refactor : 合并 data config docstring 并实现 BFD 打包策略

- 将 ProcessingConfig/OutputConfig 参数描述合并到类级 docstring

- Pipeline 支持 packing_strategy/truncation_mode，新增 bfd 打包

README.md

@@ -82,6 +82,7 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 nohup python scripts/tools/train.py \
     --nprocs=4 \
+    --parallel_mode=ddp \
     --train_type=seq \
     --data_root_path=/path/to/dataset \
     --param_path=/path/to/model \
@@ -108,8 +109,8 @@ Full reference at [Parameter Guide](assets/docs/params.md).
 ```bash
 python scripts/tools/generate.py \
     --param_path /path/to/model \
-    --input_json_file /path/to/input.json \
-    --output_json_file /path/to/output.json
+    --input_json_file /path/to/input.jsonl \
+    --output_json_file /path/to/output.jsonl
 ```
 
 #### Docker
@@ -224,6 +225,7 @@ Watch a video walkthrough on [bilibili](https://www.bilibili.com/video/BV1fuLB6y
 | [Training](./assets/docs/training.md) | Training loop, strategies & formulas |
 | [Inference](./assets/docs/inference.md) | KVCache, continuous batching, sampling & HTTP API |
 | [Data Flow](./assets/docs/dataflow.md) | Data pipeline, storage backends & dataset architecture |
+| [Preprocessing](./assets/docs/preprocessing.md) | Declarative JSON-driven data preprocessing |
 
 ### Contributing
 

assets/docs/README-zh-CN.md

@@ -88,6 +88,7 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 nohup python scripts/tools/train.py \
     --nprocs=4 \
+    --parallel_mode=ddp \
     --train_type=seq \
     --data_root_path=/path/to/dataset \
     --param_path=/path/to/model \
@@ -114,8 +115,8 @@ nohup python scripts/tools/train.py \
 ```bash
 python scripts/tools/generate.py \
     --param_path /path/to/model \
-    --input_json_file /path/to/input.json \
-    --output_json_file /path/to/output.json
+    --input_json_file /path/to/input.jsonl \
+    --output_json_file /path/to/output.jsonl
 ```
 
 #### Docker
@@ -230,6 +231,7 @@ python scripts/demo/generate_ar.py
 | [训练文档](./training.md) | 训练循环、策略与公式 |
 | [推理文档](./inference.md) | KVCache、连续批处理、采样与 HTTP API |
 | [数据流程](./dataflow.md) | 数据管道、存储后端与数据集架构 |
+| [数据预处理](./preprocessing.md) | 声明式 JSON 驱动数据预处理 |
 
 ### 贡献
 

assets/docs/architecture.md

@@ -8,6 +8,8 @@ classDiagram
         class BaseConfig {
             +to_dict() Dict
             +from_dict(d) Self
+            +from_json(path) Self
+            +to_json(path)
         }
 
         class BaseModelConfig {
@@ -17,42 +19,42 @@ classDiagram
         }
 
         class AutoRegressiveLMConfig {
-            +int vocab_size
-            +int dim
-            +int n_layers
-            +float norm_eps
-            +int dim_ffn
+            +Optional[int] vocab_size
+            +Optional[int] dim
+            +Optional[int] n_layers
+            +Optional[float] norm_eps
+            +Optional[int] dim_ffn
             +Optional[bool] tie_weight
             +Optional[dict] rope_scaling
-            +int max_len
-            +float rope_theta
+            +Optional[int] max_len
+            +Optional[float] rope_theta
             +str attn_type
-            +int n_heads
-            +int n_kv_heads
-            +bool use_qk_norm
-            +bool use_gated_attention
+            +Optional[int] n_heads
+            +Optional[int] n_kv_heads
+            +Optional[bool] use_qk_norm
+            +Optional[bool] use_gated_attention
             +Optional[int] kv_lora_rank
             +Optional[int] qk_nope_head_dim
             +Optional[int] qk_rope_head_dim
             +str ffn_type
-            +int n_routed_experts
-            +int n_shared_experts
-            +int n_activated_experts
+            +Optional[int] n_routed_experts
+            +Optional[int] n_shared_experts
+            +Optional[int] n_activated_experts
             +Optional[str] topk_method
         }
 
         class EncoderConfig {
-            +int vocab_size
-            +int dim
-            +int n_layers
-            +float norm_eps
-            +int dim_ffn
-            +int max_len
-            +float rope_theta
-            +int n_heads
-            +int n_kv_heads
-            +bool use_qk_norm
-            +bool use_gated_attention
+            +Optional[int] vocab_size
+            +Optional[int] dim
+            +Optional[int] n_layers
+            +Optional[float] norm_eps
+            +Optional[int] dim_ffn
+            +Optional[int] max_len
+            +Optional[float] rope_theta
+            +Optional[int] n_heads
+            +Optional[int] n_kv_heads
+            +Optional[bool] use_qk_norm
+            +Optional[bool] use_gated_attention
             +Optional[dict] rope_scaling
             +Optional[str] pooling_type
             +Optional[bool] normalize_embeddings
@@ -64,6 +66,38 @@ classDiagram
             +load(raw) BaseConfig
         }
 
+        class InputConfig {
+            +str type
+            +str messages_key
+            +str prompt_key
+            +str response_key
+            +str text_key
+        }
+
+        class ProcessingConfig {
+            +int max_seq_len
+            +int min_chars
+            +int max_chars
+            +bool deduplicate
+            +Optional[int] max_items
+        }
+
+        class OutputConfig {
+            +Optional[str] domain_key
+            +str storage_format
+            +int max_tokens_per_shard
+        }
+
+        class PipelineConfig {
+            +int version
+            +InputConfig input
+            +dict mask
+            +str mask_default
+            +ProcessingConfig preprocessing
+            +OutputConfig output
+            +from_dict(d) Self
+        }
+
         class TrainConfig {
             +Callable[[], nn.Module] model_fn
             +str strategy
@@ -312,10 +346,39 @@ classDiagram
         }
     }
 
+    namespace preprocessing {
+        class BaseMaskBuilder {
+            <<abstract>>
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class ChatMaskBuilder {
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class InstructionMaskBuilder {
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class TextMaskBuilder {
+            +build(item, config, tokenizer) Optional[dict]
+        }
+
+        class Pipeline {
+            +PipelineConfig config
+            +List[str] paths
+            +str output_dir
+            +str tokenizer_path
+            +BaseMaskBuilder mask_builder
+            +transform(item) Optional[dict]
+            +run()
+        }
+    }
+
     namespace tokenize {
         class AutoTokenizer {
             +vocab_size int
-            +encode(tokens, out_ids, is_pretokenized, add_special_tokens) List[int]
+            +encode(tokens, out_ids, is_pretokenized, add_special_tokens) List
             +decode(tokens, skip_special_tokens) str
             +__getattr__(name) Any (bos_id, eos_id, pad_id, stop_ids)
             +apply_chat_template(messages, system_prompt, tokenize, add_generation_prompt) Union[str, List[int]]
@@ -346,14 +409,20 @@ classDiagram
             +create(name, *args, **kwargs) T
             +list_registered() list
         }
+
+        class MaskBuilderFactory {
+            +Registry _registry
+            +register(name) decorator
+            +create(input_type, config, tokenizer) BaseMaskBuilder
+        }
     }
 
     namespace trainer {
         class Trainer {
             +TrainConfig train_config
             +List[TrainCallback] callbacks
-            +train(checkpoint)
-            +_get_default_callbacks() List[TrainCallback]
+            +train(resume_dir)
+            -_get_default_callbacks() List[TrainCallback]
         }
 
         class TrainContext {
@@ -383,8 +452,12 @@ classDiagram
         }
 
         class BaseStrategy {
-            +Union[Callable, nn.Module] model
+            +Callable model
+            +Optional[BaseExecutor] executor
+            +Optional[Callable] model_fn
+            +dict extra_kwargs
             +str device
+            +__call__(batch) Tensor
             +compute_loss(batch) Tensor
         }
 
@@ -425,6 +498,8 @@ classDiagram
         class BaseScheduler {
             +get_lr() List[float]
             +step()
+            +state_dict() dict
+            +load_state_dict(d)
         }
 
         class SchedulerFactory {
@@ -436,6 +511,7 @@ classDiagram
         class CosineScheduler {
             +int warmup_steps
             +int lr_decay_steps
+            +int total_steps
             +float min_rate
         }
 
@@ -474,11 +550,11 @@ classDiagram
             +int interval
             +bool weight_only
             +Callable save_extra_fn
-            +_save_checkpoint(context)
+            -_save_checkpoint(context)
             +on_batch_end(context)
             +on_train_end(context)
             +on_error(context)
-            +save_extra(context)$
+            +save_extra(context) dict$
         }
 
         class ProgressBarCallback {
@@ -491,7 +567,7 @@ classDiagram
         }
 
         class MetricLoggerCallback {
-            +str log_dir
+            +Path log_dir
             +int save_interval
             +int log_interval
             +List[str] metrics
@@ -501,7 +577,7 @@ classDiagram
         }
 
         class ValidationCallback {
-            +_run_validation(context)
+            -_run_validation(context)
             +on_optimizer_step(context)
         }
 
@@ -517,7 +593,7 @@ classDiagram
             +float weight_decay
             +bool nesterov
             +int ns_steps
-            +float adamw_lr
+            +Optional[float] adamw_lr
             +tuple adamw_betas
             +float adamw_eps
             +float adamw_wd
@@ -634,7 +710,7 @@ classDiagram
         class Task {
             +str task_id
             +List prompt_ids
-            +int max_tokens
+            +Optional[int] max_tokens
             +float temperature
             +float top_p
             +int top_k
@@ -643,8 +719,8 @@ classDiagram
             +int input_tokens
             +int output_tokens
             +float arrival_time
-            +float finish_time
-            +Callable stream_callback
+            +Optional[float] finish_time
+            +Optional[Callable] stream_callback
             +int next_pos
             +is_finished(stop_ids) bool
         }
@@ -671,6 +747,11 @@ classDiagram
             +activate(task)
             +return_to_waiting(tasks)
             +get_active_tasks() List[Task]
+            +has_work() bool
+            +wait_for_tasks(timeout)
+            +get_waiting_tasks() List[Task]
+            +clear_queues()
+            +wake()
             +get_stats() Dict
         }
 
@@ -760,7 +841,7 @@ classDiagram
 
         class ResponseBuilder {
             <<abstract>>
-            +prepare(request, engine) Tuple[str, GenContext, List[str]]
+            +prepare(request, tokenizer) Tuple[str, GenContext, List[str]]
             +format_stream_start(ctx) List[str]
             +format_chunk(token) str
             +format_stream_end(ctx, stop) List[str]
@@ -768,7 +849,7 @@ classDiagram
         }
 
         class OpenAIResponseBuilder {
-            +prepare(request, engine) Tuple
+            +prepare(request, tokenizer) Tuple
             +format_stream_start(ctx) List[str]
             +format_chunk(token) str
             +format_stream_end(ctx, stop) List[str]
@@ -776,7 +857,7 @@ classDiagram
         }
 
         class AnthropicResponseBuilder {
-            +prepare(request, engine) Tuple
+            +prepare(request, tokenizer) Tuple
             +format_stream_start(ctx) List[str]
             +format_chunk(token) str
             +format_stream_end(ctx, stop) List[str]
@@ -787,12 +868,13 @@ classDiagram
             +request
             +engine
             +builder: ResponseBuilder
-            +handle() Union[StreamingResponse, Dict]
-            -_handle_stream(agen, ctx, stops) StreamingResponse
-            -_handle_non_stream(agen, ctx, stops) Dict
+            +async handle() Union[StreamingResponse, Dict]
+            -_handle_stream(agen, ctx, stop_sequences) StreamingResponse
+            -async _handle_non_stream(agen, ctx, stop_sequences) Dict
         }
 
         class StopChecker {
+            +__init__(sequences)
             +check(text) Optional[str]
         }
 
@@ -804,6 +886,12 @@ classDiagram
             +int completion_tokens
         }
 
+        class StopInfo {
+            +Optional[str] matched
+            +str body
+            +str yielded
+        }
+
         class app {
             <<singleton>>
             +FastAPI app
@@ -829,14 +917,14 @@ classDiagram
     }
 
     namespace parallel {
-        class Functions {
+        class setup {
             <<module>>
             +spawn_parallel_fn(func, world_size, backend, master_addr, master_port, device_type, start_method, **kwargs)
-            +setup_parallel(rank, world_size, backend, master_addr, master_port, device_type)
+            +setup_parallel(rank, world_size, backend, master_addr, master_port, device_type) contextmanager
             +get_current_device() str
             +get_world_size() int
             +get_rank() int
-            +only_on_rank(rank, sync) decorator
+            +only_on_rank(rank, sync=False) decorator
         }
 
         class GradientState {
@@ -847,6 +935,7 @@ classDiagram
         class AccumOptimizer {
             +Optimizer optimizer
             +GradientState gradient_state
+            +param_groups (property)
             +step(closure)
             +zero_grad()
             +state_dict() dict
@@ -867,7 +956,7 @@ classDiagram
             +prepare(model, optimizer, dataloader, scheduler) tuple
             +accumulate(model) context manager
             +backward(loss)
-            +unwrap_model(model) nn.Module
+            +unwrap_model(model) dict
             +sync_gradients (property) bool
             +grad_accum_steps (property) int
         }
@@ -876,14 +965,14 @@ classDiagram
         }
 
         class DDPExecutor {
-            +_prepare_model(model) nn.Module
-            +_no_sync(model) context manager
-            +unwrap_model(model) nn.Module
+            -_prepare_model(model) nn.Module
+            -_no_sync(model) context manager
+            +unwrap_model(model) dict
         }
 
         class FSDPExecutor {
-            +_prepare_model(model) nn.Module
-            +unwrap_model(model) nn.Module
+            -_prepare_model(model) nn.Module
+            +unwrap_model(model) dict
         }
 
         class ExecutorFactory {
@@ -899,11 +988,25 @@ classDiagram
         }
 
         class ColumnParallelLinear {
+            +int in_features
+            +int out_features
+            +int out_features_per_rank
+            +bool gather_results
+            +Parameter weight
+            +Optional[Parameter] bias
             +forward(x) Tensor
+            +load_state_dict(state_dict)
         }
 
         class RowParallelLinear {
+            +int in_features
+            +int out_features
+            +int in_features_per_rank
+            +bool reduce_results
+            +Parameter weight
+            +Optional[Parameter] bias
             +forward(x) Tensor
+            +load_state_dict(state_dict)
         }
     }
 
