chore: bump version to 1.3.11

- Multi-pass MMA: encode pass in grid blockIdx.x, compute q_head0/G in-kernel
- Fixes crash for G>32 (previously block(32,G) exceeded 1024 threads)
- Fixes alloc_split_partials using uninitialized num_splits (MAX_SPLITS=32)
- __launch_bounds__ on all MMA and prefill kernels for better register allocation
- 4x vectorized combine kernel (4 head_dim per thread)
- uint4 vectorized K loads in scalar decode kernels
- cp.async .L2::128B cache hint for K/V tile streaming
- Extract warp_reduce_sum, bf16, MAX_SPLITS to attn_warp_utils.cuh

.github/workflows/release.yml

@@ -0,0 +1,92 @@
+name: Release
+
+on:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  build-pure:
+    name: Build pure-Python wheel
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Build wheel (no CUDA)
+        run: |
+          pip wheel . --no-deps -w dist/
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: pure-wheel
+          path: dist/*.whl
+          if-no-files-found: error
+
+  build-cuda-linux:
+    name: Build CUDA wheel (Linux)
+    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
+          if-no-files-found: error
+
+  release:
+    name: Attach wheels to release
+    needs: [build-pure, build-cuda-linux]
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - name: Download pure-Python wheel
+        uses: actions/download-artifact@v4
+        with:
+          name: pure-wheel
+          path: release-assets/pure
+
+      - name: Download CUDA wheel
+        uses: actions/download-artifact@v4
+        with:
+          name: cuda-wheel-linux
+          path: release-assets/cuda
+
+      - name: Verify release assets
+        shell: bash
+        run: |
+          set -euo pipefail
+          pure_wheels=(release-assets/pure/*.whl)
+          cuda_wheels=(release-assets/cuda/*.whl)
+          test "${#pure_wheels[@]}" -eq 1
+          test "${#cuda_wheels[@]}" -eq 1
+          test "$(basename "${pure_wheels[0]}")" != "$(basename "${cuda_wheels[0]}")"
+
+      - name: Create release & upload assets
+        uses: softprops/action-gh-release@v2
+        with:
+          files: |
+            release-assets/pure/*.whl
+            release-assets/cuda/*.whl
+          tag_name: ${{ github.ref_name }}
+          generate_release_notes: true

.gitignore

@@ -11,6 +11,7 @@
 !csrc/**/*.py
 
 !csrc/**/*.cu
+!csrc/**/*.h
 !csrc/**/*.cuh
 
 !scripts/**/*.sh

README.md

@@ -20,7 +20,7 @@
   <a href="assets/docs/README-zh-CN.md">中文</a> •
   <a href="https://github.com/ViperEkura/AstrAI/issues">Issue Tracker</a> •
   <a href="https://github.com/ViperEkura/AstrAI/discussions">Discussions</a> •
-  <a href="https://huggingface.co/ViperEk/">HuggingFace</a>
+  <a href="https://huggingface.co/ViperEkura">HuggingFace</a>
 </div>
 
 <br>
@@ -241,7 +241,7 @@ For major changes, please open an issue first to discuss what you would like to
 
 - **GitHub Issues**: [Issue Tracker](https://github.com/ViperEkura/AstrAI/issues)
 - **Discussions**: [GitHub Discussions](https://github.com/ViperEkura/AstrAI/discussions)
-- **HuggingFace**: [Model Hub](https://huggingface.co/ViperEk)
+- **HuggingFace**: [Model Hub](https://huggingface.co/ViperEkura)
 
 ### License
 

assets/docs/README-zh-CN.md

@@ -27,7 +27,7 @@
   <a href="#chinese">中文</a> •
   <a href="https://github.com/ViperEkura/AstrAI/issues">问题追踪</a> •
   <a href="https://github.com/ViperEkura/AstrAI/discussions">讨论区</a> •
-  <a href="https://huggingface.co/ViperEk">HuggingFace</a>
+  <a href="https://huggingface.co/ViperEkura">HuggingFace</a>
 </div>
 <br>
 
@@ -247,7 +247,7 @@ SSE 流式格式、错误码和统计端点详见[推理文档](./inference.md)
 
 - **GitHub Issues**: [问题追踪](https://github.com/ViperEkura/AstrAI/issues)
 - **Discussions**: [GitHub 讨论区](https://github.com/ViperEkura/AstrAI/discussions)
-- **HuggingFace**: [模型中心](https://huggingface.co/ViperEk)
+- **HuggingFace**: [模型中心](https://huggingface.co/ViperEkura)
 
 ### 许可证
 

assets/docs/architecture.md

@@ -28,17 +28,17 @@ classDiagram
 
         class AutoRegressiveLMConfig {
             +Optional[int] vocab_size
-            +Optional[int] dim
+            +Optional[int] hidden_size
-            +Optional[int] n_layers
+            +Optional[int] num_hidden_layers
-            +Optional[float] norm_eps
+            +Optional[float] rms_norm_eps
-            +Optional[int] dim_ffn
+            +Optional[int] intermediate_size
-            +Optional[bool] tie_weight
+            +Optional[bool] tie_word_embeddings
             +Optional[dict] rope_scaling
-            +Optional[int] max_len
+            +Optional[int] max_position_embeddings
             +Optional[float] rope_theta
             +str attn_type
-            +Optional[int] n_heads
+            +Optional[int] num_attention_heads
-            +Optional[int] n_kv_heads
+            +Optional[int] num_key_value_heads
             +Optional[bool] use_qk_norm
             +Optional[bool] use_gated_attention
             +Optional[int] kv_lora_rank
@@ -53,15 +53,15 @@ classDiagram
 
         class EncoderConfig {
             +Optional[int] vocab_size
-            +Optional[int] dim
+            +Optional[int] hidden_size
-            +Optional[int] n_layers
+            +Optional[int] num_hidden_layers
-            +Optional[float] norm_eps
+            +Optional[float] rms_norm_eps
-            +Optional[int] dim_ffn
+            +Optional[int] intermediate_size
-            +Optional[int] max_len
+            +Optional[int] max_position_embeddings
             +Optional[float] rope_theta
             +str attn_type
-            +Optional[int] n_heads
+            +Optional[int] num_attention_heads
-            +Optional[int] n_kv_heads
+            +Optional[int] num_key_value_heads
             +Optional[bool] use_qk_norm
             +str ffn_type
             +Optional[dict] rope_scaling
@@ -117,7 +117,7 @@ classDiagram
             +int n_epoch
             +int batch_per_device
             +int grad_accum_steps
-            +float max_grad_norm
+            +Optional[float] max_grad_norm
             +list gradient_checkpointing_modules
             +int start_epoch
             +int start_samples
@@ -141,6 +141,12 @@ classDiagram
             +int val_step
             +float neftune_alpha
             +str parallel_mode
+            +int rollout_interval
+            +float rollout_temperature
+            +int rollout_top_k
+            +float rollout_top_p
+            +int rollout_max_tokens
+            +Optional[Callable] reward_model_fn
             +dict executor_kwargs
             +dict extra_kwargs
             +validate()
@@ -177,13 +183,26 @@ classDiagram
         class Store {
             +Dict[str, List[Tensor]] _data
             +Dict[str, List[int]] _cum
+            +Dict[str, List[int]] _offsets
             +int _length
+            +int _num_records
             +keys (property)
             +load(path)
-            +fetch(begin, end, keys)
             +__len__()
-            -_fetch_key(key, begin, end) Tensor
+            -_normalize(raw, offsets)
-            -_normalize(raw)
+        }
+
+        class Streamable {
+            <<mixin>>
+            +fetch(begin, end, keys)
+            -_fetch_stream_key(key, begin, end) Tensor
+        }
+
+        class Recordable {
+            <<mixin>>
+            +num_records (property)
+            +fetch_record(index, keys)
+            -_fetch_record_key(key, index) Tensor
         }
 
         class H5Store {
@@ -195,7 +214,19 @@ classDiagram
             +load(path)
         }
 
-        class ResumableDistributedSampler {
+        class JsonlStore {
+            +JsonlSource _source
+            +Callable _processor
+            +load(path, transform, processor)
+            +fetch_record(index, keys)
+        }
+
+        class JsonlSource {
+            +Path path
+            +load() List[dict]
+        }
+
+        class RDSampler {
             +int epoch
             +int iter
         }
@@ -210,7 +241,7 @@ classDiagram
             +Dict _entries
             +register(name) decorator
             +create(train_type, window_size, stride) BaseDataset
-            +load(train_type, load_path, window_size, stride, storage_type) BaseDataset
+            +load(train_type, load_path, window_size, stride, storage_type, tokenizer_path, max_len, store) BaseDataset
         }
     }
 
@@ -358,19 +389,103 @@ classDiagram
             +forward(x) Tensor
             +set_neftune_alpha(alpha)
         }
+
+        class LoRAConfig {
+            +int r
+            +int alpha
+            +tuple target_modules
+        }
+
+        class LoRALinear {
+            +Linear weight
+            +Parameter lora_A, lora_B
+            +forward(x) Tensor
+            +merge()
+        }
     }
 
     namespace preprocessing {
+        class SectionRenderer {
+            +process_sections(item, sections, config, tokenizer) Tuple
+            +process_list_field(item, sections, config, tokenizer) Tuple
+        }
+
         class BaseMaskBuilder {
             <<abstract>>
             +build(item, config, tokenizer) Optional[dict]
         }
 
-        class SectionedMaskBuilder {
+        class SingleOutputMaskBuilder {
             +SectionRenderer renderer
             +build(item, config, tokenizer) Optional[dict]
-            +_build_single(item, config, tokenizer) Optional[dict]
+        }
-            +_build_multi(item, sources_spec, config, tokenizer) Optional[dict]
+
+        class MultiOutputMaskBuilder {
+            +SectionRenderer renderer
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class SectionedMaskBuilder {
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class PackingStrategy {
+            <<abstract>>
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class PackingStrategyFactory {
+            +create(name, *args, **kwargs) PackingStrategy
+        }
+
+        class SimplePacking {
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class BFDPacking {
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class BFDSplitPacking {
+            +apply(keys, max_packed_len, truncation_mode) Dict
+        }
+
+        class PositionIdStrategy {
+            <<abstract>>
+            +generate(sequences) List[int]
+        }
+
+        class PositionIdStrategyFactory {
+            +create(name, *args, **kwargs) PositionIdStrategy
+        }
+
+        class NoPositionId {
+            +generate(sequences) List[int]
+        }
+
+        class DocResetPositionId {
+            +generate(sequences) List[int]
+        }
+
+        class ContinuousPositionId {
+            +generate(sequences) List[int]
+        }
+
+        class StoreWriter {
+            <<abstract>>
+            +save(output_dir, domain, shard_idx, tensors)
+        }
+
+        class StoreWriterFactory {
+            +create(name, *args, **kwargs) StoreWriter
+        }
+
+        class BinWriter {
+            +save(output_dir, domain, shard_idx, tensors)
+        }
+
+        class H5Writer {
+            +save(output_dir, domain, shard_idx, tensors)
         }
 
         class Pipeline {
@@ -378,6 +493,7 @@ classDiagram
             +List[str] paths
             +str output_dir
             +str tokenizer_path
+            +AutoTokenizer tokenizer
             +BaseMaskBuilder mask_builder
             +PackingStrategy _packer
             +PositionIdStrategy _position_id
@@ -385,6 +501,18 @@ classDiagram
             +transform(item) Optional[dict]
             +run()
             +_flush(domains, shard_idx)
+            +_inject_doc_reset_position_ids(keys, mode, seqs) Dict
+            +_inject_continuous_position_ids(tensors, mode, seqs) Dict
+            +_to_tensors(keys) Dict
+        }
+
+        class TokenizeTransform {
+            +PipelineConfig config
+            +AutoTokenizer tokenizer
+            +BaseMaskBuilder mask_builder
+            +PositionIdStrategy position_strategy
+            +from_config_file(path) TokenizeTransform
+            +apply(records) Dict[str, list]
         }
     }
 
@@ -457,7 +585,7 @@ classDiagram
 
         class TrainContextBuilder {
             +TrainConfig config
-            +with_resume_dir(resume_dir) TrainContextBuilder
+            +with_param_path(param_path, resume) TrainContextBuilder
             +build() TrainContext
         }
 
@@ -495,14 +623,39 @@ classDiagram
         }
 
         class GRPOStrategy {
+            +nn.Module old_model
             +nn.Module ref_model
             +float clip_eps
             +float kl_coef
             +int group_size
-            +str reduction
-            +int sync_interval
             +compute_loss(batch) Tensor
-            +sync_ref_model()
+            +sync_old_model()
+        }
+
+        class RawRollout {
+            +Tensor prompts
+            +Tensor responses
+            +Tensor response_mask
+            +Tensor logprobs_old
+        }
+
+        class RolloutResult {
+            +Tensor rewards
+        }
+
+        class BaseRewardModel {
+            <<abstract>>
+            +score(prompts, responses) Tensor
+        }
+
+        class RolloutGenerator {
+            +generate(batch) RawRollout
+        }
+
+        class RolloutRunner {
+            +step()
+            +clear_cache()
+            +__call__(batch) Tuple[RolloutResult, bool]
         }
 
         class BaseScheduler {
@@ -552,7 +705,7 @@ classDiagram
         }
 
         class GradientClippingCallback {
-            +float max_grad_norm
+            +Optional[float] max_grad_norm
             +on_optimizer_step(context)
         }
 
@@ -818,12 +971,21 @@ classDiagram
             +apply(logits, filter_value) Tensor
         }
 
+        class FrequencyPenaltyStrategy {
+            +float penalty
+            +apply(logits, filter_value, input_ids, input_mask) Tensor
+        }
+
         class SamplingPipeline {
             +List[BaseSamplingStrategy] strategies
             +apply(logits, filter_value) Tensor
             +sample(logits, filter_value) Tensor
         }
 
+        class StreamDecoder {
+            +push(token_id) str
+        }
+
         class GenerateResult {
             +List[Tuple[int, str]] tokens
             +List[str] results
@@ -842,6 +1004,17 @@ classDiagram
             +Optional[str] tool_call_id
         }
 
+        class FunctionDef {
+            +str name
+            +Optional[str] description
+            +Optional[Dict] parameters
+        }
+
+        class ToolDef {
+            +str type
+            +FunctionDef function
+        }
+
         class ChatCompletionRequest {
             +str model
             +List[ChatMessage] messages
@@ -930,9 +1103,20 @@ classDiagram
             +str yielded
         }
 
-        class get_app {
+        class BaseToolParser {
-            <<module>>
+            <<abstract>>
-            +get_app() FastAPI
+            +feed(body, current_token_ids, delta_token_ids) List[Dict]
+            +parse_complete(body) Optional[Dict]
+            +has_tool_calls (property) bool
+        }
+
+        class ToolParserFactory {
+            +create(name, *args, **kwargs) BaseToolParser
+        }
+
+        class SimpleJsonToolParser {
+            +feed(body, current_token_ids, delta_token_ids) List[Dict]
+            +parse_complete(body) Optional[Dict]
         }
     }
 
@@ -955,14 +1139,17 @@ classDiagram
     }
 
     namespace parallel {
-        class setup {
+        class LaunchStrategy {
-            <<module>>
+            <<abstract>>
-            +spawn_parallel_fn(func, world_size, backend, master_addr, master_port, device_type, start_method, **kwargs)
+            +launch(func, **kwargs)
-            +setup_parallel(rank, world_size, backend, master_addr, master_port, device_type) contextmanager
+        }
-            +get_current_device() str
+
-            +get_world_size() int
+        class TorchrunStrategy {
-            +get_rank() int
+            +launch(func, **kwargs)
-            +only_on_rank(rank, sync=False) decorator
+        }
+
+        class LocalStrategy {
+            +launch(func, **kwargs)
         }
 
         class GradientState {
@@ -991,7 +1178,7 @@ classDiagram
 
         class BaseExecutor {
             +GradientState gradient_state
-            +prepare(model, optimizer, dataloader, scheduler) tuple
+            +prepare(model_fn, optimizer_fn, scheduler_fn, before_wrap) tuple
             +accumulate(model) context manager
             +backward(loss)
             +unwrap_model(model) dict
@@ -1013,6 +1200,12 @@ classDiagram
             +unwrap_model(model) dict
         }
 
+        class FSDP2Executor {
+            -_prepare_model(model) nn.Module
+            -_no_sync(model) context manager
+            +unwrap_model(model) dict
+        }
+
         class ExecutorFactory {
             +Dict _entries
             +register(name) decorator
@@ -1070,9 +1263,16 @@ classDiagram
     Store <|-- H5Store
     Store <|-- MmapStore
     Store <|-- JsonlStore
+    H5Store --|> Streamable
+    H5Store --|> Recordable
+    MmapStore --|> Streamable
+    MmapStore --|> Recordable
+    JsonlStore --|> Streamable
+    JsonlStore --|> Recordable
     BaseSamplingStrategy <|-- TemperatureStrategy
     BaseSamplingStrategy <|-- TopKStrategy
     BaseSamplingStrategy <|-- TopPStrategy
+    BaseSamplingStrategy <|-- FrequencyPenaltyStrategy
     ParallelModel <|-- RowParallelLinear
     ParallelModel <|-- ColumnParallelLinear
     AutoModel <|-- AutoRegressiveLM
@@ -1096,12 +1296,31 @@ classDiagram
     BaseFactory <|-- ExecutorFactory
     BaseFactory <|-- ConfigFactory
     BaseFactory <|-- MaskBuilderFactory
+    BaseFactory <|-- PackingStrategyFactory
+    BaseFactory <|-- PositionIdStrategyFactory
+    BaseFactory <|-- StoreWriterFactory
+    BaseFactory <|-- ToolParserFactory
     BaseExecutor <|-- NoneExecutor
     BaseExecutor <|-- DDPExecutor
     BaseExecutor <|-- FSDPExecutor
+    BaseExecutor <|-- FSDP2Executor
     ResponseBuilder <|-- OpenAIResponseBuilder
     ResponseBuilder <|-- AnthropicResponseBuilder
+    BaseToolParser <|-- SimpleJsonToolParser
     BaseMaskBuilder <|-- SectionedMaskBuilder
+    BaseMaskBuilder <|-- SingleOutputMaskBuilder
+    BaseMaskBuilder <|-- MultiOutputMaskBuilder
+    PackingStrategy <|-- SimplePacking
+    PackingStrategy <|-- BFDPacking
+    BFDPacking <|-- BFDSplitPacking
+    PositionIdStrategy <|-- NoPositionId
+    PositionIdStrategy <|-- DocResetPositionId
+    PositionIdStrategy <|-- ContinuousPositionId
+    StoreWriter <|-- BinWriter
+    StoreWriter <|-- H5Writer
+    RawRollout <|-- RolloutResult
+    LaunchStrategy <|-- TorchrunStrategy
+    LaunchStrategy <|-- LocalStrategy
     KVCache <|-- PageCache
     KVCache <|-- ContiguousCache
     CacheView <|-- PageCacheView
@@ -1123,6 +1342,8 @@ classDiagram
     EmbeddingEncoder *-- Embedding
     DecoderBlock *-- RMSNorm
     ChatCompletionRequest *-- ChatMessage
+    ChatCompletionRequest *-- ToolDef
+    ToolDef *-- FunctionDef
     MessagesRequest *-- AnthropicMessage
     BaseExecutor *-- GradientState
     AccumOptimizer o-- GradientState
@@ -1144,11 +1365,20 @@ classDiagram
     BaseDataset o-- Store
     Pipeline o-- PipelineConfig
     Pipeline o-- BaseMaskBuilder
+    Pipeline o-- AutoTokenizer
+    Pipeline o-- PackingStrategy
+    Pipeline o-- PositionIdStrategy
+    Pipeline o-- StoreWriter
+    TokenizeTransform o-- AutoTokenizer
+    TokenizeTransform o-- BaseMaskBuilder
 
     %% --- Dependency (uses temporarily) ---
     TrainConfig ..> BaseStrategy : selects
     PipelineConfig ..> MaskBuilderFactory : selects
     MaskBuilderFactory ..> BaseMaskBuilder : creates
+    PackingStrategyFactory ..> PackingStrategy : creates
+    PositionIdStrategyFactory ..> PositionIdStrategy : creates
+    StoreWriterFactory ..> StoreWriter : creates
     StrategyFactory ..> BaseStrategy : creates
     SchedulerFactory ..> BaseScheduler : creates
     DatasetFactory ..> BaseDataset : creates
@@ -1167,12 +1397,13 @@ classDiagram
     ExecutorFactory ..> NoneExecutor : creates
     ExecutorFactory ..> DDPExecutor : creates
     ExecutorFactory ..> FSDPExecutor : creates
+    ExecutorFactory ..> FSDP2Executor : creates
+    ToolParserFactory ..> BaseToolParser : creates
     TrainContextBuilder ..> ExecutorFactory : creates
     Trainer ..> TrainContextBuilder : uses
     TrainContextBuilder ..> TrainContext : creates
-    Trainer ..> Functions : spawns
     TrainContextBuilder ..> StrategyFactory : uses
-    TrainContextBuilder ..> ResumableDistributedSampler : creates
+    TrainContextBuilder ..> RDSampler : creates
     Checkpoint ..> Checkpoint : serializes
     CheckpointCallback ..> Checkpoint : creates
     PageCache ..> PageCacheView : binds
@@ -1183,11 +1414,14 @@ classDiagram
     AnthropicResponseBuilder ..> MessagesRequest : receives
     ProtocolHandler ..> StopChecker : creates
     ProtocolHandler ..> GenContext : creates
+    RolloutGenerator ..> InferenceScheduler : uses
+    RolloutRunner ..> RolloutGenerator : uses
+    RolloutRunner ..> BaseRewardModel : uses
 
     %% --- Association (general usage) ---
     Trainer --> TrainConfig
     DPOStrategy --> AutoModel
-    GRPOStrategy --> AutoModel
+    GRPOStrategy --> AutoModel : policy/old/ref
     InferenceScheduler --> Task
     InferenceScheduler --> TaskStatus
     Task --> TaskStatus
@@ -1204,14 +1438,14 @@ classDiagram
 | Module | Components | Description |
 |--------|------------|-------------|
 | **astrai.config** | BaseConfig, BaseModelConfig, AutoRegressiveLMConfig, EncoderConfig, ConfigFactory, TrainConfig, PipelineConfig, InputConfig, ProcessingConfig, OutputConfig | Configuration management (to_dict/from_dict, to_file/from_file) |
-| **astrai.preprocessing** | BaseMaskBuilder, MaskBuilderFactory, SectionedMaskBuilder, Pipeline, filter_by_length, PackingStrategy, PackingStrategyFactory, PositionIdStrategy, PositionIdStrategyFactory, StoreWriter, StoreWriterFactory | Declarative JSON-driven data preprocessing |
+| **astrai.preprocessing** | SectionRenderer, BaseMaskBuilder, MaskBuilderFactory, SectionedMaskBuilder, SingleOutputMaskBuilder, MultiOutputMaskBuilder, Pipeline, TokenizeTransform, PackingStrategy, PackingStrategyFactory, SimplePacking, BFDPacking, BFDSplitPacking, PositionIdStrategy, PositionIdStrategyFactory, NoPositionId, DocResetPositionId, ContinuousPositionId, StoreWriter, StoreWriterFactory, BinWriter, H5Writer | Declarative JSON-driven data preprocessing |
-| **astrai.dataset** | BaseDataset–GRPODataset, Store–JsonlStore/MmapStore/H5Store, StoreFactory, ResumableDistributedSampler, DatasetFactory | Dataset loading and management |
+| **astrai.dataset** | BaseDataset, SEQDataset, SFTDataset, DPODataset, GRPODataset, Store, Streamable, Recordable, H5Store, MmapStore, JsonlSource, JsonlStore, StoreFactory, RDSampler, DatasetFactory | Dataset loading and management |
 | **astrai.serialization** | Checkpoint | Model serialization |
-| **astrai.model** | AutoModel, AutoRegressiveLM, EmbeddingEncoder, DecoderBlock, GQA, MLA, MLP, DeepSeekMoE, AttnFactory, FFNFactory, RMSNorm, Linear, RotaryEmbedding, Embedding | Neural network model |
+| **astrai.model** | AutoModel, AutoRegressiveLM, EmbeddingEncoder, DecoderBlock, GQA, MLA, MLP, DeepSeekMoE, AttnFactory, FFNFactory, RMSNorm, Linear, LoRAConfig, LoRALinear, RotaryEmbedding, Embedding | Neural network model |
 | **astrai.tokenize** | AutoTokenizer, ChatTemplate | Tokenizer and chat template |
-| **astrai.trainer** | Trainer, TrainContext, TrainContextBuilder, BaseStrategy–GRPOStrategy, StrategyFactory, BaseScheduler–WSDScheduler, SchedulerFactory, TrainCallback(Protocol)–MetricCallback, CallbackFactory | Training workflow |
+| **astrai.trainer** | Trainer, TrainContext, TrainContextBuilder, BaseStrategy–GRPOStrategy, StrategyFactory, BaseScheduler–WSDScheduler, SchedulerFactory, TrainCallback(Protocol)–MetricCallback, CallbackFactory, RawRollout, RolloutResult, BaseRewardModel, RolloutGenerator, RolloutRunner | Training workflow |
-| **astrai.inference** | InferenceEngine, InferenceScheduler, Executor, 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.inference** | InferenceEngine, InferenceScheduler, Executor, KVCache–ContiguousCache/PageCache, CacheView–ContiguousCacheView/PageCacheView, Allocator–Storage, Task, TaskManager, TaskStatus, StreamDecoder, GenerationRequest, GenerateResult, BaseSamplingStrategy–SamplingPipeline, FrequencyPenaltyStrategy, ProtocolHandler, ResponseBuilder, OpenAIResponseBuilder, AnthropicResponseBuilder, StopChecker, GenContext, StopInfo, ChatMessage, FunctionDef, ToolDef, ChatCompletionRequest, AnthropicMessage, MessagesRequest, BaseToolParser, ToolParserFactory, SimpleJsonToolParser | Inference service |
-| **astrai.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.parallel** | spawn_parallel_fn, setup_parallel, get_rank/get_world_size/get_current_device, only_on_rank, LaunchStrategy, TorchrunStrategy, LocalStrategy, BaseExecutor, ExecutorFactory, NoneExecutor, DDPExecutor, FSDPExecutor, FSDP2Executor, 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 |
 
@@ -1219,7 +1453,7 @@ 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`, `ToolParserFactory` | Decorator-based component creation |
 | **Registry** | `BaseFactory` | Component registration |
 | **Strategy** | `SEQStrategy`, `SFTStrategy`, `DPOStrategy`, `GRPOStrategy` | Training strategy switching |
 | **Strategy (Sampling)** | `TemperatureStrategy`, `TopKStrategy`, `TopPStrategy`, `SamplingPipeline` | Composable logit transformations |
@@ -1228,7 +1462,7 @@ classDiagram
 | **Observer** | `TrainCallback`, callback implementations | Training process monitoring |
 | **Context** | `TrainContext` | Unified training state bag |
 | **Object Pool** | `Allocator`, `PagePool` | Page-based KV cache with LRU eviction |
-| **Executor** | `BaseExecutor`, `NoneExecutor`, `DDPExecutor`, `FSDPExecutor` | Gradient accumulation & model distribution |
+| **Executor** | `BaseExecutor`, `NoneExecutor`, `DDPExecutor`, `FSDPExecutor`, `FSDP2Executor` | Gradient accumulation & model distribution |
 | **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 |
@@ -1238,7 +1472,7 @@ classDiagram
 1. **Config → Training**: `TrainConfig` holds `model_fn`, `dataset`, `optimizer_fn`, `scheduler_fn`, `parallel_mode`, `executor_kwargs`
 2. **Training Flow**: `Trainer` → `TrainContextBuilder` → `TrainContext`, uses `BaseStrategy` for loss, `BaseExecutor` for gradient accumulation + model distribution
 3. **Strategy Selection**: `StrategyFactory` creates strategy by `train_type`
-4. **Executor Selection**: `ExecutorFactory.create(cfg.parallel_mode, grad_accum_steps=cfg.grad_accum_steps, **cfg.executor_kwargs)` → `NoneExecutor` / `DDPExecutor` / `FSDPExecutor`
+4. **Executor Selection**: `ExecutorFactory.create(cfg.parallel_mode, grad_accum_steps=cfg.grad_accum_steps, **cfg.executor_kwargs)` → `NoneExecutor` / `DDPExecutor` / `FSDPExecutor` / `FSDP2Executor`
 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/JsonlStore) loads data with explicit `_length` and multi-segment `_data`
@@ -1247,4 +1481,4 @@ classDiagram
 10. **AutoModel**: `from_pretrained()` loads `config.json` + `model.safetensors`, `_disable_random_init` replaces `nn.init.*` with no-ops
 11. **Protocols**: `OptimizerProtocol` / `SchedulerProtocol` — structural subtyping for `AccumOptimizer` / `AccumScheduler` wrappers
 
-> Document Update Time: 2026-07-09
+> Document Update Time: 2026-07-20

assets/docs/dataflow.md

@@ -61,41 +61,59 @@ StoreFactory.create("bin")   → MmapStore
 StoreFactory.create("jsonl") → JsonlStore
 ```
 
-**H5Store**: Reads HDF5 files. Tensors are loaded into host memory and normalized into segmented storage.
+All three inherit `Store` (base, owns `_data`/`_cum`/`_offsets`/`_normalize`) plus the `Streamable` and `Recordable` mixins, so every backend supports both `fetch(begin, end, keys)` (stream) and `fetch_record(index, keys)` (record) APIs.
 
-**MmapStore**: Memory-maps `.bin` files. OS page cache sharing is native — no explicit `share_memory_()` needed. Uses `torch.from_numpy(np.memmap(...))`.
+**H5Store**: Reads HDF5 files. Tensors are loaded into host memory and normalized into segmented storage. `segments_are_records=True` — each `data_i` dataset is one record.
 
-**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.
+**MmapStore**: Memory-maps `.bin` files. OS page cache sharing is native — no explicit `share_memory_()` needed. Uses `torch.from_numpy(np.memmap(...))`. `segments_are_records=False` — bin segments are contiguous streams; record access is driven by `_offsets` (written when `save_bin(..., record_keys=...)` was used at preprocessing time).
 
-All 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. Two modes: eager (default, applies `TokenizeTransform` to all records at load) and lazy (`processor=fn` given, defers tokenisation to `fetch_record` — used by DPO/GRPO).
+
+All backends normalise tensors into `Store._data[Dict[str, List[Tensor]]]` + `Store._cum[Dict[str, List[int]]]` (cumulative lengths for bisect-based stream indexing) + `Store._offsets[Dict[str, List[int]]]` (per-record offsets for record-mode indexing). Nested keys (GRPO `responses`/`masks` as `List[List[Tensor]]`) are stored as-is and excluded from both bookkeepings — they are only accessed record-by-record.
 
 ## Data Keys by Training Type
 
-| Type | Storage Keys |
+| Type | Storage Keys | Access Mode |
-|------|-------------|
+|------|-------------|-------------|
-| `seq` | `sequence` (→ input_ids, target_ids via offset-by-1) |
+| `seq` | `sequence` (→ input_ids, target_ids via offset-by-1) | stream (`fetch`) |
-| `sft` | `sequence`, `loss_mask`, `position_ids` |
+| `sft` | `sequence`, `loss_mask`, `position_ids` | stream (`fetch`) |
-| `dpo` | `chosen`, `rejected`, `chosen_mask`, `rejected_mask` |
+| `dpo` | `chosen`, `rejected`, `chosen_mask`, `rejected_mask` | record (`fetch_record`) |
-| `grpo` | `prompts`, `responses`, `masks`, `rewards` |
+| `grpo` | `prompts`, `responses`, `masks`, `rewards` | record (`fetch_record`) |
 
 ## Dataset Architecture
 
 ```
-DatasetFactory.load(train_type, load_path, window_size, stride=None, storage_type=None)
+DatasetFactory.load(train_type, load_path, window_size, stride=None,
+                    storage_type=None, tokenizer_path=None,
+                    max_position_embeddings=2048, store=None)
   → BaseDataset.load(load_path, storage_type=None)
     → detect_format(load_path)
     → StoreFactory.create(storage_type)
     → Store.load(load_path)
       → _normalize(raw)  # base Store, shared by both backends
-        → Store._data[Dict[str, List[Tensor]]] + _cum[Dict[str, List[int]]]
+        → Store._data[Dict[str, List[Tensor]]]
-          → BaseDataset.__getitem__(idx)
+          + _cum[Dict[str, List[int]]]   (stream mode)
-            → get_index(idx) → [begin, end)
+          + _offsets[Dict[str, List[int]]]  (record mode)
-            → Store.fetch(begin, end, keys) → Tensor / Dict[str, Tensor]
+
+Stream datasets (SEQ/SFT):
+  BaseDataset.__getitem__(idx)
+    → get_index(idx) → [begin, end)
+    → Store.fetch(begin, end, keys) → Tensor / Dict[str, Tensor]
+
+Record datasets (DPO/GRPO via RecordDataset):
+  RecordDataset.__getitem__(idx)
+    → Store.fetch_record(idx, keys) → Tensor / Dict[str, Tensor]
 ```
 
-`window_size` = max input length, `stride` = step between consecutive samples (defaults to `window_size`, optional). `storage_type` defaults to `None` (auto-detect via `detect_format`).
+Class hierarchy: `BaseDataset` ← `SEQDataset` / `SFTDataset` (stream); `BaseDataset` ← `RecordDataset` ← `DPODataset` / `GRPODataset` (record).
 
-`Store.fetch(begin, end, keys)` accepts a single key (`str`) returning a `Tensor`, or a list of keys returning `Dict[str, Tensor]`. Internally uses `bisect` across multi-segment tensors. Raises `RuntimeError("Store not loaded")` if called before `load()`.
+`window_size` = max input length, `stride` = step between consecutive samples (defaults to `window_size`, optional). Only meaningful for stream datasets — record datasets ignore both. `storage_type` defaults to `None` (auto-detect via `detect_format`).
+
+`tokenizer_path` triggers lazy on-the-fly tokenisation for record datasets on raw JSONL (DPO builds a `dpo_processor`; SEQ/SFT/pre-tokenised backends ignore it). `store` (pre-built `Store`) bypasses `load_path`/`storage_type`/`tokenizer_path` entirely — the caller controls Store construction.
+
+`Store.fetch(begin, end, keys)` (stream mode, on `Streamable`): accepts a single key (`str`) returning a `Tensor`, or a list of keys returning `Dict[str, Tensor]`. Internally uses `bisect` across multi-segment tensors. Raises `RuntimeError("Store not loaded")` if called before `load()`.
+
+`Store.fetch_record(index, keys)` (record mode, on `Recordable`): same key API. Uses `_offsets[key]` when present (bin layout with per-record offsets), otherwise indexes `_data[key]` directly (H5/JSONL where each segment is one record).
 
 ## Sampler
 
@@ -109,4 +127,4 @@ 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: 2026-07-09
+> Document Update Time: 2026-07-19

assets/docs/inference.md

@@ -32,7 +32,7 @@ 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.
+Created by default when no cache is passed to `InferenceScheduler`. Each task occupies a fixed slot of `[max_seq_len, num_key_value_heads, head_dim]`. Simple and efficient for small-to-medium batch sizes.
 
 ### PageCache (paged with prefix sharing)
 
@@ -42,7 +42,7 @@ PageCache (paged KV cache with prefix sharing, alternative)
   │     ├── Allocator         bitmask-based page allocator + ref-count + LRU
   │     └── PrefixCache       hash-based prefix matching (page_hash via polynomial hash)
   ├── TaskTable              maps task_id → page_table + cached token count
-  ├── Storage                k_cache / v_cache tensors (n_layers × n_pages × page_size × n_kv_heads × head_dim)
+  ├── Storage                k_cache / v_cache tensors (num_hidden_layers × n_pages × page_size × num_key_value_heads × head_dim)
   └── PageCacheView          bundles Storage + page_table + total_len for attention layers
 ```
 

assets/docs/params.md

@@ -13,7 +13,7 @@
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--train_type` | Training type (`seq`, `sft`, `dpo`, `grpo`) | required |
+| `--train_type` | Training type (`seq`, `sft`, `dpo`, `grpo`, `online_grpo`, `online_dpo`) | required |
 | `--data_root_path` | Dataset root directory | required |
 | `--param_path` | Model parameters or checkpoint path | required |
 | `--n_epoch` | Total training epochs | 1 |
@@ -26,7 +26,7 @@
 |-----------|-------------|---------|
 | `--warmup_ratio` | Fraction of total steps used for LR warmup | 0.05 |
 | `--max_lr` | Maximum learning rate (cosine decay after warmup) | 3e-4 |
-| `--max_grad_norm` | Maximum gradient norm for clipping | 1.0 |
+| `--max_grad_norm` | Maximum gradient norm for clipping (None disables) | 1.0 |
 
 ### Optimizer (MuonMix)
 
@@ -44,7 +44,7 @@ Combined optimizer: matrix parameters via **Muon**, non-matrix via **AdamW** (`f
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--window_size` | Max input sequence length | model config `max_len` |
+| `--window_size` | Max input sequence length | model config `max_position_embeddings` |
 | `--stride` | Stride for sliding window over sequences | None |
 | `--random_seed` | Random seed for reproducibility | 3407 |
 | `--num_workers` | DataLoader worker processes | 4 |
@@ -100,18 +100,31 @@ Combined optimizer: matrix parameters via **Muon**, non-matrix via **AdamW** (`f
 | `--group_size` | GRPO group size | 4 | `grpo` |
 | `--grpo_clip_eps` | GRPO clipping epsilon | 0.2 | `grpo` |
 | `--grpo_kl_coef` | GRPO KL penalty coefficient | 0.01 | `grpo` |
-| `--grpo_sync_interval` | GRPO ref_model sync interval (steps) | 200 | `grpo` |
 | `--neftune_alpha` | NEFTune noise alpha (0=disabled, typical: 5.0) | 0.0 | `sft` |
 
+### Online Rollout
+
+These options apply to `online_grpo` and `online_dpo`. Online strategies require
+a `BaseRewardModel` factory in `TrainConfig`; `train.py` does not currently
+provide a command-line option for configuring one.
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `--rollout_interval` | Optimizer steps between rollout refreshes | 512 |
+| `--rollout_temperature` | Rollout sampling temperature | 0.7 |
+| `--rollout_top_k` | Rollout top-k filtering (`0` disables) | 0 |
+| `--rollout_top_p` | Rollout nucleus sampling threshold | 0.9 |
+| `--rollout_max_tokens` | Maximum generated tokens per response | 1024 |
+
 ### Scheduler
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
 | `--schedule_type` | LR scheduler type (`cosine`, `sgdr`, `wsd`) | cosine |
-| `--min_rate` | Minimum LR as fraction of base LR | None (scheduler default: 0.01) |
+| `--min_rate` | Minimum LR as fraction of base LR | None (scheduler default: 0.05 for cosine/SGDR, 0.0 for WSD) |
 | `--cycle_length` | SGDR first cycle length in steps | None (total_steps - warmup_steps) |
 | `--t_mult` | SGDR cycle length multiplier per restart | 2 |
-| `--stable_steps` | WSD stable plateau steps | None (required for wsd) |
+| `--stable_steps` | WSD stable plateau steps | None (80% of post-warmup steps) |
 | `--decay_steps` | WSD decay steps | None (total_steps - warmup_steps - stable_steps) |
 
 ### Usage Example
@@ -173,7 +186,7 @@ See [Inference Guide](inference.md) for HTTP API documentation.
 | `--top_k` | int | `30` | Top-k filtering |
 | `--top_p` | float | `0.95` | Nucleus sampling threshold |
 | `--batch_size` | int | `1` | Batch size for generation |
-| `--max_tokens` | int | model config `max_len` | Maximum tokens to generate |
+| `--max_tokens` | int | model config `max_position_embeddings` | Maximum tokens to generate |
 
 Usage:
 ```bash
@@ -201,4 +214,4 @@ See [Preprocessing Guide](preprocessing.md) for config file format and examples.
 
 ---
 
-> Document Update Time: 2026-07-09
+> Document Update Time: 2026-07-20

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
 

assets/docs/training.md

@@ -6,7 +6,7 @@
 - [Causal Mask](#causal-mask)
 - [Rotary Position Embedding (RoPE)](#rotary-position-embedding-rope)
 - [Training Loop](#training-loop)
-- [Strategies](#strategies) — SEQ, SFT, DPO, GRPO
+- [Strategies](#strategies) — SEQ, SFT, DPO, GRPO, online rollout
 - [LR Schedulers](#lr-schedulers)
 - [Gradient Checkpointing](#gradient-checkpointing)
 - [Checkpoint](#checkpoint)
@@ -86,7 +86,7 @@ on_train_end
 | `on_error` | On exception during training | `CheckpointCallback`, `MetricCallback` |
 | `on_train_end` | Training ends (always via finally) | `CheckpointCallback`, `MetricCallback`, `GradientCheckpointingCallback` |
 
-Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `metric` (JSONL + validation, rank-0), `progress_bar` (tqdm), `gradient_clipping`.
+Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `metric` (JSONL + validation, rank-0), `progress_bar` (tqdm), `gradient_clipping` (always registered; computes grad norm, clips only when `max_grad_norm` is not `None`).
 
 ## Strategies
 
@@ -118,24 +118,47 @@ $$
 L_{\text{DPO}} = -\mathbb{E}\left[\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\text{ref}}(y_w\mid x)} - \beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\text{ref}}(y_l\mid x)}\right)\right]
 $$
 
-Parameters: `beta=0.1`, `reduction="mean"`. Keys: `chosen`, `rejected`, `chosen_mask`, `rejected_mask`.
+Parameters: `beta=0.1`, `reduction="sum"`. Keys: `chosen`, `rejected`, `chosen_mask`, `rejected_mask`.
 
 ### 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 $\rho_t = \pi_\theta(a_t|s_t) / \pi_{\text{old}}(a_t|s_t)$ is the
+per-token importance sampling ratio against the behaviour policy
+(`old_model`, synced externally between data-generation rounds) and the
+expectations are over valid response tokens. The KL term regularises
+$\pi_\theta$ towards a frozen reference model (`ref_model`, typically
+the SFT checkpoint).
+
+Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`. External sync of `old_model` weights via `sync_old_model()` between data-generation rounds.
 
 Keys: `prompts`, `responses`, `masks`, `rewards`.
 
+### Online Rollout
+
+`online_grpo` and `online_dpo` use the respective GRPO and DPO strategies with
+a `RolloutRunner`. The runner renders prompts through the tokenizer chat
+template, generates grouped responses through `InferenceScheduler`, then scores
+them with a `BaseRewardModel`. It refreshes cached rollouts every
+`rollout_interval` optimizer steps. `online_grpo` synchronizes `old_model` when
+a fresh rollout is produced.
+
+Online strategies require `TrainConfig.reward_model_fn`. `train.py` exposes the
+rollout sampling parameters but does not yet offer a CLI argument for the reward
+model factory.
+
 ## LR Schedulers
 
 | Type | Class | Description |
@@ -152,6 +175,7 @@ Trades compute for memory by recomputing activations during backward pass. Speci
 
 ```python
 from astrai.model.components.decoder_block import DecoderBlock
+
 config = TrainConfig(..., gradient_checkpointing_modules=[DecoderBlock])
 ```
 
@@ -171,18 +195,14 @@ Model config (`context.model_config`) saved into `config.json` during training v
 ## TrainContextBuilder (Builder Pattern)
 
