refactor: move sample-id indexing from dataset to store

- Store owns window_size/stride and __getitem__/__len__/sample_window
- Dataset classes become thin delegators binding a Store to a train-type key mapping
- Drop BaseDataset.get_index and the RecordDataset中间类 (window死代码)
- DatasetFactory forces window_size=0 for record datasets so record semantics never get window-tainted
- token_count/num_records split the legacy len() semantics (raw stream length vs record count)
- Update tests to the new .store/.token_count API and window/record mode switching

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

@@ -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
@@ -166,6 +166,13 @@ classDiagram
             +__getitem__(index) Dict
         }
 
+        class RecordDataset {
+            +Optional[Callable] processor
+            +load(load_path, storage_type)
+            +__getitem__(index)
+            +__len__()
+        }
+
         class DPODataset {
             +__getitem__(index) Dict
         }
@@ -177,13 +184,26 @@ 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 {
@@ -195,6 +215,13 @@ classDiagram
             +load(path)
         }
 
+        class JsonlStore {
+            +JsonlSource _source
+            +Callable _processor
+            +load(path, transform, processor)
+            +fetch_record(index, keys)
+        }
+
         class ResumableDistributedSampler {
             +int epoch
             +int iter
@@ -210,7 +237,7 @@ 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
         }
     }
 
@@ -378,6 +405,7 @@ classDiagram
             +List[str] paths
             +str output_dir
             +str tokenizer_path
+            +AutoTokenizer tokenizer
             +BaseMaskBuilder mask_builder
             +PackingStrategy _packer
             +PositionIdStrategy _position_id
@@ -385,6 +413,18 @@ 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]
         }
     }
 
@@ -495,14 +535,13 @@ 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 BaseScheduler {
@@ -552,7 +591,7 @@ classDiagram
         }
 
         class GradientClippingCallback {
-            +float max_grad_norm
+            +Optional[float] max_grad_norm
             +on_optimizer_step(context)
         }
 
@@ -1065,11 +1104,18 @@ classDiagram
     TrainCallback <|-- MetricCallback
     BaseDataset <|-- SEQDataset
     BaseDataset <|-- SFTDataset
-    BaseDataset <|-- DPODataset
+    BaseDataset <|-- RecordDataset
-    BaseDataset <|-- GRPODataset
+    RecordDataset <|-- DPODataset
+    RecordDataset <|-- GRPODataset
     Store <|-- H5Store
     Store <|-- MmapStore
     Store <|-- JsonlStore
+    H5Store --|> Streamable
+    H5Store --|> Recordable
+    MmapStore --|> Streamable
+    MmapStore --|> Recordable
+    JsonlStore --|> Streamable
+    JsonlStore --|> Recordable
     BaseSamplingStrategy <|-- TemperatureStrategy
     BaseSamplingStrategy <|-- TopKStrategy
     BaseSamplingStrategy <|-- TopPStrategy
@@ -1144,6 +1190,9 @@ classDiagram
     BaseDataset o-- Store
     Pipeline o-- PipelineConfig
     Pipeline o-- BaseMaskBuilder
+    Pipeline o-- AutoTokenizer
+    TokenizeTransform o-- AutoTokenizer
+    TokenizeTransform o-- BaseMaskBuilder
 
     %% --- Dependency (uses temporarily) ---
     TrainConfig ..> BaseStrategy : selects
@@ -1187,7 +1236,7 @@ classDiagram
     %% --- Association (general usage) ---
     Trainer --> TrainConfig
     DPOStrategy --> AutoModel
-    GRPOStrategy --> AutoModel
+    GRPOStrategy --> AutoModel : policy/old/ref
     InferenceScheduler --> Task
     InferenceScheduler --> TaskStatus
     Task --> TaskStatus
@@ -1204,8 +1253,8 @@ 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** | BaseMaskBuilder, MaskBuilderFactory, SectionedMaskBuilder, SingleOutputMaskBuilder, MultiOutputMaskBuilder, Pipeline, TokenizeTransform, filter_by_length, PackingStrategy, PackingStrategyFactory, plan_bfd, PositionIdStrategy, PositionIdStrategyFactory, StoreWriter, StoreWriterFactory, core (shared helpers) | Declarative JSON-driven data preprocessing |
-| **astrai.dataset** | BaseDataset–GRPODataset, Store–JsonlStore/MmapStore/H5Store, StoreFactory, ResumableDistributedSampler, DatasetFactory | Dataset loading and management |
+| **astrai.dataset** | BaseDataset–RecordDataset–DPO/GRPODataset, SEQDataset, SFTDataset, Store, Streamable, Recordable, H5Store, MmapStore, JsonlStore, StoreFactory, ResumableDistributedSampler, DatasetFactory | Dataset loading and management |
 | **astrai.serialization** | Checkpoint | Model serialization |
 | **astrai.model** | AutoModel, AutoRegressiveLM, EmbeddingEncoder, DecoderBlock, GQA, MLA, MLP, DeepSeekMoE, AttnFactory, FFNFactory, RMSNorm, Linear, RotaryEmbedding, Embedding | Neural network model |
 | **astrai.tokenize** | AutoTokenizer, ChatTemplate | Tokenizer and chat template |
@@ -1219,7 +1268,7 @@ classDiagram
 
 | Pattern | Classes | Purpose |
 |---------|---------|---------|
-| **Factory** | `AttnFactory`, `FFNFactory`, `StrategyFactory`, `DatasetFactory`, `SchedulerFactory`, `CallbackFactory`, `StoreFactory`, `ConfigFactory`, `ExecutorFactory` | Decorator-based component creation |
+| **Factory** | `AttnFactory`, `FFNFactory`, `StrategyFactory`, `DatasetFactory`, `SchedulerFactory`, `CallbackFactory`, `StoreFactory`, `ConfigFactory`, `ExecutorFactory`, `MaskBuilderFactory`, `StoreWriterFactory`, `PackingStrategyFactory`, `PositionIdStrategyFactory` | Decorator-based component creation |
 | **Registry** | `BaseFactory` | Component registration |
 | **Strategy** | `SEQStrategy`, `SFTStrategy`, `DPOStrategy`, `GRPOStrategy` | Training strategy switching |
 | **Strategy (Sampling)** | `TemperatureStrategy`, `TopKStrategy`, `TopPStrategy`, `SamplingPipeline` | Composable logit transformations |
@@ -1247,4 +1296,4 @@ 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: 2026-07-19

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_len=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/params.md

@@ -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) | None |
 
 ### Optimizer (MuonMix)
 
@@ -201,4 +201,4 @@ See [Preprocessing Guide](preprocessing.md) for config file format and examples.
 
 ---
 
-> Document Update Time: 2026-07-09
+> Document Update Time: 2026-07-19

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

@@ -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
 
@@ -118,21 +118,31 @@ $$
 L_{\text{DPO}} = -\mathbb{E}\left[\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\text{ref}}(y_w\mid x)} - \beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\text{ref}}(y_l\mid x)}\right)\right]
 $$
 
-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`.
 
@@ -212,4 +222,4 @@ nohup python scripts/tools/train.py \
 
 Full parameter reference at [params.md](params.md).
 
-> Document Update Time: 2026-07-09
+> Document Update Time: 2026-07-19

astrai/__init__.py

@@ -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/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=None,
+        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(

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 +37,389 @@ 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]
+      - responses: [B, G, R_max]
+      - masks:     [B, G, R_max]
+      - rewards:   [B, G]
+    """
+    B = len(batch)
+    G = len(batch[0]["responses"])
+    P_max = max(b["prompts"].size(0) for b in batch)
+    R_max = max(r.size(0) for b in batch for r in b["responses"])
+
+    prompts = torch.zeros(B, P_max, dtype=torch.long)
+    responses = torch.zeros(B, G, R_max, dtype=torch.long)
+    masks = torch.zeros(B, G, R_max, dtype=torch.bool)
+    rewards = torch.zeros(B, G, dtype=torch.float32)
+
+    for i, b in enumerate(batch):
+        p_len = b["prompts"].size(0)
+        prompts[i, :p_len] = b["prompts"]
+        rewards[i, : b["rewards"].size(0)] = b["rewards"]
+        for g in range(min(G, len(b["responses"]))):
+            r_len = b["responses"][g].size(0)
+            responses[i, g, :r_len] = b["responses"][g]
+            if g < len(b["masks"]):
+                masks[i, g, :r_len] = b["masks"][g]
+
+    return {
+        "prompts": prompts,
+        "responses": responses,
+        "masks": masks,
+        "rewards": rewards,
+    }
+
+
+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:
+            store.load(load_path, **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 +430,65 @@ 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 +51,18 @@ import json
 import logging
 from abc import ABC, abstractmethod
 from pathlib import Path
-from typing import Dict, List, Union
+from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
 from torch import Tensor
 
-from astrai.config.preprocess_config import PipelineConfig
 from astrai.factory import BaseFactory
-from astrai.preprocessing.builder import MaskBuilderFactory
+from astrai.preprocessing.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 +74,7 @@ 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 +111,526 @@ def detect_format(load_path: str) -> str:
 
 
 class Store(ABC):
-    """String keys -> segmented tensors with ``fetch(begin, end, keys)``.
+    """Common base for all storage backends.
 
-    Each key maps to one or more tensor segments (no forced concatenation).
+    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 two tokenisation modes.
+
+    A JSONL dataset is a ``.jsonl`` file or a directory of ``*.jsonl``
+    files plus (optionally) a ``dataset_config.json`` describing the
+    tokenization pipeline.
+
+    Two modes, selected at :meth:`load` time:
+
+    - **Eager** (default): applies a :class:`TokenizeTransform` to every
+      record at load time and registers per-key tensors via
+      ``_normalize``.  Both ``fetch`` (stream) and ``fetch_record``
+      (record) work.
+    - **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):
+    def __init__(
-        root = Path(path)
+        self,
-        config_path = root / self.CONFIG_NAME
+        window_size: int = 0,
-        if not config_path.exists():
+        stride: Optional[int] = None,
-            raise FileNotFoundError(
+    ):
-                f"JSONL dataset config not found: {config_path}. "
+        super().__init__(window_size=window_size, stride=stride)
-                f"Expected {self.CONFIG_NAME} alongside *.jsonl files."
+        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 None or not config_path.exists():
+                raise FileNotFoundError(
+                    f"JSONL dataset config not found. Expected "
+                    f"{self.CONFIG_NAME} alongside *.jsonl files, pass an "
+                    f"explicit transform, or pass processor= for lazy "
+                    f"on-the-fly tokenisation."
+                )
+            transform = TokenizeTransform.from_config_file(str(config_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

@@ -5,6 +5,14 @@ Public API:
     - ``attn_prefill`` — multi-query prefill attention
     - ``attn_paged_decode`` — paged decode attention (direct page-table access)
 
+Interface (shared by all wrappers):
+    causal_offset: -1 = non-causal; >=0 = absolute position of first Q token
+    mask:       2D [batch, kv_len] or 3D [batch, q_len, kv_len] (bool, True = keep)
+    scale:      0.0 = auto (1/sqrt(head_dim)); >0 = explicit
+    layout:     "bhld" (default) or "blhd"
+
+Causal and mask can coexist — both are applied simultaneously.
+
 Each wrapper dispatches to its compiled CUDA kernel (``astrai.extension.attn_*``)
 when available, otherwise falls back to ``torch.nn.functional.scaled_dot_product_attention``.
 """

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
@@ -26,20 +54,84 @@ 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(
@@ -56,23 +148,22 @@ def _gather_kv_from_pages(
         k_cache    : [n_pages, page_size, n_kv_heads, head_dim]
         v_cache    : same as k_cache
     Returns:
-        k, v : [batch, n_kv_heads, kv_len, head_dim]
+        k, v : [batch, kv_len, n_kv_heads, head_dim]  (b l h d)
     """
     batch, max_pages = page_table.shape
-    n_pages, ps, n_kv_heads, head_dim = k_cache.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}")
 
-    k = k_cache.new_empty(batch, n_kv_heads, kv_len, head_dim)
+    # Vectorized gather: build physical page + offset indices, then advanced-index
-    v = v_cache.new_empty(batch, n_kv_heads, kv_len, head_dim)
+    positions = torch.arange(kv_len, device=page_table.device)
+    logical_pages = positions // page_size  # [kv_len]
+    page_offsets = positions % page_size  # [kv_len]
 
-    for b in range(batch):
+    phys_pages = page_table[:, logical_pages]  # [batch, kv_len]
-        for pos in range(kv_len):
+    # k_cache[phys_pages, page_offsets] → [batch, kv_len, n_kv_heads, head_dim] (b l h d)
-            log_pg = pos // page_size
+    k = k_cache[phys_pages, page_offsets]
-            pg_off = pos % page_size
+    v = v_cache[phys_pages, page_offsets]
-            phys = int(page_table[b, log_pg].item())
-            k[b, :, pos, :] = k_cache[phys, pg_off, :, :]
-            v[b, :, pos, :] = v_cache[phys, pg_off, :, :]
     return k, v
 
 
@@ -81,21 +172,22 @@ def attn_decode(
     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:
+    li = _parse_layout(layout)
     if _available["attn_decode"]:
         return _modules["attn_decode"].attn_decode(
             q,
             k,
             v,
             mask=mask,
-            is_causal=is_causal,
             causal_offset=causal_offset,
             scale=scale,
+            layout=li,
         )
-    return _torch_fallback(q, k, v, mask, is_causal, scale)
+    return _torch_fallback(q, k, v, mask, causal_offset, scale, q_layout=li)
 
 
 def attn_prefill(
@@ -103,21 +195,22 @@ def attn_prefill(
     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:
+    li = _parse_layout(layout)
     if _available["attn_prefill"]:
         return _modules["attn_prefill"].attn_prefill(
             q,
             k,
             v,
             mask=mask,
-            is_causal=is_causal,
             causal_offset=causal_offset,
             scale=scale,
+            layout=li,
         )
-    return _torch_fallback(q, k, v, mask, is_causal, scale)
+    return _torch_fallback(q, k, v, mask, causal_offset, scale, q_layout=li)
 
 
 def attn_paged_decode(
@@ -128,10 +221,11 @@ def attn_paged_decode(
     page_size: int,
     kv_len: int,
     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:
+    li = _parse_layout(layout)
     if _available["attn_paged_decode"]:
         return _modules["attn_paged_decode"].attn_paged_decode(
             q,
@@ -141,9 +235,12 @@ def attn_paged_decode(
             page_size,
             kv_len,
             mask=mask,
-            is_causal=is_causal,
             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, is_causal, scale)
+    return _torch_fallback(
+        q, k, v, mask, causal_offset, scale, q_layout=li, kv_layout=1
+    )

astrai/inference/__init__.py

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

astrai/inference/api/openai.py

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

astrai/inference/api/protocol.py

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

astrai/inference/core/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)
@@ -409,23 +417,37 @@ 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)
+            for s, p in zip(indices.tolist(), pos.tolist()):
-            if new_len > cur:
+                cur = self._cache._slot_len.get(s, 0)
-                self._cache._slot_len[s] = new_len
+                if p + 1 > cur:
+                    self._cache._slot_len[s] = p + 1
+        else:
+            start_pos = self._total_len - seq_len
+            self._cache.k[layer_id, indices, start_pos : start_pos + seq_len] = k
+            self._cache.v[layer_id, indices, start_pos : start_pos + seq_len] = v
+            new_len = start_pos + seq_len
+            for s in indices.tolist():
+                cur = self._cache._slot_len.get(s, 0)
+                if new_len > cur:
+                    self._cache._slot_len[s] = new_len
 
     def gather(self, layer_id: int) -> Tuple[Tensor, Tensor]:
         max_len = max(
@@ -491,9 +513,21 @@ 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)
+        return ContiguousCacheView(
+            self, batch_indices, total_len, write_positions=write_positions
+        )

astrai/inference/core/executor.py

@@ -75,11 +75,43 @@ class Executor:
         temperatures = torch.tensor([t.temperature for t in tasks], device=self.device)
         top_ks = torch.tensor([t.top_k for t in tasks], device=self.device)
         top_ps = torch.tensor([t.top_p for t in tasks], device=self.device)
+        freq_penalties = torch.tensor(
+            [t.frequency_penalty for t in tasks], device=self.device
+        )
+
+        history_lists = []
+        mask_lists = []
+        for t in tasks:
+            window = t.rep_window
+            prompt_part = t.prompt_ids[-window:]
+            ids = prompt_part + t.output_ids
+            history_lists.append(ids)
+            mask_lists.append([True] * len(ids))
+
+        max_len = max(len(h) for h in history_lists)
+        padded_ids = torch.zeros(
+            len(tasks), max_len, dtype=torch.long, device=self.device
+        )
+        padded_mask = torch.zeros(
+            len(tasks), max_len, dtype=torch.bool, device=self.device
+        )
+        for i, (h, m) in enumerate(zip(history_lists, mask_lists)):
+            padded_ids[i, : len(h)] = torch.tensor(
+                h, dtype=torch.long, device=self.device
+            )
+            padded_mask[i, : len(m)] = torch.tensor(
+                m, dtype=torch.bool, device=self.device
+            )
 
         with torch.inference_mode():
             outputs = self.model(
                 input_ids.unsqueeze(1),
-                paged_cache=self.kv_cache.bind_tasks(task_ids, total_len, self.device),
+                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, :]
@@ -89,4 +121,7 @@ class Executor:
             temperature=temperatures,
             top_k=top_ks,
             top_p=top_ps,
+            frequency_penalty=freq_penalties,
+            input_ids=padded_ids,
+            input_mask=padded_mask,
         ).tolist()

astrai/inference/core/scheduler.py

@@ -138,36 +138,33 @@ 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)