@@ -938,6 +1041,10 @@ classDiagram
     AutoModel <|-- EmbeddingEncoder
     BaseConfig <|-- BaseModelConfig
     BaseConfig <|-- TrainConfig
+    BaseConfig <|-- InputConfig
+    BaseConfig <|-- ProcessingConfig
+    BaseConfig <|-- OutputConfig
+    BaseConfig <|-- PipelineConfig
     BaseModelConfig <|-- AutoRegressiveLMConfig
     BaseModelConfig <|-- EncoderConfig
     BaseFactory <|-- AutoModel
@@ -950,11 +1057,15 @@ classDiagram
     BaseFactory <|-- StoreFactory
     BaseFactory <|-- ExecutorFactory
     BaseFactory <|-- ConfigFactory
+    BaseFactory <|-- MaskBuilderFactory
     BaseExecutor <|-- NoneExecutor
     BaseExecutor <|-- DDPExecutor
     BaseExecutor <|-- FSDPExecutor
     ResponseBuilder <|-- OpenAIResponseBuilder
     ResponseBuilder <|-- AnthropicResponseBuilder
+    BaseMaskBuilder <|-- ChatMaskBuilder
+    BaseMaskBuilder <|-- InstructionMaskBuilder
+    BaseMaskBuilder <|-- TextMaskBuilder
 
     %% --- Composition (strong ownership, part destroyed with whole) ---
     KVCache *-- PagePool
@@ -991,9 +1102,13 @@ classDiagram
     KvcacheView o-- Storage
     SamplingPipeline o-- BaseSamplingStrategy
     BaseDataset o-- Store
+    Pipeline o-- PipelineConfig
+    Pipeline o-- BaseMaskBuilder
 
     %% --- Dependency (uses temporarily) ---
     TrainConfig ..> BaseStrategy : selects
+    PipelineConfig ..> MaskBuilderFactory : selects
+    MaskBuilderFactory ..> BaseMaskBuilder : creates
     StrategyFactory ..> BaseStrategy : creates
     SchedulerFactory ..> BaseScheduler : creates
     DatasetFactory ..> BaseDataset : creates
@@ -1046,7 +1161,8 @@ classDiagram
 
 | Module | Components | Description |
 |--------|------------|-------------|
-| **astrai.config** | BaseConfig, BaseModelConfig, AutoRegressiveLMConfig, EncoderConfig, ConfigFactory, TrainConfig | Configuration management (to_dict/from_dict, to_file/from_file) |
+| **astrai.config** | BaseConfig, BaseModelConfig, AutoRegressiveLMConfig, EncoderConfig, ConfigFactory, TrainConfig, PipelineConfig, InputConfig, ProcessingConfig, OutputConfig | Configuration management (to_dict/from_dict, to_file/from_file, from_json/to_json) |
+| **astrai.preprocessing** | BaseMaskBuilder, MaskBuilderFactory, ChatMaskBuilder, InstructionMaskBuilder, TextMaskBuilder, Pipeline, filter_by_length, dedup_signature | Declarative JSON-driven data preprocessing |
 | **astrai.dataset** | BaseDataset–GRPODataset, Store–MmapStore, 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 |
@@ -1070,14 +1186,14 @@ classDiagram
 | **Observer** | `TrainCallback`, callback implementations | Training process monitoring |
 | **Context** | `TrainContext` | Unified training state bag |
 | **Object Pool** | `Allocator`, `PagePool` | Page-based KV cache with LRU eviction |
-| **Executor** | `BaseExecutor`, `NoneExecutor`, `DDPExecutor` | Gradient accumulation & model distribution |
+| **Executor** | `BaseExecutor`, `NoneExecutor`, `DDPExecutor`, `FSDPExecutor` | Gradient accumulation & model distribution |
 | **Storage** | `Store`, `H5Store`, `MmapStore` | Format-agnostic data access with multi-segment support |
 | **Producer-Consumer** | `InferenceScheduler`, `Task`, queues | Continuous batching |
 | **AutoModel Registry** | `AutoModel`, `AutoRegressiveLM`, `EmbeddingEncoder` | Model-type dynamic loading |
 
 ## Core Relationships
 
-1. **Config → Training**: `TrainConfig` holds model, dataset, optimizer_fn, scheduler_fn, `parallel_mode`, `executor_kwargs`
+1. **Config → Training**: `TrainConfig` holds `model_fn`, `dataset`, `optimizer_fn`, `scheduler_fn`, `parallel_mode`, `executor_kwargs`
 2. **Training Flow**: `Trainer` → `TrainContextBuilder` → `TrainContext`, uses `BaseStrategy` for loss, `BaseExecutor` for gradient accumulation + model distribution
 3. **Strategy Selection**: `StrategyFactory` creates strategy by `train_type`
 4. **Executor Selection**: `ExecutorFactory.create(cfg.parallel_mode, grad_accum_steps=cfg.grad_accum_steps, **cfg.executor_kwargs)` → `NoneExecutor` / `DDPExecutor` / `FSDPExecutor`
@@ -1089,4 +1205,4 @@ 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: 2026-05-28
+> Document Update Time: 2026-05-30

assets/docs/dataflow.md

@@ -5,7 +5,7 @@ This document describes the data pipeline: from raw text to model input tensors.
 ## Overview
 
 ```
-Raw Text → AutoTokenizer → Token IDs → .h5/.bin → Dataset → Sampler → DataLoader → Training/Inference
+Raw Text → AutoTokenizer → Token IDs → .h5/.bin → Store.load() → Store.fetch() → Dataset → Sampler → DataLoader → Training/Inference
 ```
 
 ## Data Preparation
@@ -33,14 +33,21 @@ H5 backend supports shared memory via `.share_memory_()`. Bin (mmap) uses OS pag
 ## Dataset Architecture
 
 ```
-DatasetFactory.load(train_type, load_path, window_size, stride, storage_type)
-  → StoreFactory.create(detect_format(path))
-    → Store._data[Dict[str, List[Tensor]]] + _cum[Dict[str, List[int]]]
-      → BaseDataset.__getitem__(idx)
-        → sliding window [begin, end) via get_index(idx)
+DatasetFactory.load(train_type, load_path, window_size, stride=None, storage_type=None)
+  → BaseDataset.load(load_path, storage_type=None)
+    → detect_format(load_path)
+    → StoreFactory.create(storage_type)
+    → Store.load(load_path)
+      → H5Store._normalize() / MmapStore._normalize()
+        → Store._data[Dict[str, List[Tensor]]] + _cum[Dict[str, List[int]]]
+          → BaseDataset.__getitem__(idx)
+            → get_index(idx) → [begin, end)
+            → Store.fetch(begin, end, keys) → Tensor / Dict[str, Tensor]
 ```
 
-`window_size` = max input length, `stride` = step between consecutive samples (defaults to `window_size`).
+`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`).
+
+`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()`.
 
 ## Sampler
 
@@ -54,4 +61,4 @@ DatasetFactory.load(train_type, load_path, window_size, stride, storage_type)
 
 Standard PyTorch `DataLoader` with configurable `batch_size`, `num_workers`, `pin_memory`, `prefetch_factor`. Sampler produces indices; dataloader fetches tensor batches via `__getitem__`.
 
-> Document Update Time: 2026-05-28
+> Document Update Time: 2026-05-30

assets/docs/inference.md

@@ -12,7 +12,7 @@ RoPE is applied **before** KV cache write, not after — otherwise position enco
 
 ## KVCache System
 
-Six classes working together:
+Six classes (plus two helpers) working together:
 
 ```
 KVCache (facade)
@@ -43,7 +43,8 @@ KVCache (facade)
 BaseSamplingStrategy (ABC)
   ├── TemperatureStrategy
   ├── TopKStrategy
-  └── TopPStrategy
+  ├── TopPStrategy
+  └── SamplingPipeline
 ```
 
 `SamplingPipeline` composes them: Temperature → Top-K → Top-P → softmax → multinomial.  
@@ -73,7 +74,9 @@ Adding a protocol = one builder file, no handler subclassing needed.
 InferenceEngine
   ├── generate(prompt, stream, ...) → str | List[str] | Generator
   ├── generate_with_request(req)    → same
-  └── generate_async(prompt, ...)   → AsyncGenerator
+  ├── generate_async(prompt, ...)   → AsyncGenerator
+  ├── get_stats()                   → Dict
+  └── shutdown()
 ```
 
 `GenerateResult` uses `Condition` for non-streaming (`wait_completion()`) and `Event` for streaming (`wait()`). Stream callback is `cb(token)`.
@@ -124,9 +127,9 @@ Supports `stop_sequences` and streaming via `event: content_block_delta`.
 | Param | Type | Default | Description |
 |-------|------|---------|-------------|
 | `messages` | List[dict] | required | Chat messages (role, content) |
-| `temperature` | float | 1.0 | Sampling temperature (>= 0.0) |
-| `top_p` | float | 1.0 | Nucleus threshold |
 | `top_k` | int | 50 | Top-k count |
+| `top_p` | float | 1.0 | Nucleus threshold |
+| `temperature` | float | 1.0 | Sampling temperature (> 0.0) |
 | `max_tokens` | Optional[int] | None | Max generation length |
 | `stream` | bool | False | Stream output |
 
@@ -142,7 +145,8 @@ engine.generate("Hello", stream=True)           # -> Generator[str]
 engine.generate(["A", "B"], stream=True)        # -> Generator[Tuple[int, str]]
 
 # Async
-await engine.generate_async("Hello", ...)       # -> AsyncGenerator[str]
+async for token in engine.generate_async("Hello", ...):    # -> AsyncGenerator[str]
+    print(token)
 ```
 
-> Document Update Time: 2026-05-28
+> Document Update Time: 2026-05-30

assets/docs/params.md

@@ -75,6 +75,7 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 nohup python scripts/tools/train.py \
     --nprocs=4 \
+    --parallel_mode=ddp \
     --train_type=seq \
     --data_root_path=/path/to/dataset \
     --param_path=/path/to/model \

assets/docs/preprocessing.md

@@ -0,0 +1,346 @@
+# Preprocessing Pipeline
+
+Declarative JSON-driven data preprocessing. One `SectionedMaskBuilder` handles all formats via `input.sections` (single-output) or `input.sources` (multi-output).
+
+## Philosophy
+
+| Component | Responsibility |
+|-----------|---------------|
+| `tokenizer_config.json` (`chat_template`) | Formatting -- how roles become tokens |
+| `pipeline.json` (`mask`) | Masking -- which roles participate in training |
+
+A single config file captures the entire pipeline, reusable and version-controllable.
+
+## Config Structure
+
+```json
+{
+  "input":         {},   // sections (single) or sources (multi)
+  "mask":          {},   // role → "train" | "mask"
+  "mask_default":  "mask",
+  "preprocessing": {},
+  "output":        {}
+}
+```
+
+### Section Fields
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `field` | str | -- | JSONL key to read |
+| `action` | str | -- | `"train"` / `"mask"` / `"$role"` |
+| `template` | bool | `false` | Apply `chat_template` per message |
+| `add_special_tokens` | bool | `true` for first non-template section | Add special tokens during encode |
+
+### Source Fields (multi-output mode)
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `sections` | list[dict] | -- | Same as single-output section list |
+| `list_field` | bool | `false` | JSONL field holds a list; tokenise each element |
+| `mask_key` | str | `"{key}_mask"` | Explicit output key for loss mask |
+
+---
+
+## Quick Start
+
+### SFT Chat
+
+Input JSONL:
+
+```json
+{"messages": [{"role": "system", "content": "You are helpful."}, {"role": "user", "content": "Hi"}, {"role": "assistant", "content": "Hello!"}]}
+```
+
+Config:
+
+```json
+{
+  "input": {
+    "sections": [
+      {"field": "messages", "action": "$role", "template": true}
+    ]
+  },
+  "mask": {
+    "system": "mask",
+    "user": "mask",
+    "assistant": "train"
+  },
+  "mask_default": "mask",
+  "preprocessing": {
+    "max_seq_len": 2048
+  },
+  "output": {
+    "storage_format": "bin",
+    "dtype": {"loss_mask": "bool"}
+  }
+}
+```
+
+Output keys: `sequence` (int32), `loss_mask` (bool)
+
+### SFT Instruction
+
+Input JSONL:
+
+```json
+{"prompt": "Translate to French: Hello", "response": "Bonjour"}
+```
+
+Config:
+
+```json
+{
+  "input": {
+    "sections": [
+      {"field": "prompt",   "action": "mask", "add_special_tokens": true},
+      {"field": "response", "action": "train"}
+    ]
+  },
+  "mask_default": "mask",
+  "preprocessing": {
+    "max_seq_len": 2048
+  }
+}
+```
+
+Output keys: `sequence`, `loss_mask`
+
+### Pretrain
+
+Input JSONL:
+
+```json
+{"text": "Artificial Intelligence is a field of computer science..."}
+```
+
+Config:
+
+```json
+{
+  "input": {
+    "sections": [
+      {"field": "text", "action": "train"}
+    ]
+  },
+  "preprocessing": {
+    "max_seq_len": 8192,
+    "min_chars": 100
+  }
+}
+```
+
+Output keys: `sequence` (no `loss_mask` — all tokens trained)
+
+### DPO
+
+Input JSONL:
+
+```json
+{"chosen": [{"role": "user", "content": "What is 2+2?"}, {"role": "assistant", "content": "4"}], "rejected": [{"role": "user", "content": "What is 2+2?"}, {"role": "assistant", "content": "5"}]}
+```
+
+Config:
+
+```json
+{
+  "input": {
+    "sources": {
+      "chosen": {
+        "sections": [
+          {"field": "chosen", "action": "$role", "template": true}
+        ]
+      },
+      "rejected": {
+        "sections": [
+          {"field": "rejected", "action": "$role", "template": true}
+        ]
+      }
+    }
+  },
+  "mask": {
+    "user": "mask",
+    "assistant": "train"
+  },
+  "mask_default": "mask"
+}
+```
+
+Output keys: `chosen`, `chosen_mask`, `rejected`, `rejected_mask`
+
+### GRPO
+
+Input JSONL:
+
+```json
+{"prompt": [{"role": "user", "content": "What is 2+2?"}], "responses": ["4", "Five", "Four"], "rewards": [1.0, 0.3, 0.8]}
+```
+
+Config:
+
+```json
+{
+  "input": {
+    "sources": {
+      "prompts": {
+        "sections": [
+          {"field": "prompt", "action": "mask", "template": 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"
+}
+```
+
+Output keys: `prompts`, `responses`, `masks`, `rewards` (float32)
+
+- `action: "value"` — extract raw values from JSONL without tokenisation
+- `list_field: true` — tokenise each list element independently, then concatenate
+- `mask_key: "masks"` — rename the auto-generated mask key (default: `responses_mask`)
+
+---
+
+## Configuration Reference
+
+### `input`
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `sections` | list[dict] or null | `null` | Section specs for single-output mode |
+| `sources` | dict[str, dict] or null | `null` | Source specs for multi-output mode (DPO/GRPO) |
+
+When `sources` is set, `sections` is ignored.
+
+### `mask`
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `mask` | dict | `{}` | `{role: "train" \| "mask"}` |
+| `mask_default` | str | `"mask"` | Default action for unlisted roles |
+
+### `preprocessing`
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `max_seq_len` | int | `2048` | Truncate sequences to this length |
+| `min_chars` | int | `50` | Skip text-mode items shorter than this |
+| `max_chars` | int | `2000000` | Skip text-mode items longer than this |
+| `max_items` | int or null | `null` | Stop after N documents |
+
+### `output`
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `domain_key` | str or null | `null` | JSONL key for domain grouping |
+| `storage_format` | str | `"bin"` | `"bin"` (mmap) or `"h5"` |
+| `max_tokens_per_shard` | int | `100000000` | Flush threshold in cumulative tokens |
+| `dtype` | dict[str, str] | `{}` | Per-key tensor dtype override (e.g. `{"loss_mask": "bool"}`) |
+
+---
+
+## Mask Algorithm
+
+### Template mode (`template: true`)
+
+For each message in the field's array:
+
+1. Prepend BOS token (masked)
+2. Render through `chat_template` for that single message
+3. Encode rendered text
+4. Apply mask rule for the message's role
+
+### Non-template mode
+
+Encode the field value as text. Mask value is 1 (train) or 0 (mask) per the section's `action`.
+
+### Text config detection
+
+When no section uses `template` and all sections have `action: "train"`, the builder skips mask generation entirely — all tokens are trained.
+
+---
+
+## Output Layout
+
+### Single-Shard (`bin`)
+
+```
+output/
+  __default__/
+    meta.json
+    sequence.bin
+    loss_mask.bin
+  wiki/
+    meta.json
+    sequence.bin
+    loss_mask.bin
+```
+
+### Multi-Shard (`bin`)
+
+When `max_tokens_per_shard` is exceeded:
+
+```
+output/
+  __default__/
+    shard_0000/
+      meta.json
+      sequence.bin
+      loss_mask.bin
+    shard_0001/
+      meta.json
+      sequence.bin
+      loss_mask.bin
+```
+
+`MmapStore` discovers all shards under the domain directory via `rglob("meta.json")`.
+
+---
+
+## CLI
+
+```bash
+# SFT
+python scripts/tools/preprocess.py data/sft/*.jsonl -o output/sft/ -c configs/sft_chat.json
+
+# DPO
+python scripts/tools/preprocess.py data/dpo/*.jsonl -o output/dpo/ -c configs/dpo.json --tokenizer_path params
+
+# GRPO
+python scripts/tools/preprocess.py data/grpo/*.jsonl -o output/grpo/ -c configs/grpo.json
+```
+
+---
+
+## Python API
+
+```python
+from astrai.preprocessing.pipeline import Pipeline
+from astrai.config.preprocess_config import PipelineConfig
+
+config = PipelineConfig.from_json("sft.json")
+Pipeline(
+    config,
+    ["data_part1.jsonl", "data_part2.jsonl"],
+    output_dir="output/",
+    tokenizer_path="params",
+).run()
+```
+
+> Document Update Time: 2026-06-03

