["masks"][0, 0, 2] == False  # padded
+
+    # Check response content: item 0, response 1 is [6,7,8,9] no padding
+    assert result["responses"][0, 1, 3] == 9
+    assert result["masks"][0, 1, 3] == True
+
+
+def test_grpo_multiple_records(base_test_env):
+    """GRPODataset loads multiple records with correct structure."""
+    from astrai.dataset.dataset import GRPODataset
+
+    G = 4
+    n_records = 5
+
+    dummy_responses = [
+        [torch.randint(0, 100, (np.random.randint(3, 8),)) for _ in range(G)]
+        for _ in range(n_records)
+    ]
+    dataset = GRPODataset()
+    dataset.storage = type(
+        "FakeStore",
+        (),
+        {
+            "keys": ["prompts", "responses", "masks", "rewards"],
+            "_data": {
+                "prompts": [torch.randint(0, 100, (10,)) for _ in range(n_records)],
+                "responses": dummy_responses,
+                "masks": [
+                    [torch.ones(r.shape[0], dtype=torch.int64) for r in resps]
+                    for resps in dummy_responses
+                ],
+                "rewards": [
+                    torch.rand(G, dtype=torch.float32) for _ in range(n_records)
+                ],
+            },
+        },
+    )()
+    dataset._build_records()
+
+    assert len(dataset) == n_records
+
+    for i in range(n_records):
+        item = dataset[i]
+        assert len(item["responses"]) == G
+        assert len(item["masks"]) == G
+        assert item["rewards"].shape == (G,)
+        for g in range(G):
+            assert item["responses"][g].shape == item["masks"][g].shape

tests/data/test_preprocess_builder.py

@@ -8,7 +8,9 @@ from astrai.config.preprocess_config import (
 )
 from astrai.preprocessing.builder import (
     MaskBuilderFactory,
+    MultiOutputMaskBuilder,
     SectionedMaskBuilder,
+    SingleOutputMaskBuilder,
 )
 from tests.data.conftest import (
     _CHAT_SECTIONS,
@@ -272,12 +274,18 @@ def test_sectioned_text_too_short(test_tokenizer, builder):
 
 def test_factory_registered():
     names = MaskBuilderFactory.list_registered()
+    assert "single" in names
+    assert "multi" in names
     assert "sectioned" in names
 
 
 def test_factory_create():
-    builder_obj = MaskBuilderFactory.create("sectioned")
+    single = MaskBuilderFactory.create("single")
-    assert isinstance(builder_obj, SectionedMaskBuilder)
+    assert isinstance(single, SingleOutputMaskBuilder)
+    multi = MaskBuilderFactory.create("multi")
+    assert isinstance(multi, MultiOutputMaskBuilder)
+    sectioned = MaskBuilderFactory.create("sectioned")
+    assert isinstance(sectioned, SectionedMaskBuilder)
 
 
 def test_dpo_chat_basic(chat_tokenizer, builder):
@@ -341,7 +349,17 @@ def test_grpo_basic(chat_tokenizer, builder):
     assert "responses" in result
     assert "masks" in result
     assert "rewards" in result
-    assert len(result["responses"]) == len(result["masks"])
+
+    # responses is List[List[int]] — one per response
+    assert len(result["responses"]) == 4
+    assert all(isinstance(r, list) for r in result["responses"])
+    assert all(isinstance(r[0], int) for r in result["responses"])
+
+    # masks is List[List[int]] — one per response, matching length
+    assert len(result["masks"]) == 4
+    for i in range(4):
+        assert len(result["masks"][i]) == len(result["responses"][i])
+
     assert result["rewards"] == [1.0, 0.5, 0.8, 0.2]
 
 
@@ -354,8 +372,11 @@ def test_grpo_response_tokens_all_trained(chat_tokenizer, builder):
     }
     result = builder.build(item, config, chat_tokenizer)
     masks = result["masks"]
-    assert all(m == 1 for m in masks)
+    # masks is List[List[int]] — each response's mask should be all 1s
-    assert len(masks) == len(result["responses"])
+    assert len(masks) == 2
+    for m in masks:
+        assert all(v == 1 for v in m)
+        assert len(m) == len(result["responses"][masks.index(m)])
 