 ```python
-context = (
+context = TrainContextBuilder(config).with_param_path(param_path, resume=True).build()
-    TrainContextBuilder(config)
-        .with_resume_dir(resume_dir)
-        .build()
-)
 # Returns TrainContext with model, strategy, optimizer, scheduler, dataloader, checkpoint
 ```
 
-- Loads checkpoint weights if provided
+- Loads checkpoint weights before the model is wrapped
 - Creates executor via `ExecutorFactory.create(cfg.parallel_mode, grad_accum_steps=cfg.grad_accum_steps, **cfg.executor_kwargs)`
-- Calls `executor.prepare(model, optimizer, dataloader, scheduler)` for model distribution (e.g. DDP) + gradient accumulation wrappers
+- Calls `executor.prepare(model_fn, optimizer_fn, scheduler_fn, before_wrap=...)`; the executor creates, wraps, then builds the optimizer and scheduler for the wrapped model
-- Creates `ResumableDistributedSampler` for shuffle+resume
+- Creates `RDSampler` for shuffle+resume
 - Builds strategy via `StrategyFactory.create(train_type, model, device, **kwargs)`
 
 ## Training CLI
@@ -212,4 +232,4 @@ nohup python scripts/tools/train.py \
 
 Full parameter reference at [params.md](params.md).
 
-> Document Update Time: 2026-07-09
+> Document Update Time: 2026-07-20

astrai/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "1.3.8"
+__version__ = "1.3.11"
 __author__ = "ViperEkura"
 
 from astrai.config import (
@@ -12,7 +12,7 @@ from astrai.config import (
 from astrai.dataset import (
     BaseDataset,
     DatasetFactory,
-    ResumableDistributedSampler,
+    RDSampler,
     Store,
     StoreFactory,
 )
@@ -77,7 +77,7 @@ __all__ = [
     "Pipeline",
     "PipelineConfig",
     "ProtocolHandler",
-    "ResumableDistributedSampler",
+    "RDSampler",
     "SamplingPipeline",
     "SchedulerFactory",
     "Store",

astrai/config/model_config.py

@@ -29,19 +29,19 @@ class AutoRegressiveLMConfig(BaseModelConfig):
     """Configuration for autoregressive language model."""
 
     vocab_size: Optional[int] = None
-    dim: Optional[int] = None
+    hidden_size: Optional[int] = None
-    n_layers: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
+    rms_norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
+    intermediate_size: Optional[int] = None
-    tie_weight: Optional[bool] = None
+    tie_word_embeddings: Optional[bool] = None
 
-    max_len: Optional[int] = None
+    max_position_embeddings: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
 
     attn_type: str = "gqa"
-    n_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
-    n_kv_heads: Optional[int] = None
+    num_key_value_heads: Optional[int] = None
     use_qk_norm: Optional[bool] = None
     use_gated_attention: Optional[bool] = None
 
@@ -62,18 +62,18 @@ class EncoderConfig(BaseModelConfig):
     """Configuration for embedding encoder model."""
 
     vocab_size: Optional[int] = None
-    dim: Optional[int] = None
+    hidden_size: Optional[int] = None
-    n_layers: Optional[int] = None
+    num_hidden_layers: Optional[int] = None
-    norm_eps: Optional[float] = None
+    rms_norm_eps: Optional[float] = None
-    dim_ffn: Optional[int] = None
+    intermediate_size: Optional[int] = None
 
-    max_len: Optional[int] = None
+    max_position_embeddings: Optional[int] = None
     rope_theta: Optional[float] = None
     rope_scaling: Optional[dict] = None
 
     attn_type: str = "gqa"
-    n_heads: Optional[int] = None
+    num_attention_heads: Optional[int] = None
-    n_kv_heads: Optional[int] = None
+    num_key_value_heads: Optional[int] = None
     use_qk_norm: Optional[bool] = None
     use_gated_attention: Optional[bool] = None
 

astrai/config/preprocess_config.py

@@ -45,6 +45,8 @@ class ProcessingConfig(BaseConfig):
         Maximum number of characters to keep (default: 2_000_000).
     max_items : Optional[int]
         Maximum number of items to process (default: None, unlimited).
+    batch_size : int
+        Number of records tokenized together (default: 256).
     packing_strategy : str
         How to pack sequences into a contiguous stream.
 
@@ -65,6 +67,7 @@ class ProcessingConfig(BaseConfig):
     min_chars: int = 50
     max_chars: int = 2_000_000
     max_items: Optional[int] = None
+    batch_size: int = 256
     packing_strategy: str = "simple"
     max_packed_len: int = 8192
     truncation_mode: str = "keep_start"

astrai/config/train_config.py

@@ -37,8 +37,9 @@ class TrainConfig(BaseConfig):
     grad_accum_steps: int = field(
         default=1, metadata={"help": "Number of iterations between steps."}
     )
-    max_grad_norm: float = field(
+    max_grad_norm: Optional[float] = field(
-        default=1.0, metadata={"help": "Maximum gradient norm."}
+        default=1.0,
+        metadata={"help": "Maximum gradient norm. None disables clipping."},
     )
     gradient_checkpointing_modules: List[str] = field(
         default_factory=list,
@@ -87,6 +88,10 @@ class TrainConfig(BaseConfig):
     pin_memory: bool = field(
         default=False, metadata={"help": "Pin memory for dataloader."}
     )
+    collate_fn: Optional[Callable[[List[Any]], Any]] = field(
+        default=None,
+        metadata={"help": "Collate function for dataloader (e.g. dpo_collate_fn)."},
+    )
 
     # distributed training
     nprocs: int = field(
@@ -133,6 +138,32 @@ class TrainConfig(BaseConfig):
         metadata={"help": "NEFTune noise alpha (0=disabled, typical: 5.0)."},
     )
 
+    # online rollout
+    rollout_interval: int = field(
+        default=512,
+        metadata={"help": "Number of optimizer steps between online rollouts."},
+    )
+    rollout_temperature: float = field(
+        default=0.7, metadata={"help": "Sampling temperature for online rollout."}
+    )
+    rollout_top_k: int = field(
+        default=0, metadata={"help": "Top-k filtering for online rollout (0=disable)."}
+    )
+    rollout_top_p: float = field(
+        default=0.9,
+        metadata={"help": "Top-p (nucleus) filtering for online rollout."},
+    )
+    rollout_max_tokens: int = field(
+        default=1024,
+        metadata={"help": "Maximum generated tokens per response in rollout."},
+    )
+    reward_model_fn: Optional[Callable] = field(
+        default=None,
+        metadata={
+            "help": "Factory for reward model (required for online RL strategies)."
+        },
+    )
+
     executor_kwargs: Dict[str, Any] = field(
         default_factory=dict,
         metadata={"help": "Extra kwargs passed to ExecutorFactory.create()."},

astrai/dataset/__init__.py

@@ -1,14 +1,18 @@
 from astrai.dataset.dataset import (
     BaseDataset,
     DatasetFactory,
+    dpo_collate_fn,
+    grpo_collate_fn,
 )
-from astrai.dataset.sampler import ResumableDistributedSampler
+from astrai.dataset.sampler import RDSampler
 from astrai.dataset.storage import (
     H5Store,
     JsonlStore,
     MmapStore,
+    Recordable,
     Store,
     StoreFactory,
+    Streamable,
     detect_format,
 )
 from astrai.serialization import (
@@ -21,7 +25,11 @@ from astrai.serialization import (
 __all__ = [
     "BaseDataset",
     "DatasetFactory",
+    "dpo_collate_fn",
+    "grpo_collate_fn",
     "Store",
+    "Streamable",
+    "Recordable",
     "StoreFactory",
     "H5Store",
     "MmapStore",
@@ -31,5 +39,5 @@ __all__ = [
     "load_h5",
     "save_bin",
     "load_bin",
-    "ResumableDistributedSampler",
+    "RDSampler",
 ]

astrai/dataset/dataset.py

@@ -1,7 +1,31 @@
-"""Dataset implementations with factory pattern for training."""
+"""Dataset implementations for training.
+
+Composition over inheritance — every dataset is a thin wrapper that
+binds a :class:`Store` to a particular train-type's key mapping.  All
+sample-id → token/record indexing lives on the Store; datasets never
+know about window/stride math or segment layouts.
+
+Class hierarchy:
+
+    BaseDataset (ABC)         — holds a Store, exposes __len__/keys,
+                                 overrides __getitem__
+    ├── SEQDataset            — next-token prediction (stream)
+    ├── SFTDataset            — loss-mask + position_ids (stream)
+    ├── DPODataset            — chosen/rejected pairs (record)
+    └── GRPODataset           — prompt + response group (record)
+
+``DatasetFactory.load(train_type, load_path, window_size, stride, …)``
+builds the Store (auto-detecting format) before constructing the
+matching dataset.  Passing ``store=`` skips Store construction.
+
+When a record dataset (DPO) reads from raw JSONL, a *processor*
+function (pure ``record -> Dict[str, Tensor]``) is forwarded to
+:class:`JsonlStore` so tokenisation happens on the fly.
+"""
 
 from abc import ABC, abstractmethod
-from typing import Dict, List, Optional
+from functools import partial
+from typing import Callable, Dict, List, Optional
 
 import torch
 from torch import Tensor
@@ -13,202 +37,401 @@ from astrai.dataset.storage import (
     detect_format,
 )
 from astrai.factory import BaseFactory
+from astrai.tokenize import AutoTokenizer
+
+
+def dpo_tokenize(
+    record: dict,
+    tokenizer,
+    max_len: int = 2048,
+) -> Optional[dict]:
+    """Tokenize one DPO record into chosen/rejected + masks.
+
+    Applies the tokenizer's chat template so token sequences match the
+    SFT checkpoint's format.  Prompt is rendered with
+    ``add_generation_prompt=True``; chosen/rejected are appended as a
+    single assistant turn.
+
+    Accepts:
+
+    - Flat:   ``{"prompt": str, "chosen": str, "rejected": str}``
+    - Conv:   ``{"prompt": [{role, content}, ...], "chosen": [...], ...}``
+    - Legacy: ``{"input": str, "chosen": str, "rejected": str}``
+
+    No packing, no ``position_ids`` — DPO sequences are independent.
+    """
+    prompt = record.get("prompt") or record.get("input")
+    chosen = record.get("chosen")
+    rejected = record.get("rejected")
+    if prompt is None or chosen is None or rejected is None:
+        return None
+
+    prompt_messages = _to_messages(prompt)
+    chosen_text = _extract_text(chosen)
+    rejected_text = _extract_text(rejected)
+    if chosen_text is None or rejected_text is None:
+        return None
+    chosen_messages = prompt_messages + [{"role": "assistant", "content": chosen_text}]
+    rejected_messages = prompt_messages + [
+        {"role": "assistant", "content": rejected_text}
+    ]
+
+    prompt_ids = tokenizer.apply_chat_template(
+        prompt_messages, tokenize=True, add_generation_prompt=True
+    )
+    ch_ids = tokenizer.apply_chat_template(
+        chosen_messages, tokenize=True, add_generation_prompt=False
+    )
+    re_ids = tokenizer.apply_chat_template(
+        rejected_messages, tokenize=True, add_generation_prompt=False
+    )
+
+    full_ch = ch_ids[:max_len]
+    full_re = re_ids[:max_len]
+
+    prompt_len = min(len(prompt_ids), max_len)
+    ch_mask = [0] * prompt_len + [1] * max(0, len(full_ch) - prompt_len)
+    ch_mask = ch_mask[:max_len]
+    re_mask = [0] * prompt_len + [1] * max(0, len(full_re) - prompt_len)
+    re_mask = re_mask[:max_len]
+
+    return {
+        "chosen": full_ch,
+        "rejected": full_re,
+        "chosen_mask": ch_mask,
+        "rejected_mask": re_mask,
+    }
+
+
+def _to_messages(value) -> list:
+    """Accept str or conversation list; return message list."""
+    if isinstance(value, str):
+        return [{"role": "user", "content": value}]
+    if isinstance(value, list):
+        return value
+    return [{"role": "user", "content": str(value)}]
+
+
+def _extract_text(value) -> Optional[str]:
+    """Accept str or conversation list; return plain text."""
+    if value is None:
+        return None
+    if isinstance(value, str):
+        return value
+    if isinstance(value, list):
+        return "".join(m.get("content", "") for m in value if isinstance(m, dict))
+    return None
+
+
+def dpo_processor(
+    record: dict,
+    tokenizer,
+    max_len: int = 2048,
+) -> Dict[str, Tensor]:
+    """DPO processor: wraps :func:`dpo_tokenize` and returns tensors."""
+    result = dpo_tokenize(record, tokenizer, max_len=max_len)
+    if result is None:
+        raise ValueError(f"Malformed DPO record: {list(record.keys())}")
+    return {
+        "chosen": torch.tensor(result["chosen"], dtype=torch.int32),
+        "rejected": torch.tensor(result["rejected"], dtype=torch.int32),
+        "chosen_mask": torch.tensor(result["chosen_mask"], dtype=torch.bool),
+        "rejected_mask": torch.tensor(result["rejected_mask"], dtype=torch.bool),
+    }
+
+
+def dpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
+    """Collate variable-length DPO samples into padded 2-D tensors.
+
+    Input: list of dicts, each with:
+      - chosen:        [C_i]
+      - rejected:      [R_i]
+      - chosen_mask:   [C_i]
+      - rejected_mask: [R_i]
+
+    Output (padded to the max length across chosen/rejected within the batch):
+      - chosen:        [B, S_max]
+      - rejected:      [B, S_max]
+      - chosen_mask:   [B, S_max]
+      - rejected_mask: [B, S_max]
+    """
+    B = len(batch)
+    S_max = max(b["chosen"].size(0) for b in batch)
+    S_max = max(S_max, max(b["rejected"].size(0) for b in batch))
+
+    chosen = torch.zeros(B, S_max, dtype=torch.long)
+    rejected = torch.zeros(B, S_max, dtype=torch.long)
+    chosen_mask = torch.zeros(B, S_max, dtype=torch.bool)
+    rejected_mask = torch.zeros(B, S_max, dtype=torch.bool)
+
+    for i, b in enumerate(batch):
+        c_len = b["chosen"].size(0)
+        r_len = b["rejected"].size(0)
+        chosen[i, :c_len] = b["chosen"]
+        rejected[i, :r_len] = b["rejected"]
+        chosen_mask[i, :c_len] = b["chosen_mask"]
+        rejected_mask[i, :r_len] = b["rejected_mask"]
+
+    return {
+        "chosen": chosen,
+        "rejected": rejected,
+        "chosen_mask": chosen_mask,
+        "rejected_mask": rejected_mask,
+    }
+
+
+def grpo_collate_fn(batch: List[Dict[str, Tensor]]) -> Dict[str, Tensor]:
+    """Collate variable-length GRPO samples into padded 3-D tensors.
+
+    Input: list of dicts, each with:
+      - prompts:  [P_i]
+      - responses: list of G tensors, each [R_ij]
+      - masks:     list of G tensors, each [R_ij]
+      - rewards:  [G]
+
+    Output:
+      - prompts:   [B, P_max], left-padded
+      - prompt_mask: [B, P_max]
+      - responses: [B, G, R_max]
+      - masks:     [B, G, R_max]
+      - rewards:   [B, G]
+    """
+    B = len(batch)
+    G = len(batch[0]["responses"])
+    P_max = max(b["prompts"].size(0) for b in batch)
+    R_max = max(r.size(0) for b in batch for r in b["responses"])
+
+    prompts = torch.zeros(B, P_max, dtype=torch.long)
+    prompt_mask = torch.zeros(B, P_max, dtype=torch.bool)
+    responses = torch.zeros(B, G, R_max, dtype=torch.long)
+    masks = torch.zeros(B, G, R_max, dtype=torch.bool)
+    rewards = torch.zeros(B, G, dtype=torch.float32)
+
+    for i, b in enumerate(batch):
+        p_len = b["prompts"].size(0)
+        prompts[i, -p_len:] = b["prompts"]
+        prompt_mask[i, -p_len:] = True
+        rewards[i, : b["rewards"].size(0)] = b["rewards"]
+        for g in range(min(G, len(b["responses"]))):
+            r_len = b["responses"][g].size(0)
+            responses[i, g, :r_len] = b["responses"][g]
+            if g < len(b["masks"]):
+                masks[i, g, :r_len] = b["masks"][g]
+
+    return {
+        "prompts": prompts,
+        "prompt_mask": prompt_mask,
+        "responses": responses,
+        "masks": masks,
+        "rewards": rewards,
+    }
+
+
+def validate_keys(store: Store, required: List[str]) -> None:
+    """Raise ``KeyError`` if *store* is missing any *required* key."""
+    if not required:
+        return
+    actual = set(store.keys)
+    missing = [k for k in required if k not in actual]
+    if missing:
+        raise KeyError(
+            f"Store at {getattr(store, '_load_path', '?')} is missing required "
+            f"keys {missing}; available keys are {sorted(actual)}."
+        )
 
 
 class BaseDataset(Dataset, ABC):
-    """Abstract base class for all dataset types.
+    """Abstract base class for dataset types.
 
-    Implements common functionality for window-based data fetching.
+    Holds a :class:`Store`.  All sample-id indexing is delegated to the
-    Uses a storage abstraction for format-agnostic data loading.
+    store — this class exposes ``__len__`` as ``len(store)`` and the
+    ``keys`` property as ``store.keys``.  Subclasses implement
+    ``__getitem__`` with the train-type-specific key mapping and any
+    training-only index arithmetic (e.g. the next-token ``+1`` shift).
     """
 
-    def __init__(self, window_size: int, stride: int):
+    required_keys: List[str] = []
+
+    def __init__(self, store: Store):
         super().__init__()
-        self.window_size = window_size
+        self.store: Store = store
-        self.stride = stride
+        validate_keys(store, self.required_keys)
-        self.storage: Optional[Store] = None
 
-    @property
+    def __len__(self) -> int:
-    def required_keys(self) -> List[str]:
+        return len(self.store)
-        """Return required storage keys for this dataset type.
-
-        Subclasses should override to specify expected keys.
-        """
-        return []
-
-    def _validate_keys(self):
-        if not self.required_keys:
-            return
-        actual_keys = set(self.storage.keys)
-        missing = [k for k in self.required_keys if k not in actual_keys]
-        if missing:
-            raise KeyError(
-                f"Dataset {type(self).__name__} requires keys {self.required_keys}, "
-                f"but storage at {self._load_path} only has {sorted(actual_keys)}. "
-                f"Missing: {missing}"
-            )
-
-    def load(self, load_path: str, storage_type: Optional[str] = None, **kwargs):
-        """Load dataset from the given path.
-
-        Auto-detects the storage format if not specified.
-
-        Args:
-            load_path: Path to the data directory or file
-            storage_type: Force a specific storage type ("h5", "bin", "jsonl"),
-                          or None for auto-detection
-            **kwargs: Extra arguments forwarded to the store constructor and
-                      to ``store.load()``.
-
-        Raises:
-            KeyError: If the loaded storage is missing required keys.
-        """
-        if storage_type is None:
-            storage_type = detect_format(load_path)
-        self.storage = StoreFactory.create(storage_type, **kwargs)
-        self._load_path = load_path
-        self.storage.load(load_path, **kwargs)
-        self._validate_keys()
-
-    @property
-    def count(self) -> int:
-        """Return the total number of raw elements (tokens) in the dataset."""
-        if self.storage is None:
-            return 0
-        return len(self.storage)
 
     @property
     def keys(self) -> List[str]:
-        """Return the available data keys."""
+        return self.store.keys
-        if self.storage is None:
-            return []
-        return self.storage.keys
 
-    def get_index(self, index: int) -> tuple:
+    @property
-        """Calculate begin and end indices for a sample.
+    def token_count(self) -> int:
-
+        return self.store.token_count
-        Args:
-            index: Sample index
-
-        Returns:
-            Tuple of (begin_idx, end_idx)
-        """
-        if self.storage is None:
-            raise RuntimeError("Dataset not loaded, call load() first")
-        total = len(self.storage)
-        if total <= self.window_size:
-            raise ValueError(
-                f"Data too short: {total} tokens <= window_size {self.window_size}"
-            )
-
-        begin_idx = min(index * self.stride, total - 1 - self.window_size)
-        end_idx = min(begin_idx + self.window_size, total - 1)
-
-        return begin_idx, end_idx
 
     @abstractmethod
     def __getitem__(self, index: int) -> Dict[str, Tensor]:
-        """Get a single sample by index.
-
-        Must be implemented by subclasses.
-        """
         raise NotImplementedError
 
-    def __len__(self) -> int:
-        if self.storage is None:
-            return 0
-        total = len(self.storage)
-        if total <= self.window_size:
-            return 0
-        return (total - 1 - self.window_size) // self.stride + 1
-
 
 class DatasetFactory(BaseFactory["BaseDataset"]):
-    """Factory class for creating dataset instances.
+    """Factory for creating dataset instances by train-type.
 
-    Supports decorator-based registration for extensible dataset types.
+    Use :meth:`DatasetFactory.register("custom")` to register new
-    All default dataset types (seq, sft, dpo, grpo) are registered automatically
+    dataset classes; they must inherit from :class:`BaseDataset`.
-    when their classes are defined with the decorator.
-
-    Example usage:
-        @DatasetFactory.register("custom")
-        class CustomDataset(BaseDataset):
-            ...
-
-        dataset = DatasetFactory.create("custom", window_size, stride)
     """
 
     @classmethod
     def load(
         cls,
         train_type: str,
-        load_path: str,
+        load_path: Optional[str] = None,
-        window_size: int,
+        window_size: int = 0,
         stride: Optional[int] = None,
         storage_type: Optional[str] = None,
+        tokenizer_path: Optional[str] = None,
+        max_len: int = 2048,
+        store: Optional[Store] = None,
         **kwargs,
     ) -> "BaseDataset":
         """Create and load a dataset in one step.
 
+        Two entry points:
+
+        - **store given**: bind it directly — the caller fully controls
+          Store construction and processor setup.  *load_path*,
+          *storage_type*, *tokenizer_path*, *window_size*, *stride* are
+          ignored.
+        - **store is None**: build a Store from *load_path*, auto-detecting
+          format and constructing a processor when *tokenizer_path* is
+          given for a record dataset on JSONL.
+
         Args:
-            train_type: Type of training dataset
+            train_type: Registered dataset name ("seq", "sft", "dpo",
-            load_path: Path to the data file
+                "grpo", …).
-            window_size: Window size for data sampling
+            load_path: Path to the data file or directory (ignored if
-            stride: Stride between consecutive samples (default: same as window_size)
+                *store* is given).
-            storage_type: Storage type ("h5", "bin", "jsonl") or None for auto-detection
+            window_size: Stream window length — only meaningful for
-            **kwargs: Extra arguments forwarded to ``dataset.load()``.
+                stream datasets (SEQ/SFT).  Record datasets ignore it.
+            stride: Stride between consecutive stream samples
+                (default: same as *window_size*).
+            storage_type: Storage backend ("h5", "bin", "jsonl") or
+                None for auto-detection.
+            tokenizer_path: Path to tokenizer for lazy JSONL
+                tokenisation (record datasets only).
+            max_len: Max sequence length forwarded to processors.
+            store: Pre-built, already-loaded Store instance.
+            **kwargs: Extra arguments forwarded to ``store.load()``.
 
         Returns:
-            Loaded dataset instance
+            Loaded dataset instance.
         """
+        if store is not None:
+            return cls.create(train_type, store=store)
+
+        if load_path is None:
+            raise ValueError("Either load_path or store must be provided")
+
+        if storage_type is None:
+            storage_type = detect_format(load_path)
+
         if stride is None:
             stride = window_size
 
-        dataset = cls.create(train_type, window_size, stride)
+        processor = cls._maybe_build_processor(
-        dataset.load(load_path, storage_type=storage_type, **kwargs)
+            train_type, storage_type, tokenizer_path, max_len
+        )
 
-        return dataset
+        store_window = cls._store_window_for(train_type, window_size)
+        store = StoreFactory.create(
+            storage_type,
+            window_size=store_window,
+            stride=stride if stride else store_window,
+        )
+        if processor is not None:
+            store.load(load_path, processor=processor, **kwargs)
+        else:
+            load_kwargs = dict(kwargs)
+            if (
+                tokenizer_path is not None
+                and storage_type == "jsonl"
+                and train_type in ("seq", "sft")
+                and "tokenizer_path" not in load_kwargs
+            ):
+                load_kwargs["tokenizer_path"] = tokenizer_path
+            store.load(load_path, **load_kwargs)
+
+        return cls.create(train_type, store=store)
+
+    @staticmethod
+    def _store_window_for(train_type: str, window_size: int) -> int:
+        """Stream datasets consume ``window_size``; record datasets ignore it.
+
+        Record datasets (dpo/grpo) treat each record as an independent
+        training unit and never window, so the store is built with
+        ``window_size=0`` and ``len(store)`` returns the record count.
+        """
+        if train_type in ("seq", "sft"):
+            return window_size
+        return 0
+
+    @staticmethod
+    def _maybe_build_processor(
+        train_type: str,
+        storage_type: str,
+        tokenizer_path: Optional[str],
+        max_len: int,
+    ) -> Optional[Callable[[dict], Dict[str, Tensor]]]:
+        """Build an on-the-fly tokenisation processor if applicable.
+
+        Only raw JSONL + record datasets (DPO/GRPO) need a processor;
+        pre-tokenised backends (H5/bin) and stream datasets (SEQ/SFT)
+        return ``None`` so no tokenizer is loaded.
+        """
+        if tokenizer_path is None or storage_type != "jsonl":
+            return None
+        if train_type == "dpo":
+            tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+            return partial(dpo_processor, tokenizer=tokenizer, max_len=max_len)
+        return None
 
 
 @DatasetFactory.register("seq")
 class SEQDataset(BaseDataset):
-    """Dataset for sequential next-token prediction training."""
+    """Dataset for sequential next-token prediction training.
 
-    @property
+    Stream mode: ``store.fetch(begin, end, "sequence")`` returns the
-    def required_keys(self) -> List[str]:
+    input window; the +1 shifted call returns the next-token target.
-        return ["sequence"]
+    """
 
-    def _fetch_data(self, begin_idx: int, end_idx: int) -> Tensor:
+    required_keys = ["sequence"]
-        return self.storage.fetch(begin_idx, end_idx, "sequence")
 
-    def __getitem__(self, index):
+    def __getitem__(self, index: int):
-        begin_idx, end_idx = self.get_index(index)
+        begin, end = self.store.sample_window(index)
-
+        x = self.store.fetch(begin, end, "sequence")
-        x = self._fetch_data(begin_idx, end_idx).to(dtype=torch.long)
+        y = self.store.fetch(begin + 1, end + 1, "sequence")
-        y = self._fetch_data(begin_idx + 1, end_idx + 1).to(dtype=torch.long)
+        return {
-
+            "input_ids": x.to(dtype=torch.long),
-        return {"input_ids": x, "target_ids": y}
+            "target_ids": y.to(dtype=torch.long),
+        }
 
 
 @DatasetFactory.register("sft")
 class SFTDataset(BaseDataset):
-    """Dataset for supervised fine-tuning with loss masking."""
+    """Dataset for supervised fine-tuning with loss masking.
 
-    @property
+    Stream mode: ``sequence``/``loss_mask``/``position_ids`` are sliced
-    def required_keys(self) -> List[str]:
+    to the window.  ``loss_mask`` and ``target_ids`` use the +1 shifted
-        return ["sequence", "loss_mask", "position_ids"]
+    slice so they align with the predicted positions.
+    """
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
+    required_keys = ["sequence", "loss_mask", "position_ids"]
-        return self.storage.fetch(begin_idx, end_idx, key)
-
-    def __getitem__(self, index):
-        begin_idx, end_idx = self.get_index(index)
-
-        x = self._fetch_data(begin_idx, end_idx, "sequence")
-        y = self._fetch_data(begin_idx + 1, end_idx + 1, "sequence")
-        position_ids = self._fetch_data(begin_idx, end_idx, "position_ids")
-        loss_mask = self._fetch_data(begin_idx + 1, end_idx + 1, "loss_mask")
 
+    def __getitem__(self, index: int):
+        begin, end = self.store.sample_window(index)
+        x = self.store.fetch(begin, end, "sequence")
+        y = self.store.fetch(begin + 1, end + 1, "sequence")
+        position_ids = self.store.fetch(begin, end, "position_ids")
+        loss_mask = self.store.fetch(begin + 1, end + 1, "loss_mask")
         return {
             "input_ids": x.to(dtype=torch.long),
             "target_ids": y.to(dtype=torch.long),
@@ -219,59 +442,65 @@ class SFTDataset(BaseDataset):
 
 @DatasetFactory.register("dpo")
 class DPODataset(BaseDataset):
-    """Dataset for Direct Preference Optimization training."""
+    """Record-structured dataset for Direct Preference Optimization.
 
-    @property
+    Each sample is one preference pair (chosen + rejected) and is an
-    def required_keys(self) -> List[str]:
+    independent training unit — no windowing, stride, or cross-record
-        return ["chosen", "rejected", "chosen_mask", "rejected_mask"]
+    concatenation.  This keeps each sequence self-contained so attention
+    never leaks across preference pairs.
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
+    Two loading paths (handled by :class:`DatasetFactory`):
-        return self.storage.fetch(begin_idx, end_idx, key)
 
-    def __getitem__(self, index: int):
+    - **Pre-tokenized** (H5/bin): ``store.load(path)`` reads per-record
-        begin_idx, end_idx = self.get_index(index)
+      tensors; ``__getitem__`` returns them directly.
+    - **Raw JSONL** (``tokenizer_path=...``): builds a lazy processor
+      via :func:`dpo_processor` that tokenises on the fly — no packing,
+      no ``position_ids``.
+    """
 
-        chosen = self._fetch_data(begin_idx, end_idx, "chosen").to(dtype=torch.long)
+    required_keys = ["chosen", "rejected", "chosen_mask", "rejected_mask"]
-        rejected = self._fetch_data(begin_idx, end_idx, "rejected").to(dtype=torch.long)
-        chosen_mask = self._fetch_data(begin_idx, end_idx, "chosen_mask").to(
-            dtype=torch.bool
-        )
-        rejected_mask = self._fetch_data(begin_idx, end_idx, "rejected_mask").to(
-            dtype=torch.bool
-        )
 
+    def make_processor(self, tokenizer, max_len: int):
+        return partial(dpo_processor, tokenizer=tokenizer, max_len=max_len)
+
+    def __getitem__(self, index: int) -> Dict[str, Tensor]:
         return {
-            "chosen": chosen,
+            "chosen": self.store.fetch_record(index, "chosen").to(dtype=torch.long),
-            "rejected": rejected,
+            "rejected": self.store.fetch_record(index, "rejected").to(dtype=torch.long),
-            "chosen_mask": chosen_mask,
+            "chosen_mask": self.store.fetch_record(index, "chosen_mask").to(
-            "rejected_mask": rejected_mask,
+                dtype=torch.bool
+            ),
+            "rejected_mask": self.store.fetch_record(index, "rejected_mask").to(
+                dtype=torch.bool
+            ),
         }
 
 
 @DatasetFactory.register("grpo")
 class GRPODataset(BaseDataset):
-    """Dataset for Group Relative Policy Optimization training."""
+    """Dataset for offline Group Relative Policy Optimization.
 
-    @property
+    Each sample is one prompt with its group of responses and scalar
-    def required_keys(self) -> List[str]:
+    rewards — an independent training unit with no windowing or stride.
-        return ["prompts", "responses", "masks", "rewards"]
 
-    def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
+    Expected storage layout (produced by JsonlStore or pre-tokenized):
-        return self.storage.fetch(begin_idx, end_idx, key)
+
+    - ``prompts``:   List[Tensor]  — one 1-D token tensor per record
+    - ``responses``: List[List[Tensor]] — G response tensors per record
+    - ``masks``:     List[List[Tensor]] — G mask tensors per record
+    - ``rewards``:   List[Tensor]  — one 1-D float tensor (len G) per record
+    """
+
+    required_keys = ["prompts", "responses", "masks", "rewards"]
 
     def __getitem__(self, index: int) -> Dict[str, Tensor]:
-        begin_idx, end_idx = self.get_index(index)
+        prompts = self.store.fetch_record(index, "prompts")
-
+        responses = self.store.fetch_record(index, "responses")
-        prompts = self._fetch_data(begin_idx, end_idx, "prompts").to(dtype=torch.long)
+        masks = self.store.fetch_record(index, "masks")
-        responses = self._fetch_data(begin_idx, end_idx, "responses").to(
+        rewards = self.store.fetch_record(index, "rewards")
-            dtype=torch.long
-        )
-        masks = self._fetch_data(begin_idx, end_idx, "masks").to(dtype=torch.bool)
-        rewards = self._fetch_data(begin_idx, end_idx, "rewards")
-
         return {
-            "prompts": prompts,
+            "prompts": prompts.to(dtype=torch.long),
-            "responses": responses,
+            "responses": [r.to(dtype=torch.long) for r in responses],
-            "masks": masks,
+            "masks": [m.to(dtype=torch.bool) for m in masks],
-            "rewards": rewards,
+            "rewards": rewards.to(dtype=torch.float32),
         }

astrai/dataset/sampler.py

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

astrai/dataset/storage.py

@@ -1,20 +1,48 @@
 """Storage backends for different data formats.
 
-Layers:
+Architecture (composition over inheritance):
-  - I/O layer:       save_* / load_* functions, read/write raw files (HDF5/bin)
-                      return Dict[str, List[Tensor]] — format-specific, no state
-  - Store (ABC):     central abstraction, normalizes multi-segment into
-                      Dict[str, List[Tensor]] per key via _normalize(),
-                      fetch() uses bisect across segments — no forced concat
-  - Dataset layer:   BaseDataset owns a Store, only calls store.fetch(begin, end, key)
 
-Key properties:
+    Store (ABC)               — owns _data/_cum/_offsets bookkeeping
-  - Multi-segment:   segments kept as-is, no forced concatenation — safe for
+                                  + window_size/stride for sample-id
-                      datasets larger than RAM
+                                  indexing.  __getitem__/__len__ produce
-  - Explicit length: _length = min(total elements across keys), set at load,
+                                  the smallest iterable unit so Dataset
-                      __len__ returns O(1)
+                                  classes are pure delegators.
-  - Zero-copy mmap:  MmapStore wraps np.memmap(mode="r"), all DataLoader
+    Streamable (mixin)        — raw token slice  fetch(begin, end, keys)
-                       workers share OS page-cache pages
+    Recordable (mixin)        — raw record slice fetch_record(idx, keys)
+
+    H5Store(Store, Streamable, Recordable)
+    MmapStore(Store, Streamable, Recordable)
+    JsonlStore(Store, Streamable, Recordable)
+
+Each mixin is a stateless trait that relies on ``self._data`` etc.
+provided by :class:`Store`.  Concrete stores mix in whichever access
+primitives they support — ``Store`` is the sole base class, so there is
+no diamond inheritance or MRO ambiguity.
+
+Sample-id indexing lives on :class:`Store`, not on the dataset:
+
+- **Stream mode** (``window_size > 0``):  ``len(store)`` returns the number
+  of ``(window_size, stride)`` windows that fit in the token river;
+  ``store[i]`` returns the *i*-th window as a dict of per-key tensors;
+  ``store.sample_window(i)`` exposes the underlying ``(begin, end)``
+  token slice for callers (e.g. next-token trainers) that need a +1
+  shifted companion window.
+- **Record mode** (``num_records > 0``):  ``len(store)`` returns the
+  record count; ``store[i]`` returns the *i*-th record dict.
+
+Raw token/record access via :meth:`fetch` / :meth:`fetch_record`
+remains available for low-level callers that want explicit index
+control.  ``store.token_count`` is the total stream token count (what
+``len(store)`` used to mean in the legacy stream-only API).
+
+``segments_are_records`` (class attribute on each Store subclass)
+tells ``_normalize`` whether segments are inherently per-record (H5/
+JSONL) or opaque shards (bin).  Record access for bin relies on
+``_offsets`` instead.
+
+:class:`JsonlStore` supports a lazy mode (``processor=fn``) that keeps
+raw records and defers tokenisation to ``fetch_record`` — used by DPO
+to train directly from a ``.jsonl`` file without a pre-tokenised copy.
 """
 
 import bisect
@@ -23,20 +51,19 @@ import json
 import logging
 from abc import ABC, abstractmethod
 from pathlib import Path
-from typing import Dict, List, Union
+from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
 from torch import Tensor
 
 from astrai.config.preprocess_config import PipelineConfig
 from astrai.factory import BaseFactory
-from astrai.preprocessing.builder import MaskBuilderFactory
+from astrai.preprocessing.transform import TokenizeTransform
-from astrai.preprocessing.position_id import PositionIdStrategyFactory
 from astrai.serialization import (
     load_bin,
+    load_bin_offsets,
     load_h5,
 )
-from astrai.tokenize import AutoTokenizer
 
 logger = logging.getLogger(__name__)
 
@@ -48,7 +75,7 @@ def detect_format(load_path: str) -> str:
         load_path: Directory or file path
 
     Returns:
-        Format string ("h5", "bin", or "jsonl")
+        Format string ("h5", "bin", "jsonl", or "processed")
 
     Raises:
         FileNotFoundError: If no supported data files are found
@@ -85,228 +112,553 @@ def detect_format(load_path: str) -> str:
 
 
 class Store(ABC):
-    """String keys -> segmented tensors with ``fetch(begin, end, keys)``.
+    """Common base for all storage backends.
 
-    Each key maps to one or more tensor segments (no forced concatenation).
+    A Store owns both its data layout AND its sample-id → token/record
-    ``len(store)`` returns ``self._length`` (explicit, O(1)), the minimum
+    index translation.  Datasets are thin wrappers that bind a Store
-    total element count across all keys.
+    to a particular train-type's key mapping; they never know about
+    window/stride math.
 
-    Subclasses fill ``self._data`` and ``self._cum`` during ``load()``
+    Two iteration modes:
-    via ``_normalize()``.
+
+    - **Stream** (``window_size > 0``): data is treated as one long
+      token river.  ``len(store)`` returns the number of windows;
+      ``store[i]`` slices every stream-compatible key to window ``i``;
+      ``store.sample_window(i)`` returns the ``(begin, end)`` token
+      slice for callers needing a +1 shifted companion window.
+    - **Record** (``num_records > 0``): data is per-record.
+      ``len(store)`` returns ``num_records``; ``store[i]`` returns
+      the *i*-th record as a dict.
+
+    Raw token slicing is still available via :meth:`fetch` (mixed in
+    by :class:`Streamable`) when a store has stream support configured.
+    Raw record slicing via :meth:`fetch_record` (mixed in by
+    :class:`Recordable`) when a store has record support.
+
+    ``token_count`` exposes the raw total stream length — this is what
+    ``len(store)`` returned in the legacy stream-only API and what
+    stream-bound ``fetch`` uses for its bounds check.
     """
 
-    def __init__(self):
+    segments_are_records: bool = False
+
+    def __init__(
+        self,
+        window_size: int = 0,
+        stride: Optional[int] = None,
+    ):
         self._data: Dict[str, List[Tensor]] = {}
         self._cum: Dict[str, List[int]] = {}
+        self._offsets: Dict[str, List[int]] = {}
         self._length: int = 0
+        self._num_records: int = 0
+        self._window_size: int = int(window_size)
+        self._stride: int = int(stride) if stride is not None else int(window_size)
 
     @abstractmethod
-    def load(self, path: str) -> None:
+    def load(self, path: str, **kwargs) -> None:
         raise NotImplementedError
 
     @property
     def keys(self) -> List[str]:
         return list(self._data.keys())
 
-    def __len__(self) -> int:
+    @property
+    def window_size(self) -> int:
+        return self._window_size
+
+    @property
+    def stride(self) -> int:
+        return self._stride
+
+    @property
+    def token_count(self) -> int:
+        """Total tokens across all stream segments.
+
+        Useful for the bounds-checked raw :meth:`fetch` and as the
+        legacy ``len(store)`` value.
+        """
         return self._length
 
+    @property
+    def num_records(self) -> int:
+        """Number of records available via :meth:`fetch_record`.
+
+        Non-zero only when the backing layout provides per-record
+        indexing (H5/JSONL segments or bin ``_offsets``).
+        """
+        return self._num_records
+
+    @property
+    def num_samples(self) -> int:
+        """Number of items produced by ``__getitem__``.
+
+        Stream-mode wins when ``window_size > 0`` and there are tokens
+        to slice; otherwise falls back to ``num_records``.
+        """
+        if self._window_size > 0 and self._length > 0:
+            total = self._length
+            w = self._window_size
+            if total <= w:
+                return 0
+            return (total - 1 - w) // self._stride + 1
+        return self._num_records
+
+    def __len__(self) -> int:
+        return self.num_samples
+
+    def __getitem__(self, index: int) -> Dict[str, Tensor]:
+        if index < 0:
+            index += self.num_samples
+        if not 0 <= index < self.num_samples:
+            raise IndexError(
+                f"Store index out of range: {index}, num_samples={self.num_samples}"
+            )
+        if self._window_size > 0 and self._length > 0:
+            begin, end = self.sample_window(index)
+            keys = self._stream_keys()
+            return {k: self.fetch(begin, end, k) for k in keys}
+        return self.fetch_record(index, self._record_keys())
+
+    def sample_window(self, index: int) -> Tuple[int, int]:
+        """Return ``(begin, end)`` token positions for stream sample *index*.
+
+        The clipped tail keeps the last reachable window inside the
+        token river instead of overshooting.  Caller is responsible
+        for staying within :attr:`num_samples`: an out-of-range index
+        raises ``IndexError``.
+        """
+        if self._window_size <= 0:
+            raise RuntimeError("sample_window() requires window_size > 0 (stream mode)")
+        if self._window_size <= 0 or self._length <= self._window_size:
+            raise IndexError(
+                f"Data too short for window: token_count={self._length}, "
+                f"window_size={self._window_size}"
+            )
+        if not 0 <= index < self.num_samples:
+            raise IndexError(
+                f"Sample index out of range: {index}, num_samples={self.num_samples}"
+            )
+        total = self._length
+        begin = min(index * self._stride, total - 1 - self._window_size)
+        end = min(begin + self._window_size, total - 1)
+        return begin, end
+
+    def _stream_keys(self) -> List[str]:
+        out: List[str] = []
+        for k, tensors in self._data.items():
+            if tensors and isinstance(tensors[0], list):
+                continue
+            out.append(k)
+        return out
+
+    def _record_keys(self) -> List[str]:
+        return list(self._data.keys())
+
+    def _normalize(
+        self,
+        raw: Dict[str, list],
+        offsets: Optional[Dict[str, List[int]]] = None,
+    ):
+        """Register segments and pre-compute indices for both access modes.
+
+        Stream mode: ``_cum[key]`` accumulates per-segment lengths so
+        ``Streamable._fetch_stream_key`` can bisect across segments
+        without concatenation.
+
+        Record mode: if *offsets* is provided (bin layout),
+        ``_offsets[key]`` stores cumulative per-record offsets into the
+        single concatenated segment.  Otherwise, when
+        ``segments_are_records`` is True (H5/JSONL), ``_data[key]`` is
+        a per-record list and ``fetch_record`` indexes it directly.
+
+        Nested keys (GRPO ``responses``/``masks`` as
+        ``List[List[Tensor]]``) are stored as-is and excluded from both
+        cumulative bookkeepings — they are only accessed record-by-record.
+        """
+        flat_lengths = []
+        for key, tensors in raw.items():
+            self._data[key] = tensors
+            if not tensors:
+                self._cum[key] = []
+                flat_lengths.append(0)
+                continue
+            if isinstance(tensors[0], list):
+                self._cum[key] = []
+                continue
+            cum = []
+            total = 0
+            for t in tensors:
+                total += t.shape[0]
+                cum.append(total)
+            self._cum[key] = cum
+            flat_lengths.append(cum[-1] if cum else 0)
+        self._length = min(flat_lengths) if flat_lengths else 0
+
+        valid_offsets: Dict[str, List[int]] = {}
+        if offsets:
+            for key, off in offsets.items():
+                segs = self._data.get(key, [])
+                if len(segs) == 1 and len(off) > 1:
+                    valid_offsets[key] = off
+                elif len(segs) > 1:
+                    logger.warning(
+                        "Key '%s' has %d segments with offsets — record mode "
+                        "disabled for this key (multi-shard bin+offsets not "
+                        "supported). Merge shards or use H5/JSONL.",
+                        key,
+                        len(segs),
+                    )
+        self._offsets = valid_offsets
+        if valid_offsets:
+            record_counts = [len(v) - 1 for v in valid_offsets.values()]
+            self._num_records = min(record_counts) if record_counts else 0
+        elif self.segments_are_records:
+            per_record_counts = []
+            for key, tensors in self._data.items():
+                if tensors and isinstance(tensors[0], list):
+                    continue
+                per_record_counts.append(len(tensors))
+            self._num_records = min(per_record_counts) if per_record_counts else 0
+        else:
+            self._num_records = 0
+
+
+class Streamable:
+    """Mixin granting raw token-stream access via :meth:`fetch`.
+
+    Stateless trait relying on ``self._data``, ``self._cum``,
+    ``self._length`` maintained by :class:`Store`.  Stream mode is
+    active when the owning store has ``window_size > 0``; for stores
+    that can also serve record access (H5/JSONL/bin+offsets), the
+    ``fetch_record`` API from :class:`Recordable` is used instead.
+    """
+
     def fetch(
         self,
         begin: int,
         end: int,
         keys: Union[str, List[str]],
     ):
-        if not self._data:
+        return _stream_fetch(self, begin, end, keys)
-            raise RuntimeError("Store not loaded")
-        if not (0 <= begin < self._length and 0 <= end <= self._length):
-            raise ValueError(
-                f"Index out of bounds: begin={begin}, end={end}, length={self._length}"
-            )
-        if isinstance(keys, str):
-            return self._fetch_key(keys, begin, end)
-        return {k: self._fetch_key(k, begin, end) for k in keys}
 
-    def _fetch_key(self, key: str, begin: int, end: int) -> Tensor:
-        """Fetch slice [begin, end) across potentially multiple segments."""
-        segments = self._data[key]
-        cum = self._cum[key]
-        seg_start = bisect.bisect_right(cum, begin)
-        seg_end = bisect.bisect_left(cum, end)
 
-        results = []
+def _stream_fetch(self, begin: int, end: int, keys: Union[str, List[str]]):
-        for i in range(seg_start, seg_end + 1):
+    if not getattr(self, "_data", None):
-            prev = cum[i - 1] if i > 0 else 0
+        raise RuntimeError("Store not loaded")
-            s = max(begin - prev, 0)
+    if not (0 <= begin < self._length and 0 <= end <= self._length):
-            e = min(end - prev, segments[i].shape[0])
+        raise ValueError(
-            results.append(segments[i][s:e])
+            f"Index out of bounds: begin={begin}, end={end}, length={self._length}"
-
-        return results[0] if len(results) == 1 else torch.cat(results, dim=0)
-
-    def _normalize(self, raw: Dict[str, List[Tensor]]):
-        """Register segments and pre-compute cumulative lengths.
-
-        Does NOT concatenate — segments are kept as-is to avoid OOM on
-        large datasets.  Sets ``self._length`` to the minimum total
-        element count across all keys.
-        """
-        for key, tensors in raw.items():
-            self._data[key] = tensors
-            cum = []
-            total = 0
-            for t in tensors:
-                total += t.shape[0]
-                cum.append(total)
-            self._cum[key] = cum
-        self._length = (
-            min((cum[-1] if cum else 0) for cum in self._cum.values())
-            if self._cum
-            else 0
         )
+    if isinstance(keys, str):
+        return _fetch_stream_key(self, keys, begin, end)
+    return {k: _fetch_stream_key(self, k, begin, end) for k in keys}
+
+
+def _fetch_stream_key(self, key: str, begin: int, end: int) -> Tensor:
+    segments = self._data[key]
+    cum = self._cum[key]
+    seg_start = bisect.bisect_right(cum, begin)
+    seg_end = bisect.bisect_left(cum, end)
+
+    results = []
+    for i in range(seg_start, seg_end + 1):
+        prev = cum[i - 1] if i > 0 else 0
+        s = max(begin - prev, 0)
+        e = min(end - prev, segments[i].shape[0])
+        results.append(segments[i][s:e])
+
+    return results[0] if len(results) == 1 else torch.cat(results, dim=0)
+
+
+class Recordable:
+    """Mixin granting raw record access via :meth:`fetch_record`.
+
+    Stateless trait relying on ``self._data``, ``self._offsets``,
+    ``self._num_records`` maintained by :class:`Store`.
+    """
+
+    def fetch_record(
+        self,
+        index: int,
+        keys: Union[str, List[str]],
+    ):
+        return _record_fetch(self, index, keys)
+
+
+def _record_fetch(self, index: int, keys: Union[str, List[str]]):
+    if not getattr(self, "_data", None) and self._num_records == 0:
+        raise RuntimeError("Store not loaded")
+    if not 0 <= index < self._num_records:
+        raise ValueError(
+            f"Record index out of bounds: {index}, num_records={self._num_records}"
+        )
+    if isinstance(keys, str):
+        return _fetch_record_key(self, keys, index)
+    return {k: _fetch_record_key(self, k, index) for k in keys}
+
+
+def _fetch_record_key(self, key: str, index: int):
+    offsets = self._offsets.get(key)
+    if offsets:
+        start = offsets[index]
+        end = (
+            offsets[index + 1]
+            if index + 1 < len(offsets)
+            else self._data[key][0].shape[0]
+        )
+        return self._data[key][0][start:end]
+    return self._data[key][index]
 
 
 class StoreFactory(BaseFactory["Store"]):
-    """Factory for creating Store instances by type name.
+    """Factory for creating Store instances by type name."""
-
-    Example::
-
-        @StoreFactory.register("custom")
-        class CustomStore(Store):
-            ...
-    """
 
 
 @StoreFactory.register("h5")
-class H5Store(Store):
+class H5Store(Store, Streamable, Recordable):
-    """HDF5-based storage backend (pre-tokenized data)."""
+    """HDF5-based storage backend (pre-tokenized data).
 
-    def load(self, path: str):
+    Each key is stored as a group of per-record datasets (``data_0``,
+    ``data_1``, …).  Supports both access modes:
+
+    - **Stream**: ``fetch(begin, end, key)`` and ``store[i]`` slice
+      across concatenated records via ``_cum`` — used by SEQ/SFT.
+    - **Record**: ``fetch_record(i, key)`` and ``store[i]`` (when
+      ``window_size == 0``) index ``_data[key]`` directly — used by
+      DPO/GRPO.
+    """
+
+    segments_are_records = True
+
+    def __init__(
+        self,
+        window_size: int = 0,
+        stride: Optional[int] = None,
+    ):
+        super().__init__(window_size=window_size, stride=stride)
+
+    def load(self, path: str, **kwargs):
         self._normalize(load_h5(path))
 
 
 @StoreFactory.register("bin")
-class MmapStore(Store):
+class MmapStore(Store, Streamable, Recordable):
     """Memory-mapped binary storage backend.
 
     Each key is a single .bin file backed by ``np.memmap(mode="r")``.
     No per-process memory duplication — all DataLoader workers share the
     same OS page-cache pages.
 
-    Format on disk::
+    Supports both access modes:
 
-        data_root/
+    - **Stream**: always available via :meth:`fetch`.
-          meta.json          # {key: {shape, dtype}, ...}
+    - **Record** (``fetch_record(i, key)``): only when ``meta.json``
-          <key>.bin          # raw numpy array, one per key
+      contains per-record ``offsets`` (written via
+      ``save_bin(..., record_keys=...)``).  Legacy bin files without
+      offsets have ``num_records == 0`` and ``len(store)`` reflects the
+      windowed sample count when ``window_size > 0``.
+
+    ``segments_are_records`` is ``False`` here (bin segments are
+    contiguous streams, not per-record) — record access is driven
+    purely by ``_offsets``.
     """
 
-    def load(self, path: str):
+    segments_are_records = False
+
+    def __init__(
+        self,
+        window_size: int = 0,
+        stride: Optional[int] = None,
+    ):
+        super().__init__(window_size=window_size, stride=stride)
+        self._mmap_refs: List[Tensor] = []
+
+    def load(self, path: str, **kwargs):
         self._mmap_refs = []
         root = Path(path)
         all_raw: Dict[str, List[Tensor]] = {}
+        all_offsets: Dict[str, List[int]] = {}
         meta_paths = [
             Path(p) for p in glob.glob(str(root / "**" / "meta.json"), recursive=True)
         ]
         for meta_path in meta_paths:
             raw = load_bin(str(meta_path.parent))
+            off = load_bin_offsets(str(meta_path.parent))
             for key, tensors in raw.items():
                 if key not in all_raw:
                     all_raw[key] = []
                 all_raw[key].extend(tensors)
+            for key, o in off.items():
+                if key not in all_offsets:
+                    all_offsets[key] = []
+                all_offsets[key].extend(o)
         if not meta_paths:
             raise FileNotFoundError(f"No meta.json found under {path}")
-        self._normalize(all_raw)
+        self._normalize(all_raw, offsets=all_offsets or None)
         for tensors in self._data.values():
             self._mmap_refs.extend(tensors)
 
 
-@StoreFactory.register("jsonl")
+class JsonlSource:
-class JsonlStore(Store):
+    """Read raw JSON records from a ``.jsonl`` file or directory.
-    """On-the-fly tokenization store for raw JSONL files.
 
-    A JSONL dataset directory contains ``*.jsonl`` files plus a
+    A thin reader used by :class:`JsonlStore` in processor mode — holds
-    ``dataset_config.json`` file that follows the same schema as
+    no tokenizer, performs no tokenisation, just yields dicts.
-    :class:`PipelineConfig` with an additional ``tokenizer_path`` field.
+    """
-    Records are tokenized when the store is loaded and concatenated into
+
-    segmented tensors matching the key layout expected by the dataset
+    def __init__(self, path: str):
-    classes (``sequence``, ``loss_mask``, ``position_ids``, ...).
+        self.path = Path(path)
+        self._records: Optional[List[dict]] = None
+
+    def load(self) -> List[dict]:
+        if self._records is None:
+            self._records = self._read(self.path)
+        return self._records
+
+    @staticmethod
+    def _read(root: Path) -> List[dict]:
+        if root.is_file():
+            return JsonlSource._read_file(root)
+        return JsonlSource._read_dir(root)
+
+    @staticmethod
+    def _read_file(path: Path) -> List[dict]:
+        records: List[dict] = []
+        with open(path, "r", encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    records.append(json.loads(line))
+                except json.JSONDecodeError:
+                    logger.warning("Failed to parse JSON line in %s, skipping", path)
+        return records
+
+    @staticmethod
+    def _read_dir(root: Path) -> List[dict]:
+        records: List[dict] = []
+        for jsonl_path in sorted(root.glob("*.jsonl")):
+            records.extend(JsonlSource._read_file(jsonl_path))
+        return records
+
+
+@StoreFactory.register("jsonl")
+class JsonlStore(Store, Streamable, Recordable):
+    """JSONL reader with eager/lazy tokenisation modes.
+
+    A JSONL dataset is a ``.jsonl`` file or a directory of ``*.jsonl``
+    files plus (optionally) a ``dataset_config.json`` describing the
+    tokenization pipeline.
+
+    Three ways to supply an eager transform (first match wins):
+
+    - **Explicit** (``transform=``): caller-built
+      :class:`TokenizeTransform` applied eagerly.
+    - **Config file**: ``dataset_config.json`` alongside the ``*.jsonl``
+      files — loaded via :meth:`TokenizeTransform.from_config_file`.
+    - **Default messages** (``tokenizer_path=`` given, no config file):
+      a built-in chatml config that tokenises the ``messages`` field,
+      masking every role except ``assistant`` (loss on assistant only).
+      Lets SFT/SEQ train straight from a chat-style JSONL directory
+      without a hand-written config.
+
+    Two tokenisation modes, selected at :meth:`load` time:
+
+    - **Eager** (default): applies the transform to every record at load
+      time and registers per-key tensors via ``_normalize``.  Both
+      ``fetch`` (stream) and ``fetch_record`` (record) work.
+    - **Lazy** (``processor=fn`` passed): keeps raw records and defers
+      tokenisation to ``fetch_record``.  Only record access works —
+      ``len(store)`` returns ``num_records``; stream primitives raise.
     """
 
     CONFIG_NAME = "dataset_config.json"
+    segments_are_records = True
 
-    def load(self, path: str):
+    _DEFAULT_MESSAGES_CONFIG = {
-        root = Path(path)
+        "version": 1,
-        config_path = root / self.CONFIG_NAME
+        "input": {
-        if not config_path.exists():
+            "sections": [{"field": "messages", "action": "$role", "template": True}]
-            raise FileNotFoundError(
+        },
-                f"JSONL dataset config not found: {config_path}. "
+        "mask": {"system": "mask", "user": "mask", "assistant": "train"},
-                f"Expected {self.CONFIG_NAME} alongside *.jsonl files."
+        "mask_default": "mask",
+        "output": {"position_ids_mode": "doc_reset"},
+    }
+
+    def __init__(
+        self,
+        window_size: int = 0,
+        stride: Optional[int] = None,
+    ):
+        super().__init__(window_size=window_size, stride=stride)
+        self._source: Optional[JsonlSource] = None
+        self._processor: Optional[Callable[[dict], Dict[str, Tensor]]] = None
+        self._keys_cache: Optional[List[str]] = None
+
+    def load(self, path: str, transform=None, processor=None, **kwargs):
+        self._source = JsonlSource(path)
+        records = self._source.load()
+
+        if processor is not None:
+            self._processor = processor
+            self._num_records = len(records)
+            return
+
+        if transform is None:
+            root = Path(path)
+            config_path = root / self.CONFIG_NAME if root.is_dir() else None
+            if config_path is not None and config_path.exists():
+                transform = TokenizeTransform.from_config_file(str(config_path))
+            else:
+                tokenizer_path = kwargs.get("tokenizer_path")
+                if not tokenizer_path:
+                    raise FileNotFoundError(
+                        f"JSONL dataset config not found. Expected "
+                        f"{self.CONFIG_NAME} alongside *.jsonl files, pass an "
+                        f"explicit transform, pass processor= for lazy "
+                        f"on-the-fly tokenisation, or pass tokenizer_path= to "
+                        f"use the built-in messages config."
+                    )
+                config = PipelineConfig.from_dict(self._DEFAULT_MESSAGES_CONFIG)
+                transform = TokenizeTransform(config, tokenizer_path)
+
+        transformed = transform.apply(records)
+        self._normalize(transformed)
+
+    @property
+    def keys(self) -> List[str]:
+        if self._processor is not None:
+            if self._keys_cache is None and self._num_records > 0:
+                sample = self._processor(self._source.load()[0])
+                self._keys_cache = list(sample.keys())
+            return self._keys_cache or []
+        return list(self._data.keys())
+
+    def fetch_record(self, index: int, keys: Union[str, List[str]]):
+        if self._processor is not None:
+            if not 0 <= index < self._num_records:
+                raise ValueError(
+                    f"Record index out of bounds: {index}, "
+                    f"num_records={self._num_records}"
+                )
+            record = self._source.load()[index]
+            data = self._processor(record)
+            if isinstance(keys, str):
+                return data[keys]
+            return {k: data[k] for k in keys}
+        return _record_fetch(self, index, keys)
+
+    def fetch(self, begin: int, end: int, keys: Union[str, List[str]]):
+        if self._processor is not None:
+            raise RuntimeError(
+                "JsonlStore in lazy (processor) mode does not support "
+                "stream fetch(); use fetch_record() instead."
             )
+        return _stream_fetch(self, begin, end, keys)
 
-        with open(config_path, "r", encoding="utf-8") as f:
+    def __getitem__(self, index: int) -> Dict[str, Tensor]:
-            raw_config = json.load(f)
+        if self._processor is not None:
-
+            return self.fetch_record(index, self._record_keys())
-        tokenizer_path = raw_config.pop("tokenizer_path", None)
+        return super().__getitem__(index)
-        if tokenizer_path is None:
-            raise ValueError(
-                f"JSONL dataset config must specify 'tokenizer_path': {config_path}"
-            )
-
-        self.config = PipelineConfig.from_dict(raw_config)
-        tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
-        mask_builder = MaskBuilderFactory.create("sectioned")
-        position_strategy = PositionIdStrategyFactory.create(
-            self.config.output.position_ids_mode
-        )
-
-        raw: Dict[str, List[Tensor]] = {}
-        doc_sequences: List[List[int]] = []
-
-        for jsonl_path in sorted(root.glob("*.jsonl")):
-            with open(jsonl_path, "r", encoding="utf-8") as f:
-                for line in f:
-                    line = line.strip()
-                    if not line:
-                        continue
-                    try:
-                        item = json.loads(line)
-                    except json.JSONDecodeError:
-                        logger.warning(
-                            "Failed to parse JSON line in %s, skipping", jsonl_path
-                        )
-                        continue
-
-                    result = mask_builder.build(item, self.config, tokenizer)
-                    if result is None:
-                        continue
-
-                    result.pop("domain", None)
-                    primary_ids = self._primary_ids(result)
-                    if not primary_ids:
-                        continue
-
-                    doc_sequences.append(primary_ids)
-                    for key, ids in result.items():
-                        if key not in raw:
-                            raw[key] = []
-                        raw[key].append(torch.tensor(ids, dtype=self._infer_dtype(ids)))
-
-        pos_ids = position_strategy.generate(doc_sequences)
-        if pos_ids:
-            raw["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
-
-        self._normalize(raw)
-
-    @staticmethod
-    def _primary_ids(result: dict) -> List[int]:
-        """Return the first integer list in *result* as the primary id sequence."""
-        for val in result.values():
-            if isinstance(val, list) and val and isinstance(val[0], int):
-                return val
-        return []
-
-    @staticmethod
-    def _infer_dtype(ids: List) -> torch.dtype:
-        """Infer tensor dtype from the first element of a token/value list."""
-        if ids and isinstance(ids[0], float):
-            return torch.float32
-        return torch.int32

astrai/extension/__init__.py

@@ -1,19 +1,30 @@
 """CUDA attention kernel wrappers with torch fallback.
 
 Public API:
-    - ``gqa_decode_attn`` — single-query decode attention
+    - ``attn_decode`` — single-query decode attention
-    - ``gqa_prefill_attn`` — multi-query prefill attention
+    - ``attn_prefill`` — multi-query prefill attention
+    - ``attn_paged_decode`` — paged decode attention (direct page-table access)
 
-Each wrapper dispatches to its compiled CUDA kernel (``astrai.extension.gqa_*``)
+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 gqa_decode_attn, gqa_prefill_attn
+from astrai.extension.ops import attention, attn_decode, attn_paged_decode, attn_prefill
 
 __all__ = [
-    "gqa_decode_attn",
+    "attn_decode",
-    "gqa_prefill_attn",
+    "attn_paged_decode",
+    "attn_prefill",
+    "attention",
     "is_available",
     "KERNEL_NAMES",
 ]

astrai/extension/loader.py

@@ -11,7 +11,7 @@ import logging
 
 logger = logging.getLogger(__name__)
 
-KERNEL_NAMES = ["gqa_decode_attn", "gqa_prefill_attn"]
+KERNEL_NAMES = ["attn_decode", "attn_prefill", "attn_paged_decode"]
 
 _available: dict[str, bool] = {}
 _modules: dict[str, object] = {}

astrai/extension/ops.py

@@ -3,15 +3,43 @@
 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
@@ -26,61 +54,245 @@ def _expand_kv_heads(
     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,
-    is_causal: bool,
+    causal_offset: int,
-    scale: float | None,
+    scale: float,
+    q_layout: int,
+    kv_layout: int | None = None,
 ) -> torch.Tensor:
-    """Reference attention via ``scaled_dot_product_attention``."""
+    """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 = mask[:, None, None, :] if mask is not None else None
+    attn_mask, resolved_scale = _build_attn_mask(q, k, mask, causal_offset, scale)
-    return F.scaled_dot_product_attention(
+    out = F.scaled_dot_product_attention(
-        q, k, v, attn_mask=attn_mask, is_causal=is_causal and mask is None, scale=scale
+        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
     )
 
 
-def gqa_decode_attn(
+def attention(
     q: torch.Tensor,
     k: torch.Tensor,
     v: torch.Tensor,
     mask: torch.Tensor | None = None,
-    is_causal: bool = False,
+    causal_offset: int = -1,
-    causal_offset: int = 0,
+    scale: float = 0.0,
-    scale: float | None = None,
+    layout: str = "bhld",
 ) -> torch.Tensor:
-    if _available["gqa_decode_attn"]:
+    """Dispatch to decode or prefill attention based on the query length.
-        return _modules["gqa_decode_attn"].gqa_decode_attn(
-            q,
-            k,
-            v,
-            mask=mask,
-            is_causal=is_causal,
-            causal_offset=causal_offset,
-            scale=scale,
-        )
-    return _torch_fallback(q, k, v, mask, is_causal, scale)
 
+    A query length of one is the decode case; longer queries use prefill.
+    The paged-cache decode path cannot be selected here because its page-table
+    arguments are not part of this interface.
+    """
+    li = _parse_layout(layout)
 
-def gqa_prefill_attn(
+    if q.ndim not in (2, 3, 4) or k.ndim != q.ndim or v.ndim != q.ndim:
-    q: torch.Tensor,
+        raise ValueError(
-    k: torch.Tensor,
+            "q, k, and v must all have the same rank in {2, 3, 4}, "
-    v: torch.Tensor,
+            f"got {q.ndim}D, {k.ndim}D, {v.ndim}D"
-    mask: torch.Tensor | None = None,
-    is_causal: bool = False,
-    causal_offset: int = 0,
-    scale: float | None = None,
-) -> torch.Tensor:
-    if _available["gqa_prefill_attn"]:
-        return _modules["gqa_prefill_attn"].gqa_prefill_attn(
-            q,
-            k,
-            v,
-            mask=mask,
-            is_causal=is_causal,
-            causal_offset=causal_offset,
-            scale=scale,
         )
-    return _torch_fallback(q, k, v, mask, is_causal, scale)
+    if k.shape != v.shape:
+        raise ValueError(
+            f"k and v must have the same shape, got {k.shape} and {v.shape}"
+        )
+
+    original_ndim = q.ndim
+    if original_ndim == 2:
+        # [L, D] -> [1, 1, L, D] or [1, L, 1, D]
+        q = q.unsqueeze(0).unsqueeze(1 if li == 0 else 2)
+        k = k.unsqueeze(0).unsqueeze(1 if li == 0 else 2)
+        v = v.unsqueeze(0).unsqueeze(1 if li == 0 else 2)
+    elif original_ndim == 3:
+        # [B, L, D] -> single-head 4D input.
+        q = q.unsqueeze(1 if li == 0 else 2)
+        k = k.unsqueeze(1 if li == 0 else 2)
+        v = v.unsqueeze(1 if li == 0 else 2)
+
+    q_len = q.size(2 if li == 0 else 1)
+    if q_len == 1:
+        out = attn_decode(q, k, v, mask, causal_offset, scale, layout)
+    else:
+        out = attn_prefill(q, k, v, mask, causal_offset, scale, layout)
+
+    if original_ndim == 2:
+        return out.squeeze(0).squeeze(0 if li == 0 else 1)
+    if original_ndim == 3:
+        return out.squeeze(1 if li == 0 else 2)
+    return out

astrai/factory.py

@@ -67,7 +67,10 @@ class BaseFactory(ABC, Generic[T]):
             if _get_origin(orig_base) is BaseFactory:
                 (arg,) = _get_args(orig_base)
                 cls._entries = {}
-                cls._component_base = _resolve_type(arg, cls)
+                try:
+                    cls._component_base = _resolve_type(arg, cls)
+                except Exception:
+                    cls._component_base = None
                 return
 
     @classmethod

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/api/tool_parser.py

@@ -7,6 +7,7 @@ Subclasses may optionally consume ``token_ids`` for token-level parsing
 (e.g. Harmony / VLM-style parsers).
 """
 
+import json
 import re
 import uuid
 from abc import ABC, abstractmethod
@@ -117,6 +118,29 @@ def _parse_tool_call_json(json_str: str, complete: bool):
 
     Returns ``(name, args, valid)``.
     """
+    if complete:
+        try:
+            obj = json.loads(json_str)
+        except json.JSONDecodeError:
+            return None, "", False
+        name = obj.get("name")
+        if not isinstance(name, str) or not name:
+            return None, "", False
+        args = obj.get("arguments")
+        if isinstance(args, dict):
+            if not args:
+                args = ""
+            else:
+                args = json.dumps(args, ensure_ascii=False)
+                args = args[1:-1].rstrip()
+        elif isinstance(args, list):
+            args = json.dumps(args, ensure_ascii=False) if args else ""
+        elif isinstance(args, str):
+            pass
+        else:
+            args = str(args) if args is not None else ""
+        return name, args, True
+
     name_match = re.search(r'"name"\s*:\s*"([^"]*)"', json_str)
     if not name_match:
         return None, "", False
@@ -127,8 +151,6 @@ def _parse_tool_call_json(json_str: str, complete: bool):
         return name, "", True
 
     raw = args_match.group(1).rstrip()
-    if complete and raw.endswith("}"):
-        raw = raw[:-1].rstrip()
     if raw.startswith("{"):
         inner = raw[1:].rstrip()
         if inner.endswith("}"):
@@ -156,9 +178,6 @@ def _find_tool_calls(text: str, start_pos: int = 0):
             break
 
         json_str = text[brace:end]
-        if not _TOOL_CALL_HEAD_RE.search(json_str):
-            pos = end
-            continue
 
         name, args, valid = _parse_tool_call_json(json_str, complete=True)
         if not valid or name is None:
@@ -186,7 +205,7 @@ def _find_partial_tool_call(text: str, start_pos: int = 0):
         return None
 
     json_str = text[brace:]
-    if not _TOOL_CALL_HEAD_RE.search(json_str):
+    if '"name"' not in json_str:
         return None
 
     name, args, valid = _parse_tool_call_json(json_str, complete=False)

astrai/inference/core/cache.py

@@ -300,7 +300,11 @@ class KVCache(ABC):
 
     @abstractmethod
     def bind_tasks(
-        self, task_ids: List[str], total_len: int, device: torch.device
+        self,
+        task_ids: List[str],
+        total_len: int,
+        device: torch.device,
+        write_positions: Optional[Tensor] = None,
     ) -> CacheView: ...
 
     def task_cached(self, task_id: str) -> int:
@@ -399,7 +403,11 @@ class PageCache(KVCache):
             self._pool.record(page_table[i], prompt_ids, i)
 
     def bind_tasks(
-        self, task_ids: List[str], total_len: int, device: torch.device
+        self,
+        task_ids: List[str],
+        total_len: int,
+        device: torch.device,
+        write_positions: Optional[Tensor] = None,
     ) -> PageCacheView:
         page_table = self._table.table_tensor(task_ids, device)
         return PageCacheView(self._storage, page_table, total_len)
@@ -409,28 +417,31 @@ class ContiguousCacheView(CacheView):
     """Contiguous KV-cache view for attention layers."""
 
     def __init__(
-        self, cache: "ContiguousCache", batch_indices: Tensor, total_len: int = 0
+        self,
+        cache: "ContiguousCache",
+        batch_indices: Tensor,
+        total_len: int = 0,
+        write_positions: Optional[Tensor] = None,
     ):
         self._cache = cache
         self._batch_indices = batch_indices
         self._total_len = total_len
+        self._write_positions = write_positions
 
     def write(self, layer_id: int, k: Tensor, v: Tensor):
         seq_len = k.size(1)
-        start_pos = self._total_len - seq_len
         indices = self._batch_indices
-        self._cache.k[layer_id, indices, start_pos : start_pos + seq_len] = k
+        if self._write_positions is not None and seq_len == 1:
-        self._cache.v[layer_id, indices, start_pos : start_pos + seq_len] = v
+            pos = self._write_positions
-        new_len = start_pos + seq_len
+            self._cache.k[layer_id, indices, pos] = k.squeeze(1)
-        for s in indices.tolist():
+            self._cache.v[layer_id, indices, pos] = v.squeeze(1)
-            cur = self._cache._slot_len.get(s, 0)
+        else:
-            if new_len > cur:
+            start_pos = self._total_len - seq_len
-                self._cache._slot_len[s] = new_len
+            self._cache.k[layer_id, indices, start_pos : start_pos + seq_len] = k
+            self._cache.v[layer_id, indices, start_pos : start_pos + seq_len] = v
 
     def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]:
-        max_len = max(
+        max_len = self._total_len
-            self._cache._slot_len.get(int(s), 0) for s in self._batch_indices.tolist()
-        )
         indices = self._batch_indices
         k = self._cache.k[layer_id, indices, :max_len]
         v = self._cache.v[layer_id, indices, :max_len]
@@ -491,9 +502,24 @@ class ContiguousCache(KVCache):
     def task_extend(self, task_id: str, pos: int) -> bool:
         return pos < self.max_seq_len
 
+    def task_cached(self, task_id: str) -> int:
+        slot = self._task_slot.get(task_id)
+        if slot is None:
+            return 0
+        return self._slot_len.get(slot, 0)
+
     def bind_tasks(
-        self, task_ids: List[str], total_len: int, device: torch.device
+        self,
+        task_ids: List[str],
+        total_len: int,
+        device: torch.device,
+        write_positions: Optional[Tensor] = None,
     ) -> ContiguousCacheView:
         slots = [self._task_slot[tid] for tid in task_ids]
         batch_indices = torch.tensor(slots, dtype=torch.long, device=device)
-        return ContiguousCacheView(self, batch_indices, total_len)
+        for slot in slots:
+            if total_len > self._slot_len.get(slot, 0):
+                self._slot_len[slot] = total_len
+        return ContiguousCacheView(
+            self, batch_indices, total_len, write_positions=write_positions
+        )

astrai/inference/core/executor.py

@@ -43,19 +43,40 @@ class Executor:
         )
 
         task_ids = [t.task_id for t in tasks]
+        position_ids = (
+            torch.arange(start_pos, prompt_len, dtype=torch.long, device=self.device)
+            .unsqueeze(0)
+            .expand(batch_sz, -1)
+        )
+        input_mask = position_ids.unsqueeze(-1) >= torch.arange(
+            prompt_len, device=self.device
+        )
 
         with torch.inference_mode():
             self.model(
                 input_ids,
-                position_ids=torch.arange(
+                input_mask=input_mask,
-                    start_pos, prompt_len, dtype=torch.long, device=self.device
+                position_ids=position_ids,
-                )
-                .unsqueeze(0)
-                .expand(batch_sz, -1),
                 paged_cache=self.kv_cache.bind_tasks(task_ids, prompt_len, self.device),
             )
 
-    def execute_decode(self, tasks: List[Task]) -> List[int]:
+    def execute_decode(
+        self, tasks: List[Task], return_logprobs: bool = False
+    ) -> List[int]:
+        """Decode next token for each task.
+
+        Args:
+            return_logprobs: When ``True``, also record (and return)
+                the log-probability of each sampled token under the
+                post-strategy sampling distribution.  The logprob is
+                appended to ``task.output_logprobs`` and the return
+                list becomes ``List[Tuple[int, float]]``.
+
+        Returns:
+            ``List[int]`` of sampled token IDs, or
+            ``List[Tuple[int, float]]`` of ``(token_id, logprob)`` when
+            ``return_logprobs`` is ``True``.
+        """
         if not tasks:
             return []
 
@@ -68,25 +89,78 @@ class Executor:
         position_ids = torch.tensor(
             [t.next_pos for t in tasks], dtype=torch.long, device=self.device
         )
-        total_len = position_ids.max().item() + 1
+        total_len = max(t.next_pos for t in tasks) + 1
+        input_mask = position_ids[:, None, None] >= torch.arange(
+            total_len, device=self.device
+        )
 
         task_ids = [t.task_id for t in tasks]
 
         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 = []
+        history_lens = []
+        for t in tasks:
+            window = t.rep_window
+            prompt_part = t.prompt_ids[-window:]
+            ids = prompt_part + t.output_ids
+            history_lists.append(ids)
+            history_lens.append(len(ids))
+
+        max_len = max(history_lens) if history_lens else 0
+        padded_ids = torch.zeros(
+            len(tasks), max_len, dtype=torch.long, device=self.device
+        )
+        padded_mask = torch.zeros(
+            len(tasks), max_len, dtype=torch.bool, device=self.device
+        )
+        for i, h in enumerate(history_lists):
+            L = history_lens[i]
+            padded_ids[i, :L] = torch.as_tensor(h, dtype=torch.long, device=self.device)
+            padded_mask[i, :L] = True
 
         with torch.inference_mode():
             outputs = self.model(
                 input_ids.unsqueeze(1),
-                paged_cache=self.kv_cache.bind_tasks(task_ids, total_len, self.device),
+                input_mask=input_mask,
+                paged_cache=self.kv_cache.bind_tasks(
+                    task_ids,
+                    total_len,
+                    self.device,
+                    write_positions=position_ids,
+                ),
                 position_ids=position_ids.unsqueeze(1),
             )
             logits = outputs["logits"][:, -1, :]
 
+        if return_logprobs:
+            tokens, logprobs = sample(
+                logits,
+                temperature=temperatures,
+                top_k=top_ks,
+                top_p=top_ps,
+                frequency_penalty=freq_penalties,
+                input_ids=padded_ids,
+                input_mask=padded_mask,
+                return_logprobs=True,
+            )
+            tokens_list = tokens.tolist()
+            logprobs_list = logprobs.tolist()
+            for t, lp in zip(tasks, logprobs_list):
+                t.output_logprobs.append(float(lp))
+            return list(zip(tokens_list, logprobs_list))
+
         return sample(
             logits,
             temperature=temperatures,
             top_k=top_ks,
             top_p=top_ps,
+            frequency_penalty=freq_penalties,
+            input_ids=padded_ids,
+            input_mask=padded_mask,
         ).tolist()

astrai/inference/core/scheduler.py

@@ -1,5 +1,6 @@
 import logging
 import threading
+import uuid
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch
@@ -31,26 +32,26 @@ class InferenceScheduler:
 
         if max_seq_len is not None:
             self.max_seq_len = max_seq_len
-        elif config.max_len is not None:
+        elif config.max_position_embeddings is not None:
-            self.max_seq_len = config.max_len
+            self.max_seq_len = config.max_position_embeddings
         else:
             raise ValueError(
                 "max_seq_len must be provided either as argument "
-                "or in model config (config.max_len)"
+                "or in model config (config.max_position_embeddings)"
             )
         self.device = device or next(model.parameters()).device
         self.dtype = dtype or next(model.parameters()).dtype
 
-        head_dim = config.dim // config.n_heads
+        head_dim = config.hidden_size // config.num_attention_heads
 
         if cache is not None:
             self._cache = cache
         else:
             self._cache = ContiguousCache(
-                config.n_layers,
+                config.num_hidden_layers,
                 max_batch_size,
                 self.max_seq_len,
-                config.n_kv_heads,
+                config.num_key_value_heads,
                 head_dim,
                 self.device,
                 self.dtype,
@@ -138,36 +139,33 @@ class InferenceScheduler:
                                 t.task_id, t.prompt_ids, start_logical_page
                             )
 
-                pos_groups: Dict[int, List[Task]] = {}
+                decode_tasks = self._task_mgr.get_active_tasks()
-                for t in self._task_mgr.get_active_tasks():
-                    pos_groups.setdefault(t.next_pos, []).append(t)
 
-                for next_pos in sorted(pos_groups.keys()):
+                valid: List[Task] = []
-                    group = sorted(pos_groups[next_pos], key=lambda t: t.task_id)
+                for t in sorted(decode_tasks, key=lambda t: t.task_id):
+                    if cache.task_extend(t.task_id, t.next_pos):
+                        valid.append(t)
+                    else:
+                        t.status = TaskStatus.ABORTED
+                        self._task_mgr.invoke_callback(t.task_id, STOP)
 
-                    valid: List[Task] = []
+                if valid:
-                    for t in group:
+                    next_tokens = self._executor.execute_decode(valid)
-                        if cache.task_extend(t.task_id, t.next_pos):
+
-                            valid.append(t)
+                    for t, ntok in zip(valid, next_tokens):
-                        else:
+                        t.output_ids.append(ntok)
-                            t.status = TaskStatus.ABORTED
+                        t.output_tokens += 1
+                        new_text = t.decode_new_token(self._task_mgr.tokenizer)
+                        if new_text:
+                            self._task_mgr.invoke_callback(t.task_id, new_text)
+
+                    for t in valid:
+                        if t.is_finished(stop_ids):
+                            remaining = t.flush_remaining(self._task_mgr.tokenizer)
+                            if remaining:
+                                self._task_mgr.invoke_callback(t.task_id, remaining)
                             self._task_mgr.invoke_callback(t.task_id, STOP)
 
-                    if valid:
-                        next_tokens = self._executor.execute_decode(valid)
-
-                        for t, ntok in zip(valid, next_tokens):
-                            t.output_ids.append(ntok)
-                            t.output_tokens += 1
-                            self._task_mgr.invoke_callback(
-                                t.task_id,
-                                self._task_mgr.tokenizer.decode([ntok]),
-                            )
-
-                        for t in valid:
-                            if t.is_finished(stop_ids):
-                                self._task_mgr.invoke_callback(t.task_id, STOP)
-
         except Exception as e:
             self._stop_event.set()
             logger.error(f"Scheduler loop crashed: {e}", exc_info=True)
@@ -197,6 +195,117 @@ class InferenceScheduler:
             self._cache.task_free(task.task_id)
         for task in self._task_mgr.get_waiting_tasks():
             self._task_mgr.invoke_callback(task.task_id, STOP)
+            self._cache.task_free(task.task_id)
         self._task_mgr.clear_queues()
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
+
+    def run_batch(
+        self,
+        prompt_ids_list: List[List[int]],
+        *,
+        max_tokens: Optional[int] = None,
+        temperature: float = 1.0,
+        top_p: float = 1.0,
+        top_k: int = 50,
+        frequency_penalty: float = 0.0,
+        rep_window: int = 64,
+        return_logprobs: bool = False,
+    ) -> List[List[int]]:
+        """Synchronous batch generation without the scheduler thread.
+
+        Accepts already-tokenized prompts (no string round-trip) and runs
+        prefill + decode to completion on the calling thread.  Designed for
+        RL rollout, where logprobs of the behaviour policy must be collected
+        alongside generated tokens.
+
+        Args:
+            prompt_ids_list: ``B`` prompts, each a list of token IDs.
+            max_tokens: Maximum tokens to generate per prompt.  ``None``
+                uses ``self.max_seq_len - len(prompt_ids)``.
+            temperature/top_p/top_k/frequency_penalty/rep_window: Sampling
+                parameters (uniform across the batch).
+            return_logprobs: If ``True``, return ``(token_ids, logprobs)``
+                tuples per prompt (logprobs aligned 1-to-1 with token_ids).
+
+        Returns:
+            ``List[List[int]]`` of generated token IDs per prompt, or —
+            when ``return_logprobs`` is ``True`` —
+            ``List[Tuple[List[int], List[float]]]``.
+        """
+        stop_ids = self._task_mgr.tokenizer.stop_ids
+        cache = self._cache
+        seq_cap = self.max_seq_len
+
+        tasks: List[Task] = []
+        for ids in prompt_ids_list:
+            if len(ids) >= seq_cap:
+                tasks.append(None)
+                continue
+            t_max = max_tokens
+            if t_max is None:
+                t_max = seq_cap - len(ids)
+            else:
+                t_max = min(t_max, seq_cap - len(ids))
+            task = Task(
+                task_id=f"batch_{uuid.uuid4().hex[:8]}",
+                prompt_ids=list(ids),
+                max_tokens=t_max,
+                temperature=temperature,
+                top_p=top_p,
+                top_k=top_k,
+                frequency_penalty=frequency_penalty,
+                rep_window=rep_window,
+            )
+            if not cache.task_alloc(task.task_id, task.prompt_ids):
+                tasks.append(None)
+                continue
+            task.input_tokens = len(task.prompt_ids)
+            tasks.append(task)
+
+        try:
+            live = [t for t in tasks if t is not None]
+            prefill_groups: Dict[Tuple[int, int], List[Task]] = {}
+            for t in live:
+                key = (len(t.prompt_ids), cache.task_cached(t.task_id))
+                prefill_groups.setdefault(key, []).append(t)
+            for (prompt_len, start_pos), group in prefill_groups.items():
+                self._executor.execute_prefill(group, prompt_len, start_pos)
+
+            while live:
+                valid: List[Task] = []
+                for t in sorted(live, key=lambda x: x.task_id):
+                    if cache.task_extend(t.task_id, t.next_pos):
+                        valid.append(t)
+                    else:
+                        t.status = TaskStatus.ABORTED
+                if not valid:
+                    break
+
+                step_out = self._executor.execute_decode(
+                    valid, return_logprobs=return_logprobs
+                )
+                if return_logprobs:
+                    for t, (ntok, _lp) in zip(valid, step_out):
+                        t.output_ids.append(ntok)
+                        t.output_tokens += 1
+                else:
+                    for t, ntok in zip(valid, step_out):
+                        t.output_ids.append(ntok)
+                        t.output_tokens += 1
+
+                live = [t for t in valid if not t.is_finished(stop_ids)]
+        finally:
+            for t in tasks:
+                if t is not None:
+                    cache.task_free(t.task_id)
+
+        results: List[Any] = []
+        for t in tasks:
+            if t is None:
+                results.append(([], []) if return_logprobs else [])
+            elif return_logprobs:
+                results.append((list(t.output_ids), list(t.output_logprobs)))
+            else:
+                results.append(list(t.output_ids))
+        return results

astrai/inference/core/task.py

@@ -13,6 +13,40 @@ logger = logging.getLogger(__name__)
 STOP = object()
 
 
+class StreamDecoder:
+    """Incremental decoder for byte-level BPE streaming.
+
+    Byte-level BPE may split a single Unicode character (e.g. em-dash,
+    smart quotes) across multiple tokens. Decoding such a token in
+    isolation produces U+FFFD (replacement char). This decoder
+    accumulates token IDs and only emits text once the trailing
+    characters are complete, buffering incomplete multi-byte sequences
+    until the next token arrives.
+    """
+
+    __slots__ = ("_tokenizer", "_ids", "_emitted")
+
+    def __init__(self, tokenizer: AutoTokenizer):
+        self._tokenizer = tokenizer
+        self._ids: List[int] = []
+        self._emitted: str = ""
+
+    def push(self, token_id: int) -> str:
+        """Append a token ID and return newly completed text.
+
+        Returns "" while a multi-byte character is still incomplete.
+        """
+        self._ids.append(token_id)
+        full = self._tokenizer.decode(self._ids, skip_special_tokens=True)
+        if full.endswith("\ufffd"):
+            return ""
+        if len(full) > len(self._emitted):
+            diff = full[len(self._emitted) :]
+            self._emitted = full
+            return diff
+        return ""
+
+
 class TaskStatus(Enum):
     """Task lifecycle states."""
 
@@ -33,6 +67,8 @@ 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
@@ -40,13 +76,44 @@ class Task:
         self.temperature = temperature
         self.top_p = top_p
         self.top_k = top_k
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
 
         self.status = TaskStatus.PENDING
         self.output_ids: List[int] = []
+        self.output_logprobs: List[float] = []
         self.input_tokens: int = 0
         self.output_tokens: int = 0
         self.arrival_time = time.time()
         self.finish_time: Optional[float] = None
+        self._decoder: Optional[StreamDecoder] = None
+
+    def decode_new_token(self, tokenizer: AutoTokenizer) -> str:
+        """Decode the last appended output token, buffering incomplete
+        multi-byte sequences across calls.
+
+        Lazily creates a :class:`StreamDecoder` on first use.
+        """
+        if self._decoder is None:
+            self._decoder = StreamDecoder(tokenizer)
+        return self._decoder.push(self.output_ids[-1])
+
+    def flush_remaining(self, tokenizer: AutoTokenizer) -> str:
+        """Emit any text still buffered in the decoder.
+
+        Called when generation terminates (max_tokens reached, stop
+        sequence, or external removal) to avoid dropping a final
+        incomplete-looking fragment that is actually complete when
+        adjacent to the stop token.
+        """
+        if self._decoder is None or not self.output_ids:
+            return ""
+        full = tokenizer.decode(self.output_ids, skip_special_tokens=True)
+        if len(full) > len(self._decoder._emitted):
+            diff = full[len(self._decoder._emitted) :]
+            self._decoder._emitted = full
+            return diff
+        return ""
 
     @property
     def next_pos(self) -> int:
@@ -92,6 +159,8 @@ 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 +185,8 @@ 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

@@ -74,20 +74,31 @@ 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):
             raise ValueError("top_k must be a non-negative integer")
         if not (0.0 <= top_p <= 1.0):
             raise ValueError("top_p must be a float between 0.0 and 1.0")
-        if not (isinstance(temperature, (int, float)) and temperature > 0):
+        if not (isinstance(temperature, (int, float)) and temperature >= 0):
-            raise ValueError("temperature must be a positive number")
+            raise ValueError("temperature must be a non-negative number")
+        if not (
+            isinstance(frequency_penalty, (int, float))
+            and -2.0 <= frequency_penalty <= 2.0
+        ):
+            raise ValueError("frequency_penalty must be between -2.0 and 2.0")
+        if not (isinstance(rep_window, int) and rep_window > 0):
+            raise ValueError("rep_window must be a positive integer")
 
         self.messages = messages
         self.top_k = top_k
         self.top_p = top_p
         self.temperature = temperature
         self.max_tokens = max_tokens
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
         self.stream = stream
 
 
@@ -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.
 
@@ -145,25 +252,76 @@ class SamplingPipeline(BaseSamplingStrategy):
     def __init__(self, strategies: List[BaseSamplingStrategy]):
         self.strategies = strategies
 
-    def apply(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def apply(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+    ) -> Tensor:
         for strategy in self.strategies:
-            logits = strategy.apply(logits, filter_value)
+            logits = strategy.apply(logits, filter_value, input_ids, input_mask)
         return logits
 
-    @torch.no_grad()
+    @staticmethod
-    def sample(self, logits: Tensor, filter_value: float = -float("inf")) -> Tensor:
+    def _is_greedy(temperature: Union[float, Tensor]) -> bool:
+        if isinstance(temperature, Tensor):
+            return temperature.numel() == 1 and temperature.item() == 0
+        return temperature == 0
+
+    @torch.inference_mode()
+    def sample(
+        self,
+        logits: Tensor,
+        filter_value: float = -float("inf"),
+        input_ids: Optional[Tensor] = None,
+        input_mask: Optional[Tensor] = None,
+        return_logprobs: bool = False,
+    ):
         """Apply strategies then sample (softmax + multinomial).
 
+        Short-circuits to ``argmax`` when temperature is exactly 0
+        (deterministic / greedy decode).
+
         Args:
             logits: Raw logits ``[batch, vocab_size]``.
+            input_ids: Previously generated token IDs ``[batch, seq_len]``.
+            input_mask: Boolean mask for ``input_ids`` padding.
+            return_logprobs: If ``True``, return ``(tokens, logprobs)``
+                where ``logprobs[i]`` is the log-probability of
+                ``tokens[i]`` under the (post-strategy) sampling
+                distribution.
 
         Returns:
-            Sampled token IDs ``[batch]``.
+            Sampled token IDs ``[batch]``, or — when ``return_logprobs``
+            is ``True`` — a ``(token_ids, chosen_logprobs)`` tuple.
         """
-        return torch.multinomial(
+        if self._is_greedy_pipeline():
-            torch.softmax(self.apply(logits, filter_value), dim=-1),
+            tokens = logits.argmax(dim=-1)
-            num_samples=1,
+            if not return_logprobs:
+                return tokens
+            log_probs = torch.log_softmax(logits.float(), dim=-1)
+            chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
+            return tokens, chosen
+
+        transformed = self.apply(logits, filter_value, input_ids, input_mask)
+        log_probs = torch.log_softmax(transformed.float(), dim=-1)
+        tokens = torch.multinomial(
+            torch.softmax(transformed, dim=-1), num_samples=1
         ).squeeze(-1)
+        if not return_logprobs:
+            return tokens
+        chosen = torch.gather(log_probs, -1, tokens.unsqueeze(-1)).squeeze(-1)
+        return tokens, chosen
+
+    def _is_greedy_pipeline(self) -> bool:
+        """True if the first strategy is greedy temperature (temp=0)."""
+        if not self.strategies:
+            return False
+        first = self.strategies[0]
+        return isinstance(first, TemperatureStrategy) and self._is_greedy(
+            first.temperature
+        )
 
 
 @torch.inference_mode()
@@ -172,22 +330,46 @@ def sample(
     temperature: Union[float, Tensor] = 1.0,
     top_k: Union[int, Tensor] = 0,
     top_p: Union[float, Tensor] = 1.0,
+    frequency_penalty: Union[float, Tensor] = 0.0,
+    input_ids: Optional[Tensor] = None,
+    input_mask: Optional[Tensor] = None,
     filter_value: float = -float("inf"),
-) -> Tensor:
+    return_logprobs: bool = False,
+):
     """Apply sampling strategies then sample (softmax + multinomial).
 
-    Shortcut for ``SamplingPipeline(...).sample(logits)``.
+    Shortcut for ``SamplingPipeline(...).sample(logits, return_logprobs=)``.
+
+    When **temperature** is exactly 0 (scalar or single-element tensor)
+    the function short-circuits to ``argmax`` for deterministic decode.
 
     Args:
         logits: Raw logits ``[batch, vocab_size]``.
+        frequency_penalty: Penalty per occurrence for repeated tokens
+            (0.0 disables, range -2.0~2.0).
+        input_ids: Previously generated token IDs ``[batch, seq_len]``.
+        input_mask: Boolean mask for ``input_ids`` padding.
+        return_logprobs: If ``True``, also return the log-probability
+            of each sampled token under the (post-strategy) sampling
+            distribution — useful for RL rollout (PPO/GRPO importance
+            ratios).
 
     Returns:
-        Sampled token IDs ``[batch]``.
+        Sampled token IDs ``[batch]``, or — when ``return_logprobs`` is
+        ``True`` — a ``(token_ids, chosen_logprobs)`` tuple where
+        ``chosen_logprobs`` has shape ``[batch]``.
     """
     return SamplingPipeline(
         [
             TemperatureStrategy(temperature),
             TopKStrategy(top_k),
             TopPStrategy(top_p),
+            FrequencyPenaltyStrategy(frequency_penalty),
         ]
-    ).sample(logits, filter_value)
+    ).sample(
+        logits,
+        filter_value=filter_value,
+        input_ids=input_ids,
+        input_mask=input_mask,
+        return_logprobs=return_logprobs,
+    )

astrai/model/components/attention.py

@@ -76,9 +76,8 @@ class GQA(nn.Module):
         rotary_emb: Tensor,
         attn_mask: Tensor = None,
         paged_cache: Optional[CacheView] = None,
+        is_causal: bool = False,
     ) -> Tensor:
-        is_causal = attn_mask is None
-
         q = self._split_heads(self.q_proj(x), self.n_heads)
         k = self._split_heads(self.k_proj(x), self.n_kv_heads)
         v = self._split_heads(self.v_proj(x), self.n_kv_heads)
@@ -163,9 +162,9 @@ class MLA(nn.Module):
         rotary_emb: Tensor,
         attn_mask: Tensor = None,
         paged_cache: Optional[CacheView] = None,
+        is_causal: bool = False,
     ) -> Tensor:
         bsz, seq_len, _ = x.size()
-        is_causal = attn_mask is None
 
         q = self.q_proj(x)
         q = q.view(bsz, seq_len, self.n_heads, self.head_dim)

astrai/model/components/decoder_block.py

@@ -14,10 +14,18 @@ class DecoderBlock(nn.Module):
     def __init__(self, config, layer_id: int):
         super().__init__()
         cfg = asdict(config)
-        cfg["down_init_std"] = 0.02 / (2 * config.n_layers) ** 0.5
+        cfg.update(
+            dim=config.hidden_size,
+            dim_ffn=config.intermediate_size,
+            n_layers=config.num_hidden_layers,
+            n_heads=config.num_attention_heads,
+            n_kv_heads=config.num_key_value_heads,
+            norm_eps=config.rms_norm_eps,
+            down_init_std=0.02 / (2 * config.num_hidden_layers) ** 0.5,
+        )
         self.attention = AttnFactory.create(config.attn_type, **cfg, layer_id=layer_id)
-        self.input_norm = RMSNorm(config.dim, config.norm_eps)
+        self.input_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
-        self.post_attention_norm = RMSNorm(config.dim, config.norm_eps)
+        self.post_attention_norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
         self.mlp = FFNFactory.create(config.ffn_type, **cfg)
 
     def forward(
@@ -26,12 +34,14 @@ class DecoderBlock(nn.Module):
         rotary_emb: Tensor,
         attention_mask: Optional[Tensor] = None,
         paged_cache: Optional[CacheView] = None,
+        is_causal: bool = False,
     ) -> Tensor:
         attn_output = self.attention(
             self.input_norm(x),
             rotary_emb,
             attention_mask,
             paged_cache,
+            is_causal,
         )
         x = attn_output + x
         x = self.mlp(self.post_attention_norm(x)) + x

astrai/model/components/lora.py

@@ -39,8 +39,12 @@ class LoRALinear(nn.Module):
 
         self.r = r
         self.scaling = alpha / r
-        self.lora_A = nn.Parameter(torch.randn(r, self.weight.shape[1]) / r)
+        device = self.weight.device
-        self.lora_B = nn.Parameter(torch.zeros(self.weight.shape[0], r))
+        dtype = self.weight.dtype
+        lora_a = torch.randn(r, self.weight.shape[1], device=device, dtype=dtype) / r
+        lora_b = torch.zeros(self.weight.shape[0], r, device=device, dtype=dtype)
+        self.lora_A = nn.Parameter(lora_a)
+        self.lora_B = nn.Parameter(lora_b)
         self._merged = False
 
     def forward(self, x):

astrai/model/encoder.py

@@ -18,20 +18,28 @@ class EmbeddingEncoder(AutoModel):
     def __init__(self, config: EncoderConfig):
         super().__init__(config)
         self.config = config
-        rope_dim = config.dim // config.n_heads
+        rope_dim = config.hidden_size // config.num_attention_heads
         rope_base = config.rope_theta if config.rope_theta is not None else 10000
         self.rotary_embedding = RotaryEmbedding(
-            rope_dim, config.max_len, rope_base, rope_scaling=config.rope_scaling
+            rope_dim,
+            config.max_position_embeddings,
+            rope_base,
+            rope_scaling=config.rope_scaling,
         )
         self.embed_tokens = Embedding(
-            config.vocab_size, config.dim, neftune_alpha=config.neftune_alpha
+            config.vocab_size,
+            config.hidden_size,
+            neftune_alpha=config.neftune_alpha,
         )
 
         self.layers = nn.ModuleList(
-            [DecoderBlock(config, layer_id) for layer_id in range(config.n_layers)]
+            [
+                DecoderBlock(config, layer_id)
+                for layer_id in range(config.num_hidden_layers)
+            ]
         )
 
-        self.norm = RMSNorm(config.dim, config.norm_eps)
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
 
         self.pooling_type = config.pooling_type or "mean"
         self.normalize_embeddings = config.normalize_embeddings or False
@@ -59,10 +67,10 @@ class EmbeddingEncoder(AutoModel):
         x = self.embed_tokens(input_ids)
 
         rotary_emb = self.rotary_embedding(x, position_ids)
-        attn_mask = process_attention_mask(x, position_ids, input_mask, is_causal=False)
+        attn_mask = process_attention_mask(input_mask)
 
         for layer in self.layers:
-            x = layer(x, rotary_emb, attn_mask, paged_cache=None)
+            x = layer(x, rotary_emb, attn_mask)
 
         hidden_states = self.norm(x)
 

astrai/model/transformer.py

@@ -15,32 +15,15 @@ from astrai.model.components.rope import RotaryEmbedding
 
 
 def process_attention_mask(
-    input_tensor: Tensor,
+    input_mask: Optional[Tensor],
-    position_ids: Optional[Tensor],
-    input_mask: Optional[Tensor] = None,
-    is_causal: bool = False,
 ) -> Optional[Tensor]:
-    if position_ids is None:
-        return None
-    if input_mask is not None and input_mask.dim() > 2:
-        return input_mask
-
-    device = input_tensor.device
-    B = input_tensor.size(0)
-    T = position_ids.max().item() + 1
-
     if input_mask is None:
-        if position_ids.min().item() == 0 and is_causal:
+        return None
-            return None
+    if input_mask.dim() == 2:
-        attend = torch.ones(B, 1, T, dtype=torch.bool, device=device)
+        return input_mask[:, None, None, :]
-    else:
+    if input_mask.dim() == 3:
-        attend = input_mask[:, :T].to(device=device, dtype=torch.bool).unsqueeze(1)
+        return input_mask[:, None, :, :]
-
+    return input_mask
-    if is_causal:
-        causal = position_ids.unsqueeze(-1) >= torch.arange(T, device=device)
-        attend = attend & causal
-
-    return attend.unsqueeze(1)
 
 
 @AutoModel.register("autoregressive_lm")
@@ -53,24 +36,32 @@ class AutoRegressiveLM(AutoModel):
         rope_dim = (
             config.qk_rope_head_dim
             if config.attn_type == "mla"
-            else config.dim // config.n_heads
+            else config.hidden_size // config.num_attention_heads
         )
         rope_base = config.rope_theta if config.rope_theta is not None else 10000
         self.rotary_embedding = RotaryEmbedding(
-            rope_dim, config.max_len, rope_base, rope_scaling=config.rope_scaling
+            rope_dim,
+            config.max_position_embeddings,
+            rope_base,
+            rope_scaling=config.rope_scaling,
         )
         self.embed_tokens = Embedding(
-            config.vocab_size, config.dim, neftune_alpha=config.neftune_alpha
+            config.vocab_size,
+            config.hidden_size,
+            neftune_alpha=config.neftune_alpha,
         )
 
         self.layers = nn.ModuleList(
-            [DecoderBlock(config, layer_id) for layer_id in range(config.n_layers)]
+            [
+                DecoderBlock(config, layer_id)
+                for layer_id in range(config.num_hidden_layers)
+            ]
         )
 
-        self.norm = RMSNorm(config.dim, config.norm_eps)
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
-        self.lm_head = Linear(config.dim, config.vocab_size)
+        self.lm_head = Linear(config.hidden_size, config.vocab_size)
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             self.lm_head.weight = self.embed_tokens.weight
 
         self.apply(self._init_weights)
@@ -85,7 +76,7 @@ class AutoRegressiveLM(AutoModel):
 
         state_dict = dict(state_dict)
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             # same tensor for embed and lm_head
             if embed_key in state_dict:
                 state_dict[lm_head_key] = state_dict[embed_key]
@@ -101,7 +92,7 @@ class AutoRegressiveLM(AutoModel):
             destination=destination, prefix=prefix, keep_vars=keep_vars
         )
 
-        if self.config.tie_weight is True:
+        if self.config.tie_word_embeddings is True:
             lm_head_key = prefix + "lm_head.weight"
             if lm_head_key in state_dict:
                 del state_dict[lm_head_key]
@@ -119,10 +110,11 @@ class AutoRegressiveLM(AutoModel):
 
         x = self.embed_tokens(input_ids)
         rotary_emb = self.rotary_embedding(x, position_ids)
-        attn_mask = process_attention_mask(x, position_ids, input_mask, is_causal=True)
+        attn_mask = process_attention_mask(input_mask)
+        use_sdpa_causal_mask = attn_mask is None
 
         for layer in self.layers:
-            x = layer(x, rotary_emb, attn_mask, paged_cache)
+            x = layer(x, rotary_emb, attn_mask, paged_cache, use_sdpa_causal_mask)
 
         hidden_states = self.norm(x)
         logits = self.lm_head(hidden_states)

astrai/parallel/__init__.py

@@ -4,6 +4,7 @@ from astrai.parallel.executor import (
     BaseExecutor,
     DDPExecutor,
     ExecutorFactory,
+    FSDP2Executor,
     FSDPExecutor,
     GradientState,
     NoneExecutor,
@@ -35,4 +36,5 @@ __all__ = [
     "NoneExecutor",
     "DDPExecutor",
     "FSDPExecutor",
+    "FSDP2Executor",
 ]

astrai/parallel/executor.py

@@ -4,16 +4,22 @@ import contextlib
 import logging
 import os
 from contextlib import contextmanager
-from typing import Optional, Tuple
+from typing import Any, Callable, 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 (
+    FSDPModule,
+    FullStateDictConfig,
+    StateDictType,
+    fully_shard,
+)
 from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+from torch.distributed.tensor import DTensor
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
-from torch.utils.data import DataLoader
 
 from astrai.factory import BaseFactory
 from astrai.parallel.setup import get_rank, get_world_size
@@ -85,19 +91,25 @@ class BaseExecutor:
 
     def prepare(
         self,
-        model: nn.Module,
+        model_fn: Callable[[], nn.Module],
-        optimizer: Optional[Optimizer] = None,
+        optimizer_fn: Optional[Callable[[nn.Module], Optimizer]] = None,
-        dataloader: Optional[DataLoader] = None,
+        scheduler_fn: Optional[Callable[[Optimizer], LRScheduler]] = None,
-        scheduler: Optional[LRScheduler] = None,
+        before_wrap: Optional[Callable[[nn.Module], nn.Module]] = None,
-    ) -> Tuple[
+    ) -> Tuple[nn.Module, Optional[Optimizer], Optional[LRScheduler]]:
-        nn.Module, Optional[Optimizer], Optional[DataLoader], Optional[LRScheduler]
+        model = model_fn()
-    ]:
+        if before_wrap is not None:
+            model = before_wrap(model)
         model = self._prepare_model(model)
-        if optimizer is not None:
+        optimizer = None
+        scheduler = None
+        if optimizer_fn is not None:
+            optimizer = optimizer_fn(model)
+            if scheduler_fn is not None:
+                scheduler = scheduler_fn(optimizer)
             optimizer = AccumOptimizer(optimizer, self.gradient_state)
-        if scheduler is not None:
+            if scheduler is not None:
-            scheduler = AccumScheduler(scheduler, self.gradient_state)
+                scheduler = AccumScheduler(scheduler, self.gradient_state)
-        return model, optimizer, dataloader, scheduler
+        return model, optimizer, scheduler
 
     def _prepare_model(self, model: nn.Module) -> nn.Module:
         return model
@@ -120,6 +132,21 @@ 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
@@ -279,8 +306,102 @@ 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()
 
         return model.state_dict()
+
+
+@ExecutorFactory.register("fsdp2")
+class FSDP2Executor(BaseExecutor):
+    """FSDP2 executor using `torch.distributed.fsdp.fully_shard` (per-module API).
+
+    Wraps each child module individually via ``fully_shard``.
+    Skips the root model because ``ABC + Generic[T]`` in the MRO makes
+    FSDP2's dynamic ``__class__`` assignment fail at the CPython level.
+    Original ``Parameter`` objects are preserved (as DTensors) — no
+    ``FlatParameter``, no ``use_orig_params=True`` hack.
+    """
+
+    def __init__(
+        self,
+        grad_accum_steps: int = 1,
+        mesh: Optional[Any] = None,
+        mp_policy: Optional[Any] = None,
+        reshard_after_forward: bool = True,
+    ):
+        super().__init__(grad_accum_steps=grad_accum_steps)
+        self._mesh = mesh
+        self._mp_policy = mp_policy
+        self._reshard_after_forward = reshard_after_forward
+
+    def _prepare_model(self, model: nn.Module) -> nn.Module:
+        if not self.use_distributed:
+            logger.warning("FSDP2 backend selected but world_size=1, model not wrapped")
+            return model
+
+        kwargs = dict(
+            mesh=self._mesh,
+            mp_policy=self._mp_policy,
+            reshard_after_forward=self._reshard_after_forward,
+        )
+        kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+        for child in model.children():
+            if isinstance(child, nn.ModuleList):
+                for sub in child:
+                    fully_shard(sub, **kwargs)
+            else:
+                fully_shard(child, **kwargs)
+
+        logger.info(
+            "FSDP2 wrapping applied to %d direct children (root skipped for ABC compat)",
+            len(list(model.children())),
+        )
+        return model
+
+    @contextmanager
+    def _no_sync(self, model: nn.Module):
+        fsdp_modules = [m for m in model.modules() if isinstance(m, FSDPModule)]
+        if fsdp_modules:
+            for m in fsdp_modules:
+                m.set_requires_gradient_sync(False, recurse=True)
+            try:
+                yield
+            finally:
+                for m in fsdp_modules:
+                    m.set_requires_gradient_sync(True, recurse=True)
+        else:
+            yield
+
+    def clip_grad_norm(self, model: nn.Module, max_norm: float) -> float:
+        if self.use_distributed:
+            total_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
+            if isinstance(total_norm, torch.Tensor):
+                return total_norm.item()
+            return total_norm
+        return super().clip_grad_norm(model, max_norm)
+
+    def unwrap_model(self, model: nn.Module):
+        if not self.use_distributed:
+            return model.state_dict()
+
+        if get_rank() != 0:
+            return None
+
+        for module in model.modules():
+            if isinstance(module, FSDPModule):
+                module.unshard()
+
+        state_dict = model.state_dict()
+        result = {
+            k: (v.full_tensor() if isinstance(v, DTensor) else v)
+            for k, v in state_dict.items()
+        }
+
+        for module in model.modules():
+            if isinstance(module, FSDPModule):
+                module.reshard()
+
+        return result

astrai/parallel/setup.py

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

astrai/parallel/signal_handler.py

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

astrai/preprocessing/__init__.py

@@ -8,12 +8,14 @@ from astrai.preprocessing.builder import (
 from astrai.preprocessing.packing import (
     PackingStrategy,
     PackingStrategyFactory,
+    plan_bfd,
 )
 from astrai.preprocessing.pipeline import Pipeline, filter_by_length
 from astrai.preprocessing.position_id import (
     PositionIdStrategy,
     PositionIdStrategyFactory,
 )
+from astrai.preprocessing.transform import TokenizeTransform
 from astrai.preprocessing.writer import (
     StoreWriter,
     StoreWriterFactory,
@@ -32,5 +34,7 @@ __all__ = [
     "SingleOutputMaskBuilder",
     "StoreWriter",
     "StoreWriterFactory",
+    "TokenizeTransform",
     "filter_by_length",
+    "plan_bfd",
 ]

astrai/preprocessing/builder.py

@@ -94,9 +94,107 @@ class SectionRenderer:
 
         return all_ids, loss_mask
 
+    def process_sections_batch(
+        self,
+        items: list[dict],
+        sections: list,
+        config,
+        tokenizer,
+        *,
+        is_top_level=False,
+        filter_text=True,
+    ):
+        """Render and tokenize a group of records with batched Rust tokenization."""
+        has_template = any(s.get("template") for s in sections)
+        is_text_config = not has_template and all(
+            s["action"] == "train" for s in sections
+        )
+        plans: list[list[tuple[str, str, bool]]] = []
+
+        for item in items:
+            plan: list[tuple[str, str, bool]] = []
+            first_section = True
+            for sec in sections:
+                field = sec["field"]
+                action = sec["action"]
+                use_template = sec.get("template", False)
+                add_special = sec.get(
+                    "add_special_tokens", not use_template and first_section
+                )
+
+                if use_template:
+                    messages = item.get(field)
+                    if not isinstance(messages, list) or not messages:
+                        continue
+                    for msg in messages:
+                        role = msg.get("role", "")
+                        rendered = tokenizer.apply_chat_template(
+                            [msg], tokenize=False, add_generation_prompt=False
+                        )
+                        plan.append(
+                            (rendered, _resolve_action(action, role, config), False)
+                        )
+                else:
+                    text = str(item.get(field, ""))
+                    if not text.strip():
+                        continue
+                    if is_text_config and filter_text:
+                        pp = config.preprocessing
+                        if pp.min_chars > 0 and len(text) < pp.min_chars:
+                            continue
+                        if len(text) > pp.max_chars:
+                            continue
+                    plan.append((text, action, add_special))
+
+                first_section = False
+            plans.append(plan)
+
+        encoded: dict[tuple[int, int], list[int]] = {}
+        for add_special in (False, True):
+            refs = [
+                (item_idx, unit_idx, text)
+                for item_idx, plan in enumerate(plans)
+                for unit_idx, (text, _, add) in enumerate(plan)
+                if add == add_special
+            ]
+            if not refs:
+                continue
+            ids_batch = tokenizer.encode(
+                [text for _, _, text in refs], add_special_tokens=add_special
+            )
+            for (item_idx, unit_idx, _), ids in zip(refs, ids_batch):
+                encoded[(item_idx, unit_idx)] = ids
+
+        outputs = []
+        max_len = config.preprocessing.max_seq_len
+        for item_idx, plan in enumerate(plans):
+            all_ids = []
+            loss_mask = []
+            if is_top_level and has_template and tokenizer.bos_token_id is not None:
+                all_ids.append(tokenizer.bos_token_id)
+                loss_mask.append(0)
+            for unit_idx, (_, action, _) in enumerate(plan):
+                ids = encoded[(item_idx, unit_idx)]
+                all_ids.extend(ids)
+                loss_mask.extend([1 if action == "train" else 0] * len(ids))
+            all_ids = all_ids[:max_len]
+            loss_mask = loss_mask[: len(all_ids)]
+            if not all_ids or (is_top_level and has_template and len(all_ids) <= 1):
+                outputs.append((None, None))
+            else:
+                outputs.append((all_ids, loss_mask))
+        return outputs
+
     def process_list_field(self, item: dict, sections: list, config, tokenizer):
-        all_ids: list[int] = []
+        """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"]
@@ -108,17 +206,13 @@ class SectionRenderer:
                 continue
 
             for val in values:
+                ids: list[int] = []
+                mask: list[int] = []
                 if use_template:
                     if isinstance(val, list):
                         wrapper = {field: val}
                         self._append_template(
-                            wrapper,
+                            wrapper, field, action, tokenizer, config, ids, mask
-                            field,
-                            action,
-                            tokenizer,
-                            config,
-                            all_ids,
-                            loss_mask,
                         )
                 else:
                     wrapper = {field: str(val)}
@@ -130,17 +224,55 @@ class SectionRenderer:
                         False,
                         False,
                         config,
-                        all_ids,
+                        ids,
-                        loss_mask,
+                        mask,
                     )
+                if ids:
+                    max_len = config.preprocessing.max_seq_len
+                    ids = ids[:max_len]
+                    mask = mask[: len(ids)]
+                    per_item_ids.append(ids)
+                    per_item_masks.append(mask)
 
-        max_len = config.preprocessing.max_seq_len
+        if not per_item_ids:
-        all_ids = all_ids[:max_len]
-        loss_mask = loss_mask[: len(all_ids)]
-
-        if not all_ids:
             return None, None
-        return all_ids, loss_mask
+        return per_item_ids, per_item_masks
+
+    def process_list_field_batch(self, items, sections, config, tokenizer):
+        per_item_ids = [[] for _ in items]
+        per_item_masks = [[] for _ in items]
+
+        for sec in sections:
+            wrappers = []
+            owners = []
+            field = sec["field"]
+            for item_idx, item in enumerate(items):
+                values = item.get(field)
+                if not isinstance(values, list):
+                    continue
+                for val in values:
+                    if sec.get("template", False) and not isinstance(val, list):
+                        continue
+                    wrappers.append({field: val if isinstance(val, list) else str(val)})
+                    owners.append(item_idx)
+
+            rendered = self.process_sections_batch(
+                wrappers,
+                [sec],
+                config,
+                tokenizer,
+                is_top_level=False,
+                filter_text=False,
+            )
+            for owner, (ids, mask) in zip(owners, rendered):
+                if ids:
+                    per_item_ids[owner].append(ids)
+                    per_item_masks[owner].append(mask)
+
+        return [
+            (ids, masks) if ids else (None, None)
+            for ids, masks in zip(per_item_ids, per_item_masks)
+        ]
 
     @staticmethod
     def is_value_section(sections: list) -> bool:
@@ -209,6 +341,9 @@ class BaseMaskBuilder(ABC):
     @abstractmethod
     def build(self, item: dict, config, tokenizer) -> Optional[dict]: ...
 
+    def build_batch(self, items: list[dict], config, tokenizer) -> list[Optional[dict]]:
+        return [self.build(item, config, tokenizer) for item in items]
+
 
 class MaskBuilderFactory(BaseFactory["BaseMaskBuilder"]):
     pass
@@ -243,6 +378,27 @@ class SingleOutputMaskBuilder(BaseMaskBuilder):
             result["loss_mask"] = mask
         return result
 
+    def build_batch(self, items, config, tokenizer):
+        sections = config.input.sections
+        if not sections:
+            return [None] * len(items)
+        rendered = self.renderer.process_sections_batch(
+            items, sections, config, tokenizer, is_top_level=True
+        )
+        results = []
+        for item, (ids, mask) in zip(items, rendered):
+            if ids is None:
+                results.append(None)
+                continue
+            result = {
+                "sequence": ids,
+                "domain": _extract_domain(item, config.output.domain_key),
+            }
+            if not all(m == 1 for m in mask):
+                result["loss_mask"] = mask
+            results.append(result)
+        return results
+
 
 @MaskBuilderFactory.register("multi")
 class MultiOutputMaskBuilder(BaseMaskBuilder):
@@ -282,10 +438,18 @@ class MultiOutputMaskBuilder(BaseMaskBuilder):
                 ids, mask = self.renderer.process_list_field(
                     item, sections, config, tokenizer
                 )
-            else:
+                if ids is None:
-                ids, mask = self.renderer.process_sections(
+                    continue
-                    item, sections, config, tokenizer, is_top_level=True
+                # ids is List[List[int]] — preserve per-response structure
-                )
+                result[output_key] = ids
+                if mask is not None:
+                    result[mask_key] = mask
+                any_output = True
+                continue
+
+            ids, mask = self.renderer.process_sections(
+                item, sections, config, tokenizer, is_top_level=True
+            )
 
             if ids is None:
                 continue
@@ -304,6 +468,49 @@ class MultiOutputMaskBuilder(BaseMaskBuilder):
         result["domain"] = _extract_domain(item, config.output.domain_key)
         return result
 
+    def build_batch(self, items, config, tokenizer):
+        sources_spec = getattr(config.input, "sources", None)
+        if not sources_spec:
+            return [None] * len(items)
+
+        results = [{} for _ in items]
+        for output_key, spec in sources_spec.items():
+            sections = spec.get("sections", [])
+            if not sections:
+                continue
+            if self.renderer.is_value_section(sections):
+                for item, result in zip(items, results):
+                    value = self.renderer.extract_raw_value(item, sections)
+                    if value is not None:
+                        result[output_key] = value
+                continue
+
+            mask_key = spec.get("mask_key", f"{output_key}_mask")
+            if spec.get("list_field", False):
+                rendered = self.renderer.process_list_field_batch(
+                    items, sections, config, tokenizer
+                )
+            else:
+                rendered = self.renderer.process_sections_batch(
+                    items, sections, config, tokenizer, is_top_level=True
+                )
+
+            for result, (ids, mask) in zip(results, rendered):
+                if ids is None:
+                    continue
+                result[output_key] = ids
+                if spec.get("list_field", False) or not all(m == 1 for m in mask):
+                    result[mask_key] = mask
+                elif "mask_key" in spec:
+                    result[mask_key] = mask
+
+        return [
+            ({**result, "domain": _extract_domain(item, config.output.domain_key)})
+            if result
+            else None
+            for item, result in zip(items, results)
+        ]
+
 
 @MaskBuilderFactory.register("sectioned")
 class SectionedMaskBuilder(BaseMaskBuilder):
@@ -322,3 +529,9 @@ class SectionedMaskBuilder(BaseMaskBuilder):
         if sources_spec:
             return self._multi.build(item, config, tokenizer)
         return self._single.build(item, config, tokenizer)
+
+    def build_batch(self, items, config, tokenizer):
+        sources_spec = getattr(config.input, "sources", None)
+        if sources_spec:
+            return self._multi.build_batch(items, config, tokenizer)
+        return self._single.build_batch(items, config, tokenizer)

astrai/preprocessing/core.py

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

astrai/preprocessing/packing.py

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

astrai/preprocessing/pipeline.py

@@ -4,6 +4,10 @@ Composes a :class:`BaseMaskBuilder` (selected by ``input.type``) with
 sharding and flush to ``.h5`` / ``.bin`` storage.  Packing, position-id
 generation and storage writing are each delegated to pluggable strategies,
 dispatched by configuration keys.
+
+Record iteration, mask building, primary-id extraction and per-key
+accumulation are shared with :class:`TokenizeTransform` via the
+:mod:`astrai.preprocessing.core` helpers.
 """
 
 import json
@@ -17,11 +21,12 @@ import torch
 import tqdm
 
 from astrai.config.preprocess_config import PipelineConfig
-from astrai.preprocessing.builder import MaskBuilderFactory
+from astrai.preprocessing.core import (
+    build_preprocessing_components,
+    primary_ids,
+)
 from astrai.preprocessing.packing import PackingStrategyFactory
-from astrai.preprocessing.position_id import PositionIdStrategyFactory
 from astrai.preprocessing.writer import StoreWriterFactory
-from astrai.tokenize import AutoTokenizer
 
 logger = logging.getLogger(__name__)
 
@@ -64,20 +69,21 @@ class Pipeline:
         self.output_dir = output_dir
         self.tokenizer_path = tokenizer_path
 
-        self.mask_builder = MaskBuilderFactory.create("sectioned")
+        self.tokenizer, self.mask_builder, self._position_id = (
+            build_preprocessing_components(config, tokenizer_path)
+        )
         self._packer = PackingStrategyFactory.create(
             config.preprocessing.packing_strategy
         )
-        self._position_id = PositionIdStrategyFactory.create(
-            config.output.position_ids_mode
-        )
         self._writer = StoreWriterFactory.create(config.output.storage_format)
 
     def transform(self, item: dict) -> Optional[dict]:
-        return self.mask_builder.build(item, self.config, self._tokenizer)
+        return self.mask_builder.build(item, self.config, self.tokenizer)
+
+    def transform_batch(self, items: list[dict]) -> list[Optional[dict]]:
+        return self.mask_builder.build_batch(items, self.config, self.tokenizer)
 
     def run(self):
-        self._tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_path)
         domains: dict = defaultdict(lambda: defaultdict(list))
         total_tokens = 0
         shard_idx: dict[str, int] = defaultdict(int)
@@ -85,59 +91,59 @@ class Pipeline:
 
         pp = self.config.preprocessing
 
-        for item in tqdm.tqdm(
+        progress = tqdm.tqdm(desc="Tokenizing", unit="docs", mininterval=0.5)
-            self._iter_items(), desc="Tokenizing", unit="docs", mininterval=0.5
+        stop = False
-        ):
+        for items in self._iter_batches(pp.batch_size):
-            if pp.max_items and count >= pp.max_items:
+            progress.update(len(items))
-                break
-
             try:
-                result = self.transform(item)
+                results = self.transform_batch(items)
             except Exception:
                 logger.warning(
-                    "Failed to process item #%d, skipping", count + 1, exc_info=True
+                    "Failed to process batch, retrying records individually",
+                    exc_info=True,
                 )
-                continue
+                results = []
-            if result is None:
+                for item in items:
-                continue
+                    try:
+                        results.append(self.transform(item))
+                    except Exception:
+                        logger.warning(
+                            "Failed to process item, skipping", exc_info=True
+                        )
+                        results.append(None)
 
-            domain = result.pop("domain", "__default__")
+            for result in results:
+                if pp.max_items and count >= pp.max_items:
+                    stop = True
+                    break
+                if result is None:
+                    continue
 
-            is_multi = bool(getattr(self.config.input, "sources", None))
+                domain = result.pop("domain", "__default__")
-            if is_multi:
+                ids = primary_ids(result)
-                ids = self._primary_ids(result)
+                if not ids:
-            else:
+                    continue
-                ids = result.pop("sequence")
-                result["sequence"] = ids
 
-            if not ids:
+                bucket = domains[domain]
-                continue
+                self._align_bucket(bucket, result, ids)
+                for key, val in result.items():
+                    bucket[key].append(val)
 
-            bucket = domains[domain]
+                count += 1
-            self._align_bucket(bucket, result, ids)
+                total_tokens += len(ids)
-            for key, val in result.items():
-                bucket[key].append(val)
 
-            count += 1
+                if total_tokens >= self.config.output.max_tokens_per_shard:
-            total_tokens += len(ids)
+                    self._flush(domains, shard_idx)
+                    domains.clear()
+                    total_tokens = 0
+            if stop:
+                break
 
-            if total_tokens >= self.config.output.max_tokens_per_shard:
+        progress.close()
-                self._flush(domains, shard_idx)
-                domains.clear()
-                total_tokens = 0
 
         if total_tokens > 0:
             self._flush(domains, shard_idx)
 
-    @staticmethod
-    def _primary_ids(result: dict) -> list:
-        """Return the first list-valued entry in *result* as the primary id
-        sequence for token counting."""
-        for val in result.values():
-            if isinstance(val, list) and val and isinstance(val[0], int):
-                return val
-        return []
-
     @staticmethod
     def _align_bucket(bucket: dict, result: dict, ids: list):
         """Pad previously-accumulated keys that are missing from *result*."""
@@ -149,11 +155,29 @@ class Pipeline:
     def _iter_items(self):
         for path in self.paths:
             with open(path, "r", encoding="utf-8") as f:
-                for line in f:
+                if path.endswith(".json"):
-                    line = line.strip()
+                    data = json.load(f)
-                    if not line:
+                    if isinstance(data, dict):
-                        continue
+                        yield data
-                    yield json.loads(line)
+                    elif isinstance(data, list):
+                        yield from data
+                else:
+                    for line in f:
+                        line = line.strip()
+                        if not line:
+                            continue
+                        yield json.loads(line)
+
+    def _iter_batches(self, batch_size: int):
+        batch_size = max(1, batch_size)
+        batch = []
+        for item in self._iter_items():
+            batch.append(item)
+            if len(batch) >= batch_size:
+                yield batch
+                batch = []
+        if batch:
+            yield batch
 
     def _flush(self, domains, shard_idx):
         for domain, keys in domains.items():
@@ -163,24 +187,12 @@ class Pipeline:
             original_sequences = keys.get("sequence", [])
             mode = self.config.output.position_ids_mode
 
-            if mode == "doc_reset" and original_sequences:
+            keys = self._inject_doc_reset_position_ids(keys, mode, original_sequences)
-                keys["position_ids"] = [list(range(len(s))) for s in original_sequences]
-
             keys = self._packer.apply(dict(keys), pp.max_packed_len, pp.truncation_mode)
-
+            tensors = self._to_tensors(keys)
-            tensors: Dict[str, List[torch.Tensor]] = {}
+            tensors = self._inject_continuous_position_ids(
-            for key, ids_list in keys.items():
+                tensors, mode, keys.get("sequence", [])
-                dt = _STR_TO_DTYPE.get(
+            )
-                    self.config.output.dtype.get(key, "int32"), torch.int32
-                )
-                tensors[key] = [
-                    torch.tensor(list(chain.from_iterable(ids_list)), dtype=dt)
-                ]
-
-            if mode == "continuous" and original_sequences:
-                pos_ids = self._position_id.generate(keys.get("sequence", []))
-                if pos_ids:
-                    tensors["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
 
             self._writer.save(self.output_dir, domain, idx, tensors)
             shard_idx[domain] = idx + 1
@@ -190,3 +202,76 @@ class Pipeline:
                 f"  saved {domain}/shard_{idx:04d}  "
                 f"({tensors[first_key][0].numel():,} tokens)"
             )
+
+    def _inject_doc_reset_position_ids(
+        self,
+        keys: Dict[str, list],
+        mode: str,
+        original_sequences: List[List[int]],
+    ) -> Dict[str, list]:
+        """Attach per-document position_ids before packing (``doc_reset``).
+
+        ``doc_reset`` position ids must enter the packer so that each
+        packed bin concatenates the per-doc ranges in bin order.  The
+        per-record structure ``[range(len(s)) for s in seqs]`` is required
+        by the packer (it concatenates per-record lists per bin); the
+        ``PositionIdStrategy.generate`` flattens, so it cannot be used
+        directly here — it is only consulted for the ``continuous``
+        post-packing path.
+        """
+        if mode != "doc_reset" or not original_sequences:
+            return keys
+        keys["position_ids"] = [list(range(len(s))) for s in original_sequences]
+        return keys
+
+    def _inject_continuous_position_ids(
+        self,
+        tensors: Dict[str, List[torch.Tensor]],
+        mode: str,
+        packed_sequences: List[List[int]],
+    ) -> Dict[str, List[torch.Tensor]]:
+        """Attach a single continuous position_ids tensor after packing.
+
+        ``continuous`` mode spans the whole shard (post-packing), so it
+        cannot participate in bin packing — it is computed from the
+        packed sequences and appended directly to the tensor dict.
+        """
+        if mode != "continuous" or not packed_sequences:
+            return tensors
+        pos_ids = self._position_id.generate(packed_sequences)
+        if pos_ids:
+            tensors["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
+        return tensors
+
+    def _to_tensors(self, keys: Dict[str, list]) -> Dict[str, List[torch.Tensor]]:
+        """Convert packed per-key id lists to tensors.
+
+        Honours ``config.output.dtype`` overrides per key; falls back to
+        ``int32``.  Handles three shapes (see
+        :func:`astrai.preprocessing.core.to_per_record_tensors` for the
+        equivalent online-path helper):
+        - ``List[int]`` per record → one tensor per record.
+        - ``List[List[int]]`` per record (GRPO responses/masks) → one tensor
+          per record, inner lists flattened.
+        - ``List[int]`` for the whole shard (pre-packed keys) → single tensor.
+        """
+        tensors: Dict[str, List[torch.Tensor]] = {}
+        for key, ids_list in keys.items():
+            dt = _STR_TO_DTYPE.get(
+                self.config.output.dtype.get(key, "int32"), torch.int32
+            )
+            if ids_list and isinstance(ids_list[0], list):
+                tensors[key] = [
+                    torch.tensor(