assets/docs/training.md

@@ -1,38 +1,5 @@
 # Training
 
-## Model Architecture
-
-The model uses a decoder-only Transformer with **GQA** (Grouped Query Attention) and optional **MLA** (Multi-head Latent Attention). 1.0 billion parameters, Chinese–English bilingual.
-
-```mermaid
-flowchart TB
-    subgraph Layers["Transformer Layers"]
-        direction TB
-        A[Input Embedding] --> B[Transformer Block\nLayer 1]
-        B --> C[Transformer Block\nLayer ...]
-        C --> D[Transformer Block\nLayer ...]
-        D --> E[RMSNorm]
-        E --> F[Linear]
-        F --> G[SoftMax]
-    end
-
-    subgraph TransformerBlock["Transformer Block"]
-        direction TB
-        H[x] --> I[RMSNorm]
-        I --> J[Linear → Q/K/V]
-        J --> K[Q]; J --> L[K]; J --> M[V]
-        K --> N[RoPE]; L --> O[RoPE]
-        N --> P["Q @ K^T / sqrt(d)"]; O --> P
-        P --> Q[Masked SoftMax]; Q --> R[S @ V]; M --> R
-        R --> S[Linear]; S --> T[+]; H --> T
-        T --> U[RMSNorm]
-        U --> V["Linear (gate)"]; U --> W["Linear (up)"]
-        V --> X[SiLU]; X --> Y[×]; W --> Y
-        Y --> Z["Linear (down)"]; Z --> AA[+]; T --> AA
-        AA --> BB[x']
-    end
-```
-
 ### Autoregression
 
 Given a token sequence, the model predicts the probability of the next token. Each generated token is appended to the input and fed back, repeating until an end-of-sequence token or max length.
@@ -69,14 +36,16 @@ Two-level loop: **epoch** → **batch**. Optimizer step fires every `grad_accum_
 
 ```
 on_train_begin
+  model.train()
   on_epoch_begin
     for batch in dataloader:
       on_batch_begin
       with executor.accumulate(model):
-        loss = strategy(batch)
+        loss = strategy.compute_loss(batch)
+        context.loss = loss.item()
         stand_loss = loss / executor.grad_accum_steps
         executor.backward(stand_loss)
-        iteration += 1
+        context.iteration += 1
         on_batch_end
 
         if executor.sync_gradients:
@@ -94,9 +63,13 @@ on_train_end
 | Hook | Fires | Default callback |
 |------|-------|-----------------|
 | `on_train_begin` | Before training starts | `GradientCheckpointingCallback` |
+| `on_epoch_begin` | Start of each epoch | `ProgressBarCallback` |
+| `on_batch_begin` | Every batch | — |
 | `on_optimizer_step` | Every accumulation window | `GradientClippingCallback`, `ValidationCallback` |
 | `on_batch_end` | Every batch | `CheckpointCallback`, `MetricLoggerCallback`, `ProgressBarCallback` |
-| `on_train_end` | Training ends | `CheckpointCallback`, `MetricLoggerCallback` (final save) |
+| `on_epoch_end` | End of each epoch | `ProgressBarCallback` |
+| `on_error` | On exception during training | `CheckpointCallback`, `MetricLoggerCallback` |
+| `on_train_end` | Training ends (always via finally) | `CheckpointCallback`, `MetricLoggerCallback`, `GradientCheckpointingCallback` |
 
 Default callbacks (in order): `gradient_checkpointing` (activation checkpointing, optional), `checkpoint` (safetensors, rank-0), `metric_logger` (JSONL, rank-0), `progress_bar` (tqdm), `gradient_clipping`, `validation` (periodic validation on val_dataset).
 
@@ -110,7 +83,7 @@ $$
 L_{\text{PT}} = -\sum_{t=1}^{T} \log P(x_t \mid x_{\lt t}; \theta)
 $$
 
-Keys: `input_ids`, `target_ids`
+Keys: `input_ids`, `target_ids`. Optional: `label_smoothing`.
 
 ### SFT (Supervised Fine-Tuning)
 
@@ -120,7 +93,7 @@ $$
 L_{\text{SFT}} = -\sum_{t=P+1}^{P+L} \log P(s_t \mid s_{\lt t}; \theta)
 $$
 
-Keys: `input_ids`, `target_ids`, `loss_mask`
+Keys: `input_ids`, `target_ids`, `loss_mask`. Optional: `label_smoothing`.
 
 ### DPO (Direct Preference Optimization)
 
@@ -130,7 +103,7 @@ $$
 L_{\text{DPO}} = -\mathbb{E}\left[\log\sigma\left(\beta\log\frac{\pi_\theta(y_w\mid x)}{\pi_{\text{ref}}(y_w\mid x)} - \beta\log\frac{\pi_\theta(y_l\mid x)}{\pi_{\text{ref}}(y_l\mid x)}\right)\right]
 $$
 
-Parameters: `beta=0.1`. Keys: `chosen`, `rejected`, `chosen_mask`, `rejected_mask`.
+Parameters: `beta=0.1`, `reduction="mean"`. Keys: `chosen`, `rejected`, `chosen_mask`, `rejected_mask`.
 
 ### GRPO (Group Relative Policy Optimization)
 
@@ -144,7 +117,7 @@ $$
 L_{\text{GRPO}} = -\mathbb{E}\left[\min\left(\frac{\pi_\theta}{\pi_{\text{ref}}}A,\; \text{clip}\left(\frac{\pi_\theta}{\pi_{\text{ref}}}, 1-\epsilon, 1+\epsilon\right)A\right)\right] + \lambda \cdot \mathbb{E}\left[(\log\pi_\theta - \log\pi_{\text{ref}})^2\right]
 $$
 
-Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`, `sync_interval=200`.
+Parameters: `group_size=4`, `clip_eps=0.2`, `kl_coef=0.01`, `sync_interval=200`, `reduction="mean"`.
 
 Keys: `prompts`, `responses`, `masks`, `rewards`.
 
@@ -155,7 +128,7 @@ Keys: `prompts`, `responses`, `masks`, `rewards`.
 | Cosine | `CosineScheduler` | Linear warmup → cosine decay to `min_rate` |
 | SGDR | `SGDRScheduler` | Cosine annealing with warm restarts (`t_mult=2`) |
 
-Created by `SchedulerFactory.create(optimizer, schedule_type, **kwargs)`.
+Created by `SchedulerFactory.create(optimizer, schedule_type, **kwargs)`. Valid types: `"cosine"`, `"sgdr"`. Omit to use no scheduler.
 
 ## Gradient Checkpointing
 
@@ -172,8 +145,8 @@ Callback wraps each `DecoderBlock.forward` with `torch.utils.checkpoint.checkpoi
 