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
@@ -40,6 +76,8 @@ class Task:
         self.temperature = temperature
         self.top_p = top_p
         self.top_k = top_k
+        self.frequency_penalty = frequency_penalty
+        self.rep_window = rep_window
 
         self.status = TaskStatus.PENDING
         self.output_ids: List[int] = []
@@ -47,6 +85,34 @@ class Task:
         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 +158,8 @@ 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 +184,8 @@ class TaskManager:
             temperature=temperature,
             top_p=top_p,
             top_k=top_k,
+            frequency_penalty=frequency_penalty,
+            rep_window=rep_window,
         )
 
         with self._lock:

astrai/inference/engine.py

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

astrai/inference/sample.py

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

astrai/parallel/executor.py

@@ -7,6 +7,7 @@ from contextlib import contextmanager
 from typing import Optional, Tuple
 
 import torch
+import torch.distributed as dist
 import torch.nn as nn
 from torch.distributed.fsdp import FullStateDictConfig, StateDictType
 from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
@@ -120,6 +121,21 @@ class BaseExecutor:
     def unwrap_model(self, model: nn.Module):
         return model.state_dict()
 
+    @contextmanager
+    def checkpoint_context(self, model: nn.Module):
+        if self.use_distributed:
+            dist.barrier()
+        state_dict = self._gather_state_dict(model)
+        yield state_dict
+        if self.use_distributed:
+            dist.barrier()
+
+    def _gather_state_dict(self, model: nn.Module):
+        state_dict = self.unwrap_model(model)
+        if self.use_distributed and get_rank() != 0:
+            return None
+        return state_dict
+
     @property
     def use_distributed(self) -> bool:
         return get_world_size() > 1
@@ -132,7 +148,14 @@ class BaseExecutor:
     def grad_accum_steps(self) -> int:
         return self.gradient_state.num_steps
 
-    def clip_grad_norm(self, model: nn.Module, max_norm: float) -> float:
+    def clip_grad_norm(self, model: nn.Module, max_norm: Optional[float]) -> float:
+        if max_norm is None:
+            total_norm = torch.norm(
+                torch.stack(
+                    [p.grad.norm(2) for p in model.parameters() if p.grad is not None]
+                )
+            )
+            return total_norm.item()
         total_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
         if isinstance(total_norm, torch.Tensor):
             return total_norm.item()
@@ -266,7 +289,9 @@ class FSDPExecutor(BaseExecutor):
             return model.no_sync()
         return contextlib.nullcontext()
 
-    def clip_grad_norm(self, model: nn.Module, max_norm: float) -> float:
+    def clip_grad_norm(self, model: nn.Module, max_norm: Optional[float]) -> float:
+        if max_norm is None:
+            return super().clip_grad_norm(model, max_norm)
         if isinstance(model, FSDP) and self.use_distributed:
             total_norm = model.clip_grad_norm_(max_norm)
             if isinstance(total_norm, torch.Tensor):
@@ -279,7 +304,7 @@ class FSDPExecutor(BaseExecutor):
             with FSDP.state_dict_type(
                 model,
                 StateDictType.FULL_STATE_DICT,
-                FullStateDictConfig(offload_to_cpu=True, rank0_only=False),
+                FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
             ):
                 return model.state_dict()
 

astrai/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

@@ -95,8 +95,15 @@ class SectionRenderer:
         return all_ids, loss_mask
 
     def process_list_field(self, item: dict, sections: list, config, tokenizer):
-        all_ids: list[int] = []
+        """Tokenize a list-valued field, preserving per-element boundaries.
-        loss_mask: list[int] = []
+
+        Returns ``(list_of_id_lists, list_of_mask_lists)`` where each
+        inner list corresponds to one element of the source list.  This
+        is critical for GRPO where each response must stay a separate
+        sequence so the strategy can form a ``[G, R]`` tensor.
+        """
+        per_item_ids: list[list[int]] = []
+        per_item_masks: list[list[int]] = []
 
         for sec in sections:
             field = sec["field"]
@@ -108,17 +115,13 @@ class SectionRenderer:
                 continue
 
             for val in values:
+                ids: list[int] = []
+                mask: list[int] = []
                 if use_template:
                     if isinstance(val, list):
                         wrapper = {field: val}
                         self._append_template(
-                            wrapper,
+                            wrapper, field, action, tokenizer, config, ids, mask
-                            field,
-                            action,
-                            tokenizer,
-                            config,
-                            all_ids,
-                            loss_mask,
                         )
                 else:
                     wrapper = {field: str(val)}
@@ -130,17 +133,19 @@ 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
 
     @staticmethod
     def is_value_section(sections: list) -> bool:
@@ -282,10 +287,18 @@ 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

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 +21,13 @@ 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,
+    iter_raw_records,
+    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 +70,18 @@ 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 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)
@@ -102,14 +106,7 @@ class Pipeline:
                 continue
 
             domain = result.pop("domain", "__default__")
-
+            ids = primary_ids(result)
-            is_multi = bool(getattr(self.config.input, "sources", None))
-            if is_multi:
-                ids = self._primary_ids(result)
-            else:
-                ids = result.pop("sequence")
-                result["sequence"] = ids
-
             if not ids:
                 continue
 
@@ -129,15 +126,6 @@ class Pipeline:
         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 +137,18 @@ 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 _flush(self, domains, shard_idx):
         for domain, keys in domains.items():
@@ -163,24 +158,12 @@ 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 +173,76 @@ 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)
@@ -74,3 +105,19 @@ def load_bin(file_path: str) -> Dict[str, List[Tensor]]:
         )
         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/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

@@ -164,7 +164,14 @@ class AutoTokenizer:
           - tokenizer.bos_token   → returns string
           - tokenizer.bos_token_id → returns corresponding integer ID
           - tokenizer.stop_ids → returns list of corresponding integer IDs for all special tokens
+
+        Internal/private attrs are not intercepted: during unpickle
+        ``__dict__`` is empty, so probing ``self._special_token_map``
+        would recurse infinitely.
         """
+        if key.startswith("_"):
+            raise AttributeError(key)
+
         # Handle stop_ids - return IDs for all special tokens
         if key == "stop_ids":
             stop_ids = []

astrai/trainer/strategy.py

@@ -98,7 +98,6 @@ 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
 
     @abstractmethod
@@ -223,14 +222,13 @@ 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
 
@@ -267,42 +265,45 @@ class DPOStrategy(BaseStrategy):
 class GRPOStrategy(BaseStrategy):
     """Group Relative Policy Optimization strategy.
 
-    On-policy GRPO following DeepSeek-R1: the policy model is updated while
+    Implements GRPO following DeepSeek-R1 with token-level PPO clipping.
-    a frozen ref_model stores the old-policy log-probs.  ratio = exp(logπ_θ - logπ_ref),
+    Advantages are group-normalized from scalar per-response rewards and
-    clipped PPO objective.  Call ``sync_ref_model()`` after each data-generation round.
+    broadcast across all response tokens.  The loss is computed **only on
+    response tokens** — prompt tokens are masked out.
+
+    Three model roles are distinguished:
+
+    * **Policy** ``self.model`` — the model being trained.
+    * **Old policy** ``self.old_model`` — the behaviour policy that generated
+      the responses.  Used for the importance sampling ratio
+      ``ρ = π_θ / π_old``.  Synced externally after each data-generation round.
+    * **Reference model** ``self.ref_model`` — a frozen copy of the initial
+      policy (typically the SFT checkpoint) used **only** for the KL
+      regularisation term.  It is never updated during training.
     """
 
     def __init__(
         self,
         model: nn.Module,
         device: str,
+        old_model: nn.Module,
+        ref_model: nn.Module,
         clip_eps: float = 0.2,
         kl_coef: float = 0.01,
         group_size: int = 4,
-        reduction: str = "mean",
-        sync_interval: int = 200,
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(
+        self.old_model = old_model
-            self.model_fn, self.executor.unwrap_model(model)
+        self.ref_model = ref_model
-        ).to(device=self.device)
         self.clip_eps = clip_eps
         self.kl_coef = kl_coef
         self.group_size = group_size
-        self.reduction = reduction
-        self.sync_interval = sync_interval
-        self._step = 0
 
-    def sync_ref_model(self):
+    def sync_old_model(self):
-        """Copy current model weights to ref model."""
+        """Copy current policy weights to old model."""
-        self.ref_model.load_state_dict(self.executor.unwrap_model(self.model))
+        self.old_model.load_state_dict(self.executor.unwrap_model(self.model))
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
-        self._step += 1
-        if self._step % self.sync_interval == 0:
-            self.sync_ref_model()
-
         batch = move_to_device(batch, self.device)
         prompts = batch["prompts"]
         responses = batch["responses"]
@@ -313,33 +314,60 @@ class GRPOStrategy(BaseStrategy):
         responses_flat = responses.view(-1, response_len)
         masks_flat = masks.view(-1, response_len)
         prompt_expanded = prompts.unsqueeze(1).repeat(1, group_size, 1).flatten(0, 1)
+        prompt_len = prompt_expanded.size(1)
 
         full_sequences = torch.cat([prompt_expanded, responses_flat], dim=-1)
-        full_masks = torch.cat([torch.ones_like(prompt_expanded), masks_flat], dim=-1)
+        # Prompt tokens are masked out (0) so logprobs are computed only for
-
+        # response tokens.  get_logprobs shifts the mask by one position, so
-        log_probs_policy = get_logprobs(
+        # the first response token's logprob (predicted from the last prompt
-            self.model, full_sequences, full_masks, self.reduction
+        # token) is correctly included.
-        )
+        full_masks = torch.cat([torch.zeros_like(prompt_expanded), masks_flat], dim=-1)
-        log_probs_policy = log_probs_policy.view(batch_size, group_size)
 
+        # get_logprobs returns [B*G, S-1] (S = prompt_len + response_len).
+        # Response token logprobs occupy the last ``response_len`` positions
+        # (the first response token is predicted from the last prompt token).
+        token_log_probs_policy = get_logprobs(
+            self.model, full_sequences, full_masks, "none"
+        )[:, prompt_len - 1 :]
         with torch.no_grad():
-            log_probs_ref = get_logprobs(
+            token_log_probs_old = get_logprobs(
-                self.ref_model, full_sequences, full_masks, self.reduction
+                self.old_model, full_sequences, full_masks, "none"
-            )
+            )[:, prompt_len - 1 :]
-            log_probs_ref = log_probs_ref.view(batch_size, group_size)
+            token_log_probs_ref = get_logprobs(
+                self.ref_model, full_sequences, full_masks, "none"
+            )[:, prompt_len - 1 :]
 
-        eps = torch.finfo(log_probs_policy.dtype).eps
+        # Reshape to [B, G, response_len]
+        token_log_probs_policy = token_log_probs_policy.view(batch_size, group_size, -1)
+        token_log_probs_old = token_log_probs_old.view(batch_size, group_size, -1)
+        token_log_probs_ref = token_log_probs_ref.view(batch_size, group_size, -1)
+        token_masks = masks_flat.view(batch_size, group_size, -1).float()
+
+        # Group-normalized advantages from scalar per-response rewards.
+        eps = 1e-8
         mean = rewards.mean(dim=-1, keepdim=True)
-        std = rewards.std(dim=-1, keepdim=True)
+        std = rewards.std(dim=-1, keepdim=True, unbiased=False)
         advantages = (rewards - mean) / (std + eps)
+        # Broadcast scalar advantage to every response token: [B, G, 1]
+        advantages = advantages.unsqueeze(-1)
 
-        ratio = torch.exp(log_probs_policy - log_probs_ref)
+        # Token-level ratio (π_θ / π_old) and PPO clipping.
+        log_ratio = token_log_probs_policy - token_log_probs_old
+        ratio = torch.exp(log_ratio)
 
         surr1 = ratio * advantages
         surr2 = torch.clamp(ratio, 1 - self.clip_eps, 1 + self.clip_eps) * advantages
+        per_token_policy_loss = -torch.min(surr1, surr2)
+        token_count = token_masks.sum().clamp(min=1.0)
+        policy_loss = (per_token_policy_loss * token_masks).sum() / token_count
+
+        # KL penalty to frozen reference model with k1 estimator (non-negative):
+        # k1 = π_ref / π_θ - log(π_ref / π_θ) - 1, where π_ref / π_θ = exp(log_ref - log_policy).
+        log_ref_ratio = token_log_probs_ref - token_log_probs_policy
+        r = torch.exp(log_ref_ratio)
+        kl_per_token = r - torch.log(r + eps) - 1.0
+        kl_penalty = self.kl_coef * (kl_per_token * token_masks).sum() / token_count
 
-        policy_loss = -torch.min(surr1, surr2).mean()
-        kl_penalty = self.kl_coef * (log_probs_policy - log_probs_ref).square().mean()
         total_loss = policy_loss + kl_penalty
 
         return total_loss

astrai/trainer/train_callback.py

@@ -14,7 +14,7 @@ from tqdm import tqdm
 
 from astrai.factory import BaseFactory
 from astrai.parallel import only_on_rank
-from astrai.parallel.setup import get_current_device, get_rank
+from astrai.parallel.setup import get_current_device
 from astrai.serialization import Checkpoint
 from astrai.trainer.metric_util import (
     ctx_get_grad_norm,
@@ -139,28 +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

@@ -7,13 +7,13 @@ 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.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.trainer.strategy import BaseStrategy, StrategyFactory, create_ref_model
 
 
 @dataclass
@@ -54,10 +54,12 @@ class TrainContextBuilder:
         config: TrainConfig,
     ):
         self.config = config
-        self._resume_dir: Optional[str] = None
+        self._param_path: Optional[str] = None
+        self._resume: bool = False
 
-    def with_resume_dir(self, resume_dir: Optional[str]) -> Self:
+    def with_param_path(self, param_path: Optional[str], resume: bool = False) -> Self:
-        self._resume_dir = resume_dir
+        self._param_path = param_path
+        self._resume = resume
         return self
 
     def build(self) -> TrainContext:
@@ -74,8 +76,8 @@ class TrainContextBuilder:
         model = model.to(device=device)
 
         model_config = {}
-        if self._resume_dir:
+        if self._param_path:
-            config_path = Path(self._resume_dir) / "config.json"
+            config_path = Path(self._param_path) / "config.json"
             if config_path.exists():
                 model_config = load_json(config_path)
 