+                        list(chain.from_iterable(ids))
+                        if ids and isinstance(ids[0], list)
+                        else ids,
+                        dtype=dt,
+                    )
+                    for ids in ids_list
+                ]
+            else:
+                tensors[key] = [
+                    torch.tensor(list(chain.from_iterable(ids_list)), dtype=dt)
+                ]
+        return tensors

astrai/preprocessing/transform.py

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

astrai/serialization/__init__.py

@@ -19,6 +19,7 @@ from astrai.serialization.checkpoint import (
 )
 from astrai.serialization.dataset import (
     load_bin,
+    load_bin_offsets,
     load_h5,
     save_bin,
     save_h5,
@@ -37,6 +38,7 @@ __all__ = [
     "save_safetensors",
     "save_torch",
     "load_bin",
+    "load_bin_offsets",
     "load_h5",
     "save_bin",
     "save_h5",

astrai/serialization/checkpoint.py

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

astrai/serialization/dataset.py

@@ -3,7 +3,7 @@
 import json
 import os
 from pathlib import Path
-from typing import Dict, List
+from typing import Any, Dict, List, Optional
 
 import h5py
 import numpy as np
@@ -50,12 +50,43 @@ def load_h5(file_path: str, share_memory=True) -> Dict[str, List[Tensor]]:
     return tensor_group
 
 
-def save_bin(file_path: str, tensor_group: Dict[str, List[Tensor]]):
+def save_bin(
+    file_path: str,
+    tensor_group: Dict[str, List[Tensor]],
+    record_keys: Optional[List[str]] = None,
+):
+    """Save tensors as memory-mapped binary files.
+
+    When *record_keys* is provided, those keys are written with per-record
+    cumulative offsets in ``meta.json`` so that ``MmapStore.fetch_record``
+    can slice individual records from the concatenated binary without
+    cross-record concatenation.  Keys not in *record_keys* (e.g. SEQ
+    ``sequence``) are written as a single contiguous stream without
+    offsets, preserving backward compatibility.
+
+    Nested keys (``List[List[Tensor]]`` such as GRPO ``responses``) are
+    not supported in bin format — use H5 for those.
+    """
     os.makedirs(file_path, exist_ok=True)
+    record_keys = set(record_keys or [])
     meta = {}
     for key, tensors in tensor_group.items():
+        if tensors and isinstance(tensors[0], list):
+            raise ValueError(
+                f"Nested key '{key}' (List[List[Tensor]]) is not supported "
+                f"in bin format. Use H5 or JSONL storage instead."
+            )
         cat = torch.cat(tensors, dim=0)
-        meta[key] = {"shape": list(cat.shape), "dtype": str(cat.dtype).split(".")[-1]}
+        entry: Dict[str, Any] = {
+            "shape": list(cat.shape),
+            "dtype": str(cat.dtype).split(".")[-1],
+        }
+        if key in record_keys:
+            offsets = [0]
+            for t in tensors:
+                offsets.append(offsets[-1] + t.shape[0])
+            entry["offsets"] = offsets
+        meta[key] = entry
         np.asarray(cat.cpu().numpy()).tofile(os.path.join(file_path, f"{key}.bin"))
     with open(os.path.join(file_path, "meta.json"), "w") as f:
         json.dump(meta, f)
@@ -69,8 +100,24 @@ def load_bin(file_path: str) -> Dict[str, List[Tensor]]:
         arr = np.memmap(
             os.path.join(file_path, f"{key}.bin"),
             dtype=info["dtype"],
-            mode="r+",
+            mode="c",
             shape=tuple(info["shape"]),
         )
         segments[key] = [torch.from_numpy(arr)]
     return segments
+
+
+def load_bin_offsets(file_path: str) -> Dict[str, List[int]]:
+    """Read per-record cumulative offsets from ``meta.json``.
+
+    Returns an empty dict when no key has offsets (legacy bin files),
+    in which case record-mode access falls back to per-record segment
+    indexing (H5/JSONL layout).
+    """
+    with open(os.path.join(file_path, "meta.json"), "r") as f:
+        meta = json.load(f)
+    offsets: Dict[str, List[int]] = {}
+    for key, info in meta.items():
+        if "offsets" in info:
+            offsets[key] = info["offsets"]
+    return offsets

astrai/tokenize/__init__.py

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

astrai/tokenize/chat_template.py

@@ -1,3 +1,4 @@
+from functools import cached_property
 from typing import Any, Dict, List, Optional
 
 from jinja2 import Template
@@ -29,7 +30,19 @@ class ChatTemplate:
         self.description = description
         self.default_variables = default_variables or {}
         self.special_tokens = special_tokens or {}
-        self._compiled: Template = Template(template_str)
+
+    @cached_property
+    def _compiled(self) -> Template:
+        """Lazy-compiled Jinja2 template, cached on first access.
+
+        The compiled :class:`~jinja2.Template` holds a dynamically-generated
+        ``root`` render function whose ``__module__`` is ``None``; under
+        ``pickle`` it falls back to ``__main__`` and breaks ``spawn``-based
+        multiprocessing.  By deferring compilation to first access, the
+        default pickle protocol serialises only ``template_str``; each
+        worker rebuilds the cache on first render.
+        """
+        return Template(self.template_str)
 
     @classmethod
     def from_string(

astrai/tokenize/tokenizer.py

@@ -10,6 +10,12 @@ from tokenizers import Tokenizer
 
 from astrai.tokenize.chat_template import ChatTemplate
 
+Message = Dict[str, str]
+"""Single chat message with ``role`` and ``content`` keys."""
+
+Messages = List[Message]
+"""Single conversation — a list of messages."""
+
 
 class AutoTokenizer:
     """Base tokenizer class with automatic loading support"""
@@ -120,7 +126,16 @@ class AutoTokenizer:
         is_pretokenized: bool = False,
         add_special_tokens: bool = True,
     ) -> List:
-        """Encode text to tokens or token IDs."""
+        """Encode text to token IDs.
+
+        Accepts both single strings and batches:
+
+        - ``encode("hello")`` → ``[123, 456]``
+        - ``encode(["hello", "world"])`` → ``[[123, 456], [789]]``
+
+        Batches are tokenised in parallel via the Rust backend's
+        ``encode_batch`` (uses all available CPU cores).
+        """
         if self._tokenizer is None:
             raise RuntimeError(
                 "Tokenizer not initialized. Load or create a tokenizer first."
@@ -133,15 +148,13 @@ class AutoTokenizer:
                 add_special_tokens=add_special_tokens,
             )
             return encoded.ids if out_ids else encoded.tokens
-        else:
+
-            encoded_list = self._tokenizer.encode_batch(
+        encoded_list = self._tokenizer.encode_batch(
-                tokens,
+            tokens,
-                is_pretokenized=is_pretokenized,
+            is_pretokenized=is_pretokenized,
-                add_special_tokens=add_special_tokens,
+            add_special_tokens=add_special_tokens,
-            )
+        )
-            return [
+        return [encoded.ids if out_ids else encoded.tokens for encoded in encoded_list]
-                encoded.ids if out_ids else encoded.tokens for encoded in encoded_list
-            ]
 
     def decode(self, tokens: List[int], skip_special_tokens: bool = True) -> str:
         """Decode token IDs to text."""
@@ -164,7 +177,14 @@ class AutoTokenizer:
           - tokenizer.bos_token   → returns string
           - tokenizer.bos_token_id → returns corresponding integer ID
           - tokenizer.stop_ids → returns list of corresponding integer IDs for all special tokens
+
+        Internal/private attrs are not intercepted: during unpickle
+        ``__dict__`` is empty, so probing ``self._special_token_map``
+        would recurse infinitely.
         """
+        if key.startswith("_"):
+            raise AttributeError(key)
+
         # Handle stop_ids - return IDs for all special tokens
         if key == "stop_ids":
             stop_ids = []
@@ -220,45 +240,63 @@ class AutoTokenizer:
 
     def apply_chat_template(
         self,
-        messages: List[Dict[str, str]],
+        messages: Union[Messages, List[Messages]],
         system_prompt: Optional[str] = None,
         tokenize: bool = True,
         add_generation_prompt: bool = True,
         **kwargs,
-    ) -> Union[str, List[int]]:
+    ) -> Union[str, List[int], List[str], List[List[int]]]:
-        """
+        """Apply the chat template and optionally tokenize.
-        Apply the chat template to messages and optionally tokenize the result.
+
+        Accepts both single conversations and batches:
+
+        - ``apply_chat_template([msg1, msg2])`` → ``"..."`` or ``[ids]``
+        - ``apply_chat_template([[msg1, msg2], [msg3]])`` → ``["..", ".."]``
+          or ``[[ids], [ids]]``
+
+        Batches render each conversation list and tokenise all at once via
+        :meth:`encode` (``List[str]`` → Rust parallel ``encode_batch``).
 
         Args:
-            messages: List of message dicts with 'role' and 'content'.
+            messages: Single conversation (``Messages``) or batch of
-            system_prompt: Optional system prompt string (auto-converted to first message).
+                conversations (``BatchMessages``).
+            system_prompt: Optional system prompt prepended (single mode only).
             tokenize: Whether to return token IDs (True) or raw string (False).
-            add_generation_prompt: Whether to add the generation prompt (default: True).
+            add_generation_prompt: Whether to add the generation prompt.
-            **kwargs: Additional variables to pass to the template.
+            **kwargs: Additional template variables.
 
         Returns:
-            Either the rendered string or list of token IDs.
+            Single mode: ``str`` or ``List[int]``.
-
+            Batch mode: ``List[str]`` or ``List[List[int]]``.
-        Raises:
-            RuntimeError: If chat template is not set.
         """
         if self._chat_template is None:
             raise RuntimeError(
                 "Chat template not set. Use set_chat_template() to set a template first."
             )
 
-        # Auto-convert system_prompt to first message if provided
+        is_batch = bool(messages) and isinstance(messages[0], list)
+
+        if is_batch:
+            rendered = [
+                self._chat_template.render(
+                    messages=msgs,
+                    add_generation_prompt=add_generation_prompt,
+                    **kwargs,
+                )
+                for msgs in messages
+            ]
+            if tokenize:
+                return self.encode(rendered)  # List[str] → batch encode
+            return rendered
+
+        # Single conversation
         if system_prompt:
             messages = [{"role": "system", "content": system_prompt}] + list(messages)
-
-        # Render the template
         rendered = self._chat_template.render(
             messages=messages,
             add_generation_prompt=add_generation_prompt,
             **kwargs,
         )
-
         if tokenize:
             return self.encode(rendered)
-
         return rendered

astrai/trainer/rollout.py

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

astrai/trainer/strategy.py

@@ -9,6 +9,7 @@ import torch.nn.functional as F
 from torch import Tensor
 
 from astrai.factory import BaseFactory
+from astrai.trainer.rollout import RolloutResult
 
 
 def create_ref_model(
@@ -28,9 +29,10 @@ def move_to_device(batch: Dict[str, Tensor], device: str) -> Dict[str, Tensor]:
 
 
 def get_logprobs(
-    model: Union[nn.Module, Callable[..., Dict[str, Tensor]]],
+    model: nn.Module,
     input_ids: Tensor,
-    mask: Tensor,
+    attn_mask: Tensor,
+    loss_mask: Tensor,
     reduction: str,
 ) -> Tensor:
     """Compute token-wise log probabilities from model outputs.
@@ -38,7 +40,8 @@ def get_logprobs(
     Args:
         model: The language model
         input_ids: Input token IDs of shape [batch_size, seq_len]
-        mask: Attention mask of shape [batch_size, seq_len]
+        attn_mask: Attention mask passed to the model (may include causal).
+        loss_mask: Per-token mask for loss reduction.
         reduction: How to reduce over sequence dimension ("mean", "sum", "none")
 
     Returns:
@@ -51,9 +54,12 @@ def get_logprobs(
         )
 
     shifted_input_ids = input_ids[:, 1:]
-    shifted_mask = mask[:, 1:]
+    shifted_loss_mask = loss_mask[:, 1:]
 
-    logits = model(input_ids[:, :-1], mask[:, :-1])["logits"]
+    logits = model(
+        input_ids[:, :-1],
+        attn_mask[:, :, :-1, :-1] if attn_mask.dim() == 4 else attn_mask[:, :-1],
+    )["logits"]
     log_probs = torch.log_softmax(logits.float(), dim=-1)
 
     token_logprobs = torch.gather(
@@ -61,13 +67,13 @@ def get_logprobs(
     ).squeeze(-1)
 
     if reduction == "mean":
-        return (token_logprobs * shifted_mask).sum(dim=-1) / shifted_mask.sum(
+        return (token_logprobs * shifted_loss_mask).sum(dim=-1) / shifted_loss_mask.sum(
             dim=-1
         ).clamp(min=1.0)
     elif reduction == "sum":
-        return (token_logprobs * shifted_mask).sum(dim=-1)
+        return (token_logprobs * shifted_loss_mask).sum(dim=-1)
     else:
-        return token_logprobs * shifted_mask
+        return token_logprobs * shifted_loss_mask
 
 
 def make_doc_boundary_mask(position_ids: Tensor) -> Tensor:
@@ -87,7 +93,15 @@ def make_doc_boundary_mask(position_ids: Tensor) -> Tensor:
 
 
 class BaseStrategy(ABC):
-    """Abstract base class for training strategies."""
+    """Abstract base class for training strategies.
+
+    When a :class:`~astrai.trainer.rollout.RolloutRunner` is injected via
+    :meth:`set_rollout_runner`, the strategy transparently switches to
+    online mode: each ``__call__`` produces a :class:`RolloutResult`,
+    converts it to a training batch via :meth:`prepare_from_rollout`, and
+    then computes the loss.  Without a runner the strategy runs in
+    offline mode and consumes the batch directly.
+    """
 
     def __init__(
         self,
@@ -98,8 +112,8 @@ class BaseStrategy(ABC):
         self.model = model
         self.device = device
         self.executor = kwargs.pop("executor", None)
-        self.model_fn = kwargs.pop("model_fn", None)
         self.extra_kwargs = kwargs
+        self._rollout_runner = None
 
     @abstractmethod
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
@@ -113,9 +127,53 @@ class BaseStrategy(ABC):
         """
         raise NotImplementedError
 
+    def supports_online(self) -> bool:
+        """Whether this strategy can operate with a rollout runner.
+
+        Base implementation returns ``False``; strategies that implement
+        :meth:`prepare_from_rollout` should override to return ``True``.
+        """
+        return False
+
+    def set_rollout_runner(self, runner):
+        """Inject a :class:`RolloutRunner` to enable online rollout mode."""
+        self._rollout_runner = runner
+
+    def prepare_from_rollout(self, result: RolloutResult) -> Dict[str, Tensor]:
+        """Map a :class:`RolloutResult` to the batch layout expected by
+        :meth:`compute_loss`.
+
+        Strategies that return ``True`` from :meth:`supports_online` must
+        override this.  Default raises :class:`NotImplementedError`.
+        """
+        raise NotImplementedError(
+            f"{type(self).__name__} does not support online rollout"
+        )
+
+    def _on_rollout_refresh(self):
+        """Hook fired when a fresh rollout result is produced.
+
+        Override to refresh stale state (e.g. syncing the behaviour
+        policy).  Default is a no-op.
+        """
+        pass
+
+    def on_optimizer_step(self):
+        """Advance online rollout state after a successful optimizer step."""
+        if self._rollout_runner is not None:
+            self._rollout_runner.step()
+
     def __call__(self, batch: Dict[str, Tensor]) -> Tensor:
-        """Allow calling strategy directly as a callable."""
+        """Run offline or online forward depending on runner injection."""
-        return self.compute_loss(batch)
+        if self._rollout_runner is None:
+            return self.compute_loss(batch)
+
+        result, is_fresh = self._rollout_runner(batch)
+        if is_fresh:
+            self._on_rollout_refresh()
+
+        train_batch = self.prepare_from_rollout(result)
+        return self.compute_loss(train_batch)
 
 
 class StrategyFactory(BaseFactory["BaseStrategy"]):
@@ -223,14 +281,13 @@ class DPOStrategy(BaseStrategy):
         self,
         model: nn.Module,
         device: str,
+        ref_model: nn.Module,
         beta: float = 0.1,
-        reduction: str = "mean",
+        reduction: str = "sum",
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(
+        self.ref_model = ref_model
-            self.model_fn, self.executor.unwrap_model(model)
-        ).to(device=self.device)
         self.beta = beta
         self.reduction = reduction
 
@@ -240,13 +297,31 @@ class DPOStrategy(BaseStrategy):
         chosen_mask, rejected_mask = batch["chosen_mask"], batch["rejected_mask"]
 
         concat_ids = torch.cat([chosen_ids, rejected_ids], dim=0)
-        concat_mask = torch.cat([chosen_mask, rejected_mask], dim=0)
+        concat_loss_mask = torch.cat([chosen_mask, rejected_mask], dim=0)
 
-        log_pi = get_logprobs(self.model, concat_ids, concat_mask, self.reduction)
+        # Build full attention mask: key-padding + causal
+        key_pad = concat_ids.bool()[:, None, None, :]  # [B*2, 1, 1, S]
+        S = key_pad.shape[-1]
+        causal = torch.tril(
+            torch.ones(S, S, dtype=torch.bool, device=concat_ids.device)
+        )[None, None, :, :]  # [1, 1, S, S]
+        full_mask = key_pad & causal  # [B*2, 1, S, S] — composed
+
+        log_pi = get_logprobs(
+            self.model,
+            concat_ids,
+            full_mask,
+            concat_loss_mask,
+            self.reduction,
+        )
 
         with torch.no_grad():
             log_ref = get_logprobs(
-                self.ref_model, concat_ids, concat_mask, self.reduction
+                self.ref_model,
+                concat_ids,
+                full_mask,
+                concat_loss_mask,
+                self.reduction,
             )
 
         log_pi_chosen = log_pi[: chosen_ids.shape[0]]
@@ -262,47 +337,73 @@ class DPOStrategy(BaseStrategy):
 
         return dpo_loss
 
+    def supports_online(self) -> bool:
+        return True
+
+    def prepare_from_rollout(self, result: RolloutResult) -> Dict[str, Tensor]:
+        """Pick best/worst response per prompt by reward as chosen/rejected."""
+        rewards = result.rewards
+        responses = result.responses
+        masks = result.response_mask
+        best = rewards.argmax(dim=-1)
+        worst = rewards.argmin(dim=-1)
+        B = responses.shape[0]
+        idx = torch.arange(B, device=responses.device)
+        chosen = responses[idx, best]
+        chosen_mask = masks[idx, best].float()
+        rejected = responses[idx, worst]
+        rejected_mask = masks[idx, worst].float()
+        return {
+            "chosen": chosen,
+            "chosen_mask": chosen_mask,
+            "rejected": rejected,
+            "rejected_mask": rejected_mask,
+        }
+
 
 @StrategyFactory.register("grpo")
 class GRPOStrategy(BaseStrategy):
     """Group Relative Policy Optimization strategy.
 
-    On-policy GRPO following DeepSeek-R1: the policy model is updated while
+    Implements GRPO following DeepSeek-R1 with token-level PPO clipping.
-    a frozen ref_model stores the old-policy log-probs.  ratio = exp(logπ_θ - logπ_ref),
+    Advantages are group-normalized from scalar per-response rewards and
-    clipped PPO objective.  Call ``sync_ref_model()`` after each data-generation round.
+    broadcast across all response tokens.  The loss is computed **only on
+    response tokens** — prompt tokens are masked out.
+
+    Three model roles are distinguished:
+
+    * **Policy** ``self.model`` — the model being trained.
+    * **Old policy** ``self.old_model`` — the behaviour policy that generated
+      the responses.  Used for the importance sampling ratio
+      ``ρ = π_θ / π_old``.  Synced externally after each data-generation round.
+    * **Reference model** ``self.ref_model`` — a frozen copy of the initial
+      policy (typically the SFT checkpoint) used **only** for the KL
+      regularisation term.  It is never updated during training.
     """
 
     def __init__(
         self,
         model: nn.Module,
         device: str,
+        old_model: nn.Module,
+        ref_model: nn.Module,
         clip_eps: float = 0.2,
         kl_coef: float = 0.01,
         group_size: int = 4,
-        reduction: str = "mean",
-        sync_interval: int = 200,
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(
+        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 +414,101 @@ class GRPOStrategy(BaseStrategy):
         responses_flat = responses.view(-1, response_len)
         masks_flat = masks.view(-1, response_len)
         prompt_expanded = prompts.unsqueeze(1).repeat(1, group_size, 1).flatten(0, 1)
+        prompt_mask = batch.get("prompt_mask")
+        if prompt_mask is None:
+            prompt_mask = prompts.ne(0)
+        prompt_mask_expanded = (
+            prompt_mask.unsqueeze(1).expand(-1, group_size, -1).flatten(0, 1)
+        )
+        prompt_len = prompt_expanded.size(1)
 
         full_sequences = torch.cat([prompt_expanded, responses_flat], dim=-1)
-        full_masks = torch.cat([torch.ones_like(prompt_expanded), masks_flat], dim=-1)
+        # Prompt tokens are masked out (0) so logprobs are computed only for
-
+        # response tokens.  get_logprobs shifts the mask by one position, so
-        log_probs_policy = get_logprobs(
+        # the first response token's logprob (predicted from the last prompt
-            self.model, full_sequences, full_masks, self.reduction
+        # token) is correctly included.
+        full_masks = torch.cat(
+            [torch.zeros_like(prompt_expanded, dtype=torch.bool), masks_flat], dim=-1
         )
-        log_probs_policy = log_probs_policy.view(batch_size, group_size)
 
+        # Build full attention mask: key-padding + causal
+        key_pad = torch.cat([prompt_mask_expanded, masks_flat.bool()], dim=-1)[
+            :, None, None, :
+        ]
+        S = key_pad.shape[-1]
+        causal = torch.tril(
+            torch.ones(S, S, dtype=torch.bool, device=full_sequences.device)
+        )[None, None, :, :]
+        attn_mask = key_pad & causal
+
+        # get_logprobs returns [B*G, S-1] (S = prompt_len + response_len).
+        # Response token logprobs occupy the last ``response_len`` positions
+        # (the first response token is predicted from the last prompt token).
+        token_log_probs_policy = get_logprobs(
+            self.model, full_sequences, attn_mask, full_masks, "none"
+        )[:, prompt_len - 1 :]
         with torch.no_grad():
-            log_probs_ref = get_logprobs(
+            token_log_probs_old = get_logprobs(
-                self.ref_model, full_sequences, full_masks, self.reduction
+                self.old_model, full_sequences, attn_mask, full_masks, "none"
-            )
+            )[:, prompt_len - 1 :]
-            log_probs_ref = log_probs_ref.view(batch_size, group_size)
+            token_log_probs_ref = get_logprobs(
+                self.ref_model, full_sequences, attn_mask, full_masks, "none"
+            )[:, prompt_len - 1 :]
 
-        eps = torch.finfo(log_probs_policy.dtype).eps
+        # Reshape to [B, G, response_len]
+        token_log_probs_policy = token_log_probs_policy.view(batch_size, group_size, -1)
+        token_log_probs_old = token_log_probs_old.view(batch_size, group_size, -1)
+        token_log_probs_ref = token_log_probs_ref.view(batch_size, group_size, -1)
+        token_masks = masks_flat.view(batch_size, group_size, -1).float()
+
+        # Group-normalized advantages from scalar per-response rewards.
+        eps = 1e-8
         mean = rewards.mean(dim=-1, keepdim=True)
-        std = rewards.std(dim=-1, keepdim=True)
+        std = rewards.std(dim=-1, keepdim=True, unbiased=False)
         advantages = (rewards - mean) / (std + eps)
+        # Broadcast scalar advantage to every response token: [B, G, 1]
+        advantages = advantages.unsqueeze(-1)
 
-        ratio = torch.exp(log_probs_policy - log_probs_ref)
+        # Token-level ratio (π_θ / π_old) and PPO clipping.
+        log_ratio = token_log_probs_policy - token_log_probs_old
+        ratio = torch.exp(log_ratio)
 
         surr1 = ratio * advantages
         surr2 = torch.clamp(ratio, 1 - self.clip_eps, 1 + self.clip_eps) * advantages
+        per_token_policy_loss = -torch.min(surr1, surr2)
+        token_count = token_masks.sum().clamp(min=1.0)
+        policy_loss = (per_token_policy_loss * token_masks).sum() / token_count
+
+        # KL penalty to frozen reference model with k1 estimator (non-negative):
+        # k1 = π_ref / π_θ - log(π_ref / π_θ) - 1, where π_ref / π_θ = exp(log_ref - log_policy).
+        log_ref_ratio = token_log_probs_ref - token_log_probs_policy
+        r = torch.exp(log_ref_ratio)
+        kl_per_token = r - torch.log(r + eps) - 1.0
+        kl_penalty = self.kl_coef * (kl_per_token * token_masks).sum() / token_count
 
-        policy_loss = -torch.min(surr1, surr2).mean()
-        kl_penalty = self.kl_coef * (log_probs_policy - log_probs_ref).square().mean()
         total_loss = policy_loss + kl_penalty
 
         return total_loss
+
+    def supports_online(self) -> bool:
+        return True
+
+    def prepare_from_rollout(self, result: RolloutResult) -> Dict[str, Tensor]:
+        return {
+            "prompts": result.prompts,
+            "prompt_mask": result.prompt_mask,
+            "responses": result.responses,
+            "masks": result.response_mask,
+            "rewards": result.rewards,
+        }
+
+    def _on_rollout_refresh(self):
+        """Sync the behaviour policy whenever a fresh rollout arrives."""
+        self.sync_old_model()
+
+
+# Factory aliases: online variants use the same strategy class; the
+# ``RolloutRunner`` is injected by ``TrainContextBuilder`` to enable
+# online mode, so no separate subclass is needed.
+StrategyFactory._entries["online_grpo"] = GRPOStrategy
+StrategyFactory._entries["online_dpo"] = DPOStrategy

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,
@@ -139,28 +139,31 @@ class CheckpointCallback(TrainCallback):
         self.interval = interval
         self.weight_only = weight_only
         self.save_extra_fn = save_extra_fn or CheckpointCallback.save_extra
-        self.last_ckpt_step = 0
+        self.last_ckpt_step = None
 
-    def _save_checkpoint(self, context: TrainContext):
+    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):
-            save_path = os.path.join(
+        self.last_ckpt_step = context.optimizer_step
-                self.save_dir,
+
-                f"epoch_{context.epoch}_step_{context.optimizer_step}",
+        with context.executor.checkpoint_context(context.model) as state_dict:
-            )
+            if state_dict is not None:
-            extra = self.save_extra_fn(context)
+                save_path = os.path.join(
-            meta = context.config.to_dict()
+                    self.save_dir,
-            context.checkpoint = Checkpoint(
+                    f"epoch_{context.epoch}_step_{context.optimizer_step}",
-                state_dict=state_dict,
+                )
-                epoch=context.epoch,
+                extra = self.save_extra_fn(context)
-                consumed_samples=context.consumed_samples,
+                meta = context.config.to_dict()
-                config=context.model_config,
+                context.checkpoint = Checkpoint(
-                extra=extra,
+                    state_dict=state_dict,
-                meta=meta,
+                    epoch=context.epoch,
-            )
+                    consumed_samples=context.consumed_samples,
-            context.checkpoint.save(save_path)
+                    config=context.model_config,
+                    extra=extra,
+                    meta=meta,
+                )
+                context.checkpoint.save(save_path)
 
     def on_batch_end(self, context: TrainContext):
         if context.optimizer_step - self.last_ckpt_step >= self.interval:
@@ -210,7 +213,7 @@ class ProgressBarCallback(TrainCallback):
     @only_on_rank(0)
     def on_optimizer_step(self, context: TrainContext):
         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}",
         }
@@ -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

@@ -1,3 +1,4 @@
+import threading
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Dict, Optional, Self
@@ -7,13 +8,16 @@ import torch.nn as nn
 from torch.utils.data import DataLoader, random_split
 
 from astrai.config.train_config import TrainConfig
-from astrai.dataset import ResumableDistributedSampler
+from astrai.dataset import RDSampler
+from astrai.inference.core.scheduler import InferenceScheduler
 from astrai.model.components.lora import inject_lora
 from astrai.parallel.executor import BaseExecutor, ExecutorFactory
 from astrai.parallel.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.tokenize import AutoTokenizer
+from astrai.trainer.rollout import RolloutGenerator, RolloutRunner
+from astrai.trainer.strategy import BaseStrategy, StrategyFactory, create_ref_model
 
 
 @dataclass
@@ -27,7 +31,6 @@ class TrainContext:
     config: TrainConfig = field(default=None)
     model_config: dict = field(default_factory=dict)
     executor: BaseExecutor = field(default=None)
-
     epoch: int = field(default=0)
     consumed_samples: int = field(default=0)
     loss: float = field(default=0.0)
@@ -39,6 +42,15 @@ class TrainContext:
     rank: int = field(default=0)
     kwargs: Dict[str, Any] = field(default_factory=dict)
 
+    _stop_event: threading.Event = field(default_factory=threading.Event)
+
+    @property
+    def stop_requested(self) -> bool:
+        return self._stop_event.is_set()
+
+    def request_stop(self) -> None:
+        self._stop_event.set()
+
     @property
     def optimizer_step(self) -> int:
         return self.consumed_samples // (
@@ -54,10 +66,12 @@ 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:
@@ -70,50 +84,70 @@ class TrainContextBuilder:
             **cfg.executor_kwargs,
         )
 
-        model = cfg.model_fn()
-        model = model.to(device=device)
-
         model_config = {}
-        if self._resume_dir:
+        if self._param_path:
-            config_path = Path(self._resume_dir) / "config.json"
+            config_path = Path(self._param_path) / "config.json"
             if config_path.exists():
                 model_config = load_json(config_path)
 
-        if not model_config and hasattr(model, "config"):
+        preloaded_state_dict = None
-            model_config = model.config.to_dict()
+        preloaded_epoch = cfg.start_epoch
+        preloaded_consumed = cfg.start_samples * get_world_size()
+        preloaded_checkpoint = None
+        if self._param_path:
+            checkpoint = Checkpoint.load_any(self._param_path)
+            if checkpoint is not None:
+                preloaded_state_dict = checkpoint.state_dict
+                if checkpoint.config:
+                    model_config = checkpoint.config
+                if self._resume:
+                    preloaded_epoch = checkpoint.epoch or cfg.start_epoch
+                    if checkpoint.consumed_samples > 0:
+                        per_step = (
+                            cfg.batch_per_device
+                            * get_world_size()
+                            * cfg.grad_accum_steps
+                        )
+                        preloaded_consumed = (
+                            checkpoint.consumed_samples // per_step
+                        ) * per_step
+                    else:
+                        preloaded_consumed = cfg.start_samples * get_world_size()
+                    preloaded_checkpoint = checkpoint
+
+        if not model_config and hasattr(cfg.model_fn(), "config"):
+            model_config = cfg.model_fn().config.to_dict()
+
+        def _before_wrap(m):
+            m = m.to(device=device)
+            if cfg.lora is not None:
+                inject_lora(
+                    m,
+                    r=cfg.lora.r,
+                    alpha=cfg.lora.alpha,
+                    target_modules=set(cfg.lora.target_modules),
+                )
+            if preloaded_state_dict is not None:
+                m.load_state_dict(preloaded_state_dict, strict=False)
+            return m
 
         context = TrainContext(
-            model=model,
             world_size=get_world_size(),
             rank=get_rank(),
             config=cfg,
             model_config=model_config,
             executor=executor,
+            epoch=preloaded_epoch,
+            consumed_samples=preloaded_consumed,
+            checkpoint=preloaded_checkpoint,
         )
 
-        if self._resume_dir:
+        context.model, context.optimizer, context.scheduler = executor.prepare(
-            checkpoint = Checkpoint.load_any(self._resume_dir)
+            cfg.model_fn,
-            if checkpoint is not None:
+            cfg.optimizer_fn,
-                model.load_state_dict(checkpoint.state_dict, strict=False)
+            cfg.scheduler_fn,
-                if checkpoint.config:
+            before_wrap=_before_wrap,
-                    context.model_config = checkpoint.config
+        )
-                context.epoch = checkpoint.epoch or cfg.start_epoch
-                if checkpoint.consumed_samples > 0:
-                    context.consumed_samples = checkpoint.consumed_samples
-                else:
-                    context.consumed_samples = cfg.start_samples * context.world_size
-                context.checkpoint = checkpoint
-
-        if cfg.lora is not None:
-            inject_lora(
-                model,
-                r=cfg.lora.r,
-                alpha=cfg.lora.alpha,
-                target_modules=set(cfg.lora.target_modules),
-            )
-
-        context.optimizer = cfg.optimizer_fn(model)
-        context.scheduler = cfg.scheduler_fn(context.optimizer)
 
         train_dataset = cfg.dataset
         val_dataset = cfg.val_dataset
@@ -128,7 +162,7 @@ class TrainContextBuilder:
             )
 
         sampler_offset = context.consumed_samples // context.world_size
-        sampler = ResumableDistributedSampler(
+        sampler = RDSampler(
             data_source=train_dataset,
             start_epoch=context.epoch,
             start_iter=sampler_offset,
@@ -141,10 +175,11 @@ class TrainContextBuilder:
             num_workers=cfg.num_workers,
             pin_memory=cfg.pin_memory,
             prefetch_factor=cfg.prefetch_factor,
+            collate_fn=cfg.collate_fn,
         )
 
         if val_dataset is not None:
-            val_sampler = ResumableDistributedSampler(
+            val_sampler = RDSampler(
                 data_source=val_dataset,
                 start_epoch=0,
                 start_iter=0,
@@ -158,17 +193,9 @@ class TrainContextBuilder:
                 num_workers=cfg.num_workers,
                 pin_memory=cfg.pin_memory,
                 prefetch_factor=cfg.prefetch_factor,
+                collate_fn=cfg.collate_fn,
             )
 
-        context.model, context.optimizer, context.dataloader, context.scheduler = (
-            executor.prepare(
-                model,
-                context.optimizer,
-                context.dataloader,
-                context.scheduler,
-            )
-        )
-
         if context.checkpoint and context.checkpoint.extra:
             extra = context.checkpoint.extra
             for name in ("optimizer", "scheduler"):
@@ -177,13 +204,76 @@ class TrainContextBuilder:
                     if obj is not None:
                         obj.load_state_dict(extra[name])
 
+        strategy_kwargs = dict(cfg.extra_kwargs)
+
+        needs_ref = cfg.strategy in (
+            "dpo",
+            "grpo",
+            "online_grpo",
+            "online_dpo",
+        )
+        needs_old = cfg.strategy in ("grpo", "online_grpo")
+
+        if needs_ref:
+            ref_model = create_ref_model(
+                cfg.model_fn, executor.unwrap_model(context.model)
+            ).to(device=device)
+            strategy_kwargs["ref_model"] = ref_model
+
+        old_model = None
+        if needs_old:
+            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,
+            **strategy_kwargs,
-            **cfg.extra_kwargs,
         )
 
+        # Enable online rollout when the train_type is an ``online_*`` variant.
+        is_online = cfg.strategy.startswith("online_")
+        if is_online:
+            if not context.strategy.supports_online():
+                raise ValueError(
+                    f"Strategy '{cfg.strategy}' does not support online rollout"
+                )
+            if cfg.reward_model_fn is None:
+                raise ValueError("reward_model_fn is required for online RL strategies")
+
+            tokenizer = AutoTokenizer.from_pretrained(self._param_path)
+            reward_model = cfg.reward_model_fn()
+
+            group_size = strategy_kwargs.get("group_size", 1)
+            rollout_batch_size = group_size * max(1, cfg.batch_per_device)
+            max_seq_len = getattr(context.model.config, "max_position_embeddings", None)
+
+            scheduler = InferenceScheduler(
+                model=context.model,
+                tokenizer=tokenizer,
+                max_batch_size=rollout_batch_size,
+                max_seq_len=max_seq_len,
+                max_prompt_len=max_seq_len or 4096,
+            )
+
+            generator = RolloutGenerator(
+                scheduler=scheduler,
+                tokenizer=tokenizer,
+                max_tokens=cfg.rollout_max_tokens,
+                group_size=group_size,
+                temperature=cfg.rollout_temperature,
+                top_k=cfg.rollout_top_k,
+                top_p=cfg.rollout_top_p,
+            )
+            runner = RolloutRunner(
+                generator=generator,
+                reward_model=reward_model,
+                rollout_interval=cfg.rollout_interval,
+            )
+            context.strategy.set_rollout_runner(runner)
+
         return context

astrai/trainer/trainer.py

@@ -1,8 +1,14 @@
 import logging
 from typing import List, Optional
 
+import torch.distributed as dist
+
 from astrai.config import TrainConfig
 from astrai.parallel.setup import spawn_parallel_fn
+from astrai.parallel.signal_handler import (
+    register_signal_handlers,
+    unregister_signal_handlers,
+)
 from astrai.trainer.train_callback import (
     CallbackFactory,
     TrainCallback,
@@ -52,10 +58,13 @@ class Trainer:
             if method:
                 method(context)
 
-    def _trainer_loop(self, resume_dir: Optional[str] = None):
+    def _trainer_loop(self, param_path: Optional[str] = None, resume: bool = False):
         context = (
-            TrainContextBuilder(self.train_config).with_resume_dir(resume_dir).build()
+            TrainContextBuilder(self.train_config)
+            .with_param_path(param_path, resume=resume)
+            .build()
         )
+        register_signal_handlers(context)
         executor = context.executor
         self._call_callbacks("on_train_begin", context)
 
@@ -63,10 +72,14 @@ class Trainer:
             context.model.train()
 
             for epoch in range(context.epoch, context.config.n_epoch):
+                if context.stop_requested:
+                    break
                 context.epoch = epoch
                 self._call_callbacks("on_epoch_begin", context)
 
                 for batch in context.dataloader:
+                    if context.stop_requested:
+                        break
                     with executor.accumulate(context.model):
                         self._call_callbacks("on_batch_begin", context)
                         loss = context.strategy(batch)
@@ -81,6 +94,7 @@ class Trainer:
                         if executor.sync_gradients:
                             self._call_callbacks("on_optimizer_step", context)
                             context.optimizer.step()
+                            context.strategy.on_optimizer_step()
                             context.optimizer.zero_grad()
 
                             if context.scheduler:
@@ -88,14 +102,23 @@ class Trainer:
 
                 self._call_callbacks("on_epoch_end", context)
 
+            if context.stop_requested:
+                logger.warning(
+                    "Training interrupted by signal, saving emergency checkpoint..."
+                )
+                self._call_callbacks("on_error", context)
+
         except Exception as e:
             logger.error("Training failed: %s", str(e), exc_info=True)
             self._call_callbacks("on_error", context)
             raise
         finally:
             self._call_callbacks("on_train_end", context)
+            if executor.use_distributed and dist.is_initialized():
+                dist.barrier()
+            unregister_signal_handlers()
 
-    def train(self, resume_dir: Optional[str] = None):
+    def train(self, param_path: Optional[str] = None, resume: bool = False):
         cfg = self.train_config
         spawn_parallel_fn(
             self._trainer_loop,
@@ -105,5 +128,6 @@ class Trainer:
             master_port=cfg.master_port,
             device_type=cfg.device_type,
             start_method=cfg.start_method,
-            resume_dir=resume_dir,
+            param_path=param_path,
+            resume=resume,
         )

csrc/build.py

@@ -20,13 +20,14 @@ def _arch_flags() -> list[str]:
 _kernels_dir = Path("csrc/kernels")
 REGISTRY: dict[str, dict] = {}
 
-CXX_FLAGS = ["-O3", "-march=native", "-funroll-loops"]
+CXX_FLAGS = ["-O3", "-funroll-loops"]
 NVCC_FLAGS = [
     "-O3",
     "--expt-relaxed-constexpr",
     "--use_fast_math",
     "--ptxas-options=-O3,-v",
     "--extra-device-vectorization",
+    "--threads=8",
 ]
 
 
@@ -42,5 +43,6 @@ def register(name: str, sources: list[str] | None = None, **kwargs):
     }
 
 
-register("gqa_decode_attn")
+register("attn_decode")
-register("gqa_prefill_attn")
+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 +1,37 @@
+#include "attn_dispatchers.cuh"
+#include "attn_entry_utils.cuh"
+
+torch::Tensor attn_decode(
+    torch::Tensor q,
+    torch::Tensor k,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout
+) {
+    AttentionParams<bf16> p;
+    attn_pack_params(q, k, v, mask, causal_offset, scale, layout, p);
+    TORCH_CHECK(p.q_len == 1, "Q seq_len must be 1");
+    TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
+
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
+    auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    alloc_split_partials(p);
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_decode, p);
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_decode", &attn_decode,
+        py::arg("q"),
+        py::arg("k"),
+        py::arg("v"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = 0,
+        "GQA decode (tensor-core head-packing on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_decode_split_kv.cuh

@@ -0,0 +1,129 @@
+#pragma once
+#include <cuda_bf16.h>
+#include <float.h>
+#include "attn_common.h"
+#include "attn_warp_utils.cuh"
+constexpr int DC_CHUNK = 64;
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+__global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
+    int batch = blockIdx.x / p.kv_head;
+    int kv_head = blockIdx.x % p.kv_head;
+    int split = blockIdx.z;
+    int group_size = blockDim.y;
+    int q_head = kv_head * group_size + threadIdx.y;
+    int lane = threadIdx.x;
+    int hd_per_thread = p.head_dim / 32;
+
+    // Q: [batch, q_head, q_len=1, head_dim] — stride-based
+    float q_reg[8];
+    int q_off = batch * p.q_stride_b + q_head * p.q_stride_h
+              + lane * hd_per_thread * p.q_stride_d;
+    for (int i = 0; i < hd_per_thread; i++)
+        q_reg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
+
+    // KV: [batch, kv_head, kv_len, head_dim] — stride-based base
+    int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    int mask_base = batch * p.mask_b_stride;
+
+    float m = -FLT_MAX, d = 0.0f, acc_reg[8] = {0.0f};
+
+    extern __shared__ __align__(16) bf16 k_smem[];
+
+    // Split-KV: each split processes a contiguous subset of chunks
+    int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
+    int chunks_per_split = (chunks_total + p.num_splits - 1) / p.num_splits;
+    int ch_begin = split * chunks_per_split;
+    int ch_end = min(chunks_total, ch_begin + chunks_per_split);
+
+    for (int ci = ch_begin; ci < ch_end; ci++) {
+        int chunk_start = ci * DC_CHUNK;
+        int this_chunk = min(DC_CHUNK, p.kv_len - chunk_start);
+
+        // Load K into shared memory (gather from strided global)
+        int total = this_chunk * p.head_dim;
+        for (int i = threadIdx.y * 32 + lane; i < total;
+             i += blockDim.x * blockDim.y) {
+            int s = i / p.head_dim;
+            int d_dim = i % p.head_dim;
+            int kv_idx = chunk_start + s;
+            int g_off = kv_base + kv_idx * p.kv_stride_l + d_dim * p.kv_stride_d;
+            k_smem[i] = p.k[g_off];
+        }
+        __syncthreads();
+
+        for (int s = 0; s < this_chunk; s++) {
+            float partial = 0.0f;
+            for (int i = 0; i < hd_per_thread; i++)
+                partial += q_reg[i] * __bfloat162float(
+                    k_smem[s * p.head_dim + lane * hd_per_thread + i]);
+            partial = warp_reduce_sum(partial) * p.scale;
+
+            int kv_idx = chunk_start + s;
+            if constexpr (HasMask) {
+                if (!p.mask[mask_base + kv_idx])
+                    partial = -FLT_MAX;
+            }
+            if constexpr (IsCausal) {
+                if (kv_idx > p.causal_offset)
+                    partial = -FLT_MAX;
+            }
+
+            float new_m = fmaxf(m, partial);
+            float alpha = expf(m - new_m);
+            float beta  = expf(partial - new_m);
+            d = d * alpha + beta;
+
+            int v_off = kv_base + kv_idx * p.kv_stride_l
+                        + lane * hd_per_thread * p.kv_stride_d;