 ```
 Checkpoint(state_dict, epoch, iteration, extra, meta, config)
-  ├── save(save_dir)    rank-0 only: meta.json (epoch/iteration/timestamp) + config.json (model config) + state_dict.safetensors + optional {key}.pt (optimizer.pt, scheduler.pt)
-  └── load(save_dir)    broadcasts metadata from rank-0
+  ├── save(save_dir)    rank-0 only: meta.json (epoch/iteration/timestamp) + config.json (model config) + model.safetensors + optional {key}.pt (optimizer.pt, scheduler.pt)
+  └── load(save_dir, broadcast=False)    loads from local disk; set broadcast=True to broadcast metadata from rank-0
 ```
 
 Optimizer/scheduler state persisted by default via `Checkpoint.extra`.  
@@ -194,7 +167,7 @@ context = (
 - Creates executor via `ExecutorFactory.create(cfg.parallel_mode, grad_accum_steps=cfg.grad_accum_steps, **cfg.executor_kwargs)`
 - Calls `executor.prepare(model, optimizer, dataloader, scheduler)` for model distribution (e.g. DDP) + gradient accumulation wrappers
 - Creates `ResumableDistributedSampler` for shuffle+resume
-- Builds strategy via `StrategyFactory.create(train_type, ...)`
+- Builds strategy via `StrategyFactory.create(train_type, model, device, **kwargs)`
 
 ## Training CLI
 
@@ -203,6 +176,7 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 nohup python scripts/tools/train.py \
     --nprocs=4 \
+    --parallel_mode=ddp \
     --train_type=seq \
     --data_root_path=/path/to/dataset \
     --param_path=/path/to/model \
@@ -224,4 +198,4 @@ nohup python scripts/tools/train.py \
 
 Full parameter reference at [params.md](params.md).
 
-> Document Update Time: 2026-05-28
+> Document Update Time: 2026-05-30

astrai/config/__init__.py

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

astrai/config/base.py

@@ -1,6 +1,7 @@
 import json
 from dataclasses import MISSING, dataclass, fields
-from typing import Any, Dict, Optional, Self, get_type_hints
+from pathlib import Path
+from typing import Any, Dict, Optional, Self, Union, get_type_hints
 
 
 @dataclass
@@ -83,4 +84,15 @@ class BaseConfig:
             return value
         if isinstance(value, target_type):
             return value
+        if isinstance(value, dict) and issubclass(target_type, BaseConfig):
+            return target_type.from_dict(value)
         raise TypeError
+
+    @classmethod
+    def from_json(cls, path: Union[str, Path]) -> Self:
+        with open(path, "r", encoding="utf-8") as f:
+            return cls.from_dict(json.load(f))
+
+    def to_json(self, path: Union[str, Path]):
+        with open(path, "w", encoding="utf-8") as f:
+            json.dump(self.to_dict(), f, indent=2, ensure_ascii=False)

astrai/config/preprocess_config.py

@@ -0,0 +1,109 @@
+"""Pipeline configuration for JSONL preprocessing.
+
+Supports single-sequence (SFT/pretrain) and multi-output (DPO/GRPO)
+modes, both driven declaratively through ``input.sections`` or
+``input.sources``.
+"""
+
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional
+
+from astrai.config.base import BaseConfig
+
+
+@dataclass
+class InputConfig(BaseConfig):
+    """Declarative input mapping.
+
+    Single-output mode (backward-compatible)::
+
+        {"input": {"sections": [{"field": "messages", ...}]}}
+
+    Multi-output mode (DPO / GRPO)::
+
+        {"input": {"sources": {
+            "chosen": {"sections": [{"field": "chosen", ...}]},
+            "rejected": {"sections": [{"field": "rejected", ...}]},
+        }}}
+    """
+
+    sections: Optional[List[Dict]] = None
+    sources: Optional[Dict[str, Dict]] = None
+
+
+@dataclass
+class ProcessingConfig(BaseConfig):
+    """Processing configuration.
+
+    Parameters
+    ----------
+    max_seq_len : int
+        Maximum sequence length (default: 2048).
+    min_chars : int
+        Minimum number of characters to keep (default: 50).
+    max_chars : int
+        Maximum number of characters to keep (default: 2_000_000).
+    max_items : Optional[int]
+        Maximum number of items to process (default: None, unlimited).
+    packing_strategy : str
+        How to pack sequences into a contiguous stream.
+
+        - ``"simple"``: sequential concatenation (default, backward compatible).
+        - ``"bfd"``: best-fit decreasing bin packing, minimises wasted tokens.
+        - ``"bfd_split"``: BFD with over-length sequences split into chunks.
+    max_packed_len : int
+        Maximum length of a packed bin. Sequences longer than this are
+        truncated or split depending on ``packing_strategy`` (default: 8192).
+    truncation_mode : str
+        How to truncate sequences longer than ``max_packed_len``.
+
+        - ``"keep_start"``: keep the first ``max_packed_len`` tokens (default).
+        - ``"keep_end"``: keep the last ``max_packed_len`` tokens.
+    """
+
+    max_seq_len: int = 2048
+    min_chars: int = 50
+    max_chars: int = 2_000_000
+    max_items: Optional[int] = None
+    packing_strategy: str = "simple"
+    max_packed_len: int = 8192
+    truncation_mode: str = "keep_start"
+
+
+@dataclass
+class OutputConfig(BaseConfig):
+    """Output configuration.
+
+    Parameters
+    ----------
+    domain_key : Optional[str]
+        Domain key for the output store (default: None).
+    storage_format : str
+        Storage format, one of ``"bin"``, ``"jsonl"`` (default: ``"bin"``).
+    max_tokens_per_shard : int
+        Maximum tokens per shard before splitting (default: 100_000_000).
+    dtype : Dict[str, str]
+        Per-key dtype overrides, e.g. ``{"input_ids": "int32"}`` (default: {}).
+    position_ids_mode : Optional[str]
+        How to compute position_ids in packed sequences.
+
+        - ``None`` / ``"none"``: do not generate (backward compatible).
+        - ``"doc_reset"``: reset to 0 at each document boundary.
+        - ``"continuous"``: sequential 0, 1, 2, ... (pretrain, single doc).
+    """
+
+    domain_key: Optional[str] = None
+    storage_format: str = "bin"
+    max_tokens_per_shard: int = 100_000_000
+    dtype: Dict[str, str] = field(default_factory=dict)
+    position_ids_mode: Optional[str] = None
+
+
+@dataclass
+class PipelineConfig(BaseConfig):
+    version: int = 1
+    input: InputConfig = field(default_factory=InputConfig)
+    mask: Dict[str, str] = field(default_factory=dict)
+    mask_default: str = "mask"
+    preprocessing: ProcessingConfig = field(default_factory=ProcessingConfig)
+    output: OutputConfig = field(default_factory=OutputConfig)

astrai/config/train_config.py

@@ -118,6 +118,12 @@ class TrainConfig(BaseConfig):
     val_dataset: Optional[Dataset] = field(
         default=None, metadata={"help": "Dataset for validation."}
     )
+    val_split: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "Ratio to split from training dataset for validation (e.g. 0.05). Ignored if val_dataset is set."
+        },
+    )
     val_step: int = field(
         default=1000,
         metadata={"help": "Number of optimizer steps between validation runs."},

astrai/dataset/dataset.py

@@ -223,7 +223,7 @@ class SFTDataset(BaseDataset):
 
     @property
     def required_keys(self) -> List[str]:
-        return ["sequence", "loss_mask"]
+        return ["sequence", "loss_mask", "position_ids"]
 
     def _fetch_data(self, begin_idx: int, end_idx: int, key: str) -> Tensor:
         return self.storage.fetch(begin_idx, end_idx, key)
@@ -231,15 +231,17 @@ class SFTDataset(BaseDataset):
     def __getitem__(self, index):
         begin_idx, end_idx = self.get_index(index)
 
-        x = self._fetch_data(begin_idx, end_idx, "sequence").to(dtype=torch.long)
-        y = self._fetch_data(begin_idx + 1, end_idx + 1, "sequence").to(
-            dtype=torch.long
-        )
-        loss_mask = self._fetch_data(begin_idx + 1, end_idx + 1, "loss_mask").to(
-            dtype=torch.bool
-        )
+        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")
 
-        return {"input_ids": x, "target_ids": y, "loss_mask": loss_mask}
+        return {
+            "input_ids": x.to(dtype=torch.long),
+            "target_ids": y.to(dtype=torch.long),
+            "position_ids": position_ids.to(dtype=torch.long),
+            "loss_mask": loss_mask.to(dtype=torch.bool),
+        }
 
 
 @DatasetFactory.register("dpo")

astrai/dataset/storage.py

@@ -18,6 +18,7 @@ Key properties:
 """
 
 import bisect
+import glob
 import json
 import os
 from abc import ABC, abstractmethod
@@ -113,12 +114,20 @@ def detect_format(load_path: str) -> str:
             return "h5"
         raise ValueError(f"Unsupported file format: {suffix}")
 
-    h5_files = list(root.rglob("*.h5")) + list(root.rglob("*.hdf5"))
+    h5_files = [
+        Path(p)
+        for pattern in ("*.h5", "*.hdf5")
+        for p in glob.glob(str(root / "**" / pattern), recursive=True)
+    ]
     if h5_files:
         return "h5"
-    bin_files = list(root.rglob("*.bin"))
-    if bin_files and (root / "meta.json").exists():
-        return "bin"
+    bin_files = [Path(p) for p in glob.glob(str(root / "**" / "*.bin"), recursive=True)]
+    if bin_files:
+        has_meta = (root / "meta.json").exists() or len(
+            [Path(p) for p in glob.glob(str(root / "**" / "meta.json"), recursive=True)]
+        ) > 0
+        if has_meta:
+            return "bin"
     raise FileNotFoundError(f"No supported data files found at {load_path}")
 
 
@@ -244,7 +253,19 @@ class MmapStore(Store):
 
     def load(self, path: str):
         self._mmap_refs = []
-        raw = load_bin(path)
-        self._normalize(raw)
+        root = Path(path)
+        all_raw: Dict[str, List[Tensor]] = {}
+        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))
+            for key, tensors in raw.items():
+                if key not in all_raw:
+                    all_raw[key] = []
+                all_raw[key].extend(tensors)
+        if not meta_paths:
+            raise FileNotFoundError(f"No meta.json found under {path}")
+        self._normalize(all_raw)
         for tensors in self._data.values():
             self._mmap_refs.extend(tensors)

astrai/inference/__init__.py

@@ -17,7 +17,7 @@ from astrai.inference.api import (
     MessagesRequest,
     ProtocolHandler,
     StopChecker,
-    app,
+    get_app,
     run_server,
 )
 from astrai.inference.api.anthropic import AnthropicResponseBuilder
@@ -80,6 +80,6 @@ __all__ = [
     "ChatCompletionRequest",
     "AnthropicMessage",
     "MessagesRequest",
-    "app",
+    "get_app",
     "run_server",
 ]

astrai/inference/api/__init__.py

@@ -1,4 +1,8 @@
-"""Inference API: protocol handler, stop checker, and FastAPI server."""
+"""Inference API: protocol handler, stop checker, and FastAPI server.
+
+``app`` is no longer a module-level global. Use :func:`get_app` to access the
+lazy singleton FastAPI instance.
+"""
 
 from astrai.inference.api.protocol import GenContext, ProtocolHandler, StopChecker
 from astrai.inference.api.server import (
@@ -6,7 +10,7 @@ from astrai.inference.api.server import (
     ChatCompletionRequest,
     ChatMessage,
     MessagesRequest,
-    app,
+    get_app,
     run_server,
 )
 
@@ -18,6 +22,6 @@ __all__ = [
     "ChatCompletionRequest",
     "ChatMessage",
     "MessagesRequest",
-    "app",
+    "get_app",
     "run_server",
 ]

astrai/inference/api/anthropic.py

@@ -1,5 +1,6 @@
 """Anthropic message completion response builder."""
 
+import time
 import uuid
 from typing import Any, Dict, List, Tuple, Union
 
@@ -39,7 +40,7 @@ class AnthropicResponseBuilder(ResponseBuilder):
         prompt = engine.tokenizer.apply_chat_template(messages, tokenize=False)
         ctx = GenContext(
             resp_id=f"msg_{uuid.uuid4().hex[:24]}",
-            created=0,
+            created=int(time.time()),
             model=request.model,
             prompt_tokens=0,
         )

astrai/inference/api/openai.py

@@ -1,5 +1,7 @@
 """OpenAI chat completion response builder."""
 
+import logging
+import time
 import uuid
 from typing import Any, Dict, List, Tuple
 
@@ -13,6 +15,16 @@ from astrai.inference.api.protocol import (
 )
 from astrai.inference.engine import InferenceEngine
 
+logger = logging.getLogger(__name__)
+
+_UNSUPPORTED_PARAMS = (
+    "n",
+    "presence_penalty",
+    "frequency_penalty",
+    "logit_bias",
+    "user",
+)
+
 
 class OpenAIResponseBuilder(ResponseBuilder):
     def prepare(
@@ -24,9 +36,26 @@ class OpenAIResponseBuilder(ResponseBuilder):
         self._resp_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
         self._model = request.model
 
+        for param in _UNSUPPORTED_PARAMS:
+            value = getattr(request, param, None)
+            fields = getattr(type(request), "model_fields", {})
+            default = fields[param].default if param in fields else None
+            if value is not None and value != default:
+                logger.warning(
+                    "ChatCompletionRequest param '%s'=%r is not supported and will be ignored",
+                    param,
+                    value,
+                )
+            if value is not None and value != default:
+                logger.warning(
+                    "ChatCompletionRequest param '%s'=%r is not supported and will be ignored",
+                    param,
+                    value,
+                )
+
         ctx = GenContext(
             resp_id=self._resp_id,
-            created=0,
+            created=int(time.time()),
             model=self._model,
             prompt_tokens=0,
         )

astrai/inference/api/protocol.py

@@ -64,7 +64,7 @@ class StopChecker:
 class ResponseBuilder(ABC):
     """Interface for protocol-specific response formatting.
 
-    A new protocol requires one concrete builder implementing 6 methods.
+    A new protocol requires one concrete builder implementing 5 methods.
     """
 
     @abstractmethod
@@ -138,13 +138,13 @@ class ProtocolHandler:
             yielded = ""
             matched = None
             async for token in agen:
-                ctx.completion_tokens += 1
                 body += token
 
                 matched = checker.check(body)
                 if matched:
                     break
 
+                ctx.completion_tokens += 1
                 yield self.builder.format_chunk(token)
                 yielded += token
 
@@ -168,7 +168,6 @@ class ProtocolHandler:
         matched = None
 
         async for token in agen:
-            ctx.completion_tokens += 1
             chunks.append(token)
             body += token
 
@@ -176,6 +175,8 @@ class ProtocolHandler:
             if matched:
                 break
 
+            ctx.completion_tokens += 1
+
         content = "".join(chunks)
         stop = StopInfo(matched=matched, body=body)
         return self.builder.format_response(ctx, content, stop)

astrai/inference/api/server.py

@@ -3,6 +3,9 @@ OpenAI / Anthropic-compatible chat completion server backed by continuous-batchi
 
 Protocol-specific formatting is delegated to ``astrai.inference.protocol``.
 This module owns the FastAPI app, request/response schemas, and dependency wiring.
+
+``app`` is lazily constructed — importing this module does NOT create a FastAPI instance.
+Use :func:`get_app` to access the singleton.
 """
 
 import logging
@@ -12,7 +15,7 @@ from typing import Any, Dict, List, Optional, Union
 
 import torch
 import uvicorn
-from fastapi import FastAPI, HTTPException
+from fastapi import APIRouter, FastAPI, HTTPException
 from pydantic import BaseModel, Field
 
 from astrai.inference.api.anthropic import AnthropicResponseBuilder
@@ -24,7 +27,7 @@ from astrai.tokenize import AutoTokenizer
 
 logger = logging.getLogger(__name__)
 
-_project_root = Path(__file__).parent.parent.parent
+_app_instance: Optional[FastAPI] = None
 
 
 class ChatMessage(BaseModel):