 
 def test_grpo_single_reward(chat_tokenizer, builder):
@@ -367,3 +388,59 @@ def test_grpo_single_reward(chat_tokenizer, builder):
     }
     result = builder.build(item, config, chat_tokenizer)
     assert result["rewards"] == [0.9]
+
+
+def test_single_builder_matches_facade(chat_tokenizer, builder, single_builder):
+    config = make_chat_config()
+    item = {
+        "messages": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ]
+    }
+    facade_result = builder.build(item, config, chat_tokenizer)
+    single_result = single_builder.build(item, config, chat_tokenizer)
+    assert single_result == facade_result
+
+
+def test_single_builder_rejects_multi_config(chat_tokenizer, single_builder):
+    config = make_dpo_chat_config()
+    item = {
+        "chosen": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ],
+        "rejected": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "5"},
+        ],
+    }
+    assert single_builder.build(item, config, chat_tokenizer) is None
+
+
+def test_multi_builder_matches_facade(chat_tokenizer, builder, multi_builder):
+    config = make_dpo_chat_config()
+    item = {
+        "chosen": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ],
+        "rejected": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "5"},
+        ],
+    }
+    facade_result = builder.build(item, config, chat_tokenizer)
+    multi_result = multi_builder.build(item, config, chat_tokenizer)
+    assert multi_result == facade_result
+
+
+def test_multi_builder_rejects_single_config(chat_tokenizer, multi_builder):
+    config = make_chat_config()
+    item = {
+        "messages": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ]
+    }
+    assert multi_builder.build(item, config, chat_tokenizer) is None

tests/data/test_preprocess_pipeline.py

@@ -7,6 +7,7 @@ from astrai.config.preprocess_config import (
     PipelineConfig,
     ProcessingConfig,
 )
+from astrai.preprocessing.packing import PackingStrategyFactory
 from astrai.preprocessing.pipeline import Pipeline, filter_by_length
 from tests.data.conftest import (
     _CHAT_SECTIONS,
@@ -262,3 +263,69 @@ def test_grpo_pipeline(temp_dir, tokenizer_dir):
     assert "masks" in meta
     assert "rewards" in meta
     assert "sequence" not in meta
+
+
+# ---------------------------------------------------------------------------
+# BFD split packing
+# ---------------------------------------------------------------------------
+
+_TRU = "keep_start"
+
+
+def _total_tokens(keys, key="sequence"):
+    return sum(len(s) for s in keys[key])
+
+
+def test_bfd_split_preserves_all_tokens():
+    """No tokens are lost — split chunks are kept, not truncated away."""
+    packer = PackingStrategyFactory.create("bfd_split")
+    max_len = 10
+    keys = {
+        "sequence": [list(range(25)), list(range(3))],
+        "loss_mask": [[1] * 25, [1] * 3],
+    }
+    result = packer.apply(keys, max_len, _TRU)
+
+    assert _total_tokens(result) == 28
+    for seq in result["sequence"]:
+        assert len(seq) <= max_len
+
+
+def test_bfd_split_chunk_alignment():
+    """loss_mask chunks must align with sequence chunks."""
+    packer = PackingStrategyFactory.create("bfd_split")
+    max_len = 10
+    keys = {
+        "sequence": [list(range(25))],
+        "loss_mask": [[0] * 5 + [1] * 20],
+    }
+    result = packer.apply(keys, max_len, _TRU)
+
+    for seq, mask in zip(result["sequence"], result["loss_mask"]):
+        assert len(seq) == len(mask)
+
+
+def test_bfd_split_short_unchanged():
+    """Sequences under max_packed_len should not be split."""
+    packer = PackingStrategyFactory.create("bfd_split")
+    max_len = 10
+    keys = {"sequence": [list(range(5))], "loss_mask": [[1] * 5]}
+    result = packer.apply(keys, max_len, _TRU)
+
+    assert _total_tokens(result) == 5
+    assert len(result["sequence"]) >= 1
+
+
+def test_bfd_split_vs_bfd():
+    """bfd loses tokens from over-length sequences; bfd_split does not."""
+    max_len = 10
+    keys = {
+        "sequence": [list(range(25)), list(range(8))],
+        "loss_mask": [[1] * 25, [1] * 8],
+    }
+
+    bfd = PackingStrategyFactory.create("bfd").apply(keys, max_len, _TRU)
+    split = PackingStrategyFactory.create("bfd_split").apply(keys, max_len, _TRU)
+
+    assert _total_tokens(bfd) < 33
+    assert _total_tokens(split) == 33