+            for (int i = 0; i < hd_per_thread; i++)
+                acc_reg[i] = fmaf(acc_reg[i], alpha,
+                                  __bfloat162float(p.v[v_off + i * p.kv_stride_d]) * beta);
+            m = new_m;
+        }
+        __syncthreads();
+    }
+
+    // ---- write UN-normalised partials for this split ----
+    size_t bh = (size_t)batch * p.q_head + q_head;
+    size_t slot = bh * MAX_SPLITS + split;
+    int d0 = lane * hd_per_thread;
+    for (int i = 0; i < hd_per_thread; i++) {
+        int dd = d0 + i;
+        p.o_part[slot * p.head_dim + dd] = acc_reg[i];
+    }
+    if (lane == 0) {
+        p.ml_part[slot * 2] = m;
+        p.ml_part[slot * 2 + 1] = d;
+    }
+}
+
+__global__ void attn_decode_combine_kernel(AttentionParams<bf16> p) {
+    int bh = blockIdx.x;
+    int d = threadIdx.x;
+    if (d >= p.head_dim) return;
+
+    int batch = bh / p.q_head;
+    int q_head = bh % p.q_head;
+
+    size_t split_base = (size_t)bh * MAX_SPLITS;
+    const float* mlp = p.ml_part + split_base * 2;
+    const float* op = p.o_part + split_base * p.head_dim;
+
+    float m = -FLT_MAX, l = 0.0f, acc = 0.0f;
+    for (int s = 0; s < p.num_splits; s++) {
+        float mi = mlp[s * 2];
+        if (mi <= -FLT_MAX) continue;
+        float li = mlp[s * 2 + 1];
+        float nm = fmaxf(m, mi);
+        float corr = expf(m - nm);
+        float e = expf(mi - nm);
+        acc = fmaf(acc, corr, op[s * p.head_dim + d] * e);
+        l = fmaf(l, corr, li * e);
+        m = nm;
+    }
+
+    float inv = (l > 1e-20f) ? (1.0f / l) : 0.0f;
+    int o_off = batch * p.q_stride_b + q_head * p.q_stride_h + d * p.q_stride_d;
+    p.o[o_off] = __float2bfloat16(acc * inv);
+}

csrc/kernels/attn_decode_split_kv_mma.cuh

@@ -0,0 +1,161 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+#include "attn_warp_utils.cuh"
+
+// Split-K (FlashDecoding) tensor-core decode via GQA head-packing.
+// Decode has q_len == 1, so we pack G = q_head/kv_head query heads into the
+// M=16 rows of mma.sync.m16n8k16, turning G independent GEMVs into a single
+// GEMM that reuses each loaded K/V tile across all G heads.
+//
+// IsCausal and HasMask are compile-time bools — no runtime branch in the
+// inner compute loop.
+//
+// Traits = KernelTraits<HEAD_DIM, BC=32, WARPS=1, STAGES=<2 or 1>>.
+template <typename Traits, bool IsCausal, bool HasMask>
+__global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
+    const int lane = threadIdx.x;
+    const int gid = lane >> 2;
+    const int tid4 = lane & 3;
+
+    const int pass = blockIdx.x / p.kv_head;
+    const int kv_head = blockIdx.x % p.kv_head;
+    const int batch = blockIdx.y;
+    const int split = blockIdx.z;
+
+    constexpr int MAX_G = 16;
+    const int G_total = p.q_head / p.kv_head;
+    const int g_begin = pass * MAX_G;
+    const int G = min(MAX_G, G_total - g_begin);
+    const int q_head0 = kv_head * G_total + g_begin;
+
+    // Double-buffered shared memory for K/V (no sQ needed)
+    __shared__ __align__(16) bf16 sK[Traits::STAGES * Traits::BC * Traits::LD];
+    __shared__ __align__(16) bf16 sV[Traits::STAGES * Traits::BC * Traits::LD];
+
+    // Load Q directly from global into mma A-operand registers.
+    // stride_row = p.q_stride_h for decode (q_len=1).
+    const int q_base = batch * p.q_stride_b + q_head0 * p.q_stride_h;
+    const int qra = gid;
+    const int qrb = gid + 8;
+    const bool va = qra < G, vb = qrb < G;
+    unsigned Qa[Traits::KD][4];
+    load_q_mma_frags<Traits::KD>(p.q + q_base, p.q_stride_h, p.q_stride_d,
+                                  qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[Traits::DN8][4];
+    #pragma unroll
+    for (int j = 0; j < Traits::DN8; j++)
+        Oacc[j][0] = Oacc[j][1] = Oacc[j][2] = Oacc[j][3] = 0.0f;
+    float m0 = -FLT_MAX, m1 = -FLT_MAX, l0 = 0.0f, l1 = 0.0f;
+
+    const int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    const int tiles_total = (p.kv_len + Traits::BC - 1) / Traits::BC;
+    const int tiles_per_split = (tiles_total + p.num_splits - 1) / p.num_splits;
+    const int ti_begin = split * tiles_per_split;
+    const int ti_end = min(tiles_total, ti_begin + tiles_per_split);
+
+    // ---- Load tile lambda: predicated cp.async ----
+    auto load_tile = [&](int ti, int buf) {
+        int kv0 = ti * Traits::BC;
+        bf16* dK = sK + buf * Traits::BC * Traits::LD;
+        bf16* dV = sV + buf * Traits::BC * Traits::LD;
+        #pragma unroll
+        for (int i = lane * Traits::VEC; i < Traits::TOTAL;
+             i += Traits::NUM_THREADS * Traits::VEC) {
+            int r = i / Traits::HEAD_DIM, d = i % Traits::HEAD_DIM;
+            int kc = kv0 + r;
+            bool valid = kc < p.kv_len;
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            int g_off = kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
+            cp_async_16_pred(&dK[off], &p.k[g_off], valid);
+            cp_async_16_pred(&dV[off], &p.v[g_off], valid);
+        }
+        cp_async_commit();
+    };
+
+    constexpr int BUF_MASK = (Traits::STAGES > 1) ? (Traits::STAGES - 1) : 0;
+
+    // Prologue
+    if (ti_begin < ti_end) {
+        load_tile(ti_begin, 0);
+    }
+
+    for (int ti = ti_begin; ti < ti_end; ti++) {
+        int buf = (ti - ti_begin) & BUF_MASK;
+
+        cp_async_wait_group<0>();
+        __syncwarp();
+        if constexpr (Traits::STAGES > 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, (ti + 1 - ti_begin) & BUF_MASK);
+        }
+
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = ti * Traits::BC;
+
+        float Sacc[Traits::NC8][4];
+        mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
+
+        #pragma unroll
+        for (int n8 = 0; n8 < Traits::NC8; n8++)
+            Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
+            Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
+
+        // Decode: q_len=1, so qrow0=qrow1=0
+        int maxc = IsCausal ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
+        mma_softmax_tile<Traits, HasMask>(kv0, maxc, maxc,
+                                           0, 0,
+                                           p.mask_b_stride, 0,
+                                           batch,
+                                           p.mask,
+                                           Sacc, Oacc, m0, m1, l0, l1, lane);
+
+        mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+        __syncwarp();
+
+        if constexpr (Traits::STAGES == 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, 0);
+        }
+    }
+
+    // ---- write UN-normalised partials for this split ----
+    auto split_slot = [&](int h) -> size_t {
+        size_t bh = (size_t)batch * p.q_head + h;
+        return bh * MAX_SPLITS + split;
+    };
+    #pragma unroll
+    for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+        int d = dn8 * 8 + 2 * tid4;
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
+            op[d] = Oacc[dn8][0];
+            op[d + 1] = Oacc[dn8][1];
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
+            op[d] = Oacc[dn8][2];
+            op[d + 1] = Oacc[dn8][3];
+        }
+    }
+    if (tid4 == 0) {
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m0; mp[1] = l0;
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m1; mp[1] = l1;
+        }
+    }
+}

csrc/kernels/attn_dispatchers.cuh

@@ -0,0 +1,195 @@
+#pragma once
+// Shared attention dispatchers — used by both production .cu and test .cu.
+// No torch dependency; pure CUDA.
+
+#include <cuda_runtime.h>
+#include <algorithm>
+#include "attn_warp_utils.cuh"
+#include "attn_prefill_split_q.cuh"
+#include "attn_decode_split_kv.cuh"
+#include "attn_paged_decode_split_kv.cuh"
+#ifndef ASTRAI_NO_MMA
+#include "attn_prefill_split_q_mma.cuh"
+#include "attn_decode_split_kv_mma.cuh"
+#include "attn_paged_decode_split_kv_mma.cuh"
+#endif
+
+// Split-KV: compute number of splits to fill all SMs for small-batch decode.
+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, MAX_SPLITS)));
+}
+
+// ======================================================================
+// Prefill
+// ======================================================================
+
+#ifndef ASTRAI_NO_MMA
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_prefill_mma(AttentionParams<bf16>& p) {
+    constexpr int WARPS = 4;
+    constexpr int BC = (HEAD_DIM <= 128) ? 32 : 16;
+    using Traits = KernelTraits<HEAD_DIM, BC, WARPS, 2>;
+    dim3 grid((p.q_len + Traits::BR * WARPS - 1) / (Traits::BR * WARPS), p.q_head, p.batch);
+    dim3 block(Traits::NUM_THREADS);
+    attn_prefill_split_q_mma_kernel<Traits, IsCausal, HasMask><<<grid, block>>>(p);
+}
+#endif
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_prefill_scalar(AttentionParams<bf16>& p) {
+    constexpr int G = 8, ROWS = 32, P_BC = 32;
+    dim3 grid((p.q_len + ROWS - 1) / ROWS, p.q_head, p.batch);
+    dim3 block(G, ROWS);
+    attn_prefill_split_q_kernel_t<HEAD_DIM, G, ROWS, P_BC, IsCausal, HasMask><<<grid, block>>>(p);
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_prefill(AttentionParams<bf16>& p) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+
+#ifndef ASTRAI_NO_MMA
+    if (is_causal) {
+        if (has_mask)      launch_prefill_mma<HEAD_DIM, true, true>(p);
+        else               launch_prefill_mma<HEAD_DIM, true, false>(p);
+    } else {
+        if (has_mask)      launch_prefill_mma<HEAD_DIM, false, true>(p);
+        else               launch_prefill_mma<HEAD_DIM, false, false>(p);
+    }
+#else
+    if (is_causal) {
+        if (has_mask)      launch_prefill_scalar<HEAD_DIM, true, true>(p);
+        else               launch_prefill_scalar<HEAD_DIM, true, false>(p);
+    } else {
+        if (has_mask)      launch_prefill_scalar<HEAD_DIM, false, true>(p);
+        else               launch_prefill_scalar<HEAD_DIM, false, false>(p);
+    }
+#endif
+}
+
+// ======================================================================
+// Decode
+// ======================================================================
+
+#ifndef ASTRAI_NO_MMA
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_decode_mma(AttentionParams<bf16>& p, int group_size) {
+    int G = p.q_head / p.kv_head;
+    constexpr int MAX_G = 16;
+    int num_passes = (G + MAX_G - 1) / MAX_G;
+    int tiles_total = (p.kv_len + 32 - 1) / 32;
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
+    constexpr int STAGES = (HEAD_DIM <= 128) ? 2 : 1;
+    using Traits = KernelTraits<HEAD_DIM, 32, 1, STAGES>;
+    dim3 grid(p.kv_head * num_passes, p.batch, p.num_splits);
+    attn_decode_split_kv_mma_kernel<Traits, IsCausal, HasMask><<<grid, 32>>>(p);
+}
+#endif
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_decode_scalar(AttentionParams<bf16>& p, int group_size) {
+    int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
+    size_t smem = DC_CHUNK * p.head_dim * sizeof(bf16);
+    int g = min(group_size, 32);  // cap at 32 to respect 1024-thread limit
+    dim3 grid(p.batch * p.kv_head, 1, p.num_splits);
+    dim3 block(32, g);
+    attn_decode_split_kv_kernel<HEAD_DIM, IsCausal, HasMask><<<grid, block, smem>>>(p);
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_decode(AttentionParams<bf16>& p) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+    int group_size = p.q_head / p.kv_head;
+
+#ifndef ASTRAI_NO_MMA
+    if (is_causal) {
+        if (has_mask)      launch_decode_mma<HEAD_DIM, true, true>(p, group_size);
+        else               launch_decode_mma<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_decode_mma<HEAD_DIM, false, true>(p, group_size);
+        else               launch_decode_mma<HEAD_DIM, false, false>(p, group_size);
+    }
+#else
+    if (is_causal) {
+        if (has_mask)      launch_decode_scalar<HEAD_DIM, true, true>(p, group_size);
+        else               launch_decode_scalar<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_decode_scalar<HEAD_DIM, false, true>(p, group_size);
+        else               launch_decode_scalar<HEAD_DIM, false, false>(p, group_size);
+    }
+#endif
+
+    attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
+}
+
+// ======================================================================
+// Paged Decode
+// ======================================================================
+
+#ifndef ASTRAI_NO_MMA
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_paged_decode_mma(PagedAttentionParams<bf16>& p, int group_size) {
+    int G = p.q_head / p.kv_head;
+    constexpr int MAX_G = 16;
+    bool page_ok = (p.page_size >= 32);
+    if (G >= 1 && page_ok) {
+        int num_passes = (G + MAX_G - 1) / MAX_G;
+        int tiles_total = (p.kv_len + 32 - 1) / 32;
+        p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
+        constexpr int STAGES = (HEAD_DIM <= 128) ? 2 : 1;
+        using Traits = KernelTraits<HEAD_DIM, 32, 1, STAGES>;
+        dim3 grid(p.kv_head * num_passes, p.batch, p.num_splits);
+        paged_attn_decode_split_kv_mma_kernel<Traits, IsCausal, HasMask> <<<grid, 32>>>(p);
+    } else {
+        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);
+        dim3 grid(p.batch * p.kv_head, 1, p.num_splits);
+        dim3 block(32, group_size);
+        paged_attn_decode_split_kv_kernel<HEAD_DIM, IsCausal, HasMask><<<grid, block, smem>>>(p);
+    }
+}
+#endif
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+static inline void launch_paged_decode_scalar(PagedAttentionParams<bf16>& p, int group_size) {
+    int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_CHUNK;
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
+    size_t smem = PDC_CHUNK * p.head_dim * sizeof(bf16);
+    int g = min(group_size, 32);  // cap at 32 to respect 1024-thread limit
+    dim3 grid(p.batch * p.kv_head, 1, p.num_splits);
+    dim3 block(32, g);
+    paged_attn_decode_split_kv_kernel<HEAD_DIM, IsCausal, HasMask><<<grid, block, smem>>>(p);
+}
+
+template <int HEAD_DIM>
+static inline void dispatch_paged_decode(PagedAttentionParams<bf16>& p) {
+    bool is_causal = (p.causal_offset >= 0);
+    bool has_mask = (p.use_mask && p.mask);
+    int group_size = p.q_head / p.kv_head;
+
+#ifndef ASTRAI_NO_MMA
+    if (is_causal) {
+        if (has_mask)      launch_paged_decode_mma<HEAD_DIM, true, true>(p, group_size);
+        else               launch_paged_decode_mma<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_paged_decode_mma<HEAD_DIM, false, true>(p, group_size);
+        else               launch_paged_decode_mma<HEAD_DIM, false, false>(p, group_size);
+    }
+#else
+    if (is_causal) {
+        if (has_mask)      launch_paged_decode_scalar<HEAD_DIM, true, true>(p, group_size);
+        else               launch_paged_decode_scalar<HEAD_DIM, true, false>(p, group_size);
+    } else {
+        if (has_mask)      launch_paged_decode_scalar<HEAD_DIM, false, true>(p, group_size);
+        else               launch_paged_decode_scalar<HEAD_DIM, false, false>(p, group_size);
+    }
+#endif
+
+    paged_attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
+}

csrc/kernels/attn_entry_utils.cuh

@@ -0,0 +1,171 @@
+#pragma once
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+#include "attn_common.h"
+#include "attn_warp_utils.cuh"
+
+using bf16 = __nv_bfloat16;
+
+// Dispatch head_dim: shared macro — avoids C++20 lambda template syntax.
+// Usage: DISPATCH_HEAD_DIM(hd, fn, arg)
+//   Expands to: fn<32>(arg); fn<64>(arg); etc.
+#define DISPATCH_HEAD_DIM(hd, fn, arg) \
+    switch (hd) { \
+        case 32:  fn<32>(arg); break; \
+        case 64:  fn<64>(arg); break; \
+        case 128: fn<128>(arg); break; \
+        case 256: fn<256>(arg); break; \
+        default: \
+            TORCH_CHECK(false, "unsupported head_dim ", hd, \
+                         " (supported: 32, 64, 128, 256)"); \
+    }
+
+template<typename P>
+inline void alloc_split_partials(P& p) {
+    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
+    auto o_part = torch::empty({p.batch, p.q_head, MAX_SPLITS, p.head_dim}, fopt);
+    auto ml_part = torch::empty({p.batch, p.q_head, MAX_SPLITS, 2}, fopt);
+    p.o_part = (float*)o_part.data_ptr();
+    p.ml_part = (float*)ml_part.data_ptr();
+}
+
+// ---- Shared Q-dims + strides extraction ----
+template <typename P>
+inline void extract_q_dims_and_strides(torch::Tensor& q, int64_t layout, P& p) {
+    if (layout == 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 +1,291 @@
+#pragma once
+#include <cfloat>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+
+// ============================================================================
+// KernelTraits — FlashAttention-v2 style compile-time configuration bundle.
+//
+// Bundles all dimension-dependent constants so device functions only need a
+// single Traits template parameter rather than scattered <KD, NC8, KT2, ...>.
+// ============================================================================
+template <int HEAD_DIM_, int BC_, int WARPS_, int STAGES_>
+struct KernelTraits {
+    static constexpr int HEAD_DIM = HEAD_DIM_;
+    static constexpr int BC      = BC_;       // K/V tile size along seq dim
+    static constexpr int WARPS   = WARPS_;    // warps per block
+    static constexpr int STAGES  = STAGES_;   // double-buffer stages (1 or 2)
+
+    static constexpr int BR = 16;             // Q rows per warp (mma M=16)
+
+    // Derived: mma.sync.m16n8k16 tile counts
+    static constexpr int KD  = HEAD_DIM / 16;  // Q/K k-slides
+    static constexpr int NC8 = BC / 8;          // S n-tiles (N=8)
+    static constexpr int KT2 = BC / 16;         // P k-tiles (K=16)
+    static constexpr int DN8 = HEAD_DIM / 8;    // O n-tiles (N=8)
+
+    static constexpr int LD = HEAD_DIM;         // smem leading dim
+
+    // XOR swizzle chunk bits for ldmatrix bank-conflict avoidance.
+    // mask = log2(LD/8) bits, clamped to stay within LD.
+    static constexpr int SWIZ_MASK = (HEAD_DIM >= 64) ? 7 : (HEAD_DIM / 8 - 1);
+
+    static constexpr int NUM_THREADS = WARPS * 32;
+    static constexpr int VEC = 8;               // bf16 per cp.async unit (16 bytes)
+    static constexpr int TOTAL = BC * HEAD_DIM; // total elements per tile
+};
+
+// ---- PTX wrappers ----
+using bf16 = __nv_bfloat16;
+
+__device__ __forceinline__ void mma16816(float* d, const unsigned* a,
+                                          const unsigned* b, const float* c) {
+    asm volatile(
+        "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+        "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};"
+        : "=f"(d[0]), "=f"(d[1]), "=f"(d[2]), "=f"(d[3])
+        : "r"(a[0]), "r"(a[1]), "r"(a[2]), "r"(a[3]), "r"(b[0]), "r"(b[1]),
+          "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]));
+}
+
+// read two adjacent bf16 from smem as one packed .b32 (elem0 low, elem1 high)
+__device__ __forceinline__ unsigned ld2(const bf16* p) {
+    return *reinterpret_cast<const unsigned*>(p);
+}
+
+// pack two floats into one bf16x2 as .b32
+__device__ __forceinline__ unsigned pk2(float a, float b) {
+    __nv_bfloat162 v = __floats2bfloat162_rn(a, b);
+    return *reinterpret_cast<unsigned*>(&v);
+}
+
+// pack two (non-contiguous) bf16 into one .b32
+__device__ __forceinline__ unsigned pkb(bf16 a, bf16 b) {
+    __nv_bfloat162 v;
+    v.x = a;
+    v.y = b;
+    return *reinterpret_cast<unsigned*>(&v);
+}
+
+// ldmatrix: cooperatively load mma fragments from smem (one instruction per
+// 16x16 / 16x8 tile) with the exact register layout mma expects.
+__device__ __forceinline__ void ldmatrix_x4(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x4.shared.b16 {%0,%1,%2,%3}, [%4];"
+                 : "=r"(r[0]), "=r"(r[1]), "=r"(r[2]), "=r"(r[3])
+                 : "r"(a));
+}
+__device__ __forceinline__ void ldmatrix_x2(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x2.shared.b16 {%0,%1}, [%2];"
+                 : "=r"(r[0]), "=r"(r[1])
+                 : "r"(a));
+}
+__device__ __forceinline__ void ldmatrix_x2_trans(unsigned* r, const bf16* p) {
+    unsigned a = __cvta_generic_to_shared(p);
+    asm volatile("ldmatrix.sync.aligned.m8n8.x2.trans.shared.b16 {%0,%1}, [%2];"
+                 : "=r"(r[0]), "=r"(r[1])
+                 : "r"(a));
+}
+
+// XOR swizzle for shared-memory column at 8-bf16 chunk granularity.
+__device__ __forceinline__ int swiz_col(int d, int r, int mask = 7) {
+    return ((d >> 3) ^ (r & mask)) << 3 | (d & 7);
+}
+
+// cp.async: copy 16 bytes (8 bf16) from global to shared memory directly.
+__device__ __forceinline__ void cp_async_16(bf16* smem_ptr, const void* gmem_ptr) {
+    unsigned smem_addr = __cvta_generic_to_shared(smem_ptr);
+    asm volatile("cp.async.ca.shared.global [%0], [%1], 16;"
+                 :: "r"(smem_addr), "l"(gmem_ptr));
+}
+
+// Predicated cp.async: copy 16 bytes when `pred`, otherwise zero-fill.
+// src_size=0 → no bytes read from src, so out-of-bounds src address is safe.
+__device__ __forceinline__ void cp_async_16_pred(bf16* smem_ptr,
+                                                  const void* gmem_ptr,
+                                                  bool pred) {
+    unsigned smem_addr = __cvta_generic_to_shared(smem_ptr);
+    int src_size = pred ? 16 : 0;
+    asm volatile("cp.async.ca.shared.global [%0], [%1], 16, %2;"
+                 :: "r"(smem_addr), "l"(gmem_ptr), "r"(src_size));
+}
+
+__device__ __forceinline__ void cp_async_commit() {
+    asm volatile("cp.async.commit_group;");
+}
+
+__device__ __forceinline__ void cp_async_wait_all() {
+    asm volatile("cp.async.wait_all;");
+}
+
+template <int N>
+__device__ __forceinline__ void cp_async_wait_group() {
+    asm volatile("cp.async.wait_group %0;" :: "n"(N));
+}
+
+// ---------------------------------------------------------------------------
+// Q-load: load query rows directly from global memory into mma A-operand
+// register layout. One call replaces ~15 duplicated lines in each MMA kernel.
+// stride_row is p.q_stride_h for decode (q_len=1, G heads) or
+//              p.q_stride_l for prefill (multi-q rows).
+// ---------------------------------------------------------------------------
+template <int KD>
+__device__ inline void load_q_mma_frags(
+    const bf16* __restrict__ q,
+    int stride_row,
+    int stride_d,
+    int qra, int qrb,
+    bool va, bool vb,
+    int tid4,
+    unsigned Qa[KD][4])
+{
+    #pragma unroll
+    for (int kt = 0; kt < KD; kt++) {
+        int c = kt * 16 + tid4 * 2;
+        const unsigned* pau = reinterpret_cast<const unsigned*>(
+            &q[qra * stride_row + c * stride_d]);
+        const unsigned* pbu = reinterpret_cast<const unsigned*>(
+            &q[qrb * stride_row + c * stride_d]);
+        Qa[kt][0] = va ? pau[0] : 0u;
+        Qa[kt][1] = vb ? pbu[0] : 0u;
+        Qa[kt][2] = va ? pau[4] : 0u;
+        Qa[kt][3] = vb ? pbu[4] : 0u;
+    }
+}
+
+// ---------------------------------------------------------------------------
+// S = Q @ K^T  (Qa pre-loaded by the caller; scale applied post-mma in the
+// caller to avoid bf16 precision loss).
+// Traits provides KD, NC8, LD, and SWIZ_MASK.
+// ---------------------------------------------------------------------------
+template <typename Traits>
+__device__ inline void mma_compute_scores(
+    const unsigned Qa[Traits::KD][4],
+    const bf16* __restrict__ sK,
+    int lane,
+    float Sacc[Traits::NC8][4])
+{
+    #pragma unroll
+    for (int n8 = 0; n8 < Traits::NC8; n8++) {
+        Sacc[n8][0] = Sacc[n8][1] = Sacc[n8][2] = Sacc[n8][3] = 0.0f;
+        int krow_l = n8 * 8 + (lane & 7);
+        int kcol_h = (lane & 8) ? 8 : 0;
+        #pragma unroll
+        for (int kt = 0; kt < Traits::KD; kt++) {
+            unsigned b[2];
+            ldmatrix_x2(b, &sK[krow_l * Traits::LD
+                + swiz_col(kt * 16 + kcol_h, krow_l, Traits::SWIZ_MASK)]);
+            mma16816(Sacc[n8], Qa[kt], b, Sacc[n8]);
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Online softmax + Oacc rescale for one K/V tile.
+//
+// HasMask is a compile-time template bool: when false, the mask branch is
+// entirely dead-code-eliminated from the inner unrolled loop.
+// ---------------------------------------------------------------------------
+template <typename Traits, bool HasMask>
+__device__ inline void mma_softmax_tile(
+    int kv0,
+    int maxc0, int maxc1,
+    int qrow0, int qrow1,
+    int mask_b_stride, int mask_q_stride,
+    int mask_batch,
+    const bool* __restrict__ mask,
+    float Sacc[Traits::NC8][4],
+    float Oacc[Traits::DN8][4],
+    float& m0, float& m1,
+    float& l0, float& l1,
+    int lane)
+{
+    int tid4 = lane & 3;
+
+    float rmax0 = -FLT_MAX, rmax1 = -FLT_MAX;
+    int mask_base0 = mask_batch * mask_b_stride + qrow0 * mask_q_stride;
+    int mask_base1 = mask_batch * mask_b_stride + qrow1 * mask_q_stride;
+    #pragma unroll
+    for (int n8 = 0; n8 < Traits::NC8; n8++) {
+        int cc = kv0 + n8 * 8 + 2 * tid4;
+        int c1 = cc + 1;
+        bool b0 = (cc >= maxc0) || (HasMask && !mask[mask_base0 + cc]);
+        bool b1 = (c1 >= maxc0) || (HasMask && !mask[mask_base0 + c1]);
+        bool b2 = (cc >= maxc1) || (HasMask && !mask[mask_base1 + cc]);
+        bool b3 = (c1 >= maxc1) || (HasMask && !mask[mask_base1 + c1]);
+        float s0 = b0 ? -FLT_MAX : Sacc[n8][0];
+        float s1 = b1 ? -FLT_MAX : Sacc[n8][1];
+        float s2 = b2 ? -FLT_MAX : Sacc[n8][2];
+        float s3 = b3 ? -FLT_MAX : Sacc[n8][3];
+        Sacc[n8][0] = s0; Sacc[n8][1] = s1;
+        Sacc[n8][2] = s2; Sacc[n8][3] = s3;
+        rmax0 = fmaxf(rmax0, fmaxf(s0, s1));
+        rmax1 = fmaxf(rmax1, fmaxf(s2, s3));
+    }
+    rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 1));
+    rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 2));
+    rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 1));
+    rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 2));
+
+    float nm0 = fmaxf(m0, rmax0), nm1 = fmaxf(m1, rmax1);
+    float corr0 = __expf(m0 - nm0);
+    float corr1 = __expf(m1 - nm1);
+    float pn0 = (nm0 == -FLT_MAX) ? 0.0f : 1.0f;
+    float pn1 = (nm1 == -FLT_MAX) ? 0.0f : 1.0f;
+
+    float rsum0 = 0.0f, rsum1 = 0.0f;
+    #pragma unroll
+    for (int n8 = 0; n8 < Traits::NC8; n8++) {
+        float p0 = pn0 * __expf(Sacc[n8][0] - nm0);
+        float p1 = pn0 * __expf(Sacc[n8][1] - nm0);
+        float p2 = pn1 * __expf(Sacc[n8][2] - nm1);
+        float p3 = pn1 * __expf(Sacc[n8][3] - nm1);
+        Sacc[n8][0] = p0; Sacc[n8][1] = p1;
+        Sacc[n8][2] = p2; Sacc[n8][3] = p3;
+        rsum0 += p0 + p1;
+        rsum1 += p2 + p3;
+    }
+    rsum0 += __shfl_xor_sync(0xFFFFFFFF, rsum0, 1);
+    rsum0 += __shfl_xor_sync(0xFFFFFFFF, rsum0, 2);
+    rsum1 += __shfl_xor_sync(0xFFFFFFFF, rsum1, 1);
+    rsum1 += __shfl_xor_sync(0xFFFFFFFF, rsum1, 2);
+    l0 = l0 * corr0 + rsum0;
+    l1 = l1 * corr1 + rsum1;
+    m0 = nm0; m1 = nm1;
+
+    #pragma unroll
+    for (int j = 0; j < Traits::DN8; j++) {
+        Oacc[j][0] *= corr0; Oacc[j][1] *= corr0;
+        Oacc[j][2] *= corr1; Oacc[j][3] *= corr1;
+    }
+}
+
+// ---------------------------------------------------------------------------
+// O += P @ V  (Sacc must contain P = attention weights after softmax).
+// Traits provides DN8, KT2, LD, and SWIZ_MASK.
+// ---------------------------------------------------------------------------
+template <typename Traits>
+__device__ inline void mma_pv_accumulate(
+    float Sacc[][4],
+    const bf16* __restrict__ sV,
+    int lane,
+    float Oacc[Traits::DN8][4])
+{
+    #pragma unroll
+    for (int kt2 = 0; kt2 < Traits::KT2; kt2++) {
+        unsigned Pa[4];
+        Pa[0] = pk2(Sacc[kt2 * 2][0], Sacc[kt2 * 2][1]);
+        Pa[1] = pk2(Sacc[kt2 * 2][2], Sacc[kt2 * 2][3]);
+        Pa[2] = pk2(Sacc[kt2 * 2 + 1][0], Sacc[kt2 * 2 + 1][1]);
+        Pa[3] = pk2(Sacc[kt2 * 2 + 1][2], Sacc[kt2 * 2 + 1][3]);
+        int vrow_l = kt2 * 16 + (lane & 15);
+        #pragma unroll
+        for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+            unsigned b[2];
+            ldmatrix_x2_trans(b, &sV[vrow_l * Traits::LD
+                + swiz_col(dn8 * 8, vrow_l, Traits::SWIZ_MASK)]);
+            mma16816(Oacc[dn8], Pa, b, Oacc[dn8]);
+        }
+    }
+}

csrc/kernels/attn_paged_decode.cu

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

csrc/kernels/attn_paged_decode_split_kv.cuh

@@ -0,0 +1,146 @@
+#pragma once
+#include <cuda_bf16.h>
+#include <float.h>
+#include "attn_common.h"
+#include "attn_warp_utils.cuh"
+constexpr int PDC_CHUNK = 64;
+
+template <int HEAD_DIM, bool IsCausal, bool HasMask>
+__global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p) {
+    int batch = blockIdx.x / p.kv_head;
+    int kv_head = blockIdx.x % p.kv_head;
+    int split = blockIdx.z;
+    int group_size = blockDim.y;
+    int q_head = kv_head * group_size + threadIdx.y;
+    int lane = threadIdx.x;
+    int hd_per_thread = p.head_dim / 32;
+
+    float q_reg[8];
+    int q_off = batch * p.q_stride_b + q_head * p.q_stride_h
+              + lane * hd_per_thread * p.q_stride_d;
+    #pragma unroll
+    for (int i = 0; i < hd_per_thread; i++)
+        q_reg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
+
+    float m = -FLT_MAX, d = 0.0f, acc_reg[8] = {0.0f};
+
+    extern __shared__ __align__(16) bf16 k_smem[];
+
+    int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_CHUNK;
+    int chunks_per_split = (chunks_total + p.num_splits - 1) / p.num_splits;
+    int ch_begin = split * chunks_per_split;
+    int ch_end = min(chunks_total, ch_begin + chunks_per_split);
+
+    const int mask_base = batch * p.mask_b_stride;
+
+    for (int ci = ch_begin; ci < ch_end; ci++) {
+        int chunk_start = ci * PDC_CHUNK;
+        int this_chunk = min(PDC_CHUNK, p.kv_len - chunk_start);
+
+        int total = this_chunk * p.head_dim;
+        for (int i = threadIdx.y * 32 + lane; i < total;
+             i += blockDim.x * blockDim.y) {
+            int s = i / p.head_dim;
+            int d_dim = i % p.head_dim;
+            int pos = chunk_start + s;
+            int logical_page = pos / p.page_size;
+            int page_offset = pos % p.page_size;
+            int phys_page = p.page_table[batch * p.max_pages + logical_page];
+            if (phys_page >= 0) {
+                int64_t off = (int64_t)phys_page * p.page_size * p.kv_head * p.head_dim
+                            + (int64_t)page_offset * p.kv_head * p.head_dim
+                            + (int64_t)kv_head * p.head_dim
+                            + d_dim;
+                k_smem[i] = p.k_cache[off];
+            } else {
+                k_smem[i] = __float2bfloat16(0.0f);
+            }
+        }
+        __syncthreads();
+
+        for (int s = 0; s < this_chunk; s++) {
+            float partial = 0.0f;
+            #pragma unroll
+            for (int i = 0; i < hd_per_thread; i++)
+                partial += q_reg[i] * __bfloat162float(
+                    k_smem[s * p.head_dim + lane * hd_per_thread + i]);
+            partial = warp_reduce_sum(partial) * p.scale;
+
+            int kv_idx = chunk_start + s;
+            if constexpr (HasMask) {
+                if (!p.mask[mask_base + kv_idx])
+                    partial = -FLT_MAX;
+            }
+            if constexpr (IsCausal) {
+                if (kv_idx > p.causal_offset)
+                    partial = -FLT_MAX;
+            }
+
+            float new_m = fmaxf(m, partial);
+            float alpha = expf(m - new_m);
+            float beta  = expf(partial - new_m);
+            d = d * alpha + beta;
+
+            int 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] = fmaf(acc_reg[i], alpha,
+                                      __bfloat162float(p.v_cache[v_base + lane * hd_per_thread + i]) * beta);
+            } else {
+                #pragma unroll
+                for (int i = 0; i < hd_per_thread; i++)
+                    acc_reg[i] = fmaf(acc_reg[i], alpha, 0.0f);
+            }
+            m = new_m;
+        }
+        __syncthreads();
+    }
+
+    size_t bh = (size_t)batch * p.q_head + q_head;
+    size_t slot = bh * MAX_SPLITS + split;
+    int d0 = lane * hd_per_thread;
+    #pragma unroll
+    for (int i = 0; i < hd_per_thread; i++)
+        p.o_part[slot * p.head_dim + (d0 + i)] = acc_reg[i];
+    if (lane == 0) {
+        p.ml_part[slot * 2] = m;
+        p.ml_part[slot * 2 + 1] = d;
+    }
+}
+
+__global__ void paged_attn_decode_combine_kernel(PagedAttentionParams<bf16> p) {
+    int bh = blockIdx.x;
+    int d = threadIdx.x;
+    if (d >= p.head_dim) return;
+
+    int batch = bh / p.q_head;
+    int q_head = bh % p.q_head;
+
+    size_t split_base = (size_t)bh * MAX_SPLITS;
+    const float* mlp = p.ml_part + split_base * 2;
+    const float* op = p.o_part + split_base * p.head_dim;
+
+    float m = -FLT_MAX, l = 0.0f, acc = 0.0f;
+    for (int s = 0; s < p.num_splits; s++) {
+        float mi = mlp[s * 2];
+        if (mi <= -FLT_MAX) continue;
+        float li = mlp[s * 2 + 1];
+        float nm = fmaxf(m, mi);
+        float corr = expf(m - nm);
+        float e = expf(mi - nm);
+        acc = fmaf(acc, corr, op[s * p.head_dim + d] * e);
+        l = fmaf(l, corr, li * e);
+        m = nm;
+    }
+
+    float inv = (l > 1e-20f) ? (1.0f / l) : 0.0f;
+    int o_off = batch * p.q_stride_b + q_head * p.q_stride_h + d * p.q_stride_d;
+    p.o[o_off] = __float2bfloat16(acc * inv);
+}

csrc/kernels/attn_paged_decode_split_kv_mma.cuh

@@ -0,0 +1,161 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+#include "attn_warp_utils.cuh"
+
+// Paged split-KV tensor-core decode via GQA head-packing.
+// Reads K/V directly from the page pool through a page table — one tile
+// (BC=32) fits within a single page (page_size >= 32), so the page-table
+// lookup happens once per tile for cp.async.
+//
+// IsCausal and HasMask are compile-time bools.
+template <typename Traits, bool IsCausal, bool HasMask>
+__global__ void paged_attn_decode_split_kv_mma_kernel(PagedAttentionParams<bf16> p) {
+    const int lane = threadIdx.x;
+    const int gid = lane >> 2;
+    const int tid4 = lane & 3;
+
+    const int pass = blockIdx.x / p.kv_head;
+    const int kv_head = blockIdx.x % p.kv_head;
+    const int batch = blockIdx.y;
+    const int split = blockIdx.z;
+
+    constexpr int MAX_G = 16;
+    const int G_total = p.q_head / p.kv_head;
+    const int g_begin = pass * MAX_G;
+    const int G = min(MAX_G, G_total - g_begin);
+    const int q_head0 = kv_head * G_total + g_begin;
+
+    __shared__ __align__(16) bf16 sK[Traits::STAGES * Traits::BC * Traits::LD];
+    __shared__ __align__(16) bf16 sV[Traits::STAGES * Traits::BC * Traits::LD];
+
+    const int q_base = batch * p.q_stride_b + q_head0 * p.q_stride_h;
+    const int qra = gid;
+    const int qrb = gid + 8;
+    const bool va = qra < G, vb = qrb < G;
+    unsigned Qa[Traits::KD][4];
+    load_q_mma_frags<Traits::KD>(p.q + q_base, p.q_stride_h, p.q_stride_d,
+                                  qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[Traits::DN8][4];
+    #pragma unroll
+    for (int j = 0; j < Traits::DN8; j++)
+        Oacc[j][0] = Oacc[j][1] = Oacc[j][2] = Oacc[j][3] = 0.0f;
+    float m0 = -FLT_MAX, m1 = -FLT_MAX, l0 = 0.0f, l1 = 0.0f;
+
+    const int tiles_total = (p.kv_len + Traits::BC - 1) / Traits::BC;
+    const int tiles_per_split = (tiles_total + p.num_splits - 1) / p.num_splits;
+    const int ti_begin = split * tiles_per_split;
+    const int ti_end = min(tiles_total, ti_begin + tiles_per_split);
+
+    const int64_t page_stride = (int64_t)p.page_size * p.kv_head * Traits::HEAD_DIM;
+    const int64_t pos_stride  = (int64_t)p.kv_head * Traits::HEAD_DIM;
+    const int64_t head_off    = (int64_t)kv_head * Traits::HEAD_DIM;
+
+    // ---- Load tile lambda: paged addressing ----
+    auto load_tile = [&](int ti, int buf) {
+        int kv0 = ti * Traits::BC;
+        bf16* dK = sK + buf * Traits::BC * Traits::LD;
+        bf16* dV = sV + buf * Traits::BC * Traits::LD;
+        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 * Traits::VEC; i < Traits::TOTAL;
+             i += Traits::NUM_THREADS * Traits::VEC) {
+            int r = i / Traits::HEAD_DIM, d = i % Traits::HEAD_DIM;
+            int kc = kv0 + r;
+            bool valid = (kc < p.kv_len) && page_valid;
+            int page_off = kc % p.page_size;
+            int64_t gmem_base = (int64_t)phys_page * page_stride
+                              + (int64_t)page_off * pos_stride
+                              + head_off;
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            cp_async_16_pred(&dK[off], &p.k_cache[gmem_base + d], valid);
+            cp_async_16_pred(&dV[off], &p.v_cache[gmem_base + d], valid);
+        }
+        cp_async_commit();
+    };
+
+    constexpr int BUF_MASK = (Traits::STAGES > 1) ? (Traits::STAGES - 1) : 0;
+
+    if (ti_begin < ti_end) {
+        load_tile(ti_begin, 0);
+    }
+
+    for (int ti = ti_begin; ti < ti_end; ti++) {
+        int buf = (ti - ti_begin) & BUF_MASK;
+
+        cp_async_wait_group<0>();
+        __syncwarp();
+        if constexpr (Traits::STAGES > 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, (ti + 1 - ti_begin) & BUF_MASK);
+        }
+
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = ti * Traits::BC;
+
+        float Sacc[Traits::NC8][4];
+        mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
+
+        #pragma unroll
+        for (int n8 = 0; n8 < Traits::NC8; n8++)
+            Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
+            Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
+
+        int maxc = IsCausal ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
+        mma_softmax_tile<Traits, HasMask>(kv0, maxc, maxc,
+                                           0, 0,
+                                           p.mask_b_stride, 0,
+                                           batch,
+                                           p.mask,
+                                           Sacc, Oacc, m0, m1, l0, l1, lane);
+
+        mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+        __syncwarp();
+
+        if constexpr (Traits::STAGES == 1) {
+            if (ti + 1 < ti_end)
+                load_tile(ti + 1, 0);
+        }
+    }
+
+    auto split_slot = [&](int h) -> size_t {
+        size_t bh = (size_t)batch * p.q_head + h;
+        return bh * MAX_SPLITS + split;
+    };
+    #pragma unroll
+    for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+        int d = dn8 * 8 + 2 * tid4;
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
+            op[d] = Oacc[dn8][0];
+            op[d + 1] = Oacc[dn8][1];
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* op = p.o_part + split_slot(h) * Traits::HEAD_DIM;
+            op[d] = Oacc[dn8][2];
+            op[d + 1] = Oacc[dn8][3];
+        }
+    }
+    if (tid4 == 0) {
+        int r0 = gid, r1 = gid + 8;
+        if (r0 < G) {
+            int h = q_head0 + r0;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m0; mp[1] = l0;
+        }
+        if (r1 < G) {
+            int h = q_head0 + r1;
+            float* mp = p.ml_part + split_slot(h) * 2;
+            mp[0] = m1; mp[1] = l1;
+        }
+    }
+}

csrc/kernels/attn_prefill.cu

@@ -0,0 +1,35 @@
+#include "attn_dispatchers.cuh"
+#include "attn_entry_utils.cuh"
+
+torch::Tensor attn_prefill(
+    torch::Tensor q,
+    torch::Tensor k,
+    torch::Tensor v,
+    c10::optional<torch::Tensor> mask,
+    int64_t causal_offset,
+    double scale,
+    int64_t layout
+) {
+    AttentionParams<bf16> p;
+    attn_pack_params(q, k, v, mask, causal_offset, scale, layout, p);
+    TORCH_CHECK(p.head_dim % 16 == 0, "head_dim must be multiple of 16");
+
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
+    auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
+
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_prefill, p);
+    return O;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("attn_prefill", &attn_prefill,
+        py::arg("q"),
+        py::arg("k"),
+        py::arg("v"),
+        py::arg("mask") = py::none(),
+        py::arg("causal_offset") = -1,
+        py::arg("scale") = 0.0,
+        py::arg("layout") = 0,
+        "GQA prefill (tensor-core mma on sm_80+, scalar fallback)");
+}

csrc/kernels/attn_prefill_split_q.cuh

@@ -1,13 +1,14 @@
 #pragma once
-#include "gqa_common.cuh"
+#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
+// each owning HEAD_DIM/G dims of qreg[]/acc[]. IsCausal and HasMask are
-// occupancy high; the S dot product is reduced across the G-lane group with a
+// compile-time bools — the compiler eliminates dead branches.
-// short shuffle chain (log2(G) shuffles) instead of a full 32-lane warp reduce.
+// Templated on <HEAD_DIM, G, ROWS, P_BC, IsCausal, HasMask>.
-// 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) {
@@ -17,8 +18,7 @@ __device__ __forceinline__ float group_reduce_sum(float v, unsigned mask) {
     return v;
 }
 
-// load 8 contiguous bf16 from (16-byte aligned) smem as one float4, unpack to
+// load 8 contiguous bf16 from (16-byte aligned) smem as one float4
-// 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);
@@ -30,8 +30,8 @@ __device__ __forceinline__ void ld8(const bf16* p, float* o) {
     }
 }
 
-template <int HEAD_DIM, int G, int ROWS, int P_BC>
+template <int HEAD_DIM, int G, int ROWS, int P_BC, bool IsCausal, bool HasMask>
-__global__ void gqa_prefill_attn_kernel_t(GQAParams p) {
+__global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
     constexpr int DPT = HEAD_DIM / G;
 
     int q_tile = blockIdx.x;
@@ -43,16 +43,17 @@ __global__ void gqa_prefill_attn_kernel_t(GQAParams p) {
 
     int kv_head = q_head / (p.q_head / p.kv_head);
 
-    extern __shared__ __align__(16) bf16 smem[];
+    __shared__ __align__(16) bf16 sK[P_BC * HEAD_DIM];
-    bf16* sK = smem;
+    __shared__ __align__(16) bf16 sV[P_BC * HEAD_DIM];
-    bf16* sV = sK + 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_head + q_head) * p.q_len + q_row) * HEAD_DIM + gpos * DPT;
+        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.scale;
+            qreg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
     }
 
     float m = -FLT_MAX, l = 0.0f;
@@ -61,13 +62,13 @@ __global__ void gqa_prefill_attn_kernel_t(GQAParams p) {
     for (int i = 0; i < DPT; i++)
         acc[i] = 0.0f;
 
-    int kv_base = ((batch * p.kv_head + kv_head) * p.kv_len) * HEAD_DIM;
+    // 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)));
@@ -76,22 +77,29 @@ __global__ void gqa_prefill_attn_kernel_t(GQAParams p) {
         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 gidx = kv_base + (kv0 + i / HEAD_DIM) * HEAD_DIM + (i % HEAD_DIM);
+            int s = i / HEAD_DIM;
-            sK[i] = p.k[gidx];
+            int d_dim = i % HEAD_DIM;
-            sV[i] = p.v[gidx];
+            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.is_causal && q_row < p.q_len) {
+        if constexpr (IsCausal) {
-            int ep = q_row + p.causal_offset + 1;
+            if (q_row < p.q_len) {
-            if (kv0 >= ep)
+                int ep = q_row + p.causal_offset + 1;
-                lim = 0;
+                if (kv0 >= ep)
-            else if (kv0 + tlen > ep)
+                    lim = 0;
-                lim = ep - kv0;
+                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;
@@ -103,10 +111,13 @@ __global__ void gqa_prefill_attn_kernel_t(GQAParams p) {
                 for (int j = 0; j < 8; j++)
                     part = fmaf(qreg[i + j], k8[j], part);
             }
-            float dot = group_reduce_sum<G>(part, gmask);
+            float dot = group_reduce_sum<G>(part, gmask) * p.scale;
 
-            if (p.use_mask && p.mask && !p.mask[batch * p.kv_len + kv0 + s])
+            int kv_idx = kv0 + s;
-                dot = -FLT_MAX;
+            if constexpr (HasMask) {
+                if (!p.mask[mask_row_base + kv_idx])
+                    dot = -FLT_MAX;
+            }
 
             float nm = fmaxf(m, dot);
             float al = __expf(m - nm);
@@ -128,10 +139,11 @@ __global__ void gqa_prefill_attn_kernel_t(GQAParams p) {
     }
 
     if (q_row < p.q_len) {
-        int o_off = ((batch * p.q_head + q_head) * p.q_len + q_row) * HEAD_DIM + gpos * DPT;
+        int o_off = batch * p.q_stride_b + q_head * p.q_stride_h
-        float rl = (l > 1e-10f) ? (1.0f / l) : 0.0f;
+                  + q_row * p.q_stride_l + gpos * DPT * p.q_stride_d;
+        float rl = (l > 1e-20f) ? (1.0f / l) : 0.0f;
 #pragma unroll
         for (int i = 0; i < DPT; i++)
-            p.o[o_off + i] = __float2bfloat16(acc[i] * rl);
+            p.o[o_off + i * p.q_stride_d] = __float2bfloat16(acc[i] * rl);
     }
 }

csrc/kernels/attn_prefill_split_q_mma.cuh

@@ -0,0 +1,146 @@
+#pragma once
+#include <cfloat>
+#include <cuda_bf16.h>
+#include "attn_common.h"
+#include "attn_mma_utils.cuh"
+
+// Tensor-core prefill flash attention (raw mma.sync PTX).
+// One warp owns BR=16 query rows. S = Q@K^T and O = P@V run on bf16 tensor
+// cores via mma.sync.m16n8k16 (f32 accumulate).
+//
+// IsCausal and HasMask are compile-time bools — the compiler eliminates all
+// dead branches in the inner compute loop (FA2-style).
+//
+// Traits = KernelTraits<HEAD_DIM, BC, WARPS=4, STAGES=2>.
+template <typename Traits, bool IsCausal, bool HasMask>
+__global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
+    const int warp = threadIdx.x / 32;
+    const int lane = threadIdx.x % 32;
+    const int gid = lane >> 2;   // 0..7
+    const int tid4 = lane & 3;   // 0..3
+
+    const int q_head = blockIdx.y;
+    const int batch = blockIdx.z;
+    const int kv_head = q_head / (p.q_head / p.kv_head);
+    const int qrow0 = (blockIdx.x * Traits::WARPS + warp) * Traits::BR;
+
+    // Static shared memory: double-buffered K/V (no sQ — Q goes direct
+    // to registers in mma A-operand layout).
+    __shared__ __align__(16) bf16 sK[Traits::STAGES * Traits::BC * Traits::LD];
+    __shared__ __align__(16) bf16 sV[Traits::STAGES * Traits::BC * Traits::LD];
+
+    // Load Q fragments straight from global into mma A-operand layout.
+    const int q_base = batch * p.q_stride_b + q_head * p.q_stride_h;
+    const int qra = qrow0 + gid;
+    const int qrb = qrow0 + gid + 8;
+    const bool va = qra < p.q_len, vb = qrb < p.q_len;
+    unsigned Qa[Traits::KD][4];
+    load_q_mma_frags<Traits::KD>(p.q + q_base, p.q_stride_l, p.q_stride_d,
+                                  qra, qrb, va, vb, tid4, Qa);
+
+    float Oacc[Traits::DN8][4];
+    #pragma unroll
+    for (int j = 0; j < Traits::DN8; j++)