@@ -84,17 +87,15 @@ async def lifespan(app: FastAPI):
         logger.info("Inference engine shutdown complete")
 
 
-app = FastAPI(title="AstrAI Inference Server", version="0.2.0", lifespan=lifespan)
+router = APIRouter()
 
 
 def _create_engine(
-    param_path: Optional[Path] = None,
+    param_path: Path,
     device: str = "cuda",
     dtype: torch.dtype = torch.bfloat16,
     max_batch_size: int = 16,
 ) -> InferenceEngine:
-    if param_path is None:
-        param_path = _project_root / "params"
     if not param_path.exists():
         raise FileNotFoundError(f"Parameter directory not found: {param_path}")
 
@@ -112,34 +113,50 @@ def _create_engine(
     return engine
 
 
+def get_app() -> FastAPI:
+    """Return the singleton FastAPI instance (lazily created on first call)."""
+    global _app_instance
+    if _app_instance is None:
+        _app_instance = FastAPI(
+            title="AstrAI Inference Server",
+            version="0.2.0",
+            lifespan=lifespan,
+        )
+        _app_instance.include_router(router)
+        _app_instance.state.server_config = {}
+        _app_instance.state.engine = None
+    return _app_instance
+
+
 def _get_engine() -> InferenceEngine:
-    engine = app.state.engine
+    engine = get_app().state.engine
     if engine is None:
         raise HTTPException(status_code=503, detail="Engine not initialized")
     return engine
 
 
-@app.get("/health")
+@router.get("/health")
 async def health():
+    app = get_app()
     return {
         "status": "ok",
         "model_loaded": app.state.engine is not None,
     }
 
 
-@app.get("/stats")
+@router.get("/stats")
 async def get_stats():
     return _get_engine().get_stats()
 
 
-@app.post("/v1/chat/completions")
+@router.post("/v1/chat/completions")
 async def chat_completion(request: ChatCompletionRequest):
     engine = _get_engine()
     handler = ProtocolHandler(request, engine, OpenAIResponseBuilder())
     return await handler.handle()
 
 
-@app.post("/v1/messages")
+@router.post("/v1/messages")
 async def create_message(request: MessagesRequest):
     engine = _get_engine()
     handler = ProtocolHandler(request, engine, AnthropicResponseBuilder())
@@ -147,14 +164,15 @@ async def create_message(request: MessagesRequest):
 
 
 def run_server(
+    param_path: Path,
     host: str = "0.0.0.0",
     port: int = 8000,
     reload: bool = False,
     device: str = "cuda",
     dtype: torch.dtype = torch.bfloat16,
-    param_path: Optional[Path] = None,
     max_batch_size: int = 16,
 ):
+    app = get_app()
     app.state.server_config = {
         "device": device,
         "dtype": dtype,

astrai/inference/core/scheduler.py

@@ -71,6 +71,7 @@ class InferenceScheduler:
         )
 
         self._running = False
+        self._fatal_error: Optional[Exception] = None
 
     def add_task(self, prompt: str, **kwargs) -> str:
         return self._task_mgr.add_task(prompt, **kwargs)
@@ -175,6 +176,8 @@ class InferenceScheduler:
                                     t.stream_callback(STOP)
 
         except Exception as e:
+            self._fatal_error = e
+            self._running = False
             logger.error(f"Scheduler loop crashed: {e}", exc_info=True)
             for task in self._task_mgr.get_active_tasks():
                 if task.stream_callback:
@@ -184,7 +187,6 @@ class InferenceScheduler:
                 if task.stream_callback:
                     task.stream_callback(STOP)
             self._task_mgr.clear_queues()
-            raise
 
     def start(self):
         if not self._running:
@@ -199,7 +201,12 @@ class InferenceScheduler:
         if hasattr(self, "_loop_thread"):
             self._loop_thread.join(timeout=2.0)
         for task in self._task_mgr.get_active_tasks():
+            if task.stream_callback:
+                task.stream_callback(STOP)
             self._page_cache.task_free(task.task_id)
+        for task in self._task_mgr.get_waiting_tasks():
+            if task.stream_callback:
+                task.stream_callback(STOP)
         self._task_mgr.clear_queues()
         if torch.cuda.is_available():
             torch.cuda.empty_cache()

astrai/inference/core/task.py

@@ -186,7 +186,10 @@ class TaskManager:
         return bool(self.active_tasks or self.waiting_queue)
 
     def wait_for_tasks(self, timeout: float = 1.0):
-        self._task_event.clear()
+        with self._lock:
+            if self.waiting_queue or self.active_tasks:
+                return
+            self._task_event.clear()
         self._task_event.wait(timeout=timeout)
 
     def get_active_tasks(self) -> List[Task]:

astrai/inference/engine.py

@@ -79,8 +79,8 @@ class GenerationRequest:
             raise ValueError("top_k must be a non-negative integer")
         if not (0.0 <= top_p <= 1.0):
             raise ValueError("top_p must be a float between 0.0 and 1.0")
-        if not (isinstance(temperature, (int, float)) and temperature >= 0):
-            raise ValueError("temperature must be a non-negative number")
+        if not (isinstance(temperature, (int, float)) and temperature > 0):
+            raise ValueError("temperature must be a positive number")
 
         self.messages = messages
         self.top_k = top_k

astrai/inference/sample.py

@@ -44,10 +44,12 @@ class TemperatureStrategy(BaseSamplingStrategy):
     def apply(self, logits, filter_value=-float("inf")):
         t = self.temperature
         if isinstance(t, Tensor):
+            t = t.to(logits.device, non_blocking=True).view(-1, 1)
+            t = torch.clamp(t, min=1e-8)
             if (t != 1.0).any():
-                logits = logits / t.to(logits.device, non_blocking=True).view(-1, 1)
+                logits = logits / t
         elif t != 1.0:
-            logits = logits / t
+            logits = logits / max(t, 1e-8)
         return logits
 
 

astrai/model/transformer.py

@@ -26,24 +26,21 @@ def process_attention_mask(
         return input_mask
 
     device = input_tensor.device
-    dtype = input_tensor.dtype
-    B, S = input_tensor.size()[:2]
+    B = input_tensor.size(0)
     T = position_ids.max().item() + 1
 
     if input_mask is None:
         if position_ids.min().item() == 0 and is_causal:
             return None
-        pad = torch.ones(B, T, dtype=torch.bool, device=device)
+        attend = torch.ones(B, 1, T, dtype=torch.bool, device=device)
     else:
-        pad = input_mask[:, :T].to(device=device, dtype=torch.bool)
+        attend = input_mask[:, :T].to(device=device, dtype=torch.bool).unsqueeze(1)
 
-    attend = pad.view(B, 1, T).expand(B, S, T).clone()
     if is_causal:
-        attend &= position_ids.unsqueeze(-1) >= torch.arange(T, device=device)
+        causal = position_ids.unsqueeze(-1) >= torch.arange(T, device=device)
+        attend = attend & causal
 
-    return torch.full(
-        (B, 1, S, T), -torch.finfo(dtype).max / 2, dtype=dtype, device=device
-    ).masked_fill_(attend.unsqueeze(1), 0.0)
+    return attend.unsqueeze(1)
 
 
 @AutoModel.register("autoregressive_lm")

astrai/parallel/executor.py

@@ -2,11 +2,13 @@
 
 import contextlib
 import logging
+import os
 from contextlib import contextmanager
 from typing import Optional, Tuple
 
 import torch
 import torch.nn as nn
+from torch.distributed.fsdp import FullStateDictConfig, StateDictType
 from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
 from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.optim import Optimizer
@@ -115,8 +117,8 @@ class BaseExecutor:
     def backward(self, loss: torch.Tensor):
         loss.backward()
 
-    def unwrap_model(self, model: nn.Module) -> nn.Module:
-        return model
+    def unwrap_model(self, model: nn.Module):
+        return model.state_dict()
 
     @property
     def use_distributed(self) -> bool:
@@ -180,7 +182,7 @@ class DDPExecutor(BaseExecutor):
         if not self.use_distributed:
             logger.warning("DDP backend selected but world_size=1, model not wrapped")
             return model
-        local_rank = get_rank()
+        local_rank = int(os.environ.get("LOCAL_RANK", get_rank()))
         model = DDP(
             model,
             device_ids=[local_rank],
@@ -195,10 +197,10 @@ class DDPExecutor(BaseExecutor):
             return model.no_sync()
         return contextlib.nullcontext()
 
-    def unwrap_model(self, model: nn.Module) -> nn.Module:
+    def unwrap_model(self, model: nn.Module):
         if isinstance(model, DDP):
-            return model.module
-        return model
+            return model.module.state_dict()
+        return model.state_dict()
 
 
 @ExecutorFactory.register("fsdp")
@@ -217,7 +219,6 @@ class FSDPExecutor(BaseExecutor):
         sync_module_states: bool = False,
         forward_prefetch: bool = False,
         limit_all_gathers: bool = True,
-        use_orig_params: bool = False,
         ignored_states=None,
         device_mesh=None,
     ):
@@ -236,7 +237,7 @@ class FSDPExecutor(BaseExecutor):
                 sync_module_states=sync_module_states,
                 forward_prefetch=forward_prefetch,
                 limit_all_gathers=limit_all_gathers,
-                use_orig_params=use_orig_params,
+                use_orig_params=True,
                 ignored_states=ignored_states,
                 device_mesh=device_mesh,
             ).items()
@@ -259,9 +260,13 @@ class FSDPExecutor(BaseExecutor):
             return model.no_sync()
         return contextlib.nullcontext()
 
-    def unwrap_model(self, model: nn.Module) -> nn.Module:
-        if self._original_model is not None:
-            return self._original_model
-        if isinstance(model, FSDP):
-            return model._fsdp_wrapped_module
-        return model
+    def unwrap_model(self, model: nn.Module):
+        if isinstance(model, FSDP) and self.use_distributed:
+            with FSDP.state_dict_type(
+                model,
+                StateDictType.FULL_STATE_DICT,
+                FullStateDictConfig(offload_to_cpu=True, rank0_only=False),
+            ):
+                return model.state_dict()
+
+        return model.state_dict()

astrai/parallel/setup.py

@@ -1,4 +1,5 @@
 import os
+from abc import ABC, abstractmethod
 from contextlib import contextmanager
 from functools import wraps
 from typing import Callable
@@ -30,6 +31,7 @@ def get_rank() -> int:
 def setup_parallel(
     rank: int,
     world_size: int,
+    local_rank: int,
     backend: str = "nccl",
     master_addr: str = "localhost",
     master_port: str = "29500",
@@ -41,14 +43,18 @@ def setup_parallel(
         return
 
     if world_size <= 1:
+        device_id = torch.device(device_type, local_rank)
+        os.environ["LOCAL_RANK"] = str(local_rank)
+        os.environ["WORLD_SIZE"] = "1"
+        os.environ["LOCAL_DEVICE"] = str(device_id)
         yield None
         return
 
-    device_id = torch.device(device_type, rank)
+    device_id = torch.device(device_type, local_rank)
 
     os.environ["MASTER_ADDR"] = master_addr
     os.environ["MASTER_PORT"] = master_port
-    os.environ["LOCAL_RANK"] = str(rank)
+    os.environ["LOCAL_RANK"] = str(local_rank)
     os.environ["WORLD_SIZE"] = str(world_size)
     os.environ["LOCAL_DEVICE"] = str(device_id)
 
@@ -90,7 +96,7 @@ def only_on_rank(rank, sync=False):
     return decorator
 
 
-def wrapper_spawn_func(
+def _run_single_rank(
     rank: int,
     world_size: int,
     backend: str,
@@ -100,20 +106,108 @@ def wrapper_spawn_func(
     func: Callable,
     kwargs: dict,
 ):
-    try:
+    with setup_parallel(
+        rank=rank,
+        world_size=world_size,
+        local_rank=rank,
+        backend=backend,
+        master_addr=master_addr,
+        master_port=master_port,
+        device_type=device_type,
+    ):
+        func(**kwargs)
+
+
+class LaunchStrategy(ABC):
+    """Strategy for launching a function in a distributed context."""
+
+    def __init__(
+        self,
+        world_size: int,
+        backend: str,
+        master_addr: str,
+        master_port: str,
+        device_type: str,
+        start_method: str,
+    ):
+        self.world_size = world_size
+        self.backend = backend
+        self.master_addr = master_addr
+        self.master_port = master_port
+        self.device_type = device_type
+        self.start_method = start_method
+
+    @abstractmethod
+    def launch(self, func: Callable, **kwargs):
+        raise NotImplementedError
+
+
+class TorchrunStrategy(LaunchStrategy):
+    """External orchestrator (torchrun, SLURM, K8s) — env vars pre-set."""
+
+    def launch(self, func: Callable, **kwargs):
+        rank = int(os.environ["RANK"])
+        world_size = int(os.environ["WORLD_SIZE"])
+        local_rank = int(os.environ.get("LOCAL_RANK", rank))
         with setup_parallel(
             rank=rank,
             world_size=world_size,
-            backend=backend,
-            master_addr=master_addr,
-            master_port=master_port,
-            device_type=device_type,
+            local_rank=local_rank,
+            backend=self.backend,
+            master_addr=os.environ.get("MASTER_ADDR", self.master_addr),
+            master_port=os.environ.get("MASTER_PORT", self.master_port),
+            device_type=self.device_type,
         ):
             func(**kwargs)
 
-    except Exception as e:
-        print(f"Error in rank {rank}: {e}")
-        raise
+
+class LocalStrategy(LaunchStrategy):
+    """Local launcher — single-process or mp.start_processes."""
+
+    def launch(self, func: Callable, **kwargs):
+        args = (
+            self.world_size,
+            self.backend,
+            self.master_addr,
+            self.master_port,
+            self.device_type,
+            func,
+            kwargs,
+        )
+
+        if self.world_size == 1:
+            _run_single_rank(0, *args)
+            return
+
+        ctx = mp.start_processes(
+            _run_single_rank,
+            args=args,
+            nprocs=self.world_size,
+            start_method=self.start_method,
+            join=False,
+        )
+        try:
+            while not ctx.join():
+                pass
+        except BaseException:
+            for p in ctx.processes:
+                p.terminate()
+            ctx.join()
+            raise
+
+
+def _detect_launcher() -> str:
+    """Detect the distributed launcher from environment.
+
+    Returns one of: "torchelastic", "torchrun", "external", "local".
+    """
+    if dist.is_torchelastic_launched():
+        return "torchelastic"
+    if "LOCAL_WORLD_SIZE" in os.environ:
+        return "torchrun"
+    if "RANK" in os.environ and "WORLD_SIZE" in os.environ:
+        return "external"
+    return "local"
 
 
 def spawn_parallel_fn(
@@ -126,41 +220,13 @@ def spawn_parallel_fn(
     start_method: str = "spawn",
     **kwargs,
 ):
-    # clear environment variables
-    for key in [
-        "MASTER_ADDR",
-        "MASTER_PORT",
-        "RANK",
-        "WORLD_SIZE",
-        "LOCAL_RANK",
-        "LOCAL_DEVICE",
-    ]:
-        if key in os.environ:
-            del os.environ[key]
-
-    if world_size == 1:
-        device_id = torch.device(device_type, 0)
-        os.environ["LOCAL_RANK"] = "0"
-        os.environ["WORLD_SIZE"] = "1"
-        os.environ["LOCAL_DEVICE"] = str(device_id)
-
-        func(**kwargs)
-        return
-
-    wrapper_spawn_func_args = (
-        world_size,
-        backend,
-        master_addr,
-        master_port,
-        device_type,
-        func,
-        kwargs,
-    )
-
-    mp.start_processes(
-        wrapper_spawn_func,
-        args=wrapper_spawn_func_args,
-        nprocs=world_size,
-        start_method=start_method,
-        join=True,
-    )
+    launcher = _detect_launcher()