@@ -91,18 +93,29 @@ class TrainContextBuilder:
             executor=executor,
         )
 
-        if self._resume_dir:
+        if self._param_path:
-            checkpoint = Checkpoint.load_any(self._resume_dir)
+            checkpoint = Checkpoint.load_any(self._param_path)
             if checkpoint is not None:
                 model.load_state_dict(checkpoint.state_dict, strict=False)
                 if checkpoint.config:
                     context.model_config = checkpoint.config
-                context.epoch = checkpoint.epoch or cfg.start_epoch
+
-                if checkpoint.consumed_samples > 0:
+                if self._resume:
-                    context.consumed_samples = checkpoint.consumed_samples
+                    context.epoch = checkpoint.epoch or cfg.start_epoch
-                else:
+                    if checkpoint.consumed_samples > 0:
-                    context.consumed_samples = cfg.start_samples * context.world_size
+                        per_step = (
-                context.checkpoint = checkpoint
+                            cfg.batch_per_device
+                            * context.world_size
+                            * cfg.grad_accum_steps
+                        )
+                        context.consumed_samples = (
+                            checkpoint.consumed_samples // per_step
+                        ) * per_step
+                    else:
+                        context.consumed_samples = (
+                            cfg.start_samples * context.world_size
+                        )
+                    context.checkpoint = checkpoint
 
         if cfg.lora is not None:
             inject_lora(
@@ -128,7 +141,7 @@ 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 +154,11 @@ 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,
@@ -158,6 +172,7 @@ 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 = (
@@ -177,13 +192,26 @@ class TrainContextBuilder:
                     if obj is not None:
                         obj.load_state_dict(extra[name])
 
+        strategy_kwargs = dict(cfg.extra_kwargs)
+
+        if cfg.strategy in ("dpo", "grpo"):
+            ref_model = create_ref_model(
+                cfg.model_fn, executor.unwrap_model(context.model)
+            ).to(device=device)
+            strategy_kwargs["ref_model"] = ref_model
+
+        if cfg.strategy == "grpo":
+            old_model = create_ref_model(
+                cfg.model_fn, executor.unwrap_model(context.model)
+            ).to(device=device)
+            strategy_kwargs["old_model"] = old_model
+
         context.strategy = StrategyFactory.create(
             cfg.strategy,
             model=context.model,
             device=device,
             executor=executor,
-            model_fn=cfg.model_fn,
+            **strategy_kwargs,
-            **cfg.extra_kwargs,
         )
 
         return context

astrai/trainer/trainer.py

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

csrc/kernels/attn_common.h

@@ -10,11 +10,19 @@ struct AttentionParams {
     int kv_len;
     int head_dim;
     int use_mask;
-    int is_causal;
+    int causal_offset;   // -1 = non-causal; >=0 = absolute position of first Q token
-    int causal_offset;
     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;
@@ -34,7 +42,6 @@ struct PagedAttentionParams {
     int kv_len;
     int head_dim;
     int use_mask;
-    int is_causal;
     int causal_offset;
     float scale;
 
@@ -42,12 +49,19 @@ struct PagedAttentionParams {
     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

@@ -5,24 +5,12 @@
 #include "attn_decode_split_kv_mma.cuh"
 #endif
 
-static int decode_num_splits(int base_blocks, int tiles_total) {
-    int sm_count = 0;
-    cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, 0);
-    int n = (2 * sm_count + base_blocks - 1) / base_blocks;
-    return std::max(1, std::min(n, std::min(tiles_total, 32)));
-}
-
 // Scalar fallback: one warp per query head, split-KV across grid.z.
 static void launch_scalar_decode(AttentionParams<bf16>& p) {
     int group_size = p.q_head / p.kv_head;
     int chunks_total = (p.kv_len + DC_CHUNK - 1) / DC_CHUNK;
-    p.num_splits = decode_num_splits(p.batch * p.kv_head, chunks_total);
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
-
+    alloc_split_partials(p);
-    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
-    auto o_part = torch::empty({p.batch, p.q_head, p.num_splits, p.head_dim}, fopt);
-    auto ml_part = torch::empty({p.batch, p.q_head, p.num_splits, 2}, fopt);
-    p.o_part = o_part.data_ptr<float>();
-    p.ml_part = ml_part.data_ptr<float>();
 
     size_t smem = DC_CHUNK * p.head_dim * sizeof(bf16);
     attn_decode_split_kv_kernel<<<dim3(p.batch * p.kv_head, 1, p.num_splits), dim3(32, group_size), smem>>>(p);
@@ -38,13 +26,8 @@ static void launch_scalar_decode(AttentionParams<bf16>& p) {
 template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
 static void launch_mma_decode(AttentionParams<bf16>& p) {
     int tiles_total = (p.kv_len + BC - 1) / BC;
-    p.num_splits = decode_num_splits(p.batch * p.kv_head, tiles_total);
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
-
+    alloc_split_partials(p);
-    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
-    auto o_part = torch::empty({p.batch, p.q_head, p.num_splits, p.head_dim}, fopt);
-    auto ml_part = torch::empty({p.batch, p.q_head, p.num_splits, 2}, fopt);
-    p.o_part = o_part.data_ptr<float>();
-    p.ml_part = ml_part.data_ptr<float>();
 
     attn_decode_split_kv_mma_kernel<HEAD_DIM, BC, STAGES><<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
     attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
@@ -55,7 +38,7 @@ template <int HEAD_DIM>
 static void dispatch_decode(AttentionParams<bf16>& p) {
 #ifndef ASTRAI_NO_MMA
     int G = p.q_head / p.kv_head;
-    if (!p.use_mask && G >= 1 && G <= 16) {
+    if (G >= 1 && G <= 16) {
         launch_mma_decode<HEAD_DIM, 32>(p);
         return;
     }
@@ -68,35 +51,21 @@ torch::Tensor attn_decode(
     torch::Tensor k,
     torch::Tensor v,
     c10::optional<torch::Tensor> mask,
-    bool is_causal = false,
+    int64_t causal_offset,
-    int64_t causal_offset = 0,
+    double scale,
-    c10::optional<double> scale = c10::nullopt
+    int64_t layout
 ) {
     AttentionParams<bf16> p;
-    attn_pack_params(q, k, v, mask, is_causal, causal_offset, scale, 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_like(q);
+    // O matches Q's original layout
-    p.o = (bf16*)O.data_ptr();
+    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();
 
-    switch (p.head_dim) {
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_decode, p);
-        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;
 }
 
@@ -106,8 +75,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
         py::arg("k"),
         py::arg("v"),
         py::arg("mask") = py::none(),
-        py::arg("is_causal") = false,
+        py::arg("causal_offset") = -1,
-        py::arg("causal_offset") = 0,
+        py::arg("scale") = 0.0,
-        py::arg("scale") = py::none(),
+        py::arg("layout") = 0,
         "GQA decode (tensor-core head-packing on sm_80+, scalar fallback)");
 }

csrc/kernels/attn_decode_split_kv.cuh

@@ -21,13 +21,16 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
     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_head + q_head) * 1) * p.head_dim + lane * hd_per_thread;
+    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]);
+        q_reg[i] = __bfloat162float(p.q[q_off + i * p.q_stride_d]);
 
-    int kv_base = ((batch * p.kv_head + kv_head) * p.kv_len) * p.head_dim;
+    // KV: [batch, kv_head, kv_len, head_dim] — stride-based base
-    int mask_base = batch * p.kv_len;
+    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};
 
@@ -43,9 +46,15 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
         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)
+        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];
+            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++) {
@@ -54,9 +63,10 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
                 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])
+            int kv_idx = chunk_start + s;
+            if (p.use_mask && p.mask && !p.mask[mask_base + kv_idx])
                 partial = -FLT_MAX;
-            if (p.is_causal && (chunk_start + s) > p.causal_offset)
+            if (p.causal_offset >= 0 && kv_idx > p.causal_offset)
                 partial = -FLT_MAX;
 
             float new_m = fmaxf(m, partial);
@@ -64,9 +74,10 @@ __global__ void attn_decode_split_kv_kernel(AttentionParams<bf16> p) {
             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;
+            // V: stride-based read
+            int v_off = kv_base + kv_idx * p.kv_stride_l + lane * hd_per_thread * p.kv_stride_d;
             for (int i = 0; i < hd_per_thread; i++)
-                acc_reg[i] = acc_reg[i] * alpha + __bfloat162float(p.v[v_off + i]) * beta;
+                acc_reg[i] = acc_reg[i] * alpha + __bfloat162float(p.v[v_off + i * p.kv_stride_d]) * beta;
             m = new_m;
         }
         __syncthreads();
@@ -94,6 +105,9 @@ __global__ void attn_decode_combine_kernel(AttentionParams<bf16> p) {
     int d = threadIdx.x;
     if (d >= p.head_dim) return;
 
+    int batch = bh / p.q_head;
+    int q_head = bh % p.q_head;
+
     size_t split_base = (size_t)bh * p.num_splits;
     const float* mlp = p.ml_part + split_base * 2;
     const float* op = p.o_part + split_base * p.head_dim;
@@ -112,5 +126,7 @@ __global__ void attn_decode_combine_kernel(AttentionParams<bf16> p) {
     }
 
     float inv = (l > 1e-20f) ? (1.0f / l) : 0.0f;
-    p.o[(size_t)bh * p.head_dim + d] = __float2bfloat16(acc * inv);
+    // Stride-based output write (q_len=1 for decode, so stride_l not needed)
+    int o_off = batch * p.q_stride_b + q_head * p.q_stride_h + d * p.q_stride_d;
+    p.o[o_off] = __float2bfloat16(acc * inv);
 }

csrc/kernels/attn_decode_split_kv_mma.cuh

@@ -8,39 +8,21 @@ using bf16 = __nv_bfloat16;
 
 // Split-K (FlashDecoding) tensor-core decode via GQA head-packing.
 //
-// Decode has q_len == 1, so S = q @ K^T is a GEMV per head — no tensor-core work
+// Decode has q_len == 1, so S = q @ K^T is a GEMV per head — no tensor-core
-// on its own. But GQA gives us G = q_head / kv_head query heads that all share
+// work on its own. But GQA gives us G = q_head / kv_head query heads that all
-// one kv_head. We pack those G heads into the M=16 rows of mma.sync.m16n8k16,
+// share one kv_head. We pack those G heads into the M=16 rows of
-// turning G independent GEMVs into a single GEMM that reuses each loaded K/V tile
+// mma.sync.m16n8k16, turning G independent GEMVs into a single GEMM that
-// across all G heads (K/V load is the decode bottleneck, so the reuse is the win,
+// reuses each loaded K/V tile across all G heads (K/V load is the decode
-// not the flops). The KV sequence is partitioned across gridDim.z blocks so that
+// bottleneck, so the reuse is the win, not the flops). The KV sequence is
-// a decode with only batch*kv_head independent tasks can fill all SMs. Each
+// partitioned across gridDim.z blocks so that a decode with only
-// (batch, kv_head, split) block computes an UN-normalised partial (Oacc, m, l)
+// batch*kv_head independent tasks can fill all SMs. Each (batch, kv_head,
-// over its KV slice; the combine kernel below reduces across splits. Fixes the
+// split) block computes an UN-normalised partial (Oacc, m, l) over its KV
-// "grid too small" bottleneck (0.04 waves/SM → many blocks) for long-context,
+// slice; the combine kernel below reduces across splits. Fixes the "grid too
+// small" bottleneck (0.04 waves/SM → many blocks) for long-context,
 // small-batch decode.
-//
+
-// Partial layout (float, contiguous):
-//   o_part : [batch, q_head, num_splits, HEAD_DIM]
-//   ml_part: [batch, q_head, num_splits, 2]  (m, l)
-//
-// 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
-//   - Q loaded directly from global into mma A-operand registers (no sQ staging,
-//     no prologue syncwarp) — frees shared memory for double-buffering
-//   - Double-buffered KV (STAGES=2): next tile's cp.async overlaps current
-//     tile's MMA compute — hides global load latency / boosts bandwidth
-//     utilization for small-batch (low-occupancy) decode
-//   - Predicated cp.async (cp_async_16_pred) for full AND partial tiles on one
-//     uniform path — eliminates the scalar fallback branch
-//
-// Smem footprint (BC=32): STAGES=2 → 2*(sK+sV) = 2*2*32*HEAD_DIM*2 bytes.
-//   D=128: 16 KB (fits 48 KB static cap).  D=256: 32 KB (also fits).
-// STAGES=1 fallback (4/8 KB) for smem-constrained configs.
 template <int HEAD_DIM, int BC, int STAGES = 2>
 __global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
-    constexpr int BR = 16;
     constexpr int KD = HEAD_DIM / 16;
     constexpr int NC8 = BC / 8;
     constexpr int KT2 = BC / 16;
@@ -66,25 +48,13 @@ __global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
     __shared__ __align__(16) bf16 sV[STAGES * BC * LD];
 
     // ---- Load Q directly from global into mma A-operand registers ----
-    // Same layout as prefill: frag[0]/[2] = row gid, frag[1]/[3] = row gid+8
+    const int q_base = batch * p.q_stride_b + q_head0 * p.q_stride_h;
-    // cols kt*16 + tid4*2 + {0,1} / +{8,9}. pau[0]=cols c,c+1; pau[4]=c+8,c+9.
-    const int q_base = (batch * p.q_head + q_head0) * HEAD_DIM;
     const int qra = gid;
     const int qrb = gid + 8;
     const bool va = qra < G, vb = qrb < G;
     unsigned Qa[KD][4];
-#pragma unroll
+    load_q_mma_frags<KD>(p.q + q_base, p.q_stride_h, p.q_stride_d,
-    for (int kt = 0; kt < KD; kt++) {
+                          qra, qrb, va, vb, tid4, Qa);
-        int c = kt * 16 + tid4 * 2;
-        const unsigned* pau = reinterpret_cast<const unsigned*>(
-            &p.q[q_base + qra * HEAD_DIM + c]);
-        const unsigned* pbu = reinterpret_cast<const unsigned*>(
-            &p.q[q_base + qrb * HEAD_DIM + c]);
-        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;
-    }
 
     float Oacc[DN8][4];
 #pragma unroll
@@ -92,8 +62,8 @@ __global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
         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;
+    // KV: stride-based base — [batch, kv_head, kv_len, head_dim]
-    const int mask_base = batch * p.kv_len;
+    const int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
     const int tiles_total = (p.kv_len + BC - 1) / BC;
     const int tiles_per_split = (tiles_total + p.num_splits - 1) / p.num_splits;
     const int ti_begin = split * tiles_per_split;
@@ -111,8 +81,10 @@ __global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
             int kc = kv0 + r;
             bool valid = kc < p.kv_len;
             int off = r * LD + swiz_col(d, r, SWIZ_MASK);
-            cp_async_16_pred(&dK[off], &p.k[kv_base + kc * HEAD_DIM + d], valid);
+            // KV stride-based: contiguous within head_dim (stride_d == 1 typically)
-            cp_async_16_pred(&dV[off], &p.v[kv_base + kc * HEAD_DIM + d], valid);
+            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();
     };
@@ -148,9 +120,13 @@ __global__ void attn_decode_split_kv_mma_kernel(AttentionParams<bf16> p) {
             Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
             Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
 
-        int maxc = p.is_causal ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
+        // Decode: q_len=1, so qrow0=qrow1=0, mask_q_stride irrelevant
+        int maxc = (p.causal_offset >= 0) ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
         mma_softmax_tile<NC8, DN8>(kv0, maxc, maxc,
-                                    mask_base, p.mask, has_mask,
+                                    0, 0,
+                                    p.mask_b_stride, 0,
+                                    batch,
+                                    p.mask, has_mask,
                                     Sacc, Oacc, m0, m1, l0, l1, lane);
 
         mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);

csrc/kernels/attn_entry_utils.cuh

@@ -1,43 +1,177 @@
 #pragma once
 #include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
 #include "attn_common.h"
 