+        Oacc[j][0] = Oacc[j][1] = Oacc[j][2] = Oacc[j][3] = 0.0f;
+    float m0 = -FLT_MAX, m1 = -FLT_MAX, l0 = 0.0f, l1 = 0.0f;
+
+    // KV: stride-based base
+    const int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
+    const int tiles = (p.kv_len + Traits::BC - 1) / Traits::BC;
+    const int qr0 = qrow0 + gid;
+    const int qr1 = qrow0 + gid + 8;
+
+    // Causal tile-skip bounds (dead code when IsCausal == false)
+    const int max_kv = qrow0 + Traits::BR - 1 + p.causal_offset;
+    const int block_max_kv =
+        blockIdx.x * Traits::WARPS * Traits::BR + Traits::WARPS * Traits::BR - 1
+        + p.causal_offset;
+
+    int t_end = tiles - 1;
+    if constexpr (IsCausal) {
+        int bt = block_max_kv / Traits::BC;
+        if (bt < t_end) t_end = bt;
+    }
+
+    // ---- Load tile lambda: predicated cp.async ----
+    auto load_tile = [&](int ti, int buf) {
+        int kv0 = ti * Traits::BC;
+        bf16* dK = sK + buf * Traits::BC * Traits::LD;
+        bf16* dV = sV + buf * Traits::BC * Traits::LD;
+        #pragma unroll
+        for (int i = threadIdx.x * Traits::VEC; i < Traits::TOTAL;
+             i += Traits::NUM_THREADS * Traits::VEC) {
+            int r = i / Traits::HEAD_DIM, d = i % Traits::HEAD_DIM;
+            int kc = kv0 + r;
+            bool valid = kc < p.kv_len;
+            int off = r * Traits::LD + swiz_col(d, r, Traits::SWIZ_MASK);
+            int g_off = kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
+            cp_async_16_pred(&dK[off], &p.k[g_off], valid);
+            cp_async_16_pred(&dV[off], &p.v[g_off], valid);
+        }
+        cp_async_commit();
+    };
+
+    // ---- Prologue: issue first tile load ----
+    load_tile(0, 0);
+
+    for (int ti = 0; ti <= t_end; ti++) {
+        int buf = ti & 1;
+
+        // Wait for current tile, then publish cross-warp + guard buffer reuse.
+        cp_async_wait_group<0>();
+        __syncthreads();
+        if (ti < t_end) load_tile(ti + 1, (ti + 1) & 1);
+
+        const bf16* bK = sK + buf * Traits::BC * Traits::LD;
+        const bf16* bV = sV + buf * Traits::BC * Traits::LD;
+        int kv0 = ti * Traits::BC;
+
+        // Warp-level causal skip (dead branch eliminated when IsCausal == false)
+        if (!IsCausal || kv0 <= max_kv) {
+
+            float Sacc[Traits::NC8][4];
+            mma_compute_scores<Traits>(Qa, bK, lane, Sacc);
+
+            // Post-multiply scale in float (no bf16 precision loss)
+            #pragma unroll
+            for (int n8 = 0; n8 < Traits::NC8; n8++)
+                Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
+                Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
+
+            int maxc0 = IsCausal ? min(p.kv_len, qr0 + p.causal_offset + 1)
+                                 : p.kv_len;
+            int maxc1 = IsCausal ? min(p.kv_len, qr1 + p.causal_offset + 1)
+                                 : p.kv_len;
+            mma_softmax_tile<Traits, HasMask>(kv0, maxc0, maxc1,
+                                               qr0, qr1,
+                                               p.mask_b_stride, p.mask_q_stride,
+                                               batch,
+                                               p.mask,
+                                               Sacc, Oacc, m0, m1, l0, l1, lane);
+
+            mma_pv_accumulate<Traits>(Sacc, bV, lane, Oacc);
+        }
+    }
+
+    // ---- write output: packed bf16x2 stores ----
+    float rl0 = (l0 > 1e-20f) ? (1.0f / l0) : 0.0f;
+    float rl1 = (l1 > 1e-20f) ? (1.0f / l1) : 0.0f;
+    const int o_base = batch * p.q_stride_b + q_head * p.q_stride_h;
+    #pragma unroll
+    for (int dn8 = 0; dn8 < Traits::DN8; dn8++) {
+        int d = dn8 * 8 + 2 * tid4;
+        if (qr0 < p.q_len) {
+            __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][0] * rl0,
+                                                      Oacc[dn8][1] * rl0);
+            *reinterpret_cast<__nv_bfloat162*>(
+                &p.o[o_base + qr0 * p.q_stride_l + d * p.q_stride_d]) = v;
+        }
+        if (qr1 < p.q_len) {
+            __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][2] * rl1,
+                                                      Oacc[dn8][3] * rl1);
+            *reinterpret_cast<__nv_bfloat162*>(
+                &p.o[o_base + qr1 * p.q_stride_l + d * p.q_stride_d]) = v;
+        }
+    }
+}

csrc/kernels/attn_warp_utils.cuh

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

csrc/kernels/gqa_common.cuh

@@ -1,35 +0,0 @@
-#pragma once
-#include <cuda_bf16.h>
-#include <cuda_runtime.h>
-#include <cfloat>
-#include <algorithm>
-
-using bf16 = __nv_bfloat16;
-using std::min;
-
-constexpr int DC_CHUNK = 64;
-constexpr int Br = 32, Bc = 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;
-}
-
-struct GQAParams {
-    int batch;
-    int q_head;
-    int kv_head;
-    int q_len;
-    int kv_len;
-    int head_dim;
-    int use_mask;
-    int is_causal;
-    int causal_offset;
-    float scale;
-    const bf16* __restrict__ q;
-    const bf16* __restrict__ k;
-    const bf16* __restrict__ v;
-    const bool* __restrict__ mask;
-    bf16* __restrict__ o;
-};

csrc/kernels/gqa_decode_attn.cu

@@ -1,114 +0,0 @@
-#include "gqa_decode_attn.cuh"
-#include <torch/extension.h>
-
-#ifndef ASTRAI_NO_MMA
-#include "gqa_decode_attn_mma.cuh"
-#endif
-
-template <int HEAD_DIM>
-static void dispatch_decode(GQAParams& p) {
-#ifndef ASTRAI_NO_MMA
-    constexpr int BC = 32, BR = 16, LD = HEAD_DIM;  // XOR swizzle → no padding
-    int G = p.q_head / p.kv_head;
-    // head-packing tensor-core path needs 1 < G <= 16 (MMA M dim) and no mask;
-    // everything else uses the scalar kernel
-    if (!p.use_mask && G > 1 && G <= 16) {
-        dim3 grid(p.kv_head, p.batch, 1);
-        dim3 block(32, 1, 1);
-        // sK + sV + sQ, each BC/BR * LD (single buffer for high occupancy)
-        int smem = (2 * BC * LD + BR * LD) * (int)sizeof(bf16);
-        cudaFuncSetAttribute(gqa_decode_attn_mma_kernel<HEAD_DIM, BC>,
-                             cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
-        gqa_decode_attn_mma_kernel<HEAD_DIM, BC><<<grid, block, smem>>>(p);
-        return;
-    }
-    // scalar fallback (per-KV-head, one warp per query head)
-    int group_size = p.q_head / p.kv_head;
-    size_t smem = DC_CHUNK * p.head_dim * sizeof(bf16);
-    dim3 block(32, group_size);
-    dim3 grid(p.batch * p.kv_head);
-    gqa_decode_attn_kernel<<<grid, block, smem>>>(p);
-#else
-    // scalar fallback (per-KV-head, one warp per query head)
-    int group_size = p.q_head / p.kv_head;
-    size_t smem = DC_CHUNK * p.head_dim * sizeof(bf16);
-    dim3 block(32, group_size);
-    dim3 grid(p.batch * p.kv_head);
-    gqa_decode_attn_kernel<<<grid, block, smem>>>(p);
-#endif
-}
-
-torch::Tensor gqa_decode_attn(
-    torch::Tensor q,
-    torch::Tensor k, 
-    torch::Tensor v,
-    c10::optional<torch::Tensor> mask,
-    bool is_causal = false, 
-    int64_t causal_offset = 0,
-    c10::optional<double> scale = c10::nullopt
-) {
-    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(q.size(2) == 1, "Q seq_len must be 1");
-
-    GQAParams p;
-    p.batch = q.size(0);
-    p.q_head = q.size(1);
-    p.kv_head = k.size(1);
-    p.q_len = 1;
-    p.kv_len = k.size(2);
-    p.head_dim = q.size(3);
-    TORCH_CHECK(p.head_dim % 32 == 0, "head_dim must be multiple of 32");
-    p.use_mask = mask.has_value();
-    p.is_causal = (int)is_causal;
-    p.causal_offset = (int)causal_offset;
-    p.scale = scale.has_value() ? (float)scale.value() : 1.0f / sqrtf((float)p.head_dim);
-    p.q = (const bf16*)q.data_ptr();
-    p.k = (const bf16*)k.data_ptr();
-    p.v = (const bf16*)v.data_ptr();
-    if (p.use_mask) {
-        TORCH_CHECK(mask.value().dtype() == torch::kBool);
-        TORCH_CHECK(mask.value().dim() == 2);
-        TORCH_CHECK(mask.value().size(0) == p.batch);
-        TORCH_CHECK(mask.value().size(1) == p.kv_len);
-        p.mask = mask.value().data_ptr<bool>();
-    } else {
-        p.mask = nullptr;
-    }
-
-    auto O = torch::empty_like(q);
-    p.o = (bf16*)O.data_ptr();
-
-    switch (p.head_dim) {
-        case 32:
-            dispatch_decode<32>(p);
-            break;
-        case 64:
-            dispatch_decode<64>(p);
-            break;
-        case 128:
-            dispatch_decode<128>(p);
-            break;
-        case 256:
-            dispatch_decode<256>(p);
-            break;
-        default:
-            TORCH_CHECK(false, "decode: unsupported head_dim ", p.head_dim,
-                        " (supported: 32, 64, 128, 256)");
-    }
-    return O;
-}
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-    m.def("gqa_decode_attn", &gqa_decode_attn,
-        py::arg("q"),
-        py::arg("k"),
-        py::arg("v"),
-        py::arg("mask") = py::none(),
-        py::arg("is_causal") = false,
-        py::arg("causal_offset") = 0,
-        py::arg("scale") = py::none(),
-        "GQA decode (tensor-core head-packing on sm_80+, scalar fallback)");
-}

csrc/kernels/gqa_decode_attn.cuh

@@ -1,59 +0,0 @@
-#pragma once
-#include "gqa_common.cuh"
-
-__global__ void gqa_decode_attn_kernel(GQAParams p) {
-    int batch = blockIdx.x / p.kv_head;
-    int kv_head = blockIdx.x % p.kv_head;
-    int group_size = blockDim.y;
-    int q_head = kv_head * group_size + threadIdx.y;
-    int lane = threadIdx.x;
-    int hd_per_thread = p.head_dim / 32;
-
-    float q_reg[8];
-    int q_off = ((batch * p.q_head + q_head) * 1) * p.head_dim + lane * hd_per_thread;
-    for (int i = 0; i < hd_per_thread; i++)
-        q_reg[i] = __bfloat162float(p.q[q_off + i]);
-
-    int kv_base = ((batch * p.kv_head + kv_head) * p.kv_len) * p.head_dim;
-    int mask_base = batch * p.kv_len;
-
-    float m = -FLT_MAX, d = 0.0f, acc_reg[8] = {0.0f};
-
-    extern __shared__ __align__(16) bf16 k_smem[];
-
-    for (int chunk_start = 0; chunk_start < p.kv_len; chunk_start += DC_CHUNK) {
-        int this_chunk = min(DC_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)
-            k_smem[i] = p.k[kv_base + chunk_start * p.head_dim + i];
-        __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;
-
-            if (p.use_mask && p.mask && !p.mask[mask_base + chunk_start + s])
-                partial = -FLT_MAX;
-            if (p.is_causal && (chunk_start + s) > p.causal_offset)
-                partial = -FLT_MAX;
-
-            float new_m = fmaxf(m, partial);
-            float alpha = expf(m - new_m);
-            float beta  = expf(partial - new_m);
-            d = d * alpha + beta;
-
-            int v_off = kv_base + (chunk_start + s) * p.head_dim + lane * hd_per_thread;
-            for (int i = 0; i < hd_per_thread; i++)
-                acc_reg[i] = acc_reg[i] * alpha + __bfloat162float(p.v[v_off + i]) * beta;
-            m = new_m;
-        }
-        __syncthreads();
-    }
-
-    int out_off = ((batch * p.q_head + q_head) * 1) * p.head_dim + lane * hd_per_thread;
-    for (int i = 0; i < hd_per_thread; i++)
-        p.o[out_off + i] = __float2bfloat16(acc_reg[i] / d);
-}

csrc/kernels/gqa_decode_attn_mma.cuh

@@ -1,219 +0,0 @@
-#pragma once
-#include "gqa_common.cuh"
-#include "gqa_mma_utils.cuh"
-
-// Tensor-core decode via GQA head-packing with cp.async loads.
-//
-// 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). Fragment layout is identical to the prefill mma kernel; the
-// only differences are (1) the M rows come from different heads at position 0
-// instead of different sequence positions of one head, and (2) causal masking is
-// a single scalar bound shared by every row. One warp owns one (batch, kv_head);
-// requires G <= 16.
-//
-// Optimizations:
-//   - cp.async global→shared for K/V (bypasses registers, cuts instruction count)
-//   - XOR swizzle (swiz_col): LD=HEAD_DIM, zero waste, no bank conflicts
-//   - pre-scaled Q: Q scaled during load, softmax skips per-tile multiply
-//   - single-buffer: keeps smem small for high occupancy
-
-template <int HEAD_DIM, int BC>
-__global__ void gqa_decode_attn_mma_kernel(GQAParams 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 handles bank conflicts, zero waste
-    constexpr int SWIZ_MASK = (HEAD_DIM >= 64) ? 7 : (HEAD_DIM / 8 - 1);
-
-    const int lane = threadIdx.x;  // single warp
-    const int gid = lane >> 2;     // 0..7 → rows gid, gid+8
-    const int tid4 = lane & 3;
-
-    const int kv_head = blockIdx.x;
-    const int batch = blockIdx.y;
-    const int G = p.q_head / p.kv_head;
-    const int q_head0 = kv_head * G;
-
-    extern __shared__ __align__(16) bf16 smem[];
-    bf16* sK = smem;                  // [BC][LD]
-    bf16* sV = sK + BC * LD;          // [BC][LD]
-    bf16* sQ = sV + BC * LD;          // [BR][LD]
-
-    // ---- stage Q into shared (pre-scaled, swizzled) ----
-    bf16 scale_bf16 = __float2bfloat16(p.scale);
-    for (int i = lane; i < BR * HEAD_DIM; i += 32) {
-        int r = i / HEAD_DIM, d = i % HEAD_DIM;
-        bf16 val = __float2bfloat16(0.0f);
-        if (r < G) {
-            int qh = q_head0 + r;
-            val = p.q[(batch * p.q_head + qh) * HEAD_DIM + d];  // q_len == 1
-        }
-        sQ[r * LD + swiz_col(d, r, SWIZ_MASK)] = __hmul(val, scale_bf16);
-    }
-    __syncwarp();
-
-    // Q resident A-fragments
-    unsigned Qa[KD][4];
-    int qrow_l = (lane & 7) + (lane & 8);
-    int qcol_l = (lane & 16) ? 8 : 0;
-#pragma unroll
-    for (int kt = 0; kt < KD; kt++)
-        ldmatrix_x4(Qa[kt], &sQ[qrow_l * LD + swiz_col(kt * 16 + qcol_l, qrow_l, SWIZ_MASK)]);
-
-    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 kv_base = (batch * p.kv_head + kv_head) * p.kv_len * HEAD_DIM;
-    const int mask_base = batch * p.kv_len;
-    const int tiles = (p.kv_len + BC - 1) / BC;
-    const int has_mask = p.use_mask && p.mask;
-
-    for (int ti = 0; ti < tiles; ti++) {
-        int kv0 = ti * BC;
-
-        // ---- load K/V tile to shared (cp.async on full tiles) ----
-        bool full_tile = (kv0 + BC <= p.kv_len);
-        if (full_tile) {
-            constexpr int VEC = 8;  // 8 bf16 = 16 bytes per cp.async
-            int total = BC * HEAD_DIM;
-#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;
-                cp_async_16(&sK[r * LD + swiz_col(d, r, SWIZ_MASK)],
-                            &p.k[kv_base + kc * HEAD_DIM + d]);
-                cp_async_16(&sV[r * LD + swiz_col(d, r, SWIZ_MASK)],
-                            &p.v[kv_base + kc * HEAD_DIM + d]);
-            }
-            cp_async_commit();
-            cp_async_wait_all();
-        } else {
-            for (int i = lane; i < BC * HEAD_DIM; i += 32) {
-                int r = i / HEAD_DIM, d = i % HEAD_DIM;
-                int kc = kv0 + r;
-                bf16 z = __float2bfloat16(0.0f);
-                sK[r * LD + swiz_col(d, r, SWIZ_MASK)] =
-                    (kc < p.kv_len) ? p.k[kv_base + kc * HEAD_DIM + d] : z;
-                sV[r * LD + swiz_col(d, r, SWIZ_MASK)] =
-                    (kc < p.kv_len) ? p.v[kv_base + kc * HEAD_DIM + d] : z;
-            }
-        }
-        __syncwarp();
-
-        // S = Q @ K^T  (Q already pre-scaled, so Sacc includes scale)
-        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 (Q pre-scaled → no per-tile scale multiply) ----
-        float rmax0 = -FLT_MAX, rmax1 = -FLT_MAX;
-#pragma unroll
-        for (int n8 = 0; n8 < NC8; n8++) {
-            int cc = kv0 + n8 * 8 + 2 * tid4;
-            bool bc0 = (cc >= p.kv_len) ||
-                       (has_mask && !p.mask[mask_base + cc]);
-            bool bc1 = (cc + 1 >= p.kv_len) ||
-                       (has_mask && !p.mask[mask_base + cc + 1]);
-            bool cz = p.is_causal;
-            int off = p.causal_offset;
-            bool bad0 = bc0 || (cz && cc > off);
-            bool bad1 = bc1 || (cz && (cc + 1) > off);
-            float s0 = bad0 ? -FLT_MAX : Sacc[n8][0];
-            float s1 = bad1 ? -FLT_MAX : Sacc[n8][1];
-            float s2 = bad0 ? -FLT_MAX : Sacc[n8][2];
-            float s3 = bad1 ? -FLT_MAX : Sacc[n8][3];
-            Sacc[n8][0] = s0; Sacc[n8][1] = s1; Sacc[n8][2] = s2; Sacc[n8][3] = s3;
-            rmax0 = fmaxf(rmax0, fmaxf(s0, s1));
-            rmax1 = fmaxf(rmax1, fmaxf(s2, s3));
-        }
-        rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 1));
-        rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 2));
-        rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 1));
-        rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 2));
-
-        float nm0 = fmaxf(m0, rmax0), nm1 = fmaxf(m1, rmax1);
-        float corr0 = (nm0 == -FLT_MAX) ? 1.0f : __expf(m0 - nm0);
-        float corr1 = (nm1 == -FLT_MAX) ? 1.0f : __expf(m1 - nm1);
-
-        float rsum0 = 0.0f, rsum1 = 0.0f;
-#pragma unroll
-        for (int n8 = 0; n8 < NC8; n8++) {
-            float p0 = (Sacc[n8][0] == -FLT_MAX) ? 0.0f : __expf(Sacc[n8][0] - nm0);
-            float p1 = (Sacc[n8][1] == -FLT_MAX) ? 0.0f : __expf(Sacc[n8][1] - nm0);
-            float p2 = (Sacc[n8][2] == -FLT_MAX) ? 0.0f : __expf(Sacc[n8][2] - nm1);
-            float p3 = (Sacc[n8][3] == -FLT_MAX) ? 0.0f : __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
-#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]);
-            }
-        }
-        __syncwarp();  // sK/sV reused next tile
-    }
-
-    // ---- write output ----
-    float rl0 = (l0 > 1e-20f) ? (1.0f / l0) : 0.0f;
-    float rl1 = (l1 > 1e-20f) ? (1.0f / l1) : 0.0f;
-#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 o_off = (batch * p.q_head + q_head0 + r0) * HEAD_DIM + d;
-            p.o[o_off] = __float2bfloat16(Oacc[dn8][0] * rl0);
-            p.o[o_off + 1] = __float2bfloat16(Oacc[dn8][1] * rl0);
-        }
-        if (r1 < G) {
-            int o_off = (batch * p.q_head + q_head0 + r1) * HEAD_DIM + d;
-            p.o[o_off] = __float2bfloat16(Oacc[dn8][2] * rl1);
-            p.o[o_off + 1] = __float2bfloat16(Oacc[dn8][3] * rl1);
-        }
-    }
-}

csrc/kernels/gqa_mma_utils.cuh

@@ -1,93 +0,0 @@
-#pragma once
-
-// 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));
-}
-
-__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));
-}

csrc/kernels/gqa_prefill_attn.cu

@@ -1,100 +0,0 @@
-#include "gqa_prefill_attn.cuh"
-#include <torch/extension.h>
-
-#ifndef ASTRAI_NO_MMA
-#include "gqa_prefill_attn_mma.cuh"
-#endif
-
-template <int HEAD_DIM>
-static void dispatch_prefill(GQAParams& p) {
-#ifndef ASTRAI_NO_MMA
-    constexpr int WARPS = 4, BC = 32, BR = 16, LD = HEAD_DIM;
-    dim3 grid((p.q_len + BR * WARPS - 1) / (BR * WARPS), p.q_head, p.batch);
-    dim3 block(WARPS * 32, 1, 1);
-    // sK + sV (each BC*LD) + shared sQ staging (BR*LD)
-    int smem = (2 * BC * LD + BR * LD) * (int)sizeof(bf16);
-    cudaFuncSetAttribute(gqa_prefill_attn_mma_kernel<HEAD_DIM, WARPS, BC>,
-                         cudaFuncAttributeMaxDynamicSharedMemorySize, smem);
-    gqa_prefill_attn_mma_kernel<HEAD_DIM, WARPS, BC><<<grid, block, smem>>>(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);
-    size_t smem = 2 * P_BC * HEAD_DIM * sizeof(bf16);
-    gqa_prefill_attn_kernel_t<HEAD_DIM, G, ROWS, P_BC><<<grid, block, smem>>>(p);
-#endif
-}
-
-torch::Tensor gqa_prefill_attn(
-    torch::Tensor q,
-    torch::Tensor k,
-    torch::Tensor v,
-    c10::optional<torch::Tensor> mask,
-    bool is_causal = false,
-    int64_t causal_offset = 0,
-    c10::optional<double> scale = c10::nullopt
-) {
-    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);
-
-    GQAParams p;
-    p.batch = q.size(0);
-    p.q_head = q.size(1);
-    p.kv_head = k.size(1);
-    p.q_len = q.size(2);
-    p.kv_len = k.size(2);
-    p.head_dim = q.size(3);
-    TORCH_CHECK(p.head_dim % 16 == 0, "head_dim must be multiple of 16");
-    p.use_mask = mask.has_value();
-    p.is_causal = (int)is_causal;
-    p.causal_offset = (int)causal_offset;
-    p.scale = scale.has_value() ? (float)scale.value() : 1.0f / sqrtf((float)p.head_dim);
-    p.q = (const bf16*)q.data_ptr();
-    p.k = (const bf16*)k.data_ptr();
-    p.v = (const bf16*)v.data_ptr();
-    if (p.use_mask) {
-        TORCH_CHECK(mask.value().dtype() == torch::kBool);
-        TORCH_CHECK(mask.value().dim() == 2);
-        TORCH_CHECK(mask.value().size(0) == p.batch);
-        TORCH_CHECK(mask.value().size(1) == p.kv_len);
-        p.mask = mask.value().data_ptr<bool>();
-    } else {
-        p.mask = nullptr;
-    }
-
-    auto O = torch::empty_like(q);
-    p.o = (bf16*)O.data_ptr();
-
-    switch (p.head_dim) {
-        case 32:
-            dispatch_prefill<32>(p);
-            break;
-        case 64:
-            dispatch_prefill<64>(p);
-            break;
-        case 128:
-            dispatch_prefill<128>(p);
-            break;
-        case 256:
-            dispatch_prefill<256>(p);
-            break;
-        default:
-            TORCH_CHECK(false, "prefill: unsupported head_dim ", p.head_dim,
-                        " (supported: 32,64,128,256)");
-    }
-    return O;
-}
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-    m.def("gqa_prefill_attn", &gqa_prefill_attn,
-        py::arg("q"),
-        py::arg("k"),
-        py::arg("v"),
-        py::arg("mask") = py::none(),
-        py::arg("is_causal") = false,
-        py::arg("causal_offset") = 0,
-        py::arg("scale") = py::none(),
-        "GQA prefill (tensor-core mma on sm_80+, scalar fallback)");
-}

csrc/kernels/gqa_prefill_attn_mma.cuh

@@ -1,246 +0,0 @@
-#pragma once
-#include "gqa_common.cuh"
-#include "gqa_mma_utils.cuh"
-
-// Tensor-core prefill, register-resident 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 stays resident in registers;
-// S, O, and the online-softmax stats (m, l) live in registers too — nothing is
-// staged through shared memory except the cooperatively-loaded K/V tiles. 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.
-//
-// Optimizations: shared sQ staging (single area, serialized per-warp load)
-// → cuts smem; pre-scale Q by attention scale during Q load; cp.async global→
-// shared for K/V; scalar fallback only for the last partial tile; causal tile
-// skipping (block-level early break + warp-level 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 gqa_prefill_attn_mma_kernel(GQAParams 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;
-
-    extern __shared__ __align__(16) bf16 smem[];
-    bf16* sK = smem;                       // [BC][LD]
-    bf16* sV = sK + BC * LD;               // [BC][LD]
-    bf16* sQ = sV + BC * LD;               // shared staging [BR][LD]
-
-    // Q resident A-fragments (loaded once per warp via shared staging).
-    // Pre-scale by attention scale so softmax doesn't need to multiply later.
-    const int q_base = ((batch * p.q_head + q_head) * p.q_len) * HEAD_DIM;
-    unsigned Qa[KD][4];
-    bf16 scale_bf16 = __float2bfloat16(p.scale);
-    int qrow_l = (lane & 7) + (lane & 8);       // 0..15
-    int qcol_l = (lane & 16) ? 8 : 0;
-    for (int w = 0; w < WARPS; w++) {
-        if (warp == w) {
-            for (int i = lane; i < BR * HEAD_DIM; i += 32) {
-                int r = i / HEAD_DIM, d = i % HEAD_DIM;
-                int qr = qrow0 + r;
-                bf16 qv = (qr < p.q_len) ? p.q[q_base + qr * HEAD_DIM + d]
-                                         : __float2bfloat16(0.0f);
-                sQ[r * LD + swiz_col(d, r, SWIZ_MASK)] = __hmul(qv, scale_bf16);
-            }
-            __syncwarp();
-        #pragma unroll
-            for (int kt = 0; kt < KD; kt++)
-                ldmatrix_x4(Qa[kt], &sQ[qrow_l * LD + swiz_col(kt * 16 + qcol_l, qrow_l, SWIZ_MASK)]);
-        }
-        __syncthreads();  // prevent next warp from overwriting sQ prematurely
-    }
-
-    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 kv_base = ((batch * p.kv_head + kv_head) * p.kv_len) * HEAD_DIM;
-    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 is_causal == 0)
-    const int use_skip = p.is_causal;
-    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;
-    const int mb = batch * p.kv_len;
-
-    for (int ti = 0; ti < tiles; ti++) {
-        int kv0 = ti * BC;
-
-        // Block-level causal early break
-        if (use_skip && kv0 > block_max_kv) break;
-
-        // ---- load K/V tile to shared memory (cp.async on full tiles) ----
-        bool full_tile = (kv0 + BC <= p.kv_len);
-        if (full_tile) {
-            constexpr int VEC = 8;  // bf16 per cp.async unit (16 bytes)
-            int total = BC * HEAD_DIM;
-#pragma unroll
-            for (int i = threadIdx.x * VEC; i < total; i += nthreads * VEC) {
-                int r = i / HEAD_DIM;
-                int d = i % HEAD_DIM;
-                int kc = kv0 + r;
-                cp_async_16(&sK[r * LD + swiz_col(d, r, SWIZ_MASK)], &p.k[kv_base + kc * HEAD_DIM + d]);
-                cp_async_16(&sV[r * LD + swiz_col(d, r, SWIZ_MASK)], &p.v[kv_base + kc * HEAD_DIM + d]);
-            }
-            cp_async_commit();
-            cp_async_wait_all();
-        } else {
-            for (int i = threadIdx.x; i < BC * HEAD_DIM; i += nthreads) {
-                int r = i / HEAD_DIM, d = i % HEAD_DIM;
-                int kc = kv0 + r;
-                bf16 z = __float2bfloat16(0.0f);
-                sK[r * LD + swiz_col(d, r, SWIZ_MASK)] = (kc < p.kv_len)
-                                     ? p.k[kv_base + kc * HEAD_DIM + d] : z;
-                sV[r * LD + swiz_col(d, r, SWIZ_MASK)] = (kc < p.kv_len)
-                                     ? p.v[kv_base + kc * HEAD_DIM + d] : z;
-            }
-        }
-        __syncthreads();
-
-        // Warp-level causal skip
-        if (!use_skip || kv0 <= max_kv) {
-
-        // S = Q @ K^T  → Sacc[n8][0..3]   (n8: 8 kv cols each)
-        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 (in registers) ----
-        // Q is pre-scaled, so Sacc already includes the attention scale.
-        int maxc0 = p.is_causal ? min(p.kv_len, qr0 + p.causal_offset + 1)
-                                : p.kv_len;
-        int maxc1 = p.is_causal ? min(p.kv_len, qr1 + p.causal_offset + 1)
-                                : p.kv_len;
-        float rmax0 = -FLT_MAX, rmax1 = -FLT_MAX;
-#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 && !p.mask[mb + cc]);
-            bool b1 = (c1 >= maxc0) || (has_mask && !p.mask[mb + c1]);
-            bool b2 = (cc >= maxc1) || (has_mask && !p.mask[mb + cc]);
-            bool b3 = (c1 >= maxc1) || (has_mask && !p.mask[mb + c1]);
-            float s0 = b0 ? -FLT_MAX : Sacc[n8][0];
-            float s1 = b1 ? -FLT_MAX : Sacc[n8][1];
-            float s2 = b2 ? -FLT_MAX : Sacc[n8][2];
-            float s3 = b3 ? -FLT_MAX : Sacc[n8][3];
-            Sacc[n8][0] = s0; Sacc[n8][1] = s1;
-            Sacc[n8][2] = s2; Sacc[n8][3] = s3;
-            rmax0 = fmaxf(rmax0, fmaxf(s0, s1));
-            rmax1 = fmaxf(rmax1, fmaxf(s2, s3));
-        }
-        rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 1));
-        rmax0 = fmaxf(rmax0, __shfl_xor_sync(0xFFFFFFFF, rmax0, 2));
-        rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 1));
-        rmax1 = fmaxf(rmax1, __shfl_xor_sync(0xFFFFFFFF, rmax1, 2));
-
-        float nm0 = fmaxf(m0, rmax0), nm1 = fmaxf(m1, rmax1);
-        float corr0 = (nm0 == -FLT_MAX) ? 1.0f : __expf(m0 - nm0);
-        float corr1 = (nm1 == -FLT_MAX) ? 1.0f : __expf(m1 - nm1);
-
-        float rsum0 = 0.0f, rsum1 = 0.0f;
-#pragma unroll
-        for (int n8 = 0; n8 < NC8; n8++) {
-            float p0 = (Sacc[n8][0] == -FLT_MAX) ? 0.0f
-                                                 : __expf(Sacc[n8][0] - nm0);
-            float p1 = (Sacc[n8][1] == -FLT_MAX) ? 0.0f
-                                                 : __expf(Sacc[n8][1] - nm0);
-            float p2 = (Sacc[n8][2] == -FLT_MAX) ? 0.0f
-                                                 : __expf(Sacc[n8][2] - nm1);
-            float p3 = (Sacc[n8][3] == -FLT_MAX) ? 0.0f
-                                                 : __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;
-
-        // rescale O accumulator by per-row correction
-#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
-#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]);
-            }
-        }
-        }  // if active (warp-level causal skip)
-        __syncthreads();
-    }
-
-    // ---- write output ----
-    float rl0 = (l0 > 1e-20f) ? (1.0f / l0) : 0.0f;
-    float rl1 = (l1 > 1e-20f) ? (1.0f / l1) : 0.0f;
-    const int o_base = ((batch * p.q_head + q_head) * p.q_len) * HEAD_DIM;
-#pragma unroll
-    for (int dn8 = 0; dn8 < DN8; dn8++) {
-        int d = dn8 * 8 + 2 * tid4;
-        if (qr0 < p.q_len) {
-            p.o[o_base + qr0 * HEAD_DIM + d] =
-                __float2bfloat16(Oacc[dn8][0] * rl0);
-            p.o[o_base + qr0 * HEAD_DIM + d + 1] =
-                __float2bfloat16(Oacc[dn8][1] * rl0);
-        }
-        if (qr1 < p.q_len) {
-            p.o[o_base + qr1 * HEAD_DIM + d] =
-                __float2bfloat16(Oacc[dn8][2] * rl1);
-            p.o[o_base + qr1 * HEAD_DIM + d + 1] =
-                __float2bfloat16(Oacc[dn8][3] * rl1);
-        }
-    }
-}

csrc/tests/attn_decode_test.cu

@@ -0,0 +1,185 @@
+/*
+Pure-C test — uses shared dispatcher.
+nvcc -I csrc -arch=sm_89 -O3 \
+    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
+    csrc/tests/attn_decode_test.cu -o test && ./test
+*/
+
+#include "test_utils.cuh"
+#include "../kernels/attn_dispatchers.cuh"
+
+// Split-K scratch (torch-free)
+struct DecodeScratch {
+    float* o_part = nullptr;
+    float* ml_part = nullptr;
+};
+
+static void setup_scratch(AttentionParams<bf16>& p, DecodeScratch& sc) {
+    int max_splits = 32;
+    cudaMalloc(&sc.o_part, (size_t)p.batch * p.q_head * max_splits * p.head_dim * sizeof(float));
+    cudaMalloc(&sc.ml_part, (size_t)p.batch * p.q_head * max_splits * 2 * sizeof(float));
+}
+
+static void free_scratch(DecodeScratch& sc) {
+    cudaFree(sc.o_part); cudaFree(sc.ml_part);
+}
+
+// Warmed-up, CUDA-event timed sweep over the production decode MMA path.
+static void bench() {
+    const int cfgs[][5] = {
+        {1, 32, 4, 512, 128},
+        {1, 32, 4, 1024, 128},
+        {1, 32, 4, 2048, 128},
+        {1, 32, 4, 4096, 128},
+        {16, 32, 4, 2048, 128},
+        {32, 32, 4, 1024, 128},
+    };
+    const int WARMUP = 10, ITERS = 100;
+    printf("\n===== DECODE BENCH (warmup=%d iters=%d) =====\n", WARMUP, ITERS);
+    print_bench_header();
+
+    for (int ci = 0; ci < 6; ci++) {
+        int B = cfgs[ci][0], Hq = cfgs[ci][1], Hk = cfgs[ci][2];
+        int sl = cfgs[ci][3], D = cfgs[ci][4];
+        size_t nQ = (size_t)B * Hq * D;
+        size_t nKV = (size_t)B * Hk * sl * D;
+
+        bf16 *dQ, *dK, *dV, *dO;
+        cudaMalloc(&dQ, nQ*2); cudaMalloc(&dK, nKV*2);
+        cudaMalloc(&dV, nKV*2); cudaMalloc(&dO, nQ*2);
+        size_t big = nQ > nKV ? nQ : nKV; bf16* tmp = new bf16[big];
+        for (size_t i = 0; i < nQ; i++)  tmp[i] = f2bf(randf());
+        cudaMemcpy(dQ, tmp, nQ*2, cudaMemcpyHostToDevice);
+        for (size_t i = 0; i < nKV; i++) tmp[i] = f2bf(randf());
+        cudaMemcpy(dK, tmp, nKV*2, cudaMemcpyHostToDevice);
+        for (size_t i = 0; i < nKV; i++) tmp[i] = f2bf(randf());
+        cudaMemcpy(dV, tmp, nKV*2, cudaMemcpyHostToDevice);
+        delete[] tmp;
+
+        AttentionParams<bf16> p;
+        p.batch = B; p.q_head = Hq; p.kv_head = Hk; p.q_len = 1; p.kv_len = sl;
+        p.head_dim = D; p.use_mask = 0; p.causal_offset = -1;
+        p.scale = 1.0f / sqrtf((float)D);
+        set_default_strides(p);
+        p.q = dQ; p.k = dK; p.v = dV; p.mask = nullptr; p.o = dO;
+
+        DecodeScratch sc;
+        setup_scratch(p, sc);
+        p.o_part = sc.o_part; p.ml_part = sc.ml_part;
+
+        auto launch = [&]() { dispatch_by_head_dim(D, [&]<int H>() { dispatch_decode<H>(p); }); };
+        double flops = 4.0 * B * Hq * (double)sl * D;
+        double bytes = 2.0 * (2.0 * nKV * sizeof(bf16));
+        BenchResult r = bench_kernel(launch, WARMUP, ITERS, flops, bytes);
+
+        char cfg[64];
+        snprintf(cfg, sizeof(cfg),
+                 "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d causal=%d",
+                 B, Hq, Hk, 1, sl, D, 0);
+        print_bench_row(cfg, r);
+
+        cudaFree(dQ); cudaFree(dK); cudaFree(dV); cudaFree(dO);
+        free_scratch(sc);
+    }
+}
+
+static int run_test(int B, int Hq, int Hk, int sl, int D, int causal) {
+    int gs = Hq / Hk;
+    printf("=== B=%d Hq=%d Hk=%d seq=%d D=%d gs=%d causal=%d ===\n",
+           B,Hq,Hk,sl,D,gs,causal);
+
+    size_t nQ = B*Hq*1*D, nKV = B*Hk*sl*D;
+    float *hQ=new float[nQ], *hK=new float[nKV], *hV=new float[nKV];
+    for (size_t i=0;i<nQ;i++) hQ[i]=randf();
+    for (size_t i=0;i<nKV;i++){hK[i]=randf();hV[i]=randf();}
+
+    bool* hMask=new bool[B*sl];
+    for (int i=0;i<B*sl;i++) hMask[i]=true;
+
+    bf16 *dQ,*dK,*dV,*dO,*tmp;
+    bool* dMask;
+    cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+    cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+    cudaMalloc(&dMask,B*sl);
+
+    tmp=new bf16[max(nQ,nKV)];
+    for (size_t i=0;i<nQ;i++) tmp[i]=f2bf(hQ[i]);
+    cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hK[i]);
+    cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hV[i]);
+    cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+    cudaMemcpy(dMask,hMask,B*sl,cudaMemcpyHostToDevice);
+
+    AttentionParams<bf16> p;
+    p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=1; p.kv_len=sl; p.head_dim=D;
+    p.use_mask=0; p.causal_offset=causal?0:-1;
+    p.scale=1.0f/sqrtf((float)D);
+    set_default_strides(p);
+    p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
+
+    DecodeScratch sc;
+    setup_scratch(p, sc);
+    p.o_part = sc.o_part; p.ml_part = sc.ml_part;
+
+    double t0=now_ms();
+    dispatch_by_head_dim(D, [&]<int H>() { dispatch_decode<H>(p); });
+    cudaDeviceSynchronize();
+    double kms=now_ms()-t0;
+    cudaError_t err=cudaGetLastError();
+    if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return 1;}
+
+    bf16* hOut=new bf16[nQ];
+    cudaMemcpy(hOut,dO,nQ*2,cudaMemcpyDeviceToHost);
+
+    float* ref=new float[nQ];
+    cpu_attention_ref(hQ, hK, hV, hMask, ref, B, Hq, Hk, 1, sl, D, causal ? 0 : -1);
+
+    float max_abs_err=0, max_rel_err=0;
+    for (size_t i=0;i<nQ;i++){
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if(err>max_abs_err) max_abs_err=err;
+        float rel=err/fmaxf(fabsf(ref[i]), 1e-8f);
+        if(rel>max_rel_err) max_rel_err=rel;
+    }
+    const float atol=0.01f, rtol=0.01f;
+    bool pass=true;
+    for (size_t i=0;i<nQ;i++){
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if (err > atol + rtol * fabsf(ref[i])) { pass=false; break; }
+    }
+    printf("kernel: %.3f ms  max_abs_err: %.6e  max_rel_err: %.6e  %s\n\n",
+           kms, max_abs_err, max_rel_err, pass?"PASS":"FAIL");
+
+    cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);cudaFree(dMask);
+    free_scratch(sc);
+    delete[]hQ;delete[]hK;delete[]hV;delete[]hMask;delete[]hOut;delete[]ref;delete[]tmp;
+
+    return pass ? 0 : 1;
+}
+
+int main() {
+    const int configs[][6] = {
+        {1, 2, 1, 64, 32, 0},
+        {1, 32, 4, 512, 128, 0},
+        {1, 32, 4, 1024, 128, 0},
+        {1, 32, 4, 512, 128, 1},
+    };
+    int n_cfgs = sizeof(configs) / sizeof(configs[0]);
+    int fail = 0;
+
+    for (int ci = 0; ci < n_cfgs; ci++) {
+        int B = configs[ci][0], Hq = configs[ci][1], Hk = configs[ci][2];
+        int sl = configs[ci][3], D = configs[ci][4], causal = configs[ci][5];
+        fail += run_test(B, Hq, Hk, sl, D, causal);
+        if (fail) break;
+    }
+
+    if (fail) {
+        printf("FAILED\n");
+        return fail;
+    }
+    printf("All tests passed!\n");
+    bench();
+    return 0;
+}

csrc/tests/attn_paged_decode_test.cu

@@ -0,0 +1,308 @@
+// Compile:
+//   nvcc -I csrc -arch=sm_89 -O3 --use_fast_math --ptxas-options=-O3 \
+//        --extra-device-vectorization csrc/tests/attn_paged_decode_test.cu \
+//        -o /tmp/test_paged && /tmp/test_paged
+
+#include <cstring>
+#include "test_utils.cuh"
+#include "../kernels/attn_dispatchers.cuh"
+
+static void gather_kv_cpu(
+    const bf16* h_k_pool, const bf16* h_v_pool,
+    const int64_t* h_pt, int B, int Hkv, int kv_len,
+    int page_size, int head_dim,
+    bf16* h_k, bf16* h_v)
+{
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    size_t page_stride = (size_t)page_size * Hkv * head_dim;
+    for (int b = 0; b < B; b++) {
+        for (int pos = 0; pos < kv_len; pos++) {
+            int log_pg = pos / page_size;
+            int pg_off = pos % page_size;
+            int phys = (int)h_pt[b * max_pages + log_pg];
+            for (int h = 0; h < Hkv; h++) {
+                size_t src_base = (size_t)phys * page_stride
+                                + (size_t)pg_off * Hkv * head_dim
+                                + h * head_dim;
+                size_t dst_base = ((size_t)b * Hkv + h) * kv_len * head_dim
+                                  + (size_t)pos * head_dim;
+                memcpy(h_k + dst_base, h_k_pool + src_base, head_dim * sizeof(bf16));
+                memcpy(h_v + dst_base, h_v_pool + src_base, head_dim * sizeof(bf16));
+            }
+        }
+    }
+}
+
+template <int HEAD_DIM>
+static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int causal, int seed) {
+    printf("B=%d Hq=%d Hkv=%d kv_len=%d page_sz=%d head_dim=%d causal=%d ... ",
+           B, Hq, Hkv, kv_len, page_size, HEAD_DIM, causal);
+    fflush(stdout);
+
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    int n_phys_pages = B * max_pages;
+    int max_splits = 32;
+
+    size_t sz_q  = (size_t)B * Hq * 1 * HEAD_DIM * sizeof(bf16);
+    size_t sz_o  = sz_q;
+    size_t sz_kv = (size_t)n_phys_pages * page_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_pt = (size_t)B * max_pages * sizeof(int64_t);
+    size_t sz_op = (size_t)B * Hq * max_splits * HEAD_DIM * sizeof(float);
+    size_t sz_ml = (size_t)B * Hq * max_splits * 2 * sizeof(float);
+
+    bf16 *d_q, *d_o_paged;
+    bf16 *d_k_pool, *d_v_pool;
+    int64_t* d_pt;
+    float *d_op, *d_ml;
+
+    cudaMalloc(&d_q, sz_q);
+    cudaMalloc(&d_o_paged, sz_o);
+    cudaMalloc(&d_k_pool, sz_kv);
+    cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_pt, sz_pt);
+    cudaMalloc(&d_op, sz_op);
+    cudaMalloc(&d_ml, sz_ml);
+
+    srand(seed);
+    auto rnd = [&]() { return (rand() / (float)RAND_MAX) * 2.0f - 1.0f; };
+
+    bf16* h_q = (bf16*)malloc(sz_q);
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++)
+        h_q[i] = __float2bfloat16(rnd());
+    cudaMemcpy(d_q, h_q, sz_q, cudaMemcpyHostToDevice);
+
+    bf16* h_k_pool = (bf16*)malloc(sz_kv);
+    bf16* h_v_pool = (bf16*)malloc(sz_kv);
+    size_t ps = (size_t)page_size * Hkv * HEAD_DIM;
+    for (int pg = 0; pg < n_phys_pages; pg++) {
+        for (int off = 0; off < page_size; off++) {
+            for (int h = 0; h < Hkv; h++) {
+                for (int d = 0; d < HEAD_DIM; d++) {
+                    float v = sinf((float)(pg * 7919 + off * 1049 + h * 331 + d));
+                    size_t idx = (size_t)pg * ps + (size_t)off * Hkv * HEAD_DIM
+                                 + h * HEAD_DIM + d;
+                    h_k_pool[idx] = __float2bfloat16(v);
+                    h_v_pool[idx] = __float2bfloat16(v * 0.3f);
+                }
+            }
+        }
+    }
+    cudaMemcpy(d_k_pool, h_k_pool, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, h_v_pool, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_pt = (int64_t*)malloc(sz_pt);
+    int next_pg = 0;
+    for (int b = 0; b < B; b++)
+        for (int p = 0; p < max_pages; p++)
+            h_pt[b * max_pages + p] = next_pg++;
+    cudaMemcpy(d_pt, h_pt, sz_pt, cudaMemcpyHostToDevice);
+
+    bf16* h_k_cont = (bf16*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(bf16));
+    bf16* h_v_cont = (bf16*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(bf16));
+    gather_kv_cpu(h_k_pool, h_v_pool, h_pt, B, Hkv, kv_len, page_size, HEAD_DIM, h_k_cont, h_v_cont);
+
+    float* h_q_f = (float*)malloc((size_t)B * Hq * HEAD_DIM * sizeof(float));
+    float* h_k_f = (float*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(float));
+    float* h_v_f = (float*)malloc((size_t)B * kv_len * Hkv * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) h_q_f[i] = bf2f(h_q[i]);
+    for (int i = 0; i < B * kv_len * Hkv * HEAD_DIM; i++) {
+        h_k_f[i] = bf2f(h_k_cont[i]);
+        h_v_f[i] = bf2f(h_v_cont[i]);
+    }
+
+    float* h_o_ref = (float*)calloc(B * Hq * HEAD_DIM, sizeof(float));
+    cpu_attention_ref(h_q_f, h_k_f, h_v_f, nullptr, h_o_ref, B, Hq, Hkv,
+                      1, kv_len, HEAD_DIM, causal ? 0 : -1);
+
+    PagedAttentionParams<bf16> p;
+    p.batch = B; p.q_head = Hq; p.kv_head = Hkv; p.q_len = 1;
+    p.kv_len = kv_len; p.head_dim = HEAD_DIM;
+    p.use_mask = 0; p.causal_offset = causal ? 0 : -1;
+    set_default_paged_strides(p);
+    p.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    p.page_size = page_size; p.max_pages = max_pages;
+    p.page_table = d_pt;
+    p.k_cache = d_k_pool; p.v_cache = d_v_pool;
+    p.q = d_q; p.mask = nullptr; p.o = d_o_paged;
+    p.o_part = d_op; p.ml_part = d_ml;
+
+    dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(p); });
+    cudaDeviceSynchronize();
+
+    bf16* h_o_bf16 = (bf16*)malloc(sz_o);
+    cudaMemcpy(h_o_bf16, d_o_paged, sz_o, cudaMemcpyDeviceToHost);
+    float* h_o_paged = (float*)malloc(B * Hq * HEAD_DIM * sizeof(float));
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++)
+        h_o_paged[i] = __bfloat162float(h_o_bf16[i]);
+
+    float max_abs_err = 0.0f, max_rel_err = 0.0f;
+    int bad_idx = -1;
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_paged[i] - h_o_ref[i]);
+        if (e > max_abs_err) { max_abs_err = e; bad_idx = i; }
+        float rel = e / fmaxf(fabsf(h_o_ref[i]), 1e-8f);
+        if (rel > max_rel_err) max_rel_err = rel;
+    }
+
+    const float atol = 0.01f, rtol = 0.01f;
+    bool pass = true;
+    for (int i = 0; i < B * Hq * HEAD_DIM; i++) {
+        float e = fabsf(h_o_paged[i] - h_o_ref[i]);
+        if (e > atol + rtol * fabsf(h_o_ref[i])) { pass = false; break; }
+    }
+
+    if (pass) {
+        printf("PASS (max_abs_err=%.4e max_rel_err=%.4e)\n", max_abs_err, max_rel_err);
+    } else {
+        int b = bad_idx / (Hq * HEAD_DIM);
+        int h = (bad_idx / HEAD_DIM) % Hq;
+        int d = bad_idx % HEAD_DIM;
+        printf("FAIL (max_abs_err=%.4e max_rel_err=%.4e at [%d,%d,%d]: ref=%.4f got=%.4f)\n",
+               max_abs_err, max_rel_err, b, h, d, h_o_ref[bad_idx], h_o_paged[bad_idx]);
+        printf("  ref[0..7]:");
+        for (int i = 0; i < 8 && i < HEAD_DIM; i++)
+            printf(" %.4f", h_o_ref[i]);
+        printf("\n  got[0..7]:");
+        for (int i = 0; i < 8 && i < HEAD_DIM; i++)
+            printf(" %.4f", h_o_paged[i]);
+        printf("\n");
+    }
+
+    free(h_q); free(h_k_pool); free(h_v_pool); free(h_pt);
+    free(h_k_cont); free(h_v_cont);
+    free(h_q_f); free(h_k_f); free(h_v_f);
+    free(h_o_ref); free(h_o_bf16); free(h_o_paged);
+    cudaFree(d_q); cudaFree(d_o_paged);
+    cudaFree(d_k_pool); cudaFree(d_v_pool); cudaFree(d_pt);
+    cudaFree(d_op); cudaFree(d_ml);
+
+    return pass ? 0 : 1;
+}
+
+struct TestCase {
+    int head_dim;
+    int B, Hq, Hkv, kv_len, page_size, causal, seed;
+};
+
+static const TestCase TESTS[] = {
+    {128, 1, 1, 1, 8, 128, 0, 1},
+    {128, 1, 4, 4, 128, 128, 0, 2},
+    {128, 2, 4, 4, 256, 128, 0, 3},
+    {128, 1, 4, 1, 64, 64, 0, 4},
+    {128, 1, 8, 2, 64, 128, 0, 5},
+    {128, 2, 16, 4, 128, 128, 0, 6},
+    {64, 1, 4, 2, 32, 128, 0, 7},
+    {256, 1, 2, 1, 16, 128, 0, 8},
+    {32, 1, 4, 2, 32, 64, 0, 9},
+    {128, 3, 8, 2, 256, 128, 0, 10},
+    {128, 2, 32, 8, 512, 128, 0, 11},
+    {128, 1, 16, 2, 256, 128, 0, 12},
+    {128, 2, 32, 4, 512, 128, 0, 13},
+    {128, 2, 8, 2, 128, 128, 1, 14},   // causal
+};
+
+static int dispatch_test(const TestCase& tc) {
+    int r = 0;
+    dispatch_by_head_dim(tc.head_dim, [&]<int D>() {
+        r = run_test<D>(tc.B, tc.Hq, tc.Hkv, tc.kv_len, tc.page_size, tc.causal, tc.seed);
+    });
+    return r;
+}
+
+template <int HEAD_DIM>
+static void bench_config(int B, int Hq, int Hkv, int kv_len, int page_size) {
+    int max_pages = (kv_len + page_size - 1) / page_size;
+    int n_phys_pages = B * max_pages;
+    int max_splits = 32;
+
+    size_t sz_q  = (size_t)B * Hq * 1 * HEAD_DIM * sizeof(bf16);
+    size_t sz_kv = (size_t)n_phys_pages * page_size * Hkv * HEAD_DIM * sizeof(bf16);
+    size_t sz_pt = (size_t)B * max_pages * sizeof(int64_t);
+    size_t sz_op = (size_t)B * Hq * max_splits * HEAD_DIM * sizeof(float);
+    size_t sz_ml = (size_t)B * Hq * max_splits * 2 * sizeof(float);
+
+    bf16 *d_q, *d_o, *d_k_pool, *d_v_pool;
+    int64_t* d_pt;
+    float *d_op, *d_ml;
+    cudaMalloc(&d_q, sz_q); cudaMalloc(&d_o, sz_q);
+    cudaMalloc(&d_k_pool, sz_kv); cudaMalloc(&d_v_pool, sz_kv);
+    cudaMalloc(&d_pt, sz_pt);
+    cudaMalloc(&d_op, sz_op); cudaMalloc(&d_ml, sz_ml);
+
+    bf16* tmp = (bf16*)malloc(sz_kv > sz_q ? sz_kv : sz_q);
+    for (size_t i = 0; i < sz_q / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_q, tmp, sz_q, cudaMemcpyHostToDevice);
+    for (size_t i = 0; i < sz_kv / sizeof(bf16); i++) tmp[i] = f2bf(randf());
+    cudaMemcpy(d_k_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+    cudaMemcpy(d_v_pool, tmp, sz_kv, cudaMemcpyHostToDevice);
+
+    int64_t* h_pt = (int64_t*)malloc(sz_pt);
+    int next_pg = 0;
+    for (int b = 0; b < B; b++)
+        for (int p = 0; p < max_pages; p++)
+            h_pt[b * max_pages + p] = next_pg++;
+    cudaMemcpy(d_pt, h_pt, sz_pt, cudaMemcpyHostToDevice);
+    free(h_pt);
+
+    PagedAttentionParams<bf16> pa;
+    pa.batch = B; pa.q_head = Hq; pa.kv_head = Hkv; pa.q_len = 1;
+    pa.kv_len = kv_len; pa.head_dim = HEAD_DIM;
+    pa.use_mask = 0; pa.causal_offset = -1;
+    set_default_paged_strides(pa);
+    pa.scale = 1.0f / sqrtf((float)HEAD_DIM);
+    pa.page_size = page_size; pa.max_pages = max_pages;
+    pa.page_table = d_pt;
+    pa.k_cache = d_k_pool; pa.v_cache = d_v_pool;
+    pa.q = d_q; pa.mask = nullptr; pa.o = d_o;
+    pa.o_part = d_op; pa.ml_part = d_ml;
+
+    const int WARMUP = 10, ITERS = 100;
+    auto launch = [&]() {
+        dispatch_by_head_dim(HEAD_DIM, [&]<int H>() { dispatch_paged_decode<H>(pa); });
+    };
+    double flops = 4.0 * B * Hq * (double)kv_len * HEAD_DIM;