+    if launcher in ("torchelastic", "torchrun", "external"):
+        strategy = TorchrunStrategy(
+            world_size, backend, master_addr, master_port, device_type, start_method
+        )
+    else:
+        strategy = LocalStrategy(
+            world_size, backend, master_addr, master_port, device_type, start_method
+        )
+    strategy.launch(func, **kwargs)

astrai/preprocessing/__init__.py

@@ -0,0 +1,14 @@
+from astrai.preprocessing.builder import (
+    BaseMaskBuilder,
+    MaskBuilderFactory,
+    SectionedMaskBuilder,
+)
+from astrai.preprocessing.pipeline import Pipeline, filter_by_length
+
+__all__ = [
+    "BaseMaskBuilder",
+    "MaskBuilderFactory",
+    "SectionedMaskBuilder",
+    "Pipeline",
+    "filter_by_length",
+]

astrai/preprocessing/builder.py

@@ -0,0 +1,338 @@
+"""Mask building strategies for preprocessing pipeline.
+
+The single :class:`SectionedMaskBuilder` handles all input formats
+(single-sequence / DPO / GRPO) via declarative config: ``input.sections``
+for single-output or ``input.sources`` for multi-output.
+"""
+
+from abc import ABC, abstractmethod
+from typing import Optional
+
+from astrai.factory import BaseFactory
+
+
+class BaseMaskBuilder(ABC):
+    """Convert a JSONL item into token ids and optional loss_mask."""
+
+    @abstractmethod
+    def build(self, item: dict, config, tokenizer) -> Optional[dict]:
+        """Build ``{ids, loss_mask?, domain}`` from a JSONL record.
+
+        Returns ``None`` to skip the item entirely.
+        """
+        ...
+
+
+class MaskBuilderFactory(BaseFactory["BaseMaskBuilder"]):
+    @classmethod
+    def _validate_component(cls, component_cls: type):
+        if not issubclass(component_cls, BaseMaskBuilder):
+            raise TypeError(
+                f"{component_cls.__name__} must inherit from BaseMaskBuilder"
+            )
+
+
+def _extract_domain(item: dict, domain_key: Optional[str]) -> str:
+    if not domain_key:
+        return "__default__"
+    val = item.get(domain_key, "__default__")
+    return val if isinstance(val, str) else "__default__"
+
+
+def _resolve_action(action: str, role: str, config) -> str:
+    """Resolve action to "train" or "mask".
+
+    - ``"train"`` / ``"mask"`` → literal
+    - ``"$role"`` → look up ``role`` in ``config.mask``, fall back to ``config.mask_default``
+    """
+    if action == "$role":
+        return config.mask.get(role, config.mask_default)
+    return action
+
+
+@MaskBuilderFactory.register("sectioned")
+class SectionedMaskBuilder(BaseMaskBuilder):
+    """Config-driven builder supporting single and multi-output modes.
+
+    Single-output (backward-compatible)::
+
+        {"input": {"sections": [
+            {"field": "messages", "action": "$role", "template": true}
+        ]}}
+        → {"sequence": [...], "loss_mask": [...], "domain": "..."}
+
+    Multi-output (DPO / GRPO)::
+
+        {"input": {"sources": {
+            "chosen": {"sections": [
+                {"field": "chosen", "action": "$role", "template": true}
+            ]},
+            "rejected": {"sections": [
+                {"field": "rejected", "action": "$role", "template": true}
+            ]}
+        }}}
+        → {"chosen": [...], "chosen_mask": [...],
+           "rejected": [...], "rejected_mask": [...], "domain": "..."}
+
+    Output spec fields::
+
+        sections      – list of section specs (same format as single-output)
+        list_field    – True when the JSONL field holds a list of values to
+                        tokenise individually and concatenate (GRPO responses)
+        mask_key      – explicit output key for the loss mask
+                        (default: ``"{output_key}_mask"``)
+        dtype         – explicit tensor dtype for this output key
+                        (default: "int32")
+    """
+
+    def build(self, item: dict, config, tokenizer) -> Optional[dict]:
+        sources_spec = getattr(config.input, "sources", None)
+        if sources_spec:
+            return self._build_multi(item, sources_spec, config, tokenizer)
+        return self._build_single(item, config, tokenizer)
+
+    def _build_single(self, item: dict, config, tokenizer) -> Optional[dict]:
+        sections = config.input.sections
+        if not sections:
+            return None
+
+        ids, mask = self._process_sections(
+            item, sections, config, tokenizer, is_top_level=True
+        )
+        if ids is None:
+            return None
+
+        result: dict = {
+            "sequence": ids,
+            "domain": _extract_domain(item, config.output.domain_key),
+        }
+        if not all(m == 1 for m in mask):
+            result["loss_mask"] = mask
+        return result
+
+    def _build_multi(
+        self, item: dict, sources_spec: dict, config, tokenizer
+    ) -> Optional[dict]:
+        result: dict = {}
+        any_output = False
+
+        for output_key, spec in sources_spec.items():
+            sections = spec.get("sections", [])
+            if not sections:
+                continue
+
+            if self._is_value_section(sections):
+                ids = self._extract_raw_value(item, sections)
+                if ids is None:
+                    continue
+                result[output_key] = ids
+                any_output = True
+                continue
+
+            list_field = spec.get("list_field", False)
+            mask_key = spec.get("mask_key", f"{output_key}_mask")
+
+            if list_field:
+                ids, mask = self._process_list_field(item, sections, config, tokenizer)
+            else:
+                ids, mask = self._process_sections(
+                    item, sections, config, tokenizer, is_top_level=True
+                )
+
+            if ids is None:
+                continue
+
+            result[output_key] = ids
+            if not all(m == 1 for m in mask):
+                result[mask_key] = mask
+            elif "mask_key" in spec:
+                result[mask_key] = mask
+
+            any_output = True
+
+        if not any_output:
+            return None
+
+        result["domain"] = _extract_domain(item, config.output.domain_key)
+        return result
+
+    @staticmethod
+    def _is_value_section(sections: list) -> bool:
+        return len(sections) == 1 and sections[0].get("action") == "value"
+
+    @staticmethod
+    def _extract_raw_value(item: dict, sections: list):
+        """Extract a raw value from a JSONL field without tokenisation.
+
+        Used for GRPO rewards where the field contains float values.
+        """
+        sec = sections[0]
+        field = sec["field"]
+        raw = item.get(field)
+        if raw is None:
+            return None
+        if isinstance(raw, list):
+            return [float(v) for v in raw]
+        return [float(raw)]
+
+    def _process_sections(
+        self,
+        item: dict,
+        sections: list,
+        config,
+        tokenizer,
+        *,
+        is_top_level: bool = False,
+    ):
+        """Process a list of sections into ``(ids, loss_mask)``.
+
+        Returns ``(None, None)`` if the item should be skipped.
+        """
+        all_ids: list[int] = []
+        loss_mask: list[int] = []
+
+        has_template = any(s.get("template") for s in sections)
+        is_text_config = not has_template and all(
+            s["action"] == "train" for s in sections
+        )
+
+        if is_top_level and has_template and tokenizer.bos_token_id is not None:
+            all_ids.append(tokenizer.bos_token_id)
+            loss_mask.append(0)
+
+        first_section = True
+        for sec in sections:
+            field = sec["field"]
+            action = sec["action"]
+            use_template = sec.get("template", False)
+            add_special = sec.get(
+                "add_special_tokens", not use_template and first_section
+            )
+
+            if use_template:
+                success = self._append_template_section(
+                    item, field, action, tokenizer, config, all_ids, loss_mask
+                )
+                if not success:
+                    continue
+            else:
+                success = self._append_text_section(
+                    item,
+                    field,
+                    action,
+                    tokenizer,
+                    add_special,
+                    is_text_config,
+                    config,
+                    all_ids,
+                    loss_mask,
+                )
+                if not success:
+                    continue
+
+            first_section = False
+
+        max_len = config.preprocessing.max_seq_len
+        all_ids = all_ids[:max_len]
+        loss_mask = loss_mask[: len(all_ids)]
+
+        if not all_ids:
+            return None, None
+
+        if is_top_level and has_template and len(all_ids) <= 1:
+            return None, None
+
+        return all_ids, loss_mask
+
+    def _append_template_section(
+        self, item, field, action, tokenizer, config, all_ids, loss_mask
+    ):
+        messages = item.get(field)
+        if not isinstance(messages, list) or not messages:
+            return False
+        for msg in messages:
+            role = msg.get("role", "")
+            act = _resolve_action(action, role, config)
+            rendered = tokenizer.apply_chat_template(
+                [msg], tokenize=False, add_generation_prompt=False
+            )
+            ids = tokenizer.encode(rendered, add_special_tokens=False)
+            all_ids.extend(ids)
+            val = 1 if act == "train" else 0
+            loss_mask.extend([val] * len(ids))
+        return True
+
+    def _append_text_section(
+        self,
+        item,
+        field,
+        action,
+        tokenizer,
+        add_special,
+        is_text_config,
+        config,
+        all_ids,
+        loss_mask,
+    ):
+        text = str(item.get(field, ""))
+        if not text.strip():
+            return False
+        if is_text_config:
+            pp = config.preprocessing
+            if pp.min_chars > 0 and len(text) < pp.min_chars:
+                return False
+            if len(text) > pp.max_chars:
+                return False
+        ids = tokenizer.encode(text, add_special_tokens=add_special)
+        all_ids.extend(ids)
+        val = 1 if action == "train" else 0
+        loss_mask.extend([val] * len(ids))
+        return True
+
+    def _process_list_field(self, item: dict, sections: list, config, tokenizer):
+        all_ids: list[int] = []
+        loss_mask: list[int] = []
+
+        for sec in sections:
+            field = sec["field"]
+            action = sec["action"]
+            use_template = sec.get("template", False)
+
+            values = item.get(field)
+            if not isinstance(values, list):
+                continue
+
+            for val in values:
+                if use_template:
+                    if isinstance(val, list):
+                        wrapper = {field: val}
+                        self._append_template_section(
+                            wrapper,
+                            field,
+                            action,
+                            tokenizer,
+                            config,
+                            all_ids,
+                            loss_mask,
+                        )
+                else:
+                    wrapper = {field: str(val)}
+                    self._append_text_section(
+                        wrapper,
+                        field,
+                        action,
+                        tokenizer,
+                        False,
+                        False,
+                        config,
+                        all_ids,
+                        loss_mask,
+                    )
+
+        max_len = config.preprocessing.max_seq_len
+        all_ids = all_ids[:max_len]
+        loss_mask = loss_mask[: len(all_ids)]
+
+        if not all_ids:
+            return None, None
+        return all_ids, loss_mask

astrai/preprocessing/pipeline.py

@@ -0,0 +1,257 @@
+"""Config-driven JSONL preprocessing pipeline.
+
+Composes a :class:`BaseMaskBuilder` (selected by ``input.type``) with
+sharding and flush to ``.h5`` / ``.bin`` storage.
+"""
+
+import json
+import os
+from collections import defaultdict
+from itertools import chain
+from typing import List, Optional, Tuple
+
+import torch
+import tqdm
+
+from astrai.config.preprocess_config import PipelineConfig
+from astrai.dataset.storage import save_bin, save_h5
+from astrai.preprocessing.builder import SectionedMaskBuilder
+from astrai.tokenize import AutoTokenizer
+
+_STR_TO_DTYPE: dict[str, torch.dtype] = {
+    "bool": torch.bool,
+    "uint8": torch.uint8,
+    "int8": torch.int8,
+    "int16": torch.int16,
+    "int32": torch.int32,
+    "int64": torch.int64,
+    "float16": torch.float16,
+    "float32": torch.float32,
+    "float64": torch.float64,
+}
+
+
+def filter_by_length(text: str, min_len: int = 50, max_len: int = 2_000_000) -> bool:
+    return min_len <= len(text) <= max_len
+
+
+def _truncate(seq: list, max_len: int, mode: str) -> list:
+    if len(seq) <= max_len:
+        return seq
+    if mode == "keep_end":
+        return seq[-max_len:]
+    return seq[:max_len]
+
+
+def pack_sequences(
+    sequences: List[list],
+    max_packed_len: int,
+    strategy: str,
+    truncation_mode: str,
+) -> List[Tuple[int, int]]:
+    """Pack *sequences* into bins and return a reorder plan.
+
+    Returns a list of ``(orig_idx, truncated_length)`` in flush order.
+    All keys (sequence, loss_mask, …) must be reordered and truncated
+    identically according to this plan.
+
+    Supported *strategy* values:
+
+    - ``"simple"``: sequential, no reordering.
+    - ``"bfd"``: best-fit decreasing bin packing.
+    """
+    n = len(sequences)
+    if strategy == "simple":
+        return [(i, min(len(sequences[i]), max_packed_len)) for i in range(n)]
+
+    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 = min(len(sequences[orig_idx]), max_packed_len)
+
+        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)
+
+    plan: List[Tuple[int, int]] = []
+    for bin_indices in bins:
+        for orig_idx in bin_indices:
+            plan.append((orig_idx, min(len(sequences[orig_idx]), max_packed_len)))
+
+    return plan
+
+
+class Pipeline:
+    """Tokenization pipeline driven by a declarative :class:`PipelineConfig`.
+
+    Usage::
+
+        config = PipelineConfig.from_json("sft_pipeline.json")
+        Pipeline(config, ["data.jsonl"], output_dir="out", tokenizer_path="params").run()
+    """
+
+    def __init__(
+        self,
+        config: PipelineConfig,
+        input_paths: list[str],
+        output_dir: str,
+        tokenizer_path: str,
+    ):
+        os.makedirs(output_dir, exist_ok=True)
+        self.config = config
+        self.paths = input_paths
+        self.output_dir = output_dir
+        self.tokenizer_path = tokenizer_path
+
+        self.mask_builder = SectionedMaskBuilder()
+
+    def transform(self, item: dict) -> Optional[dict]:
+        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)
+        count = 0
+
+        pp = self.config.preprocessing
+
+        for item in tqdm.tqdm(
+            self._iter_items(), desc="Tokenizing", unit="docs", mininterval=0.5
+        ):
+            if pp.max_items and count >= pp.max_items:
+                break
+
+            result = self.transform(item)
+            if result is None:
+                continue
+
+            domain = result.pop("domain", "__default__")
+
+            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
+
+            bucket = domains[domain]