+using bf16 = __nv_bfloat16;
+
+inline int compute_num_splits(int base_blocks, int tiles_total) {
+    int sm_count = 0;
+    cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, 0);
+    int n = (2 * sm_count + base_blocks - 1) / base_blocks;
+    return std::max(1, std::min(n, std::min(tiles_total, 32)));
+}
+
+// Dispatch head_dim: shared macro — avoids C++20 lambda template syntax.
+// Usage: DISPATCH_HEAD_DIM(hd, fn, arg)
+//   Expands to: fn<32>(arg); fn<64>(arg); etc.
+#define DISPATCH_HEAD_DIM(hd, fn, arg) \
+    switch (hd) { \
+        case 32:  fn<32>(arg); break; \
+        case 64:  fn<64>(arg); break; \
+        case 128: fn<128>(arg); break; \
+        case 256: fn<256>(arg); break; \
+        default: \
+            TORCH_CHECK(false, "unsupported head_dim ", hd, \
+                         " (supported: 32, 64, 128, 256)"); \
+    }
+
+template<typename P>
+inline void alloc_split_partials(P& p) {
+    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
+    auto o_part = torch::empty({p.batch, p.q_head, p.num_splits, p.head_dim}, fopt);
+    auto ml_part = torch::empty({p.batch, p.q_head, p.num_splits, 2}, fopt);
+    p.o_part = (float*)o_part.data_ptr();
+    p.ml_part = (float*)ml_part.data_ptr();
+}
+
+// ---- Shared Q-dims + strides extraction ----
+template <typename P>
+inline void extract_q_dims_and_strides(torch::Tensor& q, int64_t layout, P& p) {
+    if (layout == 1) q = q.transpose(1, 2);
+    p.batch = (int)q.size(0);
+    p.q_head = (int)q.size(1);
+    p.q_len = (int)q.size(2);
+    p.head_dim = (int)q.size(3);
+    p.q_stride_b = (int)q.stride(0);
+    p.q_stride_h = (int)q.stride(1);
+    p.q_stride_l = (int)q.stride(2);
+    p.q_stride_d = (int)q.stride(3);
+}
+
+// ---- Shared mask packing ----
+template <typename P>
+inline void pack_mask(const c10::optional<torch::Tensor>& mask, P& p) {
+    if (p.use_mask) {
+        auto m = mask.value();
+        TORCH_CHECK(m.is_cuda(), "mask must be on CUDA");
+        TORCH_CHECK(m.dtype() == torch::kBool, "mask must be bool");
+        TORCH_CHECK(m.size(0) == p.batch, "mask batch mismatch");
+        TORCH_CHECK(m.size(m.dim() - 1) == p.kv_len, "mask kv_len mismatch");
+        if (m.dim() == 2) {
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_q_stride = 0;
+        } else if (m.dim() == 3) {
+            TORCH_CHECK(m.size(1) == p.q_len, "mask q_len mismatch");
+            p.mask_b_stride = (int)m.stride(0);
+            p.mask_q_stride = (int)m.stride(1);
+        } else {
+            TORCH_CHECK(false, "mask must be 2D [batch, kv_len] or 3D [batch, q_len, kv_len]");
+        }
+        p.mask = m.data_ptr<bool>();
+    } else {
+        p.mask = nullptr;
+        p.mask_b_stride = 0;
+        p.mask_q_stride = 0;
+    }
+}
+
+// ---- attn_pack_params (contiguous KV) ----
 template<typename T>
 inline void attn_pack_params(
     torch::Tensor q,
     torch::Tensor k,
     torch::Tensor v,
     c10::optional<torch::Tensor> mask,
-    bool is_causal,
     int64_t causal_offset,
-    c10::optional<double> scale,
+    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.batch = (int)q.size(0);
-    p.q_head = (int)q.size(1);
     p.kv_head = (int)k.size(1);
-    p.q_len = (int)q.size(2);
     p.kv_len = (int)k.size(2);
-    p.head_dim = (int)q.size(3);
+    TORCH_CHECK(k.size(3) == p.head_dim, "K/V head_dim must match Q");
-    p.use_mask = mask.has_value() ? 1 : 0;
+
-    p.is_causal = is_causal ? 1 : 0;
+    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.scale = scale.has_value() ? (float)scale.value() : 1.0f / sqrtf((float)p.head_dim);
+    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();
-    if (p.use_mask) {
+    p.o = nullptr;
-        TORCH_CHECK(mask.value().dtype() == torch::kBool);
+    p.o_part = nullptr;
-        TORCH_CHECK(mask.value().dim() == 2);
+    p.ml_part = nullptr;
-        TORCH_CHECK(mask.value().size(0) == p.batch);
+
-        TORCH_CHECK(mask.value().size(1) == p.kv_len);
+    pack_mask(mask, p);
-        p.mask = mask.value().data_ptr<bool>();
+}
-    } else {
+
-        p.mask = nullptr;
+// ---- 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

@@ -108,6 +108,36 @@ __device__ __forceinline__ void cp_async_wait_group() {
     asm volatile("cp.async.wait_group %0;" :: "n"(N));
 }
 
+// ---------------------------------------------------------------------------
+// Q-load: load query rows directly from global memory into mma A-operand
+// register layout. One call replaces ~15 duplicated lines in each MMA kernel.
+// stride_row is p.q_stride_h for decode (q_len=1, G heads) or
+//              p.q_stride_l for prefill (multi-q rows).
+// ---------------------------------------------------------------------------
+template <int KD>
+__device__ inline void load_q_mma_frags(
+    const bf16* __restrict__ q,
+    int stride_row,
+    int stride_d,
+    int qra, int qrb,
+    bool va, bool vb,
+    int tid4,
+    unsigned Qa[KD][4])
+{
+    #pragma unroll
+    for (int kt = 0; kt < KD; kt++) {
+        int c = kt * 16 + tid4 * 2;
+        const unsigned* pau = reinterpret_cast<const unsigned*>(
+            &q[qra * stride_row + c * stride_d]);
+        const unsigned* pbu = reinterpret_cast<const unsigned*>(
+            &q[qrb * stride_row + c * stride_d]);
+        Qa[kt][0] = va ? pau[0] : 0u;
+        Qa[kt][1] = vb ? pbu[0] : 0u;
+        Qa[kt][2] = va ? pau[4] : 0u;
+        Qa[kt][3] = vb ? pbu[4] : 0u;
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Shared MMA compute functions — used by both decode and prefill MMA kernels.
 // Extracted because S=Q@K^T, online softmax, and P@V are structurally identical
@@ -122,7 +152,9 @@ template <int KD, int NC8>
 __device__ inline void mma_compute_scores(
     const unsigned Qa[KD][4],
     const bf16* __restrict__ sK,
-    int LD, int SWIZ_MASK, int lane,
+    int LD, 
+    int SWIZ_MASK, 
+    int lane,
     float Sacc[NC8][4])
 {
     #pragma unroll
@@ -142,13 +174,20 @@ __device__ inline void mma_compute_scores(
 // Online softmax + Oacc rescale for one K/V tile.
 //   maxc0/maxc1: per-row KV column bounds (prefill: per-query-row causal limits;
 //     decode: same value for both rows since q_len==1).
+//   qrow0/qrow1: query row indices (for 3D mask indexing; decode passes 0).
+//   mask_b_stride/mask_q_stride: mask layout (2D: mask_q_stride=0; 3D: =kv_len).
 // Reads Sacc (Q@K^T scores), applies causal/mask, computes P = exp(S - nm),
 // rescales Oacc by exp(m_old - nm), and updates m/l — all in place.
 template <int NC8, int DN8>
 __device__ inline void mma_softmax_tile(
     int kv0,
-    int maxc0, int maxc1,
+    int maxc0, 
-    int mask_base,
+    int maxc1,
+    int qrow0, 
+    int qrow1,
+    int mask_b_stride, 
+    int mask_q_stride,
+    int mask_batch,
     const bool* __restrict__ mask,
     bool has_mask,
     float Sacc[NC8][4],
@@ -162,14 +201,16 @@ __device__ inline void mma_softmax_tile(
     // Mask out-of-bounds / masked columns: set -FLT_MAX so expf → 0 downstream
     // without per-element sentinel checks. Compute tile-local row maxima.
     float rmax0 = -FLT_MAX, rmax1 = -FLT_MAX;
+    int mask_base0 = mask_batch * mask_b_stride + qrow0 * mask_q_stride;
+    int mask_base1 = mask_batch * mask_b_stride + qrow1 * mask_q_stride;
     #pragma unroll
     for (int n8 = 0; n8 < NC8; n8++) {
         int cc = kv0 + n8 * 8 + 2 * tid4;
         int c1 = cc + 1;
-        bool b0 = (cc >= maxc0) || (has_mask && !mask[mask_base + cc]);
+        bool b0 = (cc >= maxc0) || (has_mask && !mask[mask_base0 + cc]);
-        bool b1 = (c1 >= maxc0) || (has_mask && !mask[mask_base + c1]);
+        bool b1 = (c1 >= maxc0) || (has_mask && !mask[mask_base0 + c1]);
-        bool b2 = (cc >= maxc1) || (has_mask && !mask[mask_base + cc]);
+        bool b2 = (cc >= maxc1) || (has_mask && !mask[mask_base1 + cc]);
-        bool b3 = (c1 >= maxc1) || (has_mask && !mask[mask_base + c1]);
+        bool b3 = (c1 >= maxc1) || (has_mask && !mask[mask_base1 + c1]);
         float s0 = b0 ? -FLT_MAX : Sacc[n8][0];
         float s1 = b1 ? -FLT_MAX : Sacc[n8][1];
         float s2 = b2 ? -FLT_MAX : Sacc[n8][2];

csrc/kernels/attn_paged_decode.cu

@@ -3,26 +3,13 @@
 #include "attn_paged_decode_split_kv_mma.cuh"
 #endif
 
-#include <torch/extension.h>
+#include "attn_entry_utils.cuh"
-#include <c10/cuda/CUDAGuard.h>
-
-static int paged_decode_num_splits(int base_blocks, int tiles_total) {
-    int sm_count = 0;
-    cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, 0);
-    int n = (2 * sm_count + base_blocks - 1) / base_blocks;
-    return std::max(1, std::min(n, std::min(tiles_total, 32)));
-}
 
 static void launch_paged_scalar_decode(PagedAttentionParams<bf16>& p) {
     int group_size = p.q_head / p.kv_head;
     int chunks_total = (p.kv_len + PDC_CHUNK - 1) / PDC_CHUNK;
-    p.num_splits = paged_decode_num_splits(p.batch * p.kv_head, chunks_total);
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, chunks_total);
-
+    alloc_split_partials(p);
-    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
-    auto o_part = torch::empty({p.batch, p.q_head, p.num_splits, p.head_dim}, fopt);
-    auto ml_part = torch::empty({p.batch, p.q_head, p.num_splits, 2}, fopt);
-    p.o_part = o_part.data_ptr<float>();
-    p.ml_part = ml_part.data_ptr<float>();
 
     size_t smem = PDC_CHUNK * p.head_dim * sizeof(bf16);
     dim3 grid = dim3(p.batch * p.kv_head, 1, p.num_splits);
@@ -35,13 +22,8 @@ static void launch_paged_scalar_decode(PagedAttentionParams<bf16>& p) {
 template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
 static void launch_paged_mma_decode(PagedAttentionParams<bf16>& p) {
     int tiles_total = (p.kv_len + BC - 1) / BC;
-    p.num_splits = paged_decode_num_splits(p.batch * p.kv_head, tiles_total);
+    p.num_splits = compute_num_splits(p.batch * p.kv_head, tiles_total);
-
+    alloc_split_partials(p);
-    auto fopt = torch::TensorOptions().dtype(torch::kFloat32).device(torch::kCUDA);
-    auto o_part = torch::empty({p.batch, p.q_head, p.num_splits, p.head_dim}, fopt);
-    auto ml_part = torch::empty({p.batch, p.q_head, p.num_splits, 2}, fopt);
-    p.o_part = o_part.data_ptr<float>();
-    p.ml_part = ml_part.data_ptr<float>();
 
     paged_attn_decode_split_kv_mma_kernel<HEAD_DIM, BC, STAGES><<<dim3(p.kv_head, p.batch, p.num_splits), 32>>>(p);
     paged_attn_decode_combine_kernel<<<p.batch * p.q_head, p.head_dim>>>(p);
@@ -52,7 +34,7 @@ template <int HEAD_DIM>
 static void dispatch_paged_decode(PagedAttentionParams<bf16>& p) {
 #ifndef ASTRAI_NO_MMA
     int G = p.q_head / p.kv_head;
-    if (!p.use_mask && G >= 1 && G <= 16 && p.page_size >= 32) {
+    if (G >= 1 && G <= 16 && p.page_size >= 32) {
         launch_paged_mma_decode<HEAD_DIM, 32>(p);
         return;
     }
@@ -68,68 +50,19 @@ torch::Tensor attn_paged_decode(
     int64_t page_size,
     int64_t kv_len,
     c10::optional<torch::Tensor> mask,
-    bool is_causal = false,
+    int64_t causal_offset,
-    int64_t causal_offset = 0,
+    double scale,
-    c10::optional<double> scale = c10::nullopt
+    int64_t layout
 ) {
-    const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
+    PagedAttentionParams<bf16> p;
+    attn_pack_paged_params(q, page_table, k_cache, v_cache,
+                           page_size, kv_len, mask, causal_offset, scale, layout, p);
 
-    int batch = q.size(0);
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
-    int q_head = q.size(1);
+    auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
-    int head_dim = q.size(3);
+    p.o = (bf16*)O_view.data_ptr();
-    int kv_head = k_cache.size(2);
-    int max_pages = page_table.size(1);
-
-    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(q.size(2) == 1, "Q seq_len must be 1 (decode)");
-    TORCH_CHECK(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);
-    TORCH_CHECK(k_cache.size(0) >= 0, "k_cache must have at least 0 pages");
-
-    float scale_val = scale.has_value()
-        ? static_cast<float>(scale.value())
-        : 1.0f / std::sqrt(static_cast<float>(head_dim));
-
-    auto O = torch::empty_like(q);
-
-    PagedAttentionParams<bf16, float> p;
-    p.batch = batch;
-    p.q_head = q_head;
-    p.kv_head = kv_head;
-    p.q_len = static_cast<int>(q.size(2));
-    p.kv_len = static_cast<int>(kv_len);
-    p.head_dim = head_dim;
-    p.use_mask = (mask.has_value() && mask.value().defined()) ? 1 : 0;
-    p.is_causal = is_causal ? 1 : 0;
-    p.causal_offset = static_cast<int>(causal_offset);
-    p.scale = scale_val;
-    p.page_size = static_cast<int>(page_size);
-    p.max_pages = max_pages;
-    p.page_table = page_table.data_ptr<int64_t>();
-    p.k_cache = reinterpret_cast<const bf16*>(k_cache.data_ptr());
-    p.v_cache = reinterpret_cast<const bf16*>(v_cache.data_ptr());
-    p.q = reinterpret_cast<const bf16*>(q.data_ptr());
-    p.mask = p.use_mask ? mask.value().data_ptr<bool>() : nullptr;
-    p.o = reinterpret_cast<bf16*>(O.data_ptr());
-    p.o_part = nullptr;
-    p.ml_part = nullptr;
-
-    switch (p.head_dim) {
-        case 32:  dispatch_paged_decode<32>(p);  break;
-        case 64:  dispatch_paged_decode<64>(p);  break;
-        case 128: dispatch_paged_decode<128>(p); break;
-        case 256: dispatch_paged_decode<256>(p); break;
-        default:
-            TORCH_CHECK(false, "paged_decode: unsupported head_dim ", p.head_dim,
-                        " (supported: 32, 64, 128, 256)");
-    }
 