tests/inference/test_sample.py

@@ -3,6 +3,7 @@
 import torch
 
 from astrai.inference.sample import (
+    FrequencyPenaltyStrategy,
     SamplingPipeline,
     TemperatureStrategy,
     TopKStrategy,
@@ -125,3 +126,108 @@ def test_module_sample_batch():
     assert tokens.shape == (2,)
     for t in tokens:
         assert 0 <= t < logits.size(-1)
+
+
+def test_frequency_penalty_noop_when_zero():
+    logits = torch.tensor([[1.0, 2.0, 3.0]])
+    input_ids = torch.tensor([[0, 2]])
+    s = FrequencyPenaltyStrategy(penalty=0.0)
+    result = s.apply(logits.clone(), input_ids=input_ids)
+    assert torch.equal(result, logits)
+
+
+def test_frequency_penalty_noop_when_no_input_ids():
+    logits = torch.tensor([[1.0, 2.0, 3.0]])
+    s = FrequencyPenaltyStrategy(penalty=0.5)
+    result = s.apply(logits.clone())
+    assert torch.equal(result, logits)
+
+
+def test_frequency_penalty_single_occurrence():
+    logits = torch.tensor([[4.0, 1.0, 2.0]])
+    input_ids = torch.tensor([[0, 2]])
+    input_mask = torch.tensor([[True, True]])
+    s = FrequencyPenaltyStrategy(penalty=0.5)
+    result = s.apply(logits.clone(), input_ids=input_ids, input_mask=input_mask)
+    assert result[0, 0] == 3.5
+    assert result[0, 1] == 1.0
+    assert result[0, 2] == 1.5
+
+
+def test_frequency_penalty_multiple_occurrences():
+    logits = torch.tensor([[4.0, 1.0, 2.0]])
+    input_ids = torch.tensor([[0, 2, 0]])
+    input_mask = torch.tensor([[True, True, True]])
+    s = FrequencyPenaltyStrategy(penalty=0.5)
+    result = s.apply(logits.clone(), input_ids=input_ids, input_mask=input_mask)
+    assert result[0, 0] == 3.0
+    assert result[0, 1] == 1.0
+    assert result[0, 2] == 1.5
+
+
+def test_frequency_penalty_respects_padding_mask():
+    logits = torch.tensor([[4.0, 1.0, 2.0]])
+    input_ids = torch.tensor([[0, 2, 0]])
+    input_mask = torch.tensor([[True, True, False]])
+    s = FrequencyPenaltyStrategy(penalty=0.5)
+    result = s.apply(logits.clone(), input_ids=input_ids, input_mask=input_mask)
+    assert result[0, 0] == 3.5
+    assert result[0, 1] == 1.0
+    assert result[0, 2] == 1.5
+
+
+def test_frequency_penalty_batch_tensor():
+    logits = torch.tensor(
+        [
+            [4.0, 1.0, 2.0],
+            [3.0, 5.0, 1.0],
+        ]
+    )
+    input_ids = torch.tensor([[0, 2, 0], [1, 1, 0]])
+    input_mask = torch.tensor([[True, True, True], [True, True, False]])
+    s = FrequencyPenaltyStrategy(penalty=torch.tensor([0.5, 1.0]))
+    result = s.apply(logits.clone(), input_ids=input_ids, input_mask=input_mask)
+    assert result[0, 0] == 3.0
+    assert result[0, 2] == 1.5
+    assert result[1, 1] == 3.0
+
+
+def test_frequency_penalty_negative_penalty_boosts_repeats():
+    logits = torch.tensor([[4.0, 1.0, 2.0]])
+    input_ids = torch.tensor([[0, 0]])
+    input_mask = torch.tensor([[True, True]])
+    s = FrequencyPenaltyStrategy(penalty=-0.5)
+    result = s.apply(logits.clone(), input_ids=input_ids, input_mask=input_mask)
+    assert result[0, 0] == 5.0
+
+
+def test_frequency_penalty_in_pipeline():
+    logits = torch.tensor([[5.0, 1.0, 2.0, 3.0]])
+    input_ids = torch.tensor([[0, 2, 0]])
+    input_mask = torch.tensor([[True, True, True]])
+    pipeline = SamplingPipeline(
+        [
+            TemperatureStrategy(1.0),
+            FrequencyPenaltyStrategy(0.5),
+        ]
+    )
+    result = pipeline.apply(logits.clone(), input_ids=input_ids, input_mask=input_mask)
+    assert result[0, 0] == 4.0
+    assert result[0, 2] == 1.5
+
+
+def test_sample_with_frequency_penalty():
+    logits = torch.tensor([[5.0, 1.0, 2.0, 3.0]])
+    input_ids = torch.tensor([[0, 2, 0]])
+    input_mask = torch.tensor([[True, True, True]])
+    tokens = sample(
+        logits,
+        temperature=1.0,
+        top_k=0,
+        top_p=1.0,
+        frequency_penalty=0.5,
+        input_ids=input_ids,
+        input_mask=input_mask,
+    )
+    assert tokens.shape == (1,)
+    assert 0 <= tokens[0] < logits.size(-1)