+            self._align_bucket(bucket, result, ids, is_multi)
+            for key, val in result.items():
+                bucket[key].append(val)
+
+            count += 1
+            total_tokens += len(ids)
+
+            if total_tokens >= self.config.output.max_tokens_per_shard:
+                self._flush(domains, shard_idx)
+                domains.clear()
+                total_tokens = 0
+
+        if total_tokens > 0:
+            self._flush(domains, shard_idx)
+
+        print(f"Done. {count} documents tokenized.")
+
+    @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, is_multi: bool):
+        """Pad previously-accumulated keys that are missing from *result*."""
+        for key in list(bucket.keys()):
+            if key in result:
+                continue
+            if is_multi:
+                pad = bucket[key][-1] if bucket[key] else [1] * len(ids)
+                bucket[key].append(pad)
+            else:
+                bucket[key].append([1] * len(ids))
+
+    def _iter_items(self):
+        for path in self.paths:
+            with open(path, "r", encoding="utf-8") as f:
+                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():
+            idx = shard_idx[domain]
+            chunk_dir = os.path.join(self.output_dir, domain)
+
+            pp = self.config.preprocessing
+            if pp.packing_strategy != "simple" and "sequence" in keys:
+                plan = pack_sequences(
+                    keys["sequence"],
+                    pp.max_packed_len,
+                    pp.packing_strategy,
+                    pp.truncation_mode,
+                )
+                reordered = defaultdict(list)
+                for orig_idx, truncated_len in plan:
+                    for k, vals in keys.items():
+                        reordered[k].append(
+                            _truncate(
+                                vals[orig_idx], pp.max_packed_len, pp.truncation_mode
+                            )
+                        )
+                keys = reordered
+
+            tensors = {}
+            for key, ids_list in keys.items():
+                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)
+                ]
+
+            pid_mode = self.config.output.position_ids_mode
+            if pid_mode and pid_mode != "none" and "sequence" in tensors:
+                pos_ids = []
+                if pid_mode == "doc_reset":
+                    for item in keys["sequence"]:
+                        pos_ids.extend(range(len(item)))
+                else:
+                    total = sum(len(item) for item in keys["sequence"])
+                    pos_ids = list(range(total))
+                tensors["position_ids"] = [torch.tensor(pos_ids, dtype=torch.int32)]
+
+            shard_path = os.path.join(chunk_dir, f"shard_{idx:04d}")
+            fmt = self.config.output.storage_format
+            if fmt == "bin":
+                save_bin(shard_path, tensors)
+            else:
+                save_h5(chunk_dir, f"data_{idx:04d}", tensors)
+            shard_idx[domain] = idx + 1
+            first_key = "sequence" if "sequence" in tensors else next(iter(tensors))
+            tqdm.tqdm.write(
+                f"  saved {domain}/shard_{idx:04d}  "
+                f"({tensors[first_key][0].numel():,} tokens)"
+            )

astrai/tokenize/chat_template.py

@@ -1,13 +1,10 @@
-from dataclasses import dataclass
 from typing import Any, Dict, List, Optional
 
 from jinja2 import Template
 
-# Message type for chat messages
 type MessageType = Dict[str, Any]
 
 
-@dataclass
 class ChatTemplate:
     """A chat template with Jinja2 rendering support.
 
@@ -15,23 +12,24 @@ class ChatTemplate:
         name: Unique identifier for the template.
         template_str: Jinja2 template string.
         description: Optional description.
-        default_variables: Optional dictionary of default variable values
-            that will be passed to the template if not overridden during rendering.
+        default_variables: Optional dictionary of default variable values.
         special_tokens: Optional dictionary mapping token names to their string values.
-            These tokens are automatically added to the template variables.
     """
 
-    name: str
-    template_str: str
-    description: str = ""
-    default_variables: Dict[str, Any] = None
-    special_tokens: Dict[str, str] = None
-
-    def __post_init__(self):
-        if self.default_variables is None:
-            self.default_variables = {}
-        if self.special_tokens is None:
-            self.special_tokens = {}
+    def __init__(
+        self,
+        name: str = "",
+        template_str: str = "",
+        description: str = "",
+        default_variables: Optional[Dict[str, Any]] = None,
+        special_tokens: Optional[Dict[str, str]] = None,
+    ):
+        self.name = name
+        self.template_str = template_str
+        self.description = description
+        self.default_variables = default_variables or {}
+        self.special_tokens = special_tokens or {}
+        self._compiled: Template = Template(template_str)
 
     @classmethod
     def from_string(
@@ -43,7 +41,7 @@ class ChatTemplate:
     ) -> "ChatTemplate":
         """Create a ChatTemplate instance directly from a template string."""
         return cls(
-            name="",  # empty name for ad‑hoc templates
+            name="",
             template_str=template_str,
             description=description,
             default_variables=default_variables,
@@ -73,5 +71,4 @@ class ChatTemplate:
         if system_prompt is not None:
             variables["system_prompt"] = system_prompt
 
-        jinja_template = Template(self.template_str)
-        return jinja_template.render(**variables)
+        return self._compiled.render(**variables)

astrai/trainer/strategy.py

@@ -1,6 +1,5 @@
 """Training strategy implementations with factory pattern."""
 
-import copy
 from abc import ABC, abstractmethod
 from typing import Any, Callable, Dict, Union
 
@@ -8,28 +7,14 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from torch import Tensor
-from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
-from torch.nn.parallel import DistributedDataParallel as DDP
 
 from astrai.factory import BaseFactory
 
 
-def unwrap_model(model: nn.Module) -> nn.Module:
-    if isinstance(model, DDP):
-        return model.module
-    if isinstance(model, FSDP):
-        return model._fsdp_wrapped_module
-    return model
-
-
-def create_ref_model(model: nn.Module) -> nn.Module:
-    """Create a reference model for DPO/GRPO training.
-
-    Handles DDP-wrapped models safely by unwrapping first,
-    then creating a deep copy with frozen gradients.
-    """
-    original_model = unwrap_model(model)
-    ref_model = copy.deepcopy(original_model)
+def create_ref_model(model_fn, state_dict: dict) -> nn.Module:
+    """Create a frozen reference model from model_fn + full state dict."""
+    ref_model = model_fn()
+    ref_model.load_state_dict(state_dict)
     ref_model.requires_grad_(False)
     ref_model.eval()
     return ref_model
@@ -83,6 +68,22 @@ def get_logprobs(
         return token_logprobs * shifted_mask
 
 
+def make_doc_boundary_mask(position_ids: Tensor) -> Tensor:
+    S = position_ids.size(1)
+    device = position_ids.device
+    boundaries = position_ids[:, 1:] <= position_ids[:, :-1]
+    doc_ids = torch.cat(
+        [
+            torch.zeros(position_ids.size(0), 1, dtype=torch.long, device=device),
+            boundaries.long().cumsum(dim=1),
+        ],
+        dim=1,
+    )
+    same_doc = doc_ids.unsqueeze(-1) == doc_ids.unsqueeze(-2)
+    causal = torch.tril(torch.ones(S, S, dtype=torch.bool, device=device))
+    return (same_doc & causal).unsqueeze(1)
+
+
 class BaseStrategy(ABC):
     """Abstract base class for training strategies."""
 
@@ -91,6 +92,8 @@ class BaseStrategy(ABC):
     ):
         self.model = model
         self.device = device
+        self.executor = kwargs.pop("executor", None)
+        self.model_fn = kwargs.pop("model_fn", None)
         self.extra_kwargs = kwargs
 
     @abstractmethod
@@ -193,15 +196,19 @@ class SFTStrategy(BaseStrategy):
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
         batch = move_to_device(batch, self.device)
-        input_ids, target_ids, loss_mask = (
+        input_ids, target_ids, position_ids, loss_mask = (
             batch["input_ids"],
             batch["target_ids"],
+            batch["position_ids"],
             batch["loss_mask"],
         )
 
         ignore_index = -100
-        logits = self.model(input_ids=input_ids)["logits"]
+        input_mask = make_doc_boundary_mask(position_ids)
         target_ids = target_ids.masked_fill(loss_mask == 0, ignore_index)
+        logits = self.model(
+            input_ids=input_ids, position_ids=position_ids, input_mask=input_mask
+        )["logits"]
 
         loss = F.cross_entropy(
             input=logits.flatten(0, 1).float(),
@@ -230,7 +237,9 @@ class DPOStrategy(BaseStrategy):
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(model)
+        self.ref_model = create_ref_model(
+            self.model_fn, self.executor.unwrap_model(model)
+        ).to(device=self.device)
         self.beta = beta
         self.reduction = reduction
 
@@ -284,7 +293,9 @@ class GRPOStrategy(BaseStrategy):
         **kwargs,
     ):
         super().__init__(model, device, **kwargs)
-        self.ref_model = create_ref_model(model)
+        self.ref_model = create_ref_model(
+            self.model_fn, self.executor.unwrap_model(model)
+        ).to(device=self.device)
         self.clip_eps = clip_eps
         self.kl_coef = kl_coef
         self.group_size = group_size
@@ -294,8 +305,7 @@ class GRPOStrategy(BaseStrategy):
 
     def sync_ref_model(self):
         """Copy current model weights to ref model."""
-        ref_state = self.model.state_dict()
-        self.ref_model.load_state_dict(ref_state)
+        self.ref_model.load_state_dict(self.executor.unwrap_model(self.model))
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
         self._step += 1

astrai/trainer/train_callback.py

@@ -146,8 +146,7 @@ class CheckpointCallback(TrainCallback):
         self.last_ckpt_iter = 0
 
     def _save_checkpoint(self, context: TrainContext):
-        unwrapped = context.executor.unwrap_model(context.model)
-        state_dict = unwrapped.state_dict()
+        state_dict = context.executor.unwrap_model(context.model)
         self.last_ckpt_iter = context.iteration
 
         if get_rank() == 0:

astrai/trainer/train_context.py

@@ -2,8 +2,9 @@ from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Optional, Self
 
+import torch
 import torch.nn as nn
-from torch.utils.data import DataLoader
+from torch.utils.data import DataLoader, random_split
 
 from astrai.config.train_config import TrainConfig
 from astrai.dataset import ResumableDistributedSampler
@@ -108,15 +109,27 @@ class TrainContextBuilder:
         context.optimizer = cfg.optimizer_fn(model)
         context.scheduler = cfg.scheduler_fn(context.optimizer)
 
+        train_dataset = cfg.dataset
+        val_dataset = cfg.val_dataset
+
+        if val_dataset is None and cfg.val_split is not None:
+            n_total = len(cfg.dataset)
+            n_val = max(1, int(n_total * cfg.val_split))
+            n_train = n_total - n_val
+            generator = torch.Generator().manual_seed(cfg.random_seed)
+            train_dataset, val_dataset = random_split(
+                cfg.dataset, [n_train, n_val], generator=generator
+            )
+
         sampler_offset = context.iteration * cfg.batch_per_device
         sampler = ResumableDistributedSampler(
-            data_source=cfg.dataset,
+            data_source=train_dataset,
             start_epoch=context.epoch,
             start_iter=sampler_offset,
             seed=cfg.random_seed,
         )
         context.dataloader = DataLoader(
-            cfg.dataset,
+            train_dataset,
             batch_size=cfg.batch_per_device,
             sampler=sampler,
             num_workers=cfg.num_workers,
@@ -124,16 +137,16 @@ class TrainContextBuilder:
             prefetch_factor=cfg.prefetch_factor,
         )
 
-        if cfg.val_dataset is not None:
+        if val_dataset is not None:
             val_sampler = ResumableDistributedSampler(
-                data_source=cfg.val_dataset,
+                data_source=val_dataset,
                 start_epoch=0,
                 start_iter=0,
                 seed=cfg.random_seed,
                 shuffle=False,
             )
             context.val_dataloader = DataLoader(
-                cfg.val_dataset,
+                val_dataset,
                 batch_size=cfg.batch_per_device,
                 sampler=val_sampler,
                 num_workers=cfg.num_workers,
@@ -162,6 +175,8 @@ class TrainContextBuilder:
             model=context.model,
             train_type=cfg.strategy,
             device=device,
+            executor=executor,
+            model_fn=cfg.model_fn,
             **cfg.extra_kwargs,
         )
 

scripts/tools/evaluate_humaneval.py

@@ -0,0 +1,336 @@
+"""HumanEval code generation benchmark.
+
+Generates n completions per problem, extracts function bodies, executes
+against hidden tests, and computes pass@k.
+
+Usage::
+
+    python scripts/tools/evaluate_humaneval.py --param_path ./params \
+        --data_path HumanEval.jsonl.gz --output results.json \
+        --num_samples 200 --temperature 0.8 --max_tokens 512
+"""
+
+import argparse
+import json
+import os
+import re
+import signal
+import sys
+from math import prod
+from multiprocessing import Process, Queue
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+import tqdm
+
+from astrai.inference import InferenceEngine
+from astrai.model import AutoModel
+from astrai.tokenize import AutoTokenizer
+
+HUMANEVAL_URL = (
+    "https://github.com/openai/human-eval/raw/master/data/HumanEval.jsonl.gz"
+)
+
+_STOP_SEQUENCES = [
+    "\nclass ",
+    "\ndef ",
+    "\n# ",
+    "\nif __name__",
+    "\nprint(",
+    "\n\n\n",
+]
+
+
+def _download_humaneval(data_path: str):
+    if os.path.exists(data_path):
+        return
+    import gzip
+    import urllib.request
+
+    os.makedirs(os.path.dirname(data_path) or ".", exist_ok=True)
+    print(f"Downloading HumanEval from {HUMANEVAL_URL} ...")
+    tmp = data_path + ".tmp"
+    urllib.request.urlretrieve(HUMANEVAL_URL, tmp)
+    with gzip.open(tmp, "rb") as f_in:
+        with open(data_path, "wb") as f_out:
+            f_out.write(f_in.read())
+    os.remove(tmp)
+    print(f"  saved to {data_path}")
+
+
+def _load_problems(data_path: str) -> List[dict]:
+    problems = []
+    with open(data_path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                problems.append(json.loads(line))
+    return problems
+
+
+def _extract_function_body(code: str, entry_point: str) -> Optional[str]:
+    """Extract the function body from a completion."""
+    pattern = rf"def\s+{re.escape(entry_point)}\b[^:]*:"
+    match = re.search(pattern, code)
+    if not match:
+        # Use the full code as-is if we can't find the function
+        return code
+
+    body_start = match.end()
+    lines = code[body_start:].split("\n")
+    body_lines = []
+    started = False
+
+    for line in lines:
+        stripped = line.rstrip()
+        if not stripped and not started:
+            continue
+        if not stripped and started:
+            body_lines.append("")
+            continue
+        if not started:
+            started = True
+        if stripped.lstrip() == stripped and started:
+            break
+        body_lines.append(stripped)