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_decode, p);
     return O;
 }
 
@@ -142,8 +75,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
         py::arg("page_size"),
         py::arg("kv_len"),
         py::arg("mask") = py::none(),
-        py::arg("is_causal") = false,
+        py::arg("causal_offset") = -1,
-        py::arg("causal_offset") = 0,
+        py::arg("scale") = 0.0,
-        py::arg("scale") = py::none(),
+        py::arg("layout") = 0,
         "Paged GQA decode — split-KV with direct page-table access.");
 }

csrc/kernels/attn_paged_decode_split_kv.cuh

@@ -22,11 +22,13 @@ __global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p)
     int lane = threadIdx.x;
     int hd_per_thread = p.head_dim / 32;
 
+    // Q: stride-based [batch, q_head, q_len=1, head_dim]
     float q_reg[8];
-    int q_off = ((batch * p.q_head + q_head) * 1) * p.head_dim + lane * hd_per_thread;
+    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]);
+        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};
 
@@ -37,7 +39,7 @@ __global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p)
     int ch_begin = split * chunks_per_split;
     int ch_end = min(chunks_total, ch_begin + chunks_per_split);
 
-    const int mask_base = batch * p.kv_len;
+    const int mask_base = batch * p.mask_b_stride;
 
     for (int ci = ch_begin; ci < ch_end; ci++) {
         int chunk_start = ci * PDC_CHUNK;
@@ -70,9 +72,10 @@ __global__ void paged_attn_decode_split_kv_kernel(PagedAttentionParams<bf16> p)
                 partial += q_reg[i] * __bfloat162float(k_smem[s * p.head_dim + lane * hd_per_thread + i]);
             partial = paged_warp_reduce_sum(partial) * p.scale;
 
-            if (p.use_mask && p.mask && !p.mask[mask_base + chunk_start + s])
+            int kv_idx = chunk_start + s;
+            if (p.use_mask && p.mask && !p.mask[mask_base + kv_idx])
                 partial = -FLT_MAX;
-            if (p.is_causal && (chunk_start + s) > p.causal_offset)
+            if (p.causal_offset >= 0 && kv_idx > p.causal_offset)
                 partial = -FLT_MAX;
 
             float new_m = fmaxf(m, partial);
@@ -118,6 +121,9 @@ __global__ void paged_attn_decode_combine_kernel(PagedAttentionParams<bf16> p) {
     int d = threadIdx.x;
     if (d >= p.head_dim) return;
 
+    int batch = bh / p.q_head;
+    int q_head = bh % p.q_head;
+
     size_t split_base = (size_t)bh * p.num_splits;
     const float* mlp = p.ml_part + split_base * 2;
     const float* op = p.o_part + split_base * p.head_dim;
@@ -136,5 +142,6 @@ __global__ void paged_attn_decode_combine_kernel(PagedAttentionParams<bf16> p) {
     }
 
     float inv = (l > 1e-20f) ? (1.0f / l) : 0.0f;
-    p.o[(size_t)bh * p.head_dim + d] = __float2bfloat16(acc * inv);
+    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

@@ -11,14 +11,9 @@ using bf16 = __nv_bfloat16;
 // directly from the page pool through a page table, eliminating the gather
 // copy.  Each tile (BC=32) fits within a single page (page_size >= 32), so
 // the page-table lookup happens once per tile for cp.async.
-//
+
-// Optimizations mirror attn_decode_split_kv_mma_kernel:
-//   - Q loaded directly from global into mma A-operand registers (no sQ)
-//   - Double-buffered KV (STAGES=2) for D<=128, single-buffer for D=256
-//   - Predicated cp.async for unified full/partial tile path
 template <int HEAD_DIM, int BC, int STAGES = (HEAD_DIM <= 128) ? 2 : 1>
 __global__ void paged_attn_decode_split_kv_mma_kernel(PagedAttentionParams<bf16> p) {
-    constexpr int BR = 16;
     constexpr int KD = HEAD_DIM / 16;
     constexpr int NC8 = BC / 8;
     constexpr int KT2 = BC / 16;
@@ -32,33 +27,23 @@ __global__ void paged_attn_decode_split_kv_mma_kernel(PagedAttentionParams<bf16>
     const int gid = lane >> 2;
     const int tid4 = lane & 3;
 
-    const int kv_head_idx = blockIdx.x;
+    const int kv_head = blockIdx.x;
     const int batch = blockIdx.y;
     const int split = blockIdx.z;
     const int G = p.q_head / p.kv_head;
-    const int q_head0 = kv_head_idx * G;
+    const int q_head0 = kv_head * G;
 
     __shared__ __align__(16) bf16 sK[STAGES * BC * LD];
     __shared__ __align__(16) bf16 sV[STAGES * BC * LD];
 
     // ---- Load Q directly from global into mma A-operand registers ----
-    const int q_base = (batch * p.q_head + q_head0) * HEAD_DIM;
+    const int q_base = batch * p.q_stride_b + q_head0 * p.q_stride_h;
     const int qra = gid;
     const int qrb = gid + 8;
     const bool va = qra < G, vb = qrb < G;
     unsigned Qa[KD][4];
-#pragma unroll
+    load_q_mma_frags<KD>(p.q + q_base, p.q_stride_h, p.q_stride_d,
-    for (int kt = 0; kt < KD; kt++) {
+                          qra, qrb, va, vb, tid4, Qa);
-        int c = kt * 16 + tid4 * 2;
-        const unsigned* pau = reinterpret_cast<const unsigned*>(
-            &p.q[q_base + qra * HEAD_DIM + c]);
-        const unsigned* pbu = reinterpret_cast<const unsigned*>(
-            &p.q[q_base + qrb * HEAD_DIM + c]);
-        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;
-    }
 
     float Oacc[DN8][4];
 #pragma unroll
@@ -66,7 +51,6 @@ __global__ void paged_attn_decode_split_kv_mma_kernel(PagedAttentionParams<bf16>
         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 mask_base = batch * p.kv_len;
     const int tiles_total = (p.kv_len + BC - 1) / BC;
     const int tiles_per_split = (tiles_total + p.num_splits - 1) / p.num_splits;
     const int ti_begin = split * tiles_per_split;
@@ -76,7 +60,7 @@ __global__ void paged_attn_decode_split_kv_mma_kernel(PagedAttentionParams<bf16>
     // Paged strides (constant for the block)
     const int64_t page_stride = (int64_t)p.page_size * p.kv_head * HEAD_DIM;
     const int64_t pos_stride  = (int64_t)p.kv_head * HEAD_DIM;
-    const int64_t head_off    = (int64_t)kv_head_idx * HEAD_DIM;
+    const int64_t head_off    = (int64_t)kv_head * HEAD_DIM;
 
     // ---- Load tile lambda: predicated cp.async, paged addressing ----
     auto load_tile = [&](int ti, int buf) {
@@ -130,9 +114,13 @@ __global__ void paged_attn_decode_split_kv_mma_kernel(PagedAttentionParams<bf16>
             Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
             Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
 
-        int maxc = p.is_causal ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
+        // Decode: q_len=1, so qrow0=qrow1=0, mask_q_stride irrelevant
+        int maxc = (p.causal_offset >= 0) ? min(p.kv_len, p.causal_offset + 1) : p.kv_len;
         mma_softmax_tile<NC8, DN8>(kv0, maxc, maxc,
-                                    mask_base, p.mask, has_mask,
+                                    0, 0,
+                                    p.mask_b_stride, 0,
+                                    batch,
+                                    p.mask, has_mask,
                                     Sacc, Oacc, m0, m1, l0, l1, lane);
 
         mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);

csrc/kernels/attn_prefill.cu

@@ -35,34 +35,19 @@ torch::Tensor attn_prefill(
     torch::Tensor k,
     torch::Tensor v,
     c10::optional<torch::Tensor> mask,
-    bool is_causal = false,
+    int64_t causal_offset,
-    int64_t causal_offset = 0,
+    double scale,
-    c10::optional<double> scale = c10::nullopt
+    int64_t layout
 ) {
     AttentionParams<bf16> p;
-    attn_pack_params(q, k, v, mask, is_causal, causal_offset, scale, 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_like(q);
+    auto O = torch::empty_strided(q.sizes(), q.strides(), q.options());
-    p.o = (bf16*)O.data_ptr();
+    auto O_view = (layout == 1) ? O.transpose(1, 2) : O;
+    p.o = (bf16*)O_view.data_ptr();
 
-    switch (p.head_dim) {
+    DISPATCH_HEAD_DIM(p.head_dim, dispatch_prefill, p);
-        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;
 }
 
@@ -72,8 +57,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
         py::arg("k"),
         py::arg("v"),
         py::arg("mask") = py::none(),
-        py::arg("is_causal") = false,
+        py::arg("causal_offset") = -1,
-        py::arg("causal_offset") = 0,
+        py::arg("scale") = 0.0,
-        py::arg("scale") = py::none(),
+        py::arg("layout") = 0,
         "GQA prefill (tensor-core mma on sm_80+, scalar fallback)");
 }

csrc/kernels/attn_prefill_split_q.cuh

@@ -50,12 +50,14 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
     __shared__ __align__(16) bf16 sK[P_BC * HEAD_DIM];
     __shared__ __align__(16) bf16 sV[P_BC * HEAD_DIM];
 
+    // Q: stride-based load [batch, q_head, q_len, head_dim]
     float qreg[DPT];
     if (q_row < p.q_len) {
-        int q_off = ((batch * p.q_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]) * p.scale;
     }
 
     float m = -FLT_MAX, l = 0.0f;
@@ -64,7 +66,9 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> 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;
@@ -79,15 +83,19 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> 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 (p.causal_offset >= 0 && q_row < p.q_len) {
             int ep = q_row + p.causal_offset + 1;
             if (kv0 >= ep)
                 lim = 0;
@@ -95,6 +103,7 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
                 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;
@@ -108,7 +117,8 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> p) {
             }
             float dot = group_reduce_sum<G>(part, gmask);
 
-            if (p.use_mask && p.mask && !p.mask[batch * p.kv_len + kv0 + s])
+            int kv_idx = kv0 + s;
+            if (p.use_mask && p.mask && !p.mask[mask_row_base + kv_idx])
                 dot = -FLT_MAX;
 
             float nm = fmaxf(m, dot);
@@ -131,10 +141,12 @@ __global__ void attn_prefill_split_q_kernel_t(AttentionParams<bf16> 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;
+        // O: stride-based write
+        int o_off = batch * p.q_stride_b + q_head * p.q_stride_h
+                  + q_row * p.q_stride_l + gpos * DPT * p.q_stride_d;
         float rl = (l > 1e-10f) ? (1.0f / l) : 0.0f;
 #pragma unroll
         for (int i = 0; i < DPT; i++)
-            p.o[o_off + i] = __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

@@ -57,7 +57,7 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
     const int kv_head = q_head / (p.q_head / p.kv_head);
     const int qrow0 = (blockIdx.x * WARPS + warp) * BR;
 
-    // Static shared memory — sized by template parameters at compile time.
+    // ---- Static shared memory: double-buffered K/V ----
     // K/V are double-buffered (STAGES=2): the next tile's cp.async load runs
     // while the current tile's tensor-core math executes, hiding global-load
     // latency (FA2-style software pipeline). No dynamic smem / carveout opt-in.
@@ -65,30 +65,16 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
     __shared__ __align__(16) bf16 sK[STAGES * BC * LD];
     __shared__ __align__(16) bf16 sV[STAGES * BC * LD];
 
-    // Load the Q fragments straight from global into the mma A-operand layout
+    // Load Q fragments straight from global into mma A-operand layout.
-    // (m16n8k16, row-major): no sQ staging area and no serialized per-warp
+    // stride_row = p.q_stride_l for prefill (multi-q rows across q_len).
-    // prologue barriers. Each lane reads exactly the 8 Q elements ldmatrix
+    // See attn_mma_utils.cuh for the shared template.
-    // would have produced, pre-scaled by the attention scale. Kept resident in
+    const int q_base = batch * p.q_stride_b + q_head * p.q_stride_h;
-    // registers across the tile loop.
-    //   frag[0]/[2]: row = qrow0 + gid ;  frag[1]/[3]: row = qrow0 + gid + 8
-    //   frag[0]/[1]: cols kt*16 + tid4*2 + {0,1} ; frag[2]/[3]: + 8
-    const int q_base = ((batch * p.q_head + q_head) * p.q_len) * HEAD_DIM;
     const int qra = qrow0 + gid;
     const int qrb = qrow0 + gid + 8;
     const bool va = qra < p.q_len, vb = qrb < p.q_len;
     unsigned Qa[KD][4];
-#pragma unroll
+    load_q_mma_frags<KD>(p.q + q_base, p.q_stride_l, p.q_stride_d,
-    for (int kt = 0; kt < KD; kt++) {
+                          qra, qrb, va, vb, tid4, Qa);
-        int c = kt * 16 + tid4 * 2;
-        const unsigned* pau = reinterpret_cast<const unsigned*>(
-            &p.q[q_base + qra * HEAD_DIM + c]);
-        const unsigned* pbu = reinterpret_cast<const unsigned*>(
-            &p.q[q_base + qrb * HEAD_DIM + c]);
-        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;
-    }
 
     float Oacc[DN8][4];
 #pragma unroll
@@ -96,18 +82,18 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
         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;
+    // KV: stride-based base
+    const int kv_base = batch * p.kv_stride_b + kv_head * p.kv_stride_h;
     const int tiles = (p.kv_len + BC - 1) / BC;
     const int qr0 = qrow0 + gid;      // row for c0/c1
     const int qr1 = qrow0 + gid + 8;  // row for c2/c3
 
-    // Causal tile-skip bounds (no-op when is_causal == 0)
+    // Causal tile-skip bounds (no-op when causal_offset < 0)
-    const int use_skip = p.is_causal;
+    const int use_skip = (p.causal_offset >= 0) ? 1 : 0;
     const int max_kv = qrow0 + BR - 1 + p.causal_offset;
     const int block_max_kv =
         blockIdx.x * WARPS * BR + WARPS * BR - 1 + p.causal_offset;
     const int has_mask = p.use_mask && p.mask;
-    const int mb = batch * p.kv_len;
 
     // Last active tile: block-level causal bound (all warps in the block share
     // the K/V load, so the prefetch range is the block max, not per-warp).
@@ -120,6 +106,7 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
     constexpr int VEC = 8;  // bf16 per cp.async unit (16 bytes)
     constexpr int TOTAL = BC * HEAD_DIM;
 
+    // ---- Load tile lambda: predicated cp.async ----
     // Issue cp.async loads for tile `ti` into shared buffer `buf`. Predicated
     // loads zero-fill rows past kv_len, so partial tiles need no scalar path.
     auto load_tile = [&](int ti, int buf) {
@@ -132,13 +119,14 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
             int kc = kv0 + r;
             bool valid = kc < p.kv_len;
             int off = r * LD + swiz_col(d, r, SWIZ_MASK);
-            cp_async_16_pred(&dK[off], &p.k[kv_base + kc * HEAD_DIM + d], valid);
+            int g_off = kv_base + kc * p.kv_stride_l + d * p.kv_stride_d;
-            cp_async_16_pred(&dV[off], &p.v[kv_base + kc * HEAD_DIM + d], valid);
+            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: kick off the first tile's load.
+    // ---- Prologue: issue first tile load ----
     load_tile(0, 0);
 
     for (int ti = 0; ti <= t_end; ti++) {
@@ -170,12 +158,15 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
             Sacc[n8][0] *= p.scale, Sacc[n8][1] *= p.scale,
             Sacc[n8][2] *= p.scale, Sacc[n8][3] *= p.scale;
 