+    size_t nKV = (size_t)B * Hkv * kv_len * HEAD_DIM;
+    double bytes = 2.0 * (2.0 * nKV * sizeof(bf16));
+    BenchResult r = bench_kernel(launch, WARMUP, ITERS, flops, bytes);
+
+    char cfg[64];
+    snprintf(cfg, sizeof(cfg),
+             "B=%2d Hq=%2d Hk=%d q=%4d kv=%4d D=%3d page=%3d",
+             B, Hq, Hkv, 1, kv_len, HEAD_DIM, page_size);
+    print_bench_row(cfg, r);
+
+    free(tmp);
+    cudaFree(d_q); cudaFree(d_o);
+    cudaFree(d_k_pool); cudaFree(d_v_pool); cudaFree(d_pt);
+    cudaFree(d_op); cudaFree(d_ml);
+}
+
+static void bench() {
+    printf("\n===== PAGED DECODE BENCH =====\n");
+    print_bench_header();
+    bench_config<128>(1, 32, 4, 512, 128);
+    bench_config<128>(1, 32, 4, 1024, 128);
+    bench_config<128>(1, 32, 4, 2048, 128);
+    bench_config<128>(1, 32, 4, 4096, 128);
+    bench_config<128>(16, 32, 4, 2048, 128);
+    bench_config<128>(32, 32, 4, 1024, 128);
+}
+
+int main() {
+    int n = sizeof(TESTS) / sizeof(TESTS[0]);
+    int fail = 0;
+    printf("=== Paged Decode vs CPU reference (%d cases) ===\n\n", n);
+
+    for (int i = 0; i < n; i++) {
+        fail += dispatch_test(TESTS[i]);
+        if (fail) break;
+    }
+
+    if (fail) {
+        printf("\nFAILED (%d/%d tests failed)\n", fail, n);
+        return fail;
+    }
+    printf("\nAll %d tests passed!\n", n);
+    bench();
+    return 0;
+}

csrc/tests/attn_prefill_test.cu

@@ -0,0 +1,169 @@
+/*
+Pure-C test — uses shared dispatcher.
+nvcc -I csrc -arch=sm_89 -O3 \
+    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
+    csrc/tests/attn_prefill_test.cu -o test && ./test
+*/
+
+#include "test_utils.cuh"
+#include "../kernels/attn_dispatchers.cuh"
+
+// Warmed-up, CUDA-event timed throughput sweep over the production MMA path.
+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;
+
+        auto launch = [&]() { dispatch_by_head_dim(D, [&]<int H>() { dispatch_prefill<H>(p); }); };
+        for (int i=0;i<WARMUP;i++) launch();
+        cudaDeviceSynchronize();
+        cudaError_t err=cudaGetLastError();
+        if (err!=cudaSuccess){printf("CUDA err: %s\n",cudaGetErrorString(err));return;}
+
+        cudaEvent_t s,e; cudaEventCreate(&s); cudaEventCreate(&e);
+        cudaEventRecord(s);
+        for (int i=0;i<ITERS;i++) launch();
+        cudaEventRecord(e); cudaEventSynchronize(e);
+        float ms=0; cudaEventElapsedTime(&ms,s,e); ms/=ITERS;
+
+        double flops = 4.0*B*Hq*(double)ql*kl*D;
+        if (causal) flops *= 0.5;
+        double tflops = flops/(ms*1e-3)/1e12;
+        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);
+    }
+}
+
+static int run_test(int B, int Hq, int Hk, int ql, int kl, int D, int causal) {
+    printf("=== B=%d Hq=%d Hk=%d q=%d kv=%d D=%d causal=%d ===\n",
+           B,Hq,Hk,ql,kl,D,causal);
+
+    size_t nQ = B*Hq*ql*D, nKV = B*Hk*kl*D;
+    float *hQ=new float[nQ], *hK=new float[nKV], *hV=new float[nKV];
+    for (size_t i=0;i<nQ;i++) hQ[i]=randf();
+    for (size_t i=0;i<nKV;i++){hK[i]=randf();hV[i]=randf();}
+
+    bf16 *dQ,*dK,*dV,*dO,*tmp;
+    cudaMalloc(&dQ,nQ*2); cudaMalloc(&dK,nKV*2);
+    cudaMalloc(&dV,nKV*2); cudaMalloc(&dO,nQ*2);
+    tmp=new bf16[max(nQ,nKV)];
+    for (size_t i=0;i<nQ;i++) tmp[i]=f2bf(hQ[i]);
+    cudaMemcpy(dQ,tmp,nQ*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hK[i]);
+    cudaMemcpy(dK,tmp,nKV*2,cudaMemcpyHostToDevice);
+    for (size_t i=0;i<nKV;i++) tmp[i]=f2bf(hV[i]);
+    cudaMemcpy(dV,tmp,nKV*2,cudaMemcpyHostToDevice);
+
+    AttentionParams<bf16> p;
+    p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=ql; p.kv_len=kl; p.head_dim=D;
+    p.use_mask=0; p.causal_offset=causal?0:-1;
+    set_default_strides(p);
+    p.scale=1.0f/sqrtf((float)D);
+    p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
+
+    double t0=now_ms();
+    dispatch_by_head_dim(D, [&]<int H>() { dispatch_prefill<H>(p); });
+    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_abs_err=0, max_rel_err=0;
+    for (size_t i=0;i<nQ;i++) {
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if(err>max_abs_err) max_abs_err=err;
+        float rel=err/fmaxf(fabsf(ref[i]), 1e-8f);
+        if(rel>max_rel_err) max_rel_err=rel;
+    }
+    const float atol=0.01f, rtol=0.01f;
+    bool pass=true;
+    for (size_t i=0;i<nQ;i++) {
+        float err=fabsf(bf2f(hOut[i])-ref[i]);
+        if (err > atol + rtol * fabsf(ref[i])) { pass=false; break; }
+    }
+    printf("kernel: %.3f ms  max_abs_err: %.6e  max_rel_err: %.6e  %s\n\n",
+           kms, max_abs_err, max_rel_err, pass?"PASS":"FAIL");
+
+    cudaFree(dQ);cudaFree(dK);cudaFree(dV);cudaFree(dO);
+    delete[]hQ;delete[]hK;delete[]hV;delete[]hOut;delete[]ref;delete[]tmp;
+
+    return pass ? 0 : 1;
+}
+
+int main() {
+    const int configs[][7] = {
+        {1,2,1,64,128,64,0},     // tiny: B,Hq,Hk,q,kv,D,causal
+        {1,32,4,512,512,128,0},  // standard
+        {1,32,4,128,256,128,0},  // medium
+        {1,4,2,256,256,128,1},   // causal
+    };
+    int n_configs = sizeof(configs) / sizeof(configs[0]);
+    int fail = 0;
+
+    for (int ci = 0; ci < n_configs; ci++) {
+        int B=configs[ci][0], Hq=configs[ci][1], Hk=configs[ci][2];
+        int ql=configs[ci][3], kl=configs[ci][4], D=configs[ci][5];
+        int causal=configs[ci][6];
+        fail += run_test(B, Hq, Hk, ql, kl, D, causal);
+        if (fail) break;
+    }
+
+    if (fail) {
+        printf("FAILED\n");
+        return fail;
+    }
+    printf("All tests passed!\n");
+    bench();
+    return 0;
+}

csrc/tests/gqa_decode_test.cu

@@ -1,130 +0,0 @@
-/*
-Pure-C test:
-nvcc -I csrc -arch=sm_89 -O3 \
-    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
-    csrc/tests/gqa_decode_test.cu -o test && ./test
-*/
-
-#include <cstdio>
-#include <cstdlib>
-#include <cmath>
-#include <sys/time.h>
-#include "../kernels/gqa_decode_attn.cuh"
-
-static double now_ms() {
-    struct timeval tv;
-    gettimeofday(&tv, NULL);
-    return tv.tv_sec * 1000.0 + tv.tv_usec / 1000.0;
-}
-
-static void cpu_decode(const float* Q, const float* K, const float* V,
-                        const bool* mask, float* O,
-                        int B, int Hq, int Hk, int seq_len, int D) {
-    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;
-            float mv = -INFINITY, sv = 0.0f;
-            float accum[256] = {0};
-            for (int s = 0; s < seq_len; s++) {
-                if (!mask[b * seq_len + s]) continue;
-                float dot = 0.0f;
-                for (int d = 0; d < D; d++)
-                    dot += Q[((b * Hq + h) * 1 + 0) * D + d]
-                         * K[((b * Hk + kv_h) * seq_len + s) * D + d];
-                dot *= scale;
-                float nm = fmaxf(mv, dot);
-                float al = expf(mv - nm);
-                float be = expf(dot - nm);
-                sv = sv * al + be;
-                for (int d = 0; d < D; d++)
-                    accum[d] = accum[d] * al
-                             + V[((b * Hk + kv_h) * seq_len + s) * D + d] * be;
-                mv = nm;
-            }
-            float inv = 1.0f / sv;
-            for (int d = 0; d < D; d++)
-                O[((b * Hq + h) * 1 + 0) * D + d] = accum[d] * inv;
-        }
-    }
-}
-
-static bf16 f2bf(float x) { return __float2bfloat16(x); }
-static float bf2f(bf16 x) { return __bfloat162float(x); }
-static float randf() { return (float)rand() / (float)RAND_MAX - 0.5f; }
-
-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);
-
-        GQAParams 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=1; p.is_causal=0; p.causal_offset=0;
-        p.scale=1.0f/sqrtf((float)D);
-        p.q=dQ; p.k=dK; p.v=dV; p.mask=dMask; p.o=dO;
-
-        size_t smem=DC_CHUNK*D*sizeof(bf16);
-        dim3 block(32, gs);
-        dim3 grid(B*Hk);
-        printf("grid=(%d,1,1) block=(%d,%d,1) smem=%zu\n",
-               grid.x, block.x, block.y, smem);
-
-        double t0=now_ms();
-        gqa_decode_attn_kernel<<<grid,block,smem>>>(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_decode(hQ,hK,hV,hMask,ref,B,Hq,Hk,sl,D);
-
-        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);
-        delete[]hQ;delete[]hK;delete[]hV;delete[]hMask;delete[]hOut;delete[]ref;delete[]tmp;
-    }
-    printf("All tests passed!\n");
-    return 0;
-}

csrc/tests/gqa_prefill_test.cu

@@ -1,133 +0,0 @@
-/*
-Pure-C test:
-nvcc -I csrc -arch=sm_89 -O3 \
-    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
-    csrc/tests/gqa_prefill_test.cu -o test && ./test
-*/
-
-#include <cstdio>
-#include <cstdlib>
-#include <cmath>
-#include <sys/time.h>
-#include "../kernels/gqa_prefill_attn.cuh"
-
-static double now_ms() {
-    struct timeval tv;
-    gettimeofday(&tv, NULL);
-    return tv.tv_sec * 1000.0 + tv.tv_usec / 1000.0;
-}
-
-static void cpu_attention(const float* Q, const float* K, const float* V, float* O,
-                          int B, int Hq, int Hk, int q_len, int kv_len, int D,
-                          int is_causal, int causal_off) {
-    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++) {
-            for (int qi = 0; qi < q_len; qi++) {
-                int kv_h = h / n_rep;
-                float mv = -INFINITY, sv = 0.0f;
-                float accum[256] = {0};
-                int lim = is_causal ? min(kv_len, qi + causal_off + 1) : kv_len;
-                for (int kj = 0; kj < lim; kj++) {
-                    float dot = 0.0f;
-                    for (int d = 0; d < D; d++)
-                        dot += Q[((b*Hq + h)*q_len + qi)*D + d]
-                             * K[((b*Hk + kv_h)*kv_len + kj)*D + d];
-                    dot *= scale;
-                    float nm = fmaxf(mv, dot);
-                    float al = expf(mv - nm);
-                    float be = expf(dot - nm);
-                    sv = sv * al + be;
-                    for (int d = 0; d < D; d++)
-                        accum[d] = accum[d] * al
-                                 + V[((b*Hk + kv_h)*kv_len + kj)*D + d] * be;
-                    mv = nm;
-                }
-                float inv = 1.0f / sv;
-                for (int d = 0; d < D; d++)
-                    O[((b*Hq + h)*q_len + qi)*D + d] = accum[d] * inv;
-            }
-        }
-    }
-}
-
-static __nv_bfloat16 f2bf(float x) { return __float2bfloat16(x); }
-static float bf2f(__nv_bfloat16 x) { return __bfloat162float(x); }
-static float randf() { return (float)rand() / (float)RAND_MAX - 0.5f; }
-
-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);
-
-        GQAParams 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.is_causal=causal; p.causal_offset=0;
-        p.scale=1.0f/sqrtf((float)D);
-        p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
-
-        constexpr int G=8, ROWS=32, P_BC=32;
-        dim3 grid((ql+ROWS-1)/ROWS, Hq, B);
-        dim3 block(G, ROWS, 1);
-        size_t smem=2*P_BC*D*sizeof(bf16);
-        printf("grid=(%d,%d,%d) block=(%d,%d,%d) smem=%zu\n",
-               grid.x,grid.y,grid.z, block.x,block.y,block.z, smem);
-
-        double t0=now_ms();
-        switch (D) {
-            case 64:  gqa_prefill_attn_kernel_t<64, G,ROWS,P_BC><<<grid,block,smem>>>(p); break;
-            case 128: gqa_prefill_attn_kernel_t<128,G,ROWS,P_BC><<<grid,block,smem>>>(p); break;
-            default: printf("unsupported D=%d\n",D); return 1;
-        }
-        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(hQ,hK,hV,ref,B,Hq,Hk,ql,kl,D,causal,0);
-
-        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");
-    return 0;
-}

csrc/tests/test_utils.cuh

@@ -0,0 +1,171 @@
+#pragma once
+
+#include <cstdio>
+#include <cstdlib>
+#include <cmath>
+#include <chrono>
+#include <cuda_bf16.h>
+
+using bf16 = __nv_bfloat16;
+
+inline bf16 f2bf(float x) { return __float2bfloat16(x); }
+inline float bf2f(bf16 x) { return __bfloat162float(x); }
+
+inline float randf() { return (float)rand() / (float)RAND_MAX - 0.5f; }
+
+inline double now_ms() {
+    using namespace std::chrono;
+    return duration_cast<milliseconds>(steady_clock::now().time_since_epoch()).count();
+}
+
+#define CUDA_CHECK(call) \
+    do { \
+        cudaError_t _e = (call); \
+        if (_e != cudaSuccess) { \
+            printf("CUDA error %s at %s:%d\n", cudaGetErrorString(_e), __FILE__, __LINE__); \
+            exit(1); \
+        } \
+    } while (0)
+
+struct BenchResult {
+    float ms;
+    double 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;
+            }
+        }
+    }
+}

pyproject.toml

@@ -49,4 +49,5 @@ target-version = "py312"
 quote-style = "double"
 indent-style = "space"
 skip-magic-trailing-comma = false
-line-ending = "auto"
+line-ending = "auto"
+exclude = ["*.md", "*.json", "*.yml", "*.yaml"]

scripts/demo/download.py

@@ -5,7 +5,7 @@ from huggingface_hub import snapshot_download
 
 PROJECT_ROOT = Path(__file__).resolve().parents[2]
 DEFAULT_LOCAL_DIR = Path(PROJECT_ROOT, "params")
-DEFAULT_REPO_ID = "ViperEk/KHAOSZ"
+DEFAULT_REPO_ID = "ViperEkura/AstrAI-V1-instruct"
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(

scripts/demo/generate_batch.py

@@ -26,11 +26,9 @@ def batch_generate():
 
     prompts = [
         tokenizer.apply_chat_template(
-            [
+            [{"role": "user", "content": q}],
-                {"role": "system", "content": "You are a helpful assistant."},
-                {"role": "user", "content": q},
-            ],
             tokenize=False,
+            add_generation_prompt=True,
         )
         for q in inputs
     ]

scripts/demo/stream_chat.py

@@ -42,11 +42,24 @@ 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,
-        default="You are a helpful assistant.",
+        default="",
-        help="Optional system prompt",
+        help="Optional system prompt (default: empty, model not SFT-trained on system role)",
     )
     return parser.parse_args()
 
@@ -60,18 +73,20 @@ def chat():
     model.to(device="cuda", dtype=torch.bfloat16)
     engine = InferenceEngine(model=model, tokenizer=tokenizer)
 
-    messages = [{"role": "system", "content": args.system_prompt}]
-
     while True:
         query = input(">> ")
         if query == "!exit":
             break
 
-        messages.append({"role": "user", "content": query})
+        msgs = []
+        if args.system_prompt:
+            msgs.append({"role": "system", "content": args.system_prompt})
+        msgs.append({"role": "user", "content": query})
+        prompt = tokenizer.apply_chat_template(
+            msgs, tokenize=False, add_generation_prompt=True
+        )
 
         full_response = ""
-        prompt = tokenizer.apply_chat_template(messages, tokenize=False)
-
         for token in engine.generate(
             prompt=prompt,
             stream=True,
@@ -79,12 +94,13 @@ 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
 
         print()
-        messages.append({"role": "assistant", "content": full_response.strip()})
 
 
 if __name__ == "__main__":

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)
@@ -143,7 +143,7 @@ def print_layer_grid(results: dict[str, dict]):
     widths = [6] + [10] * len(comps)
     metric = "er_99_norm"
 
-    print(f"\n--- Per-Layer Effective Rank (99% energy) ---")
+    print("\n--- Per-Layer Effective Rank (99% energy) ---")
     print(format_header(["Layer"] + comps, widths))
     print("-" * sum(widths))
 
@@ -173,7 +173,7 @@ def print_layer_grid(results: dict[str, dict]):
 def print_weight_stats(results: dict[str, dict]):
     groups = group_by_component(results)
     widths = [20, 12, 12, 12, 12]
-    print(f"\n--- Weight Value Statistics ---")
+    print("\n--- Weight Value Statistics ---")
     print(format_header(["Component", "Mean", "Std", "Min", "Max"], widths))
     print("-" * sum(widths))
 
@@ -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"
@@ -257,7 +265,7 @@ def main():
             )
         print(f"{'=' * 70}")
 
-        print(f"Loading weights...")
+        print("Loading weights...")
         sd = safetensors.torch.load_file(str(weights_path))
         print(f"  {len(sd)} keys loaded")
 
@@ -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
@@ -21,6 +20,7 @@ from typing import Dict, Iterator, List, Optional, Sequence, Tuple
 import numpy as np
 import torch
 import tqdm
+from datasets import load_dataset
 
 from astrai.inference import InferenceEngine
 from astrai.model import AutoModel
@@ -30,9 +30,7 @@ from astrai.tokenize import AutoTokenizer
 # Config
 # ---------------------------------------------------------------------------
 
-HUMANEVAL_URL = (
+HUMANEVAL_HF_DATASET = "openai/openai_humaneval"
-    "https://github.com/openai/human-eval/raw/master/data/HumanEval.jsonl.gz"
-)
 
 STOP_SEQUENCES = [
     "\nclass ",
@@ -65,21 +63,16 @@ class EvalConfig:
     problem_indices: Optional[List[int]] = None
 
 
-def download(url: str, path: str):
+def download(path: str):
     if os.path.exists(path):
         return
-    import gzip
-    import urllib.request
-
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
-    print(f"Downloading {url} ...")
+    print(f"Downloading HumanEval from HuggingFace ({HUMANEVAL_HF_DATASET}) ...")
-    tmp = path + ".tmp"
+    ds = load_dataset(HUMANEVAL_HF_DATASET, split="test")
-    urllib.request.urlretrieve(url, tmp)
+    with open(path, "w", encoding="utf-8") as f:
-    with gzip.open(tmp, "rb") as f_in:
+        for item in ds:
-        with open(path, "wb") as f_out:
+            f.write(json.dumps(item, ensure_ascii=False) + "\n")
-            f_out.write(f_in.read())
+    print(f"  saved {len(ds)} problems to {path}")
-    os.remove(tmp)
-    print(f"  saved to {path}")
 
 
 def load_jsonl(path: str) -> List[dict]:
@@ -233,7 +226,7 @@ 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 +240,16 @@ 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 +257,14 @@ def test_all(
     items: Sequence[dict],
     cfg: EvalConfig,
 ) -> Iterator[Tuple[str, int, int]]:
-    for item in tqdm.tqdm(items, desc="Testing", unit="problem"):
+    from concurrent.futures import ProcessPoolExecutor
-        yield test_one(item, cfg)
+
+    pool = ProcessPoolExecutor(max_workers=cfg.test_workers)
+    try:
+        for item in tqdm.tqdm(items, desc="Testing", unit="problem"):
+            yield test_one(item, cfg, pool)
+    finally:
+        pool.shutdown(wait=True)
 
 
 def pass_at_k(n: int, c: int, k: int) -> float:
@@ -273,26 +277,32 @@ 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:
-            pk = round(pass_at_k(n, passed, k), 4)
+            if k <= n:
-            entry[f"pass@{k}"] = pk
+                pk = round(pass_at_k(n, passed, k), 4)
-            scores[k].append(pk)
+                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]
-        summary[f"pass@{k}"] = round(float(np.mean(vals)), 4)
+        if vals:
+            summary[f"pass@{k}"] = round(float(np.mean(vals)), 4)
+        else:
+            summary[f"pass@{k}"] = None
     output["_summary"] = summary
     return output
 
@@ -302,7 +312,7 @@ def run_pipeline(cfg: EvalConfig) -> Dict:
         with open(cfg.test_only, encoding="utf-8") as f:
             generated = json.load(f)
     else:
-        download(HUMANEVAL_URL, cfg.data_path)
+        download(cfg.data_path)
 
         problems = load_jsonl(cfg.data_path)
         if cfg.problem_indices:
@@ -375,7 +385,10 @@ def report(scored: Dict):
     summary = scored.pop("_summary", {})
     print(f"\n{'=' * 60}")
     for k, v in summary.items():
-        print(f"  {k}: {v:.2%}")
+        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
@@ -24,28 +26,22 @@ import torch.nn.functional as F
 import tqdm
 
 from astrai.model import AutoModel
+from astrai.preprocessing.packing import plan_bfd
 from astrai.tokenize import AutoTokenizer
 
 
 def _pack_bins(pairs, max_len):
-    """BFD bin packing: pack (c+r) into bins of max total length."""
+    """BFD bin packing: pack (c+r) into bins of max total length.
-    indexed = sorted(enumerate(pairs), key=lambda x: -(len(x[1][0]) + len(x[1][1])))
+
-    bins = []
+    Reuses :func:`plan_bfd` so the BFD heuristic stays single-sourced.
-    lengths = []
+    """
-    for orig_idx, (c, r) in indexed:
+    # Treat each pair as a single sequence of length len(c)+len(r) for
-        size = len(c) + len(r)
+    # planning purposes; plan_bfd works on pure lengths.
-        best_bin = -1
+    fake_sequences = [[0] * (len(c) + len(r)) for c, r in pairs]
-        for bi, rem in enumerate(lengths):
+    plan = plan_bfd(fake_sequences, max_len)
-            if rem >= size:
+    return [
-                if best_bin < 0 or rem < lengths[best_bin]:
+        [(i, pairs[i][0], pairs[i][1]) for i in bin_indices] for bin_indices in plan
-                    best_bin = bi
+    ]
-        if best_bin >= 0:
-            bins[best_bin].append((orig_idx, c, r))
-            lengths[best_bin] -= size
-        else:
-            bins.append([(orig_idx, c, r)])
-            lengths.append(max_len - size)
-    return bins
 
 
 def _resolve_sentinel_ids(tokenizer, sentinel_text):
@@ -65,6 +61,29 @@ 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 +223,9 @@ 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 +234,31 @@ 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 +318,22 @@ 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)}")
@@ -368,7 +344,8 @@ def process_file(
         print(f"  Min IFD:       {min(valid_ifd):.4f}")
         print(f"  Max IFD:       {max(valid_ifd):.4f}")
         print(f"{'=' * 50}")
-    print(f"Results saved to {output_file}")
+    print(f"  Results saved to {output_file}")
+    return stats
 
 
 def _flush_buffer(
@@ -422,8 +399,18 @@ 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 +429,12 @@ 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 +446,60 @@ 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)
-        data_format=args.format,
+    model.to(device=args.device, dtype=dtype)
-        batch_size=args.batch_size,
+    model.eval()
-        device=args.device,
+
-        sentinel_text=args.sentinel_text,
+    sentinel_ids = _resolve_sentinel_ids(tokenizer, args.sentinel_text)
-        per_token=args.per_token,
+
-    )
+    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_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
 """
@@ -14,21 +14,17 @@ import argparse
 import json
 import os
 import re
-import urllib.request
 from typing import Callable, Dict, List, Optional
 
 import torch
 import tqdm
+from datasets import load_dataset
 
 from astrai.inference import InferenceEngine
 from astrai.model import AutoModel
 from astrai.tokenize import AutoTokenizer
 
-IFEVAL_URL = (
+IFEVAL_HF_DATASET = "google/IFEval"
-    "https://raw.githubusercontent.com/google-research/"
-    "google-research/master/instruction_following_eval/data/input_data.jsonl"
-)
-
 CONSTRAINT_VERIFIERS: Dict[str, Callable[[str, dict], bool]] = {}
 
 
@@ -310,15 +306,12 @@ def download_ifeval(data_path: str):
     if os.path.exists(data_path):
         return
     os.makedirs(os.path.dirname(data_path) or ".", exist_ok=True)
-    print(f"Downloading IFEval from {IFEVAL_URL} ...")
+    print(f"Downloading IFEval from HuggingFace ({IFEVAL_HF_DATASET}) ...")
-    tmp = data_path + ".tmp"
+    ds = load_dataset(IFEVAL_HF_DATASET, split="train")
-    urllib.request.urlretrieve(IFEVAL_URL, tmp)
+    with open(data_path, "w", encoding="utf-8") as f:
-    with open(tmp, "rb") as f_in:
+        for item in ds:
-        content = f_in.read()
+            f.write(json.dumps(item, ensure_ascii=False) + "\n")
-    with open(data_path, "wb") as f_out:
+    print(f"  saved {len(ds)} items to {data_path}")
-        f_out.write(content)
-    os.remove(tmp)
-    print(f"  saved to {data_path}")
 
 
 def load_problems(data_path: str) -> List[dict]:

scripts/eval/evaluate_mmlu.py

@@ -4,18 +4,18 @@ import argparse
 import csv
 import json
 import os
-import shutil
+import random
-import tarfile
+from collections import defaultdict
 
-import requests
 import torch
 import torch.nn.functional as F
 import tqdm
+from datasets import load_dataset
 
 from astrai.model import AutoModel
 from astrai.tokenize import AutoTokenizer
 
-MMLU_URL = "https://people.eecs.berkeley.edu/~hendrycks/data.tar"
+MMLU_HF_DATASET = "cais/mmlu"
 MMLU_SUBJECTS = [
     "abstract_algebra",
     "anatomy",
@@ -77,38 +77,40 @@ MMLU_SUBJECTS = [
 ]
 
 
-def _download_and_extract(url: str, data_dir: str):
+def _write_subject_csv(data_dir: str, split: str, subject: str, rows: list[dict]):
-    tar_path = os.path.join(data_dir, "data.tar")
+    split_dir = os.path.join(data_dir, split)
-    os.makedirs(data_dir, exist_ok=True)
+    os.makedirs(split_dir, exist_ok=True)
-    print(f"Downloading MMLU data from {url}...")
+    path = os.path.join(split_dir, f"{subject}_{split}.csv")
-    resp = requests.get(url, stream=True, timeout=300)
+    with open(path, "w", encoding="utf-8", newline="") as f:
-    resp.raise_for_status()
+        writer = csv.writer(f)
-    total = int(resp.headers.get("content-length", 0))
+        for row in rows:
-    with tqdm.tqdm(total=total, unit="B", unit_scale=True, desc="  Download") as bar:
+            writer.writerow(row)
-        with open(tar_path, "wb") as f:
-            for chunk in resp.iter_content(chunk_size=8192):
-                f.write(chunk)
-                bar.update(len(chunk))
-    print("Extracting...")
-    with tarfile.open(tar_path, "r") as tf:
-        tf.extractall(data_dir)
-    os.remove(tar_path)
 
 
 def download_mmlu(data_dir: str):
-    _download_and_extract(MMLU_URL, data_dir)
+    print(f"Downloading MMLU from HuggingFace ({MMLU_HF_DATASET}) ...")
-    src = os.path.join(data_dir, "data")
+    letters = ("A", "B", "C", "D")
-    if os.path.exists(src):
+    split_map = {"dev": "dev", "val": "validation", "test": "test"}
-        for item in os.listdir(src):
+    for local_split, hf_split in split_map.items():
-            src_item = os.path.join(src, item)
+        ds = load_dataset(MMLU_HF_DATASET, "all", split=hf_split)
-            dst_item = os.path.join(data_dir, item)
+        grouped: dict[str, list[dict]] = defaultdict(list)
-            if os.path.exists(dst_item):
+        for item in tqdm.tqdm(ds, desc=f"  {local_split}", leave=False):
-                if os.path.isdir(dst_item):
+            subject = item["subject"]
-                    shutil.rmtree(dst_item)
+            choices = item["choices"]
-                else:
+            ans_letter = letters[item["answer"]]
-                    os.remove(dst_item)
+            grouped[subject].append(
-            os.rename(src_item, dst_item)
+                [
-        os.rmdir(src)
+                    item["question"],
+                    f"A){choices[0]}",
+                    f"B){choices[1]}",
+                    f"C){choices[2]}",
+                    f"D){choices[3]}",
+                    ans_letter,
+                ]
+            )
+        for subject, rows in grouped.items():
+            _write_subject_csv(data_dir, local_split, subject, rows)
+        print(f"  {local_split}: {len(ds)} items, {len(grouped)} subjects")
     print(f"MMLU data saved to {data_dir}")
 
 
@@ -139,17 +141,12 @@ 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"
+    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"
@@ -158,19 +155,22 @@ def build_prompt(
 
 
 def apply_chat(
-    tokenizer, raw_prompt: str, n_shot: int, dev_data: list[dict] | None
+    tokenizer,
+    raw_prompt: str,
+    n_shot: int,
+    dev_data: list[dict] | None,
+    subject: str = "",
 ) -> str:
     """Wrap raw MMLU prompt in the model's chat template format.
 
-    For few-shot, prepend example Q&A pairs as a second user/assistant exchange.
+    For few-shot, prepend example Q&A pairs as user/assistant exchanges.
+    Few-shot examples use the same subject preamble as the test question to
+    keep the format consistent.
     """
     messages = []
     if n_shot > 0 and dev_data:
         for item in dev_data[:n_shot]:
-            q = f"Question: {item['question']}\n"
+            q = build_prompt(item["question"], item, subject)
-            for k in ("A", "B", "C", "D"):
-                q += f"{k}. {item[k]}\n"
-            q += "Answer:"
             messages.append({"role": "user", "content": q})
             messages.append({"role": "assistant", "content": item["answer"]})
     messages.append({"role": "user", "content": raw_prompt})
@@ -185,7 +185,7 @@ def choice_logprob(
     choice_text = choice_letter
     choice_ids = tokenizer.encode(choice_text, add_special_tokens=False)
     input_ids = context_ids + choice_ids
-    max_len = model.config.max_len
+    max_len = model.config.max_position_embeddings
     if len(input_ids) > max_len:
         overflow = len(input_ids) - max_len
         input_ids = input_ids[overflow:]
@@ -206,6 +206,24 @@ def choice_logprob(
     return score
 
 
+def _permute_choices(item: dict, rng: random.Random) -> tuple[dict, str]:
+    """Shuffle the option order of a question.
+
+    Returns ``(permuted_item, new_answer_letter)``. The question text and
+    the *content* of each choice are unchanged; only which letter (A/B/C/D)
+    maps to which content is shuffled. This neutralises the model's
+    positional bias (e.g. always picking B).
+    """
+    letters = ("A", "B", "C", "D")
+    perm = list(letters)
+    rng.shuffle(perm)
+    permuted = {"question": item["question"]}
+    for new_letter, orig_letter in zip(letters, perm):
+        permuted[new_letter] = item[orig_letter]
+    new_answer = letters[perm.index(item["answer"])]
+    return permuted, new_answer
+
+
 def evaluate_subject(
     model,
     tokenizer,
@@ -214,20 +232,24 @@ def evaluate_subject(
     dev_data: list[dict] | None,
     device: str,
     n_shot: int,
+    seed: int = 0,
 ) -> tuple[float, int, int]:
+    rng = random.Random(seed) if seed >= 0 else None
     correct = 0
     total = 0
     for item in tqdm.tqdm(test_data, desc=f"{subject:40s}", leave=False):
-        raw_prompt = build_prompt(
+        if rng is not None:
-            item["question"], item, subject, n_shot, dev_data or []
+            permuted, answer = _permute_choices(item, rng)
-        )
+        else:
-        context = apply_chat(tokenizer, raw_prompt, n_shot, dev_data or [])
+            permuted, answer = item, item["answer"]
+        raw_prompt = build_prompt(permuted["question"], permuted, subject)
+        context = apply_chat(tokenizer, raw_prompt, n_shot, dev_data or [], subject)
         context_ids = tokenizer.encode(context)
         scores = {
             c: choice_logprob(model, tokenizer, context_ids, c, device)
             for c in ("A", "B", "C", "D")
         }
-        if max(scores, key=scores.get) == item["answer"]:
+        if max(scores, key=scores.get) == answer:
             correct += 1
         total += 1
     return correct / total, correct, total
@@ -262,6 +284,12 @@ def main():
         default="bfloat16" if torch.cuda.is_available() else "float32",
         help="Torch dtype",
     )
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=0,
+        help="Seed for option permutation (0 to enable, -1 to disable)",
+    )
     args = parser.parse_args()
 
     if args.download or not os.path.exists(args.data_dir):
@@ -293,7 +321,14 @@ def main():
         test_data = load_csv(test_path)
 
         acc, corr, tot = evaluate_subject(
-            model, tokenizer, subject, test_data, dev_data, device, args.n_shot
+            model,
+            tokenizer,
+            subject,
+            test_data,
+            dev_data,
+            device,
+            args.n_shot,
+            seed=args.seed,
         )
         results[subject] = {"accuracy": round(acc, 4), "correct": corr, "total": tot}
         total_correct += corr

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 _collect_input_files(input_path: str) -> List[str]:
+    """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[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()]
+
+
+def _encode_batch(
+    tokenizer: AutoTokenizer, texts: List[str], max_length: int
+) -> Tuple[List[List[int]], List[List[int]]]:
+    """Encode *texts* and return (token_ids, attention_masks).
+
+    Each sequence is left-aligned and padded to the batch max length.
+    """
+    encoded = [tokenizer.encode(t)[:max_length] for t in texts]
+    if not encoded:
+        return [], []
+    max_len = max(len(seq) for seq in encoded)
+    padded_ids = []
+    masks = []
+    for seq in encoded:
+        pad_len = max_len - len(seq)
+        padded_ids.append(seq + [tokenizer.pad_id] * pad_len)
+        masks.append([1] * len(seq) + [0] * pad_len)
+    return padded_ids, masks
+
+
+def _compute_batch(
+    model,
+    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].
+
+    log_probs[i, j] = log P(token j+1 | tokens 0..j)
+    """
+    output = model(input_ids, input_mask=attention_mask)
+    logits = output["logits"][:, :-1, :]  # [B, S-1, V]
+    targets = input_ids[:, 1:]  # [B, S-1]
+    valid = attention_mask[:, 1:].float()  # [B, S-1]
+
+    log_probs = F.log_softmax(logits.float(), dim=-1)  # [B, S-1, V]
+    token_log_probs = log_probs.gather(2, targets.unsqueeze(-1)).squeeze(-1)  # [B, S-1]
+
+    return token_log_probs, valid
+
+
+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(v, 0.0), self._HIST_MAX) for v 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)
+
+    def add_typed(self, ttype: str, losses: List[float]):
+        if not self.stream:
+            self.by_type.setdefault(ttype, []).extend(losses)
+        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)
+
+    def stats(self) -> Dict:
+        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))
+
+        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(
-    param_path: str, input_file: str, output_file: str, batch_size: int, text_key: str
+    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 per_sample:
+                f.write(json.dumps(item, ensure_ascii=False) + "\n")
+
+    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("\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",
 ):
-    # Load model and tokenizer
+    print(f"Loading model from {param_path} ...")
     model = AutoModel.from_pretrained(param_path)
     tokenizer = AutoTokenizer.from_pretrained(param_path)
-    model.to(device="cuda", dtype=torch.bfloat16)
+    torch_dtype = getattr(torch, dtype)
+    model.to(device=device, dtype=torch_dtype)
+    model.eval()
 
-    with open(input_file, "r", encoding="utf-8") as f:
+    input_files = _collect_input_files(input_path)
-        input_data = [json.loads(line) for line in f]
+    if not input_files:
+        print(f"No input files found at {input_path}")
+        return
 
-    texts = [item[text_key] for item in input_data]
+    print(f"Found {len(input_files)} file(s) to evaluate")
+    os.makedirs(output_dir, exist_ok=True)
 
-    # Encode all texts
+    all_stats = {}
-    print(f"Encoding {len(texts)} texts...")
+    for filepath in input_files:
-    encoded_texts = [tokenizer.encode(text) for text in texts]
+        label = os.path.splitext(os.path.basename(filepath))[0]
+        items = _load_items(filepath)
+        if not items:
+            print(f"  [{label}] empty, skipping")
+            continue
 
-    output_data = []
+        token_output = (
-    total_batches = (len(encoded_texts) + batch_size - 1) // batch_size
+            os.path.join(output_dir, f"{label}_tokens.jsonl") if token_level else None
-
-    for i in tqdm.tqdm(
-        range(0, len(encoded_texts), batch_size),
-        total=total_batches,
-        desc="Computing perplexity",
-    ):
-        batch_encoded = encoded_texts[i : i + batch_size]
-        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 batch_encoded:
-            pad_len = max_len - len(seq)
-            padded_seq = seq + [tokenizer.pad_id] * pad_len
-            mask = [True] * len(seq) + [False] * pad_len
-            padded_ids.append(padded_seq)
-            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
-        output = model(input_ids, input_mask=input_mask)
-        logits = output["logits"]
-
-        # Shift for causal language modeling
-        shifted_logits = logits[:, :-1, :]  # [batch_size, seq_len-1, vocab_size]
-        shifted_input_ids = input_ids[:, 1:]  # [batch_size, seq_len-1]
-        shifted_mask = input_mask[:, 1:]  # [batch_size, seq_len-1]
-
-        # Compute cross entropy loss
-        loss = F.cross_entropy(
-            shifted_logits.flatten(0, 1),
-            shifted_input_ids.flatten(0, 1),
-            reduction="none",
         )
 
-        loss = loss.view(shifted_input_ids.shape)  # [batch_size, seq_len-1]
+        stats = process_file(
-        loss = loss * shifted_mask
+            model=model,
-        sentence_loss = loss.sum(dim=1) / shifted_mask.sum(dim=1).clamp(min=1)
+            tokenizer=tokenizer,
-        perplexity = torch.exp(sentence_loss)  # [batch_size]
+            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)
 
-        for text, ppl in zip(batch_texts, perplexity):
+        if token_output:
-            output_data.append({text_key: text, "ppl": float(ppl.item())})
+            print(f"  token-level output: {token_output}")
 
-    # Write results
+    summary_path = os.path.join(output_dir, "summary.json")
-    with open(output_file, "w", encoding="utf-8") as f:
+    with open(summary_path, "w", encoding="utf-8") as f:
-        for item in output_data:
+        json.dump(all_stats, f, ensure_ascii=False, indent=2)
-            f.write(json.dumps(item, ensure_ascii=False) + "\n")
+    print(f"\nSummary saved to {summary_path}")
-
-    print(f"Perplexity computation complete. Results saved to {output_file}")
 
 
 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/eval/evaluate_rouge.py

@@ -15,7 +15,7 @@ Usage::
 import argparse
 import json
 from collections import Counter
-from typing import Dict, List, Tuple
+from typing import Dict, List
 
 
 def _tokenize(text: str) -> List[str]:

scripts/tools/benchmark.py

@@ -119,15 +119,15 @@ class GenerationBenchmark:
                 dtype=torch.long,
             )
 
-            head_dim = self.config.dim // self.config.n_heads
+            head_dim = self.config.hidden_size // self.config.num_attention_heads
             max_seq = prompt_length + gen_length
 
             if self.cache_type == "contiguous":
                 cache = ContiguousCache(
-                    self.config.n_layers,
+                    self.config.num_hidden_layers,
                     batch_size,
                     max_seq,
-                    self.config.n_kv_heads,
+                    self.config.num_key_value_heads,
                     head_dim,
                     self.device,
                     self.dtype,
@@ -136,10 +136,10 @@ class GenerationBenchmark:
                 page_size = 128
                 n_pages = (max_seq + page_size - 1) // page_size * batch_size
                 cache = PageCache(
-                    self.config.n_layers,
+                    self.config.num_hidden_layers,
                     n_pages,
                     page_size,
-                    self.config.n_kv_heads,
+                    self.config.num_key_value_heads,
                     head_dim,
                     self.device,
                     self.dtype,
@@ -262,13 +262,13 @@ if __name__ == "__main__":
 
     config = AutoRegressiveLMConfig(
         vocab_size=10000,
-        dim=1536,
+        hidden_size=1536,
-        n_heads=24,
+        num_attention_heads=24,
-        n_kv_heads=4,
+        num_key_value_heads=4,
-        dim_ffn=6912,
+        intermediate_size=6912,
-        max_len=2048,
+        max_position_embeddings=2048,
-        n_layers=24,
+        num_hidden_layers=24,
-        norm_eps=1e-5,
+        rms_norm_eps=1e-5,
     )
 
     benchmark = GenerationBenchmark(

scripts/tools/generate.py

@@ -1,8 +1,10 @@
 import argparse
 import json
+import time
 from typing import Optional
 
 import torch
+from tqdm import tqdm
 
 from astrai.inference import InferenceEngine
 from astrai.model import AutoModel
@@ -20,55 +22,102 @@ def processor(
     response_key: str,
     max_tokens: Optional[int],
     batch_size: int,
+    num_samples: int = 1,
+    cache_len: int = 2048,
+    frequency_penalty: float = 0.0,
+    rep_window: int = 64,
 ):
-    # Load model and tokenizer
+    print(f"Loading model from {param_path} ...")
+    t0 = time.time()
     model = AutoModel.from_pretrained(param_path)
     tokenizer = AutoTokenizer.from_pretrained(param_path)
     model.to(device="cuda", dtype=torch.bfloat16)
+    print(f"  model loaded in {time.time() - t0:.1f}s")
 
-    # Create inference engine
     engine = InferenceEngine(
-        model=model, tokenizer=tokenizer, max_batch_size=batch_size
+        model=model,
+        tokenizer=tokenizer,
+        max_batch_size=batch_size * num_samples,
+        max_seq_len=cache_len,
+        max_prompt_len=cache_len,
     )
 
+    print(f"Reading {input_json_file} ...")
     with open(input_json_file, "r", encoding="utf-8") as f:
         input_data = [json.loads(line) for line in f]
 
-    # Check input format: chat messages or raw text
     if input_data and "messages" in input_data[0]:
-        # Chat format: [{"messages": [...]}]
         prompts = [
             tokenizer.apply_chat_template(item["messages"], tokenize=False)
             for item in input_data
         ]
     else:
-        # Raw text format: [{"question": "..."}]
         prompts = [item[question_key] for item in input_data]
+    print(f"  {len(prompts)} prompts loaded\n")
 
-    # Use provided max_tokens or default to model config max_len
     if max_tokens is None:
-        max_tokens = model.config.max_len
+        max_tokens = model.config.max_position_embeddings
 
-    # Generate responses (batch)
+    chunk_size = max(1, batch_size)
-    responses = engine.generate(
-        prompt=prompts,
-        stream=False,
-        max_tokens=max_tokens,
-        temperature=temperature,
-        top_p=top_p,
-        top_k=top_k,
-    )
 
-    # Write results
     with open(output_json_file, "w", encoding="utf-8") as f:
-        for prompt, response in zip(prompts, responses):
+        pbar = tqdm(
-            if input_data and "messages" in input_data[0]:
+            total=len(prompts) * num_samples,
-                output_item = {"response": response}
+            unit="gen",
+            desc=f" Generating ({num_samples}x/prompt)",
+        )
+        for chunk_start in range(0, len(prompts), chunk_size):
+            chunk = prompts[chunk_start : chunk_start + chunk_size]
+
+            if num_samples > 1:
+                chunk_expanded = [p for p in chunk for _ in range(num_samples)]
+                resp_chunk = engine.generate(
+                    prompt=chunk_expanded,
+                    stream=False,
+                    max_tokens=max_tokens,
+                    temperature=temperature,
+                    top_p=top_p,
+                    top_k=top_k,
+                    frequency_penalty=frequency_penalty,
+                    rep_window=rep_window,
+                )
+                resp_chunk = [
+                    resp_chunk[i * num_samples : (i + 1) * num_samples]
+                    for i in range(len(chunk))
+                ]
             else:
-                output_item = {question_key: prompt, response_key: response}
+                resp_chunk = engine.generate(
-            f.write(json.dumps(output_item, ensure_ascii=False) + "\n")
+                    prompt=chunk,
+                    stream=False,
+                    max_tokens=max_tokens,
+                    temperature=temperature,
+                    top_p=top_p,
+                    top_k=top_k,
+                    frequency_penalty=frequency_penalty,
+                    rep_window=rep_window,
+                )
+
+            for i, prompt in enumerate(chunk):
+                if input_data and "messages" in input_data[0]:
+                    orig = input_data[chunk_start + i]
+                    output_item = {**orig, response_key: resp_chunk[i]}
+                else:
+                    output_item = {
+                        question_key: prompt,
+                        response_key: resp_chunk[i],
+                    }
+                f.write(json.dumps(output_item, ensure_ascii=False) + "\n")
+
+            pbar.update(len(chunk) * num_samples)
+
+        pbar.close()
+
+    elapsed = time.time() - t0
+    print(
+        f"\nDone! {len(prompts)} prompts x {num_samples} samples -> {output_json_file}"
+    )
+    print(f"Total time: {elapsed:.1f}s ({elapsed / len(prompts):.2f}s/prompt)")
 
-    # Cleanup
     engine.shutdown()
 
 
@@ -126,11 +175,38 @@ if __name__ == "__main__":
         default=1,
         help="Batch size for generating responses (default: 1).",
     )
+    parser.add_argument(
+        "--num_samples",
+        type=int,
+        default=1,
+        help="Number of responses per prompt (expands batch internally, default: 1).",
+    )
     parser.add_argument(
         "--max_tokens",
         type=int,
         default=None,
-        help="Maximum tokens to generate (default: model config max_len).",
+        help=(
+            "Maximum tokens to generate "
+            "(default: model config max_position_embeddings)."
+        ),
+    )
+    parser.add_argument(
+        "--cache_len",
+        type=int,
+        default=2048,
+        help="KV cache & prompt truncation length (default: 2048, lower = less memory).",
+    )
+    parser.add_argument(
+        "--frequency_penalty",
+        type=float,
+        default=0.0,
+        help="Frequency penalty to reduce repetition (default: 0.0, try 0.5-1.0).",
+    )
+    parser.add_argument(
+        "--rep_window",
+        type=int,
+        default=64,
+        help="Window size for frequency penalty (default: 64).",
     )
 
     args = parser.parse_args()

scripts/tools/preprocess.py

@@ -22,9 +22,19 @@ def main():
         default="params",
         help="Path to tokenizer directory (default: params)",
     )
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=None,
+        help="Number of records tokenized together (default: config value)",
+    )
     args = parser.parse_args()
 
     config = PipelineConfig.from_file(args.config)
+    if args.batch_size is not None:
+        if args.batch_size < 1:
+            parser.error("--batch_size must be at least 1")
+        config.preprocessing.batch_size = args.batch_size
 
     Pipeline(
         config=config,