+
+    body = "\n".join(body_lines)
+    if not body.strip():
+        return None
+    return body
+
+
+def _trim_stop_sequences(text: str) -> str:
+    for stop in _STOP_SEQUENCES:
+        idx = text.find(stop)
+        if idx != -1:
+            text = text[:idx]
+    return text
+
+
+def _execute_code(problem: dict, completion: str, timeout: float = 3.0) -> bool:
+    """Run the completion against hidden tests in a subprocess."""
+
+    def _worker(queue, full_code):
+        try:
+            namespace = {}
+            exec(full_code, namespace)
+            check = namespace.get("check")
+            if check is None:
+                queue.put(False)
+                return
+            check(namespace.get(problem["entry_point"]))
+            queue.put(True)
+        except Exception:
+            queue.put(False)
+
+    full_code = problem["prompt"] + completion + "\n" + problem["test"]
+
+    queue: Queue = Queue()
+    proc = Process(target=_worker, args=(queue, full_code))
+    proc.start()
+    proc.join(timeout)
+
+    if proc.is_alive():
+        proc.terminate()
+        proc.join()
+        return False
+
+    try:
+        return queue.get_nowait()
+    except Exception:
+        return False
+
+
+def _pass_at_k(n: int, c: int, k: int) -> float:
+    """Unbiased estimator of pass@k."""
+    if n - c < k:
+        return 1.0
+    return 1.0 - float(prod(1.0 - k / np.arange(n - c + 1, n + 1)))
+
+
+def _deduplicate(completions: List[str]) -> List[str]:
+    seen = set()
+    unique = []
+    for c in completions:
+        if c not in seen:
+            seen.add(c)
+            unique.append(c)
+    return unique
+
+
+def _generate(
+    engine: InferenceEngine,
+    prompt: str,
+    num_samples: int,
+    max_tokens: int,
+    temperature: float,
+    top_p: float,
+    top_k: int,
+    batch_size: int,
+) -> List[str]:
+    batches = [prompt] * min(batch_size, num_samples)
+    completions = []
+    remaining = num_samples
+
+    while remaining > 0:
+        current = min(batch_size, remaining)
+        batch_prompts = batches[:current]
+        outputs = engine.generate(
+            prompt=batch_prompts,
+            stream=False,
+            max_tokens=max_tokens,
+            temperature=temperature,
+            top_p=top_p,
+            top_k=top_k,
+        )
+        if isinstance(outputs, str):
+            outputs = [outputs]
+        completions.extend(outputs)
+        remaining -= current
+
+    return _deduplicate(completions)
+
+
+def evaluate(
+    engine: InferenceEngine,
+    problems: List[dict],
+    num_samples: int,
+    max_tokens: int,
+    temperature: float,
+    top_p: float,
+    top_k: int,
+    batch_size: int,
+    k_values: Tuple[int, ...] = (1, 10, 100),
+) -> Dict:
+    results = {}
+    all_pass_at_k = {k: [] for k in k_values}
+
+    for problem in tqdm.tqdm(problems, desc="HumanEval", unit="problem"):
+        task_id = problem["task_id"]
+        prompt = problem["prompt"]
+        entry_point = problem["entry_point"]
+
+        raw_completions = _generate(
+            engine,
+            prompt,
+            num_samples,
+            max_tokens,
+            temperature,
+            top_p,
+            top_k,
+            batch_size,
+        )
+
+        completions = []
+        for raw in raw_completions:
+            trimmed = _trim_stop_sequences(raw)
+            body = _extract_function_body(trimmed, entry_point)
+            if body:
+                completions.append(body)
+
+        passed = 0
+        for comp in completions:
+            if _execute_code(problem, comp):
+                passed += 1
+
+        n = len(completions)
+        c = passed
+        result = {"task_id": task_id, "n": n, "passed": c}
+        for k in k_values:
+            result[f"pass@{k}"] = round(_pass_at_k(n, c, k), 4)
+            all_pass_at_k[k].append(_pass_at_k(n, c, k))
+        results[task_id] = result
+
+    summary = {}
+    for k in k_values:
+        vals = all_pass_at_k[k]
+        summary[f"pass@{k}"] = round(float(np.mean(vals)), 4)
+    results["_summary"] = summary
+
+    return results
+
+
+def main():
+    parser = argparse.ArgumentParser(description="HumanEval benchmark")
+    parser.add_argument(
+        "--param_path", type=str, default="./params", help="Model directory"
+    )
+    parser.add_argument(
+        "--data_path",
+        type=str,
+        default="./humaneval/HumanEval.jsonl",
+        help="HumanEval JSONL file (auto-download if missing)",
+    )
+    parser.add_argument("--output", type=str, default=None, help="Output JSON path")
+    parser.add_argument(
+        "--num_samples",
+        type=int,
+        default=200,
+        help="Completions per problem",
+    )
+    parser.add_argument(
+        "--max_tokens", type=int, default=512, help="Max generation tokens"
+    )
+    parser.add_argument(
+        "--temperature", type=float, default=0.8, help="Sampling temperature"
+    )
+    parser.add_argument("--top_p", type=float, default=0.95, help="Top-p sampling")
+    parser.add_argument("--top_k", type=int, default=50, help="Top-k sampling")
+    parser.add_argument(
+        "--batch_size", type=int, default=1, help="Inference batch size"
+    )
+    parser.add_argument(
+        "--problems",
+        type=int,
+        nargs="+",
+        default=None,
+        help="Specific problem indices (0-based)",
+    )
+    args = parser.parse_args()
+
+    _download_humaneval(args.data_path)
+    problems = _load_problems(args.data_path)
+    if args.problems:
+        problems = [problems[i] for i in args.problems if i < len(problems)]
+
+    model = AutoModel.from_pretrained(args.param_path)
+    tokenizer = AutoTokenizer.from_pretrained(args.param_path)
+    model.to(device="cuda", dtype=torch.bfloat16)
+
+    engine = InferenceEngine(
+        model=model,
+        tokenizer=tokenizer,
+        max_batch_size=args.batch_size,
+    )
+
+    results = evaluate(
+        engine=engine,
+        problems=problems,
+        num_samples=args.num_samples,
+        max_tokens=args.max_tokens,
+        temperature=args.temperature,
+        top_p=args.top_p,
+        top_k=args.top_k,
+        batch_size=args.batch_size,
+        k_values=(1, 10, 100),
+    )
+
+    summary = results.pop("_summary")
+    print(f"\n{'=' * 60}")
+    for k, v in summary.items():
+        print(f"  {k}: {v:.2%}")
+    print(f"{'=' * 60}")
+
+    if args.output:
+        results["_summary"] = summary
+        with open(args.output, "w", encoding="utf-8") as f:
+            json.dump(results, f, indent=2, ensure_ascii=False)
+        print(f"Results saved to {args.output}")
+
+    engine.shutdown()
+
+
+if __name__ == "__main__":
+    main()

scripts/tools/evaluate_mmlu.py

@@ -5,9 +5,9 @@ import csv
 import json
 import os
 import shutil
-import urllib.request
-import zipfile
+import tarfile
 
+import requests
 import torch
 import torch.nn.functional as F
 import tqdm
@@ -15,7 +15,7 @@ import tqdm
 from astrai.model import AutoModel
 from astrai.tokenize import AutoTokenizer
 
-MMLU_URL = "https://github.com/hendrycks/test/archive/refs/heads/master.zip"
+MMLU_URL = "https://people.eecs.berkeley.edu/~hendrycks/data.tar"
 MMLU_SUBJECTS = [
     "abstract_algebra",
     "anatomy",
@@ -78,23 +78,37 @@ MMLU_SUBJECTS = [
 
 
 def _download_and_extract(url: str, data_dir: str):
-    zip_path = os.path.join(data_dir, "mmlu.zip")
+    tar_path = os.path.join(data_dir, "data.tar")
     os.makedirs(data_dir, exist_ok=True)
     print(f"Downloading MMLU data from {url}...")
-    urllib.request.urlretrieve(url, zip_path)
+    resp = requests.get(url, stream=True, timeout=300)
+    resp.raise_for_status()
+    total = int(resp.headers.get("content-length", 0))
+    with tqdm.tqdm(total=total, unit="B", unit_scale=True, desc="  Download") as bar:
+        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 zipfile.ZipFile(zip_path, "r") as zf:
-        zf.extractall(data_dir)
-    os.remove(zip_path)
+    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)
-    src = os.path.join(data_dir, "test-master", "data")
+    src = os.path.join(data_dir, "data")
     if os.path.exists(src):
         for item in os.listdir(src):
-            os.rename(os.path.join(src, item), os.path.join(data_dir, item))
-        shutil.rmtree(os.path.join(data_dir, "test-master"))
+            src_item = os.path.join(src, item)
+            dst_item = os.path.join(data_dir, item)
+            if os.path.exists(dst_item):
+                if os.path.isdir(dst_item):
+                    shutil.rmtree(dst_item)
+                else:
+                    os.remove(dst_item)
+            os.rename(src_item, dst_item)
+        os.rmdir(src)
     print(f"MMLU data saved to {data_dir}")
 
 
@@ -143,10 +157,32 @@ def build_prompt(
     return prompt
 
 
+def apply_chat(
+    tokenizer, raw_prompt: str, n_shot: int, dev_data: list[dict] | None
+) -> 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.
+    """
+    messages = []
+    if n_shot > 0 and dev_data:
+        for item in dev_data[:n_shot]:
+            q = f"Question: {item['question']}\n"
+            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})
+    return tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+
+
 def choice_logprob(
     model, tokenizer, context_ids: list[int], choice_letter: str, device: str
 ) -> float:
-    choice_text = f" {choice_letter}"
+    choice_text = choice_letter
     choice_ids = tokenizer.encode(choice_text, add_special_tokens=False)
     input_ids = context_ids + choice_ids
     max_len = model.config.max_len
@@ -182,8 +218,11 @@ def evaluate_subject(
     correct = 0
     total = 0
     for item in tqdm.tqdm(test_data, desc=f"{subject:40s}", leave=False):
-        prompt = build_prompt(item["question"], item, subject, n_shot, dev_data or [])
-        context_ids = tokenizer.encode(prompt)
+        raw_prompt = build_prompt(
+            item["question"], item, subject, n_shot, dev_data or []
+        )
+        context = apply_chat(tokenizer, raw_prompt, n_shot, dev_data or [])
+        context_ids = tokenizer.encode(context)
         scores = {
             c: choice_logprob(model, tokenizer, context_ids, c, device)
             for c in ("A", "B", "C", "D")
@@ -233,6 +272,7 @@ def main():
     device = args.device
     dtype = getattr(torch, args.dtype)
     model.to(device=device, dtype=dtype)
+    model.eval()
 
     subjects = args.subjects or MMLU_SUBJECTS
     results = {}

scripts/tools/preprocess.py

@@ -0,0 +1,38 @@
+"""CLI: JSONL → tokenized .h5/.bin via config-driven Pipeline."""
+
+import argparse
+
+from astrai.config.preprocess_config import PipelineConfig
+from astrai.preprocessing.pipeline import Pipeline
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Raw JSONL → tokenized .h5/.bin via config-driven Pipeline"
+    )
+    parser.add_argument(
+        "inputs", nargs="+", metavar="JSONL", help="One or more JSONL files"
+    )
+    parser.add_argument("--output_dir", "-o", required=True, help="Output directory")
+    parser.add_argument(
+        "--config", "-c", required=True, help="Path to pipeline config JSON"
+    )
+    parser.add_argument(
+        "--tokenizer_path",
+        default="params",
+        help="Path to tokenizer directory (default: params)",
+    )
+    args = parser.parse_args()
+
+    config = PipelineConfig.from_json(args.config)
+
+    Pipeline(
+        config=config,
+        input_paths=args.inputs,
+        output_dir=args.output_dir,
+        tokenizer_path=args.tokenizer_path,
+    ).run()
+
+
+if __name__ == "__main__":
+    main()

scripts/tools/train.py

@@ -8,6 +8,7 @@ import torch.optim as optim
 from astrai.config import AutoRegressiveLMConfig, TrainConfig
 from astrai.dataset import DatasetFactory
 from astrai.model import AutoRegressiveLM
+from astrai.model.components.decoder_block import DecoderBlock
 from astrai.trainer import SchedulerFactory, Trainer
 
 
@@ -115,6 +116,12 @@ def parse_args() -> argparse.Namespace:
         default=0.05,
         help="cross_entropy function label smoothing parameter",
     )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action=argparse.BooleanOptionalAction,
+        default=False,
+        help="Enable activation checkpointing for DecoderBlock modules.",
+    )
 
     parser.add_argument(
         "--ckpt_interval",
@@ -128,6 +135,36 @@ def parse_args() -> argparse.Namespace:
         default="checkpoint",
         help="Directory to save checkpoints.",
     )
+    parser.add_argument(
+        "--val_split",
+        type=float,
+        default=None,
+        help="Ratio to split from training dataset for validation (e.g. 0.05).",
+    )
+    parser.add_argument(
+        "--val_step",
+        type=int,
+        default=1000,
+        help="Number of optimizer steps between validation runs.",
+    )
+    parser.add_argument(
+        "--metrics",
+        nargs="*",
+        default=["loss", "lr"],
+        help="Metrics to log (e.g. --metrics loss lr val_loss). Default: loss lr.",
+    )
+    parser.add_argument(
+        "--log_dir",
+        type=str,
+        default="checkpoint/logs",
+        help="Directory for metric logs.",
+    )
+    parser.add_argument(
+        "--log_interval",
+        type=int,
+        default=100,
+        help="Number of batch iterations between metric logs.",
+    )
     parser.add_argument(
         "--grpo_sync_interval",
         type=int,
@@ -141,6 +178,24 @@ def parse_args() -> argparse.Namespace:
         "--start_batch", type=int, default=0, help="Start batch for training."
     )
 
+    parser.add_argument(
+        "--master_addr",
+        type=str,
+        default="localhost",
+        help="Master node address for distributed training.",
+    )
+    parser.add_argument(
+        "--master_port",
+        type=str,
+        default="29500",
+        help="Master node port for distributed training.",
+    )
+    parser.add_argument(
+        "--backend",
+        type=str,
+        default="nccl",
+        help="Distributed training backend.",
+    )
     parser.add_argument("--nprocs", type=int, default=1, help="Number of GPUs to use.")
     parser.add_argument(
         "--parallel_mode",
@@ -209,6 +264,11 @@ def train(
     warmup_ratio: float,
     ckpt_interval: int,
     ckpt_dir: str,
+    val_split: float,
+    val_step: int,
+    metrics: list[str],
+    log_dir: str,
+    log_interval: int,
     dpo_beta: float,
     grpo_clip_eps: float,
     grpo_kl_coef: float,
@@ -222,11 +282,15 @@ def train(
     random_seed: int,
     num_workers: int,
     pin_memory: bool,
+    gradient_checkpointing: bool,
     window_size: int,
     stride: int,
     nprocs: int,
     parallel_mode: str,
     device_type: str,
+    backend: str,
+    master_addr: str,
+    master_port: str,
     start_method: str,
 ):
     assert train_type in ["seq", "sft", "dpo", "grpo"]
@@ -286,6 +350,8 @@ def train(
         },
     )
 
+    grad_ckpt_modules = [DecoderBlock] if gradient_checkpointing else []
+
     train_config = TrainConfig(
         model_fn=model_fn,
         strategy=train_type,
@@ -304,9 +370,18 @@ def train(
         num_workers=num_workers,
         pin_memory=pin_memory,
         nprocs=nprocs,
+        backend=backend,
+        master_addr=master_addr,
+        master_port=master_port,
         parallel_mode=parallel_mode,
         device_type=device_type,
         start_method=start_method,
+        val_split=val_split,
+        val_step=val_step,
+        metrics=metrics,
+        log_dir=log_dir,
+        log_interval=log_interval,
+        gradient_checkpointing_modules=grad_ckpt_modules,
         executor_kwargs=executor_kwargs,