-        int maxc0 = p.is_causal ? min(p.kv_len, qr0 + p.causal_offset + 1)
+        int maxc0 = (p.causal_offset >= 0) ? min(p.kv_len, qr0 + p.causal_offset + 1)
-                                : p.kv_len;
+                                        : p.kv_len;
-        int maxc1 = p.is_causal ? min(p.kv_len, qr1 + p.causal_offset + 1)
+        int maxc1 = (p.causal_offset >= 0) ? min(p.kv_len, qr1 + p.causal_offset + 1)
-                                : p.kv_len;
+                                        : p.kv_len;
         mma_softmax_tile<NC8, DN8>(kv0, maxc0, maxc1,
-                                    mb, p.mask, has_mask,
+                                    qr0, qr1,
+                                    p.mask_b_stride, p.mask_q_stride,
+                                    batch,
+                                    p.mask, has_mask,
                                     Sacc, Oacc, m0, m1, l0, l1, lane);
 
         mma_pv_accumulate<DN8, KT2>(Sacc, bV, LD, SWIZ_MASK, lane, Oacc);
@@ -186,19 +177,20 @@ __global__ void attn_prefill_split_q_mma_kernel(AttentionParams<bf16> p) {
     // halves store count and removes the uncoalesced scalar-store penalty)
     float rl0 = (l0 > 1e-20f) ? (1.0f / l0) : 0.0f;
     float rl1 = (l1 > 1e-20f) ? (1.0f / l1) : 0.0f;
-    const int o_base = ((batch * p.q_head + q_head) * p.q_len) * HEAD_DIM;
+    // O: stride-based write
+    const int o_base = batch * p.q_stride_b + q_head * p.q_stride_h;
 #pragma unroll
     for (int dn8 = 0; dn8 < DN8; dn8++) {
         int d = dn8 * 8 + 2 * tid4;
         if (qr0 < p.q_len) {
             __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][0] * rl0,
                                                      Oacc[dn8][1] * rl0);
-            *reinterpret_cast<__nv_bfloat162*>(&p.o[o_base + qr0 * HEAD_DIM + d]) = v;
+            *reinterpret_cast<__nv_bfloat162*>(&p.o[o_base + qr0 * p.q_stride_l + d * p.q_stride_d]) = v;
         }
         if (qr1 < p.q_len) {
             __nv_bfloat162 v = __floats2bfloat162_rn(Oacc[dn8][2] * rl1,
                                                      Oacc[dn8][3] * rl1);
-            *reinterpret_cast<__nv_bfloat162*>(&p.o[o_base + qr1 * HEAD_DIM + d]) = v;
+            *reinterpret_cast<__nv_bfloat162*>(&p.o[o_base + qr1 * p.q_stride_l + d * p.q_stride_d]) = v;
         }
     }
 }

csrc/tests/attn_decode_test.cu

@@ -98,8 +98,9 @@ static void bench() {
 
         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.is_causal = 0; p.causal_offset = 0;
+        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;
@@ -160,8 +161,9 @@ int main() {
 
         AttentionParams<bf16> p;
         p.batch=B; p.q_head=Hq; p.kv_head=Hk; p.q_len=1; p.kv_len=sl; p.head_dim=D;
-        p.use_mask=0; p.is_causal=0; p.causal_offset=0;
+        p.use_mask=0; p.causal_offset=-1;
         p.scale=1.0f/sqrtf((float)D);
+        set_default_strides(p);
         p.q=dQ; p.k=dK; p.v=dV; p.mask=nullptr; p.o=dO;
 
         // Split-K scratch (max 32 splits), sized for the production MMA path.
@@ -180,7 +182,7 @@ int main() {
         cudaMemcpy(hOut,dO,nQ*2,cudaMemcpyDeviceToHost);
 
         float* ref=new float[nQ];
-        cpu_attention_ref(hQ, hK, hV, hMask, ref, B, Hq, Hk, 1, sl, D, 0, 0);
+        cpu_attention_ref(hQ, hK, hV, hMask, ref, B, Hq, Hk, 1, sl, D, -1);
 
         float max_err=0;
         for (size_t i=0;i<nQ;i++){

csrc/tests/attn_paged_decode_test.cu

@@ -138,13 +138,14 @@ static int run_test(int B, int Hq, int Hkv, int kv_len, int page_size, int seed)
     }
 
     float* h_o_ref = (float*)calloc(B * Hq * HEAD_DIM, sizeof(float));
-    cpu_attention_ref(h_q_f, h_k_f, h_v_f, nullptr, h_o_ref, B, Hq, Hkv, 1, kv_len, HEAD_DIM, 0, 0);
+    cpu_attention_ref(h_q_f, h_k_f, h_v_f, nullptr, h_o_ref, B, Hq, Hkv, 1, kv_len, HEAD_DIM, -1);
 
     float scale_val = 1.0f / sqrtf((float)HEAD_DIM);
     PagedAttentionParams<bf16, float> p;
     p.batch = B; p.q_head = Hq; p.kv_head = Hkv; p.q_len = 1;
     p.kv_len = kv_len; p.head_dim = HEAD_DIM;
-    p.use_mask = 0; p.is_causal = 0; p.causal_offset = 0;
+    p.use_mask = 0; p.causal_offset = -1;
+    set_default_paged_strides(p);
     p.num_splits = 1; p.scale = scale_val;
     p.page_size = page_size; p.max_pages = max_pages;
     p.page_table = d_pt;
@@ -272,7 +273,8 @@ static void bench_config(int B, int Hq, int Hkv, int kv_len, int page_size) {
     PagedAttentionParams<bf16, float> pa;
     pa.batch = B; pa.q_head = Hq; pa.kv_head = Hkv; pa.q_len = 1;
     pa.kv_len = kv_len; pa.head_dim = HEAD_DIM;
-    pa.use_mask = 0; pa.is_causal = 0; pa.causal_offset = 0;
+    pa.use_mask = 0; pa.causal_offset = -1;
+    set_default_paged_strides(pa);
     pa.num_splits = 1; pa.scale = scale_val;
     pa.page_size = page_size; pa.max_pages = max_pages;
     pa.page_table = d_pt;

csrc/tests/attn_prefill_test.cu

@@ -75,7 +75,8 @@ static void bench() {
 
         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.is_causal=causal; p.causal_offset=0;
+        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;
 
@@ -143,7 +144,8 @@ int main() {
 
         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.is_causal=causal; p.causal_offset=0;
+        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;
 
@@ -158,7 +160,7 @@ int main() {
         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);
+        cpu_attention_ref(hQ, hK, hV, nullptr, ref, B, Hq, Hk, ql, kl, D, causal ? 0 : -1);
 
         float max_err=0;
         for (size_t i=0;i<nQ;i++) {

csrc/tests/test_utils.cuh

@@ -101,6 +101,32 @@ 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]
@@ -108,10 +134,11 @@ void dispatch_by_head_dim(int head_dim, Fn&& fn) {
 //   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 is_causal, int causal_offset
+    int causal_offset
 ) {
     float scale = 1.0f / sqrtf((float)D);
     int n_rep = Hq / Hk;
@@ -122,7 +149,7 @@ static void cpu_attention_ref(
                 float mv = -INFINITY, sv = 0.0f;
                 float accum[256] = {0.0f};
                 int lim = kv_len;
-                if (is_causal) {
+                if (causal_offset >= 0) {
                     int c = qi + causal_offset + 1;
                     lim = (c < kv_len) ? c : kv_len;
                 }

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)
@@ -232,8 +232,16 @@ def main():
         action="store_true",
         help="Skip SVD analysis, only show weight statistics (mean/std/min/max).",
     )
+    parser.add_argument(
+        "--output",
+        type=str,
+        default=None,
+        help="Save results as JSON to this path.",
+    )
     args = parser.parse_args()
 
+    all_results = {}
+
     def analyze_one(ckpt_dir: str, label: str):
         ckpt_dir = Path(ckpt_dir)
         weights_path = ckpt_dir / "model.safetensors"
@@ -294,13 +302,19 @@ def main():
             )
             print_layer_grid(results)
         print_weight_stats(results)
+        all_results[label] = results
         return results
 
     analyze_one(args.ckpt_dir, "Primary")
 
     if args.compare:
         for cdir in args.compare:
-            analyze_one(cdir, "Compare")
+            analyze_one(cdir, f"Compare_{cdir}")
+
+    if args.output:
+        with open(args.output, "w", encoding="utf-8") as f:
+            json.dump(all_results, f, indent=2)
+        print(f"\nResults saved to {args.output}")
 
 
 if __name__ == "__main__":

scripts/eval/evaluate_humaneval.py

@@ -8,7 +8,6 @@ Config is a single dataclass; side effects are isolated at pipeline boundaries.
 """
 
 import argparse
-import itertools
 import json
 import os
 import re
@@ -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})
@@ -206,6 +206,25 @@ 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")
+    contents = [item[k] for k in letters]
+    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 +233,24 @@ 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 +285,12 @@ 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 +322,14 @@ 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(l, 0.0), self._HIST_MAX) for l in losses]
+            idx = torch.tensor(clamped) / self._HIST_MAX * (self._HIST_BINS - 1)
+            self.hist += torch.bincount(
+                idx.long().clamp(0, self._HIST_BINS - 1),
+                minlength=self._HIST_BINS,
+            )
+        else:
+            self.losses.extend(losses)
+
+    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(f"\n  by token type:")
+        print(f"  {'type':<12} {'count':>8} {'mean_loss':>10} {'ppl':>8}")
+        print(f"  {'-' * 12} {'-' * 8} {'-' * 10} {'-' * 8}")
+        for ttype, s in by_type.items():
+            print(
+                f"  {ttype:<12} {s['num_tokens']:>8,} "
+                f"{s['mean_loss']:>10.4f} {s['ppl']:>8.2f}"
+            )
+
+
+def main(
+    param_path: str,
+    input_path: str,
+    output_dir: str,
+    text_key: str,
+    batch_size: int,
+    max_length: int,
+    token_level: bool,
+    max_samples: Optional[int],
+    device: str = "cuda",
+    dtype: str = "bfloat16",
 ):
-    # 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/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
 
-    # 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()
 
 
@@ -126,12 +175,36 @@ 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).",
     )
+    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/train.py

@@ -8,7 +8,7 @@ import torch.optim as optim
 from torch import Tensor, nn
 
 from astrai.config import AutoRegressiveLMConfig, TrainConfig
-from astrai.dataset import DatasetFactory
+from astrai.dataset import DatasetFactory, dpo_collate_fn, grpo_collate_fn
 from astrai.model import AutoRegressiveLM
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.trainer import SchedulerFactory, Trainer
@@ -90,6 +90,7 @@ class MuonMix(optim.Optimizer):
     def load_state_dict(self, state_dict: Dict[str, Any]):
         self.muon.load_state_dict(state_dict["muon"])
         self.adamw.load_state_dict(state_dict["adamw"])
+        self.param_groups = [*self.muon.param_groups, *self.adamw.param_groups]
 
 
 def parse_args() -> argparse.Namespace:
@@ -115,6 +116,13 @@ def parse_args() -> argparse.Namespace:
         required=True,
         help="Path to the model parameters or resume checkpoint.",
     )
+    parser.add_argument(
+        "--resume",
+        action="store_true",
+        default=False,
+        help="Resume training from checkpoint at --param_path "
+        "(restore epoch, consumed_samples, optimizer & scheduler state).",
+    )
 
     parser.add_argument(
         "--n_epoch", type=int, default=1, help="Number of epochs to train."
@@ -140,8 +148,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--max_grad_norm",
         type=float,
-        default=1.0,
+        default=None,
-        help="Max gradient norm for clipping.",
+        help="Max gradient norm for clipping. None disables clipping.",
     )
     parser.add_argument(
         "--weight_decay",
@@ -252,12 +260,6 @@ def parse_args() -> argparse.Namespace:
         default="checkpoint/logs",
         help="Directory for metric logs.",
     )
-    parser.add_argument(
-        "--grpo_sync_interval",
-        type=int,
-        default=200,
-        help="GRPO ref model sync interval (steps).",
-    )
     parser.add_argument(
         "--start_epoch", type=int, default=0, help="Start epoch for training."
     )
@@ -390,6 +392,7 @@ def train(
     train_type: str,
     param_path: str,
     data_root_path: str,
+    resume: bool,
     n_epoch: int,
     batch_per_device: int,
     start_epoch: int,
@@ -444,7 +447,6 @@ def train(
         "clip_eps": kwargs.pop("grpo_clip_eps"),
         "kl_coef": kwargs.pop("grpo_kl_coef"),
         "group_size": kwargs.pop("group_size"),
-        "sync_interval": kwargs.pop("grpo_sync_interval"),
     }
 
     executor_kwargs = {
@@ -458,6 +460,7 @@ def train(
         load_path=data_root_path,
         window_size=window_size,
         stride=stride,
+        tokenizer_path=param_path,
     )
 
     optimizer_fn = partial(
@@ -502,6 +505,12 @@ def train(
 
     grad_ckpt_modules = [DecoderBlock] if gradient_checkpointing else []
 
+    collate_fn = None
+    if train_type == "dpo":
+        collate_fn = dpo_collate_fn
+    elif train_type == "grpo":
+        collate_fn = grpo_collate_fn
+
     train_config = TrainConfig(
         model_fn=model_fn,
         strategy=train_type,
@@ -534,10 +543,11 @@ def train(
         executor_kwargs=executor_kwargs,
         extra_kwargs=strategy_kwargs,
         neftune_alpha=neftune_alpha,
+        collate_fn=collate_fn,
     )
 
     trainer = Trainer(train_config)
-    trainer.train(resume_dir=param_path)
+    trainer.train(param_path=param_path, resume=resume)
 
 
 if __name__ == "__main__":

tests/data/test_dataset.py

@@ -1,14 +1,16 @@
 import json
 import os
+import tempfile
 
 import numpy as np
 import pytest
 import torch
 
 from astrai.config.preprocess_config import PipelineConfig
-from astrai.dataset.dataset import DatasetFactory, SEQDataset
+from astrai.dataset.dataset import DatasetFactory, dpo_tokenize
 from astrai.dataset.storage import (
     H5Store,
+    JsonlStore,
     StoreFactory,
     detect_format,
 )
@@ -56,6 +58,13 @@ def _write_jsonl_dataset(test_dir, tokenizer_path, records, config_overrides=Non
     return data_dir
 
 
+def _fake_fetch_record(self, idx, keys):
+    """FakeStore.fetch_record matching real Store semantics."""
+    if isinstance(keys, str):
+        return self._data[keys][idx]
+    return {k: self._data[k][idx] for k in keys}
+
+
 def _make_seq_dataset(
     test_dir, name="data", seq_length=200, train_type="seq", data=None, **load_kwargs
 ):
@@ -109,7 +118,7 @@ def test_dpo_strategy_with_random_data(base_test_env):
     )
 
     assert dpo_dataset is not None
-    assert dpo_dataset.storage is not None
+    assert dpo_dataset.store is not None
     assert len(dpo_dataset) > 0
 
     # Test that we can get DPO items without errors
@@ -138,7 +147,7 @@ def test_sft_dataset_with_random_data(base_test_env):
     )
 
     assert sft_dataset is not None
-    assert sft_dataset.storage is not None
+    assert sft_dataset.store is not None
     assert len(sft_dataset) > 0
 
     # Test that we can get SFT items without errors
@@ -169,39 +178,37 @@ def test_dataset_with_custom_stride(base_test_env):
     assert len(dataset) > len(default_stride_dataset)
 
 
-def test_dataset_count_property(base_test_env):
+def test_dataset_token_count_property(base_test_env):
+    """dataset.token_count exposes the raw stream token length."""
     test_dir = base_test_env["test_dir"]
     dataset = _make_seq_dataset(test_dir, "count_test_data")
-    assert dataset.count == 200
+    assert dataset.token_count == 200
-    assert dataset.count > len(dataset)
+    assert dataset.token_count > len(dataset)
     assert len(dataset) == (200 - 1 - 64) // 64 + 1
 