tests/trainer/test_early_stopping.py

@@ -46,7 +46,7 @@ def test_early_stopping_simulation(base_test_env, early_stopping_dataset):
     # Resume from latest checkpoint
     load_dir = os.path.join(base_test_env["test_dir"], "epoch_0_step_1")
     trainer = Trainer(train_config)
-    trainer.train(resume_dir=load_dir)
+    trainer.train(param_path=load_dir, resume=True)
 
     # Verify checkpoint was saved at expected step
     load_dir = os.path.join(base_test_env["test_dir"], "epoch_1_step_5")

tests/trainer/test_grpo_strategy.py

@@ -0,0 +1,224 @@
+import pytest
+import torch
+
+from astrai.config.model_config import AutoRegressiveLMConfig
+from astrai.model.transformer import AutoRegressiveLM
+from astrai.trainer.strategy import GRPOStrategy
+
+
+class _FakeExecutor:
+    """Minimal executor stub providing ``unwrap_model`` for ref model creation."""
+
+    def unwrap_model(self, model):
+        return model.state_dict()
+
+
+def _make_config(vocab_size=200, max_len=64):
+    return AutoRegressiveLMConfig(
+        vocab_size=vocab_size,
+        dim=16,
+        n_heads=2,
+        n_kv_heads=1,
+        dim_ffn=32,
+        max_len=max_len,
+        n_layers=2,
+        norm_eps=1e-5,
+    )
+
+
+def _make_model(device):
+    config = _make_config()
+    model = AutoRegressiveLM(config).to(device=device)
+    return model, config
+
+
+def _make_batch(
+    batch_size=2, group_size=4, prompt_len=8, response_len=12, device="cpu"
+):
+    """Construct a GRPO batch with deterministic shapes.
+
+    Returns dict with prompts [B, P], responses [B, G, R], masks [B, G, R],
+    rewards [B, G].
+    """
+    prompts = torch.randint(0, 200, (batch_size, prompt_len), device=device)
+    responses = torch.randint(
+        0, 200, (batch_size, group_size, response_len), device=device
+    )
+    # All response tokens valid.
+    masks = torch.ones(batch_size, group_size, response_len, device=device)
+    # Distinct rewards per group member so std > 0.
+    rewards = torch.randn(batch_size, group_size, device=device)
+    return {
+        "prompts": prompts,
+        "responses": responses,
+        "masks": masks,
+        "rewards": rewards,
+    }
+
+
+def _make_frozen_copy(model, device):
+    """Create a frozen copy of ``model`` with independent weights loaded."""
+    config = _make_config()
+    copy = AutoRegressiveLM(config).to(device=device)
+    copy.load_state_dict(model.state_dict())
+    copy.requires_grad_(False)
+    copy.eval()
+    return copy
+
+
+@pytest.fixture
+def grpo_strategy():
+    """Build a GRPOStrategy with a small real model and fake executor."""
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model, config = _make_model(device)
+    old_model = _make_frozen_copy(model, device)
+    ref_model = _make_frozen_copy(model, device)
+
+    strategy = GRPOStrategy(
+        model=model,
+        device=device,
+        old_model=old_model,
+        ref_model=ref_model,
+        clip_eps=0.2,
+        kl_coef=0.01,
+        group_size=4,
+        model_fn=lambda c=config: AutoRegressiveLM(c).to(device=device),
+        executor=_FakeExecutor(),
+    )
+    return strategy, device
+
+
+def test_grpo_loss_is_finite(grpo_strategy):
+    """compute_loss returns a finite scalar."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(device=device)
+    loss = strategy.compute_loss(batch)
+    assert loss.dim() == 0
+    assert torch.isfinite(loss).item()
+
+
+def test_grpo_loss_backward(grpo_strategy):
+    """Loss is differentiable w.r.t. policy model parameters."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(device=device)
+    loss = strategy.compute_loss(batch)
+    loss.backward()
+    # At least some parameter should receive a gradient.
+    has_grad = any(
+        p.grad is not None and p.grad.abs().sum().item() > 0
+        for p in strategy.model.parameters()
+    )
+    assert has_grad
+
+
+def test_grpo_ref_model_not_updated(grpo_strategy):
+    """Backward should not populate gradients on ref_model."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(device=device)