 
-def test_empty_dataset_count():
-    """Test count returns 0 when no data is loaded"""
-    dataset = SEQDataset(window_size=64, stride=32)
-    assert dataset.count == 0
-    assert dataset.keys == []
-
-
 def test_dataset_too_short_for_window(base_test_env):
     test_dir = base_test_env["test_dir"]
     dataset = _make_seq_dataset(test_dir, "short", seq_length=30)
     assert len(dataset) == 0
-    assert dataset.count == 30
+    assert dataset.token_count == 30
 
 
-def test_unloaded_dataset_getitem_raises():
+def test_unloaded_sample_window_raises():
-    """__getitem__ without load() should fail clearly"""
+    """Store.sample_window before load raises RuntimeError."""
-    dataset = SEQDataset(window_size=64, stride=32)
+    from astrai.dataset.storage import H5Store
-    with pytest.raises(RuntimeError, match="not loaded"):
+
-        dataset.get_index(0)
+    store = H5Store(window_size=64, stride=64)
+    with pytest.raises(IndexError, match="Data too short"):
+        store.sample_window(0)
 
 
 def test_unloaded_dataset_len():
-    """__len__ without load() returns 0"""
+    """__len__ on a store with no data returns 0."""
-    dataset = SEQDataset(window_size=64, stride=32)
+    from astrai.dataset.storage import H5Store
-    assert len(dataset) == 0
+
+    store = H5Store(window_size=64, stride=64)
+    assert len(store) == 0
 
 
 def test_store_unloaded_len():
@@ -214,7 +221,7 @@ def test_store_unloaded_len():
 def test_store_fetch_begin_equals_end(base_test_env):
     test_dir = base_test_env["test_dir"]
     dataset = _make_seq_dataset(test_dir, "empty_fetch", seq_length=100, window_size=32)
-    result = dataset.storage.fetch(10, 10, "sequence")
+    result = dataset.store.fetch(10, 10, "sequence")
     assert result.numel() == 0
 
 
@@ -264,7 +271,7 @@ def test_store_multi_segment_concat(base_test_env):
 
     store = StoreFactory.create("h5")
     store.load(data_dir)
-    assert len(store) == 9
+    assert store.token_count == 9
     result = store.fetch(2, 7, "sequence")
     assert result.tolist() == [3, 4, 5, 6, 7]
 
@@ -293,7 +300,9 @@ def test_mmap_store_load_and_fetch(base_test_env):
 
     store = StoreFactory.create("bin")
     store.load(test_dir)
-    assert len(store) == 200
+    assert store.token_count == 200
+    assert store.num_records == 0
+    assert len(store) == 0  # no window configured, no records → 0 samples
     assert "sequence" in store.keys
 
     result = store.fetch(10, 20, "sequence")
@@ -306,64 +315,110 @@ def test_mmap_dataset_load(base_test_env):
     save_bin(test_dir, data)
     dataset = DatasetFactory.load("seq", test_dir, window_size=64)
     assert len(dataset) > 0
-    assert dataset.count == 200
+    assert dataset.token_count == 200
     assert dataset[0]["input_ids"].shape[0] == 64
 
 
 def test_normalize_empty_key():
-    """_normalize with empty tensor list does not crash"""
+    """_normalize with empty tensor list does not crash."""
     store = H5Store()
     store._normalize({"sequence": []})
     assert len(store) == 0
+    assert store.num_records == 0  # empty key forces num_records=0
     assert store.keys == ["sequence"]
 
 
 def test_normalize_mixed_empty_key():
-    """_normalize with empty + non-empty keys returns min=0"""
+    """_normalize with empty + non-empty keys returns min=0 records."""
     store = H5Store()
     store._normalize({"sequence": [torch.tensor([1, 2, 3])], "loss_mask": []})
     assert len(store) == 0
+    assert store.num_records == 0
+    assert store.token_count == 0  # min() over keys
     assert set(store.keys) == {"sequence", "loss_mask"}
 
 
 def test_grpo_dataset_dtype(base_test_env):
-    test_dir = base_test_env["test_dir"]
+    """GRPO dataset returns correct dtypes for per-record structured data."""
-    dummy_data = {
+    from astrai.dataset.dataset import GRPODataset
-        "prompts": [torch.randint(0, 100, (100,), dtype=torch.int32)],
+
-        "responses": [torch.randint(0, 100, (100,), dtype=torch.int32)],
+    G = 4
-        "masks": [torch.ones(100, dtype=torch.int32)],
+    store = type(
-        "rewards": [torch.ones(100, dtype=torch.float32)],
+        "FakeStore",
-    }
+        (),
-    dataset = _make_seq_dataset(
+        {
-        test_dir, "grpo_dtype", train_type="grpo", data=dummy_data, window_size=32
+            "keys": ["prompts", "responses", "masks", "rewards"],
-    )
+            "num_records": 1,
+            "token_count": 0,
+            "_data": {
+                "prompts": [torch.randint(0, 100, (10,), dtype=torch.int32)],
+                "responses": [
+                    [torch.randint(0, 100, (5,), dtype=torch.int32) for _ in range(G)]
+                ],
+                "masks": [[torch.ones(5, dtype=torch.int32) for _ in range(G)]],
+                "rewards": [torch.rand(G, dtype=torch.float32)],
+            },
+            "fetch_record": _fake_fetch_record,
+            "__len__": lambda self: self.num_records,
+        },
+    )()
+    dataset = GRPODataset(store=store)
     item = dataset[0]
 
     assert item["prompts"].dtype == torch.long
-    assert item["responses"].dtype == torch.long
+    assert all(r.dtype == torch.long for r in item["responses"])
-    assert item["masks"].dtype == torch.bool
+    assert all(m.dtype == torch.bool for m in item["masks"])
     assert item["rewards"].dtype == torch.float32
 
 
 def test_grpo_dataset_load(base_test_env):
-    test_dir = base_test_env["test_dir"]
+    """GRPO dataset loads record-structured data with per-response boundaries."""
-    dummy_data = {
+    from astrai.dataset.dataset import GRPODataset
-        "prompts": [_rand_seq(200)],
+
-        "responses": [_rand_seq(200)],
+    G = 3
-        "masks": [torch.ones(200, dtype=torch.int64)],
+    prompt_len = 8
-        "rewards": [torch.rand(200, dtype=torch.float32)],
+    resp_lens = [5, 7, 4]
-    }
+    store = type(
-    dataset = _make_seq_dataset(
+        "FakeStore",
-        test_dir, "grpo_test", train_type="grpo", data=dummy_data
+        (),
-    )
+        {
-    assert len(dataset) > 0
+            "keys": ["prompts", "responses", "masks", "rewards"],
+            "num_records": 1,
+            "token_count": 0,
+            "_data": {
+                "prompts": [torch.randint(0, 100, (prompt_len,))],
+                "responses": [[torch.randint(0, 100, (rl,)) for rl in resp_lens]],
+                "masks": [[torch.ones(rl, dtype=torch.int64) for rl in resp_lens]],
+                "rewards": [torch.tensor([0.9, 0.3, 0.7], dtype=torch.float32)],
+            },
+            "fetch_record": _fake_fetch_record,
+            "__len__": lambda self: self.num_records,
+        },
+    )()
+    dataset = GRPODataset(store=store)
+
+    assert len(dataset) == 1
     item = dataset[0]
     assert "prompts" in item
     assert "responses" in item
     assert "masks" in item
     assert "rewards" in item
-    assert item["prompts"].shape[0] == 64
+
-    assert item["responses"].shape[0] == 64
+    # Prompts is 1-D
+    assert item["prompts"].shape == (prompt_len,)
+
+    # Responses is a list of G tensors with correct lengths
+    assert len(item["responses"]) == G
+    for i, r in enumerate(item["responses"]):
+        assert r.shape == (resp_lens[i],)
+
+    # Masks align with responses
+    assert len(item["masks"]) == G
+    for i, m in enumerate(item["masks"]):
+        assert m.shape == (resp_lens[i],)
+
+    # Rewards has G elements
+    assert item["rewards"].shape == (G,)
 
 
 def test_detect_format_bin_dir(base_test_env):
@@ -408,7 +463,34 @@ def test_dataset_load_explicit_storage_type(base_test_env):
     test_dir = base_test_env["test_dir"]
     dataset = _make_seq_dataset(test_dir, "explicit", storage_type="h5")
     assert len(dataset) > 0
-    assert dataset.count == 200
+    assert dataset.token_count == 200
+
+
+def _write_json_dataset(test_dir, tokenizer_path, records, config_overrides=None):
+    """Write JSONL dataset — one JSON object per line."""
+    data_dir = os.path.join(test_dir, "json_data")
+    os.makedirs(data_dir, exist_ok=True)
+
+    with open(os.path.join(data_dir, "data.jsonl"), "w", encoding="utf-8") as f:
+        for rec in records:
+            f.write(json.dumps(rec, ensure_ascii=False) + "\n")
+
+    config = {
+        "tokenizer_path": tokenizer_path,
+        "version": 1,
+        "input": {"sections": [{"field": "text", "action": "train"}]},
+        "preprocessing": {"max_seq_len": 128, "min_chars": 0},
+        "output": {"position_ids_mode": "continuous"},
+    }
+    if config_overrides:
+        config.update(config_overrides)
+
+    with open(
+        os.path.join(data_dir, "dataset_config.json"), "w", encoding="utf-8"
+    ) as f:
+        json.dump(config, f, ensure_ascii=False, indent=2)
+
+    return data_dir
 
 
 def test_detect_format_jsonl_dir(base_test_env):
@@ -422,6 +504,90 @@ def test_detect_format_jsonl_dir(base_test_env):
     assert detect_format(data_dir) == "jsonl"
 
 
+def test_detect_format_json_dir(base_test_env):
+    """detect_format returns 'jsonl' for directory with .json files."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
+    data_dir = _write_json_dataset(
+        test_dir,
+        tokenizer_path,
+        [{"text": "hello world"}, {"text": "foo bar baz qux"}],
+    )
+    assert detect_format(data_dir) == "jsonl"
+
+
+def test_json_store_seq(base_test_env):
+    """JsonlStore loads .json array correctly."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
+    data_dir = _write_json_dataset(
+        test_dir,
+        tokenizer_path,
+        [{"text": "hello world"}, {"text": "foo bar baz qux"}],
+    )
+
+    store = StoreFactory.create("jsonl")
+    store.load(data_dir)
+    assert len(store) > 0
+    assert "sequence" in store.keys
+
+    dataset = DatasetFactory.load("seq", data_dir, window_size=8)
+    assert len(dataset) > 0
+    item = dataset[0]
+    assert "input_ids" in item
+    assert "target_ids" in item
+
+
+def test_json_store_no_tokenizer_path(base_test_env):
+    """JsonlStore uses dataset dir as tokenizer_path when omitted."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer = base_test_env["tokenizer"]
+    tokenizer.set_chat_template(
+        "{% for message in messages %}{{ message['role'] }}:{{ message['content'] }}\n{% endfor %}"
+    )
+
+    data_dir = os.path.join(test_dir, "self_contained")
+    os.makedirs(data_dir, exist_ok=True)
+
+    # Save tokenizer files directly in the dataset directory
+    tokenizer.save_pretrained(data_dir)
+
+    # Write .jsonl data
+    records = [
+        {
+            "messages": [
+                {"role": "user", "content": "hi"},
+                {"role": "assistant", "content": "hello"},
+            ]
+        }
+    ]
+    with open(os.path.join(data_dir, "data.jsonl"), "w", encoding="utf-8") as f:
+        for rec in records:
+            f.write(json.dumps(rec, ensure_ascii=False) + "\n")
+
+    # dataset_config.json WITHOUT tokenizer_path
+    config = {
+        "version": 1,
+        "input": {
+            "sections": [{"field": "messages", "action": "$role", "template": True}]
+        },
+        "mask": {"user": "mask", "assistant": "train"},
+        "mask_default": "mask",
+        "preprocessing": {"max_seq_len": 128, "min_chars": 0},
+        "output": {"position_ids_mode": "continuous"},
+    }
+    with open(
+        os.path.join(data_dir, "dataset_config.json"), "w", encoding="utf-8"
+    ) as f:
+        json.dump(config, f, ensure_ascii=False, indent=2)
+
+    store = StoreFactory.create("jsonl")
+    store.load(data_dir)
+    assert len(store) > 0
+    assert "sequence" in store.keys
+    assert "loss_mask" in store.keys
+
+
 def test_jsonl_store_seq(base_test_env):
     test_dir = base_test_env["test_dir"]
     tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
@@ -512,3 +678,425 @@ def test_jsonl_store_pipeline_config_roundtrip(base_test_env):
     config = PipelineConfig.from_dict(raw)
     assert config.output.position_ids_mode == "doc_reset"
     assert config.preprocessing.max_seq_len == 64
+
+
+# ---------------------------------------------------------------------------
+# GRPO end-to-end: builder → JsonlStore → GRPODataset → collate_fn
+# ---------------------------------------------------------------------------
+
+
+def _write_grpo_jsonl(test_dir, tokenizer_path, records):
+    """Write a GRPO JSONL dataset directory with config."""
+    data_dir = os.path.join(test_dir, "grpo_jsonl")
+    os.makedirs(data_dir, exist_ok=True)
+
+    with open(os.path.join(data_dir, "data.jsonl"), "w", encoding="utf-8") as f:
+        for rec in records:
+            f.write(json.dumps(rec, ensure_ascii=False) + "\n")
+
+    config = {
+        "tokenizer_path": tokenizer_path,
+        "version": 1,
+        "input": {
+            "sources": {
+                "prompts": {
+                    "sections": [
+                        {
+                            "field": "prompt",
+                            "action": "mask",
+                            "add_special_tokens": True,
+                        }
+                    ]
+                },
+                "responses": {
+                    "sections": [{"field": "responses", "action": "train"}],
+                    "list_field": True,
+                    "mask_key": "masks",
+                },
+                "rewards": {
+                    "sections": [{"field": "rewards", "action": "value"}],
+                },
+            }
+        },
+        "mask": {"user": "mask", "assistant": "train"},
+        "mask_default": "mask",
+        "preprocessing": {"max_seq_len": 128},
+        "output": {"position_ids_mode": "none"},
+    }
+
+    with open(
+        os.path.join(data_dir, "dataset_config.json"), "w", encoding="utf-8"
+    ) as f:
+        json.dump(config, f, ensure_ascii=False, indent=2)
+
+    return data_dir
+
+
+def test_grpo_builder_preserves_response_boundaries(base_test_env):
+    """MultiOutputMaskBuilder with list_field returns List[List[int]] for responses."""
+    from astrai.preprocessing.builder import SectionedMaskBuilder
+    from tests.data.conftest import make_grpo_no_template_config
+
+    tokenizer = base_test_env["tokenizer"]
+    tokenizer_path = _save_test_tokenizer(base_test_env["test_dir"], tokenizer)
+
+    builder = SectionedMaskBuilder()
+    config = make_grpo_no_template_config()
+    config.preprocessing.max_seq_len = 128
+
+    item = {
+        "prompt": "What is 2+2?",
+        "responses": ["4", "four", "2+2=4"],
+        "rewards": [0.9, 0.1, 0.5],
+    }
+
+    result = builder.build(item, config, tokenizer)
+    assert result is not None
+
+    # prompts should be flat list of ints
+    assert isinstance(result["prompts"], list)
+    assert isinstance(result["prompts"][0], int)
+
+    # responses should be list of lists (one per response)
+    assert isinstance(result["responses"], list)
+    assert isinstance(result["responses"][0], list)
+    assert isinstance(result["responses"][0][0], int)
+    assert len(result["responses"]) == 3
+
+    # masks should match responses structure
+    assert isinstance(result["masks"], list)
+    assert len(result["masks"]) == 3
+    for i in range(3):
+        assert len(result["masks"][i]) == len(result["responses"][i])
+
+    # rewards should be flat list of floats
+    assert isinstance(result["rewards"], list)
+    assert all(isinstance(r, float) for r in result["rewards"])
+    assert len(result["rewards"]) == 3
+
+
+def test_grpo_end_to_end_jsonl(base_test_env):
+    """Full GRPO pipeline: JSONL → JsonlStore → GRPODataset → collate_fn."""
+    from astrai.dataset.dataset import grpo_collate_fn
+
+    test_dir = base_test_env["test_dir"]
+    tokenizer = base_test_env["tokenizer"]
+    tokenizer_path = _save_test_tokenizer(test_dir, tokenizer)
+
+    records = [
+        {
+            "prompt": "What is 2+2?",
+            "responses": ["4", "four", "The answer is 4"],
+            "rewards": [0.9, 0.1, 0.5],
+        },
+        {
+            "prompt": "Write a haiku",
+            "responses": ["Leaves fall", "Cherry blossoms bloom in spring"],
+            "rewards": [0.3, 0.8],
+        },
+    ]
+
+    data_dir = _write_grpo_jsonl(test_dir, tokenizer_path, records)
+
+    dataset = DatasetFactory.load("grpo", data_dir, window_size=0)
+    assert len(dataset) == 2
+
+    # Item 0: 3 responses
+    item0 = dataset[0]
+    assert item0["prompts"].ndim == 1
+    assert len(item0["responses"]) == 3
+    assert len(item0["masks"]) == 3
+    assert item0["rewards"].shape == (3,)
+    for r, m in zip(item0["responses"], item0["masks"]):
+        assert r.shape == m.shape
+
+    # Item 1: 2 responses (different group size)
+    item1 = dataset[1]
+    assert len(item1["responses"]) == 2
+    assert item1["rewards"].shape == (2,)
+
+    # Collate: batch records with same G (item0 has G=3)
+    batch = grpo_collate_fn([item0, item0])
+    assert batch["prompts"].shape[0] == 2
+    assert batch["responses"].ndim == 3
+    assert batch["responses"].shape[0] == 2
+    assert batch["responses"].shape[1] == 3  # G=3
+    assert batch["masks"].shape == batch["responses"].shape
+    assert batch["rewards"].shape == (2, 3)
+
+
+def test_grpo_collate_variable_lengths():
+    """collate_fn pads variable-length responses to [B, G, R_max]."""
+    from astrai.dataset.dataset import grpo_collate_fn
+
+    batch = [
+        {
+            "prompts": torch.tensor([1, 2, 3]),
+            "responses": [torch.tensor([4, 5]), torch.tensor([6, 7, 8, 9])],
+            "masks": [torch.tensor([1, 1]), torch.tensor([1, 1, 1, 1])],
+            "rewards": torch.tensor([0.9, 0.1]),
+        },
+        {
+            "prompts": torch.tensor([10, 11]),
+            "responses": [torch.tensor([12]), torch.tensor([13, 14, 15])],
+            "masks": [torch.tensor([1]), torch.tensor([1, 1, 1])],
+            "rewards": torch.tensor([0.5, 0.5]),
+        },
+    ]
+
+    result = grpo_collate_fn(batch)
+
+    assert result["prompts"].shape == (2, 3)  # B=2, P_max=3
+    assert result["responses"].shape == (2, 2, 4)  # B=2, G=2, R_max=4
+    assert result["masks"].shape == (2, 2, 4)
+    assert result["rewards"].shape == (2, 2)
+
+    # Check padding: item 1 prompt is length 2, padded to 3
+    assert result["prompts"][1, 2] == 0
+
+    # Check response content: item 0, response 0 is [4,5] padded to 4
+    assert result["responses"][0, 0, 0] == 4
+    assert result["responses"][0, 0, 1] == 5
+    assert result["responses"][0, 0, 2] == 0  # padded
+    assert not result["masks"][0, 0, 2]  # padded
+
+    # Check response content: item 0, response 1 is [6,7,8,9] no padding
+    assert result["responses"][0, 1, 3] == 9
+    assert result["masks"][0, 1, 3]
+
+
+def test_grpo_multiple_records(base_test_env):
+    """GRPODataset loads multiple records with correct structure."""
+    from astrai.dataset.dataset import GRPODataset
+
+    G = 4
+    n_records = 5
+
+    dummy_responses = [
+        [torch.randint(0, 100, (np.random.randint(3, 8),)) for _ in range(G)]
+        for _ in range(n_records)
+    ]
+    store = type(
+        "FakeStore",
+        (),
+        {
+            "keys": ["prompts", "responses", "masks", "rewards"],
+            "num_records": n_records,
+            "token_count": 0,
+            "_data": {
+                "prompts": [torch.randint(0, 100, (10,)) for _ in range(n_records)],
+                "responses": dummy_responses,
+                "masks": [
+                    [torch.ones(r.shape[0], dtype=torch.int64) for r in resps]
+                    for resps in dummy_responses
+                ],
+                "rewards": [
+                    torch.rand(G, dtype=torch.float32) for _ in range(n_records)
+                ],
+            },
+            "fetch_record": _fake_fetch_record,
+            "__len__": lambda self: self.num_records,
+        },
+    )()
+    dataset = GRPODataset(store=store)
+
+    assert len(dataset) == n_records
+
+    for i in range(n_records):
+        item = dataset[i]
+        assert len(item["responses"]) == G
+        assert len(item["masks"]) == G
+        assert item["rewards"].shape == (G,)
+        for g in range(G):
+            assert item["responses"][g].shape == item["masks"][g].shape
+
+
+def _write_dpo_jsonl(test_dir, records):
+    """Write a raw DPO JSONL file (no dataset_config.json)."""
+    path = os.path.join(test_dir, "dpo.jsonl")
+    with open(path, "w", encoding="utf-8") as f:
+        for rec in records:
+            f.write(json.dumps(rec, ensure_ascii=False) + "\n")
+    return path
+
+
+def test_dpo_tokenize_pure_function():
+    """dpo_tokenize returns flat lists with correct mask alignment."""
+
+    class FakeTokenizer:
+        def apply_chat_template(
+            self, messages, tokenize=True, add_generation_prompt=True
+        ):
+            ids = []
+            for m in messages:
+                ids.append(len(m["content"]))
+                ids.append(-1)
+            if add_generation_prompt:
+                ids.append(99)
+            return ids
+
+    record = {"prompt": "ab", "chosen": "xyz", "rejected": "w"}
+    result = dpo_tokenize(record, FakeTokenizer(), max_len=64)
+
+    assert set(result.keys()) == {"chosen", "rejected", "chosen_mask", "rejected_mask"}
+    assert len(result["chosen"]) == len(result["chosen_mask"])
+    assert len(result["rejected"]) == len(result["rejected_mask"])
+
+    assert result["chosen_mask"][0] == 0
+    assert any(m == 1 for m in result["chosen_mask"])
+    assert result["rejected_mask"][0] == 0
+
+
+def test_dpo_tokenize_malformed_record():
+    """dpo_tokenize returns None for missing fields."""
+
+    class FakeTokenizer:
+        def apply_chat_template(
+            self, messages, tokenize=True, add_generation_prompt=True
+        ):
+            return [1]
+
+    assert dpo_tokenize({}, FakeTokenizer()) is None
+    assert dpo_tokenize({"prompt": "a"}, FakeTokenizer()) is None
+    assert dpo_tokenize({"prompt": "a", "chosen": "b"}, FakeTokenizer()) is None
+
+
+def test_dpo_jsonl_lazy_load(base_test_env):
+    """DPODataset loads raw JSONL with tokenizer_path → lazy processor."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
+
+    records = [
+        {"input": "Hello", "chosen": "world", "rejected": "earth"},
+        {"input": "Foo", "chosen": "bar", "rejected": "baz"},
+    ]
+    path = _write_dpo_jsonl(test_dir, records)
+
+    ds = DatasetFactory.load(
+        train_type="dpo",
+        load_path=path,
+        window_size=0,
+        tokenizer_path=tokenizer_path,
+    )
+
+    assert len(ds) == 2
+    assert ds.store.num_records == 2
+    assert ds.store._processor is not None
+
+    item = ds[0]
+    assert set(item.keys()) == {"chosen", "rejected", "chosen_mask", "rejected_mask"}
+    assert item["chosen"].dtype == torch.long
+    assert item["chosen_mask"].dtype == torch.bool
+    assert item["chosen"].shape == item["chosen_mask"].shape
+    assert item["chosen"].shape == item["rejected"].shape
+
+
+def test_dpo_jsonl_lazy_no_tokenizer():
+    """DPODataset on jsonl without tokenizer_path falls back to eager
+    (which requires dataset_config.json, so it should raise)."""
+    with tempfile.TemporaryDirectory() as d:
+        path = os.path.join(d, "dpo.jsonl")
+        with open(path, "w") as f:
+            f.write(json.dumps({"input": "a", "chosen": "b", "rejected": "c"}) + "\n")
+
+        with pytest.raises(FileNotFoundError, match="dataset_config.json"):
+            DatasetFactory.load(
+                train_type="dpo",
+                load_path=path,
+                window_size=0,
+            )
+
+
+def test_jsonl_store_lazy_len_returns_record_count(base_test_env):
+    """JsonlStore in lazy mode: len() returns record count, not tokens."""
+    test_dir = base_test_env["test_dir"]
+    records = [{"input": str(i), "chosen": "c", "rejected": "r"} for i in range(5)]
+    path = _write_dpo_jsonl(test_dir, records)
+
+    store = JsonlStore()
+    store.load(path, processor=lambda r: {"chosen": torch.tensor([1, 2])})
+
+    assert len(store) == 5
+    assert store.num_records == 5
+
+
+def test_jsonl_store_eager_len_returns_token_count(base_test_env):
+    """JsonlStore in eager mode: num_records reflects per-record count."""
+    test_dir = base_test_env["test_dir"]
+    tokenizer_path = _save_test_tokenizer(test_dir, base_test_env["tokenizer"])
+    data_dir = _write_jsonl_dataset(
+        test_dir,
+        tokenizer_path,
+        [{"text": "hello world"}, {"text": "foo bar"}],
+        config_overrides={
+            "preprocessing": {"max_seq_len": 128, "min_chars": 0},
+            "output": {"position_ids_mode": "none"},
+        },
+    )
+
+    store = JsonlStore()
+    store.load(data_dir)
+
+    assert store.num_records == 2
+    assert len(store.keys) > 0
+
+
+def test_h5_store_dual_mode(base_test_env):
+    """H5Store supports both fetch (stream) and fetch_record (record).
+
+    No window configured → ``len(store)`` reflects the record count
+    (2).  ``token_count`` retains the legacy stream length (128), and
+    token-stream access via :meth:`fetch` is still available for
+    callers that want explicit begin/end control.
+    """
+    test_dir = base_test_env["test_dir"]
+
+    seq_length = 64
+    dummy_data = {
+        "chosen": [_rand_seq(seq_length), _rand_seq(seq_length)],
+        "rejected": [_rand_seq(seq_length), _rand_seq(seq_length)],
+    }
+    save_h5(test_dir, "dpo_data", dummy_data)
+
+    store = H5Store()
+    store.load(test_dir)
+
+    assert store.token_count == seq_length * 2
+    assert store.num_records == 2
+    assert len(store) == 2  # no window configured → record count
+
+    rec0 = store.fetch_record(0, "chosen")
+    assert rec0.shape == (seq_length,)
+
+    stream = store.fetch(0, 10, "chosen")
+    assert stream.shape == (10,)
+
+    # Window-configured view of the same data uses stream sample count:
+    #   token_count=128, window_size=64 → num_samples = (128-1-64)//64 + 1 = 1
+    stream_view = H5Store(window_size=seq_length, stride=seq_length)
+    stream_view.load(test_dir)
+    assert len(stream_view) == 1
+
+
+def test_mmap_store_stream_only_no_offsets(base_test_env):
+    """MmapStore without offsets: num_records == 0, stream works.
+
+    No window configured → ``len(store)`` is 0 (no iterate units).
+    ``token_count`` remains 128 for raw token slicing, and ``fetch``
+    provides direct token-range access.
+    """
+    test_dir = base_test_env["test_dir"]
+
+    seq_length = 128
+    dummy_data = {"sequence": [_rand_seq(seq_length)]}
+    save_bin(test_dir, dummy_data)
+
+    store = StoreFactory.create("bin")
+    store.load(test_dir)
+
+    assert store.token_count == seq_length
+    assert store.num_records == 0
+    assert len(store) == 0
+
+    chunk = store.fetch(0, 32, "sequence")
+    assert chunk.shape == (32,)

tests/data/test_preprocess_builder.py

@@ -349,7 +349,17 @@ def test_grpo_basic(chat_tokenizer, builder):
     assert "responses" in result
     assert "masks" in result
     assert "rewards" in result
-    assert len(result["responses"]) == len(result["masks"])
+
+    # responses is List[List[int]] — one per response
+    assert len(result["responses"]) == 4
+    assert all(isinstance(r, list) for r in result["responses"])
+    assert all(isinstance(r[0], int) for r in result["responses"])
+
+    # masks is List[List[int]] — one per response, matching length
+    assert len(result["masks"]) == 4
+    for i in range(4):
+        assert len(result["masks"][i]) == len(result["responses"][i])
+
     assert result["rewards"] == [1.0, 0.5, 0.8, 0.2]
 
 
@@ -362,8 +372,11 @@ def test_grpo_response_tokens_all_trained(chat_tokenizer, builder):
     }
     result = builder.build(item, config, chat_tokenizer)
     masks = result["masks"]
-    assert all(m == 1 for m in masks)
+    # masks is List[List[int]] — each response's mask should be all 1s
-    assert len(masks) == len(result["responses"])
+    assert len(masks) == 2
+    for m in masks:
+        assert all(v == 1 for v in m)
+        assert len(m) == len(result["responses"][masks.index(m)])
 
 
 def test_grpo_single_reward(chat_tokenizer, builder):

tests/data/test_preprocess_pipeline.py

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

tests/data/test_sampler.py

@@ -1,4 +1,4 @@
-from astrai.dataset import ResumableDistributedSampler
+from astrai.dataset import RDSampler
 
 
 def test_random_sampler_consistency(random_dataset):
@@ -6,8 +6,8 @@ def test_random_sampler_consistency(random_dataset):
     dataset = random_dataset
 
     # Create two samplers with same seed
-    sampler1 = ResumableDistributedSampler(dataset, seed=42)
+    sampler1 = RDSampler(dataset, seed=42)
-    sampler2 = ResumableDistributedSampler(dataset, seed=42)
+    sampler2 = RDSampler(dataset, seed=42)
 
     indices1 = list(iter(sampler1))
     indices2 = list(iter(sampler2))
@@ -20,8 +20,8 @@ def test_random_sampler_different_seeds(random_dataset):
     dataset = random_dataset
 
     # Create two samplers with different seeds
-    sampler1 = ResumableDistributedSampler(dataset, seed=42)
+    sampler1 = RDSampler(dataset, seed=42)
-    sampler2 = ResumableDistributedSampler(dataset, seed=123)
+    sampler2 = RDSampler(dataset, seed=123)
 
     indices1 = list(iter(sampler1))
     indices2 = list(iter(sampler2))
@@ -35,7 +35,7 @@ def test_sampler_across_epochs(random_dataset):
     dataset = random_dataset
     n = len(dataset)
 
-    sampler = ResumableDistributedSampler(dataset, seed=42)
+    sampler = RDSampler(dataset, seed=42)
 
     # Get indices for first epoch
     epoch1_indices = list(iter(sampler))

tests/inference/test_sample.py

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

tests/trainer/test_early_stopping.py

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

tests/trainer/test_grpo_strategy.py

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