+    loss = strategy.compute_loss(batch)
+    loss.backward()
+    for p in strategy.ref_model.parameters():
+        assert p.grad is None
+
+
+def test_grpo_old_model_not_updated(grpo_strategy):
+    """Backward should not populate gradients on old_model."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(device=device)
+    loss = strategy.compute_loss(batch)
+    loss.backward()
+    for p in strategy.old_model.parameters():
+        assert p.grad is None
+
+
+def test_grpo_prompt_tokens_masked(grpo_strategy):
+    """When only prompt-equivalent tokens are unmasked (response mask all 0),
+    the policy loss should be zero (no valid tokens contribute)."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(device=device)
+    # Zero out all response masks → no response token contributes.
+    batch["masks"] = torch.zeros_like(batch["masks"])
+    loss = strategy.compute_loss(batch)
+    # With no valid tokens, policy_loss term is 0 and KL term is 0.
+    assert loss.item() == pytest.approx(0.0, abs=1e-6)
+
+
+def test_grpo_identical_rewards_zero_advantage(grpo_strategy):
+    """When all group rewards are identical, advantage is 0 → policy_loss is 0.
+    Only the KL term remains (which is 0 when policy == ref at init)."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(device=device)
+    batch["rewards"] = torch.ones(batch["rewards"].shape, device=device)
+    loss = strategy.compute_loss(batch)
+    # At init policy == old == ref, so ratio == 1, KL == 0; advantage == 0.
+    assert loss.item() == pytest.approx(0.0, abs=1e-5)
+
+
+def test_grpo_sync_old_model(grpo_strategy):
+    """sync_old_model copies current policy weights into old_model."""
+    strategy, device = grpo_strategy
+    # Perturb policy model so it differs from old.
+    with torch.no_grad():
+        for p in strategy.model.parameters():
+            p.add_(0.05)
+    # old_model should still hold original weights (differ from policy).
+    policy_sd = strategy.model.state_dict()
+    old_sd = strategy.old_model.state_dict()
+    differs_before = any(
+        not torch.allclose(policy_sd[k], old_sd[k]) for k in policy_sd if k in old_sd
+    )
+    assert differs_before
+
+    strategy.sync_old_model()
+
+    old_sd_after = strategy.old_model.state_dict()
+    matches = all(
+        torch.allclose(policy_sd[k], old_sd_after[k])
+        for k in policy_sd
+        if k in old_sd_after
+    )
+    assert matches
+
+
+def test_grpo_partial_mask(grpo_strategy):
+    """Only the first half of response tokens are valid."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(device=device)
+    B, G, R = batch["masks"].shape
+    half = R // 2
+    batch["masks"][:, :, half:] = 0.0
+    loss = strategy.compute_loss(batch)
+    assert torch.isfinite(loss).item()
+
+
+def test_grpo_clipping_effect(grpo_strategy):
+    """After diverging policy from ref, ratio should be clipped to [1-eps, 1+eps]
+    on the surrogate. Verify loss is finite and non-zero for distinct rewards."""
+    strategy, device = grpo_strategy
+    # Diverge policy from ref.
+    with torch.no_grad():
+        for p in strategy.model.parameters():
+            p.add_(0.3)
+    batch = _make_batch(device=device)
+    loss = strategy.compute_loss(batch)
+    assert torch.isfinite(loss).item()
+    # With distinct rewards and diverged policy, loss should be non-trivial.
+    assert loss.abs().item() > 1e-4
+
+
+def test_grpo_no_reduction_param():
+    """GRPOStrategy.__init__ must not accept ``reduction`` (removed)."""
+    import inspect
+
+    sig = inspect.signature(GRPOStrategy.__init__)
+    assert "reduction" not in sig.parameters
+
+
+def test_grpo_shapes_3d_batch(grpo_strategy):
+    """Verify compute_loss handles non-square prompt/response lengths."""
+    strategy, device = grpo_strategy
+    batch = _make_batch(
+        batch_size=3, group_size=4, prompt_len=10, response_len=8, device=device
+    )
+    loss = strategy.compute_loss(batch)
+    assert torch.isfinite(loss).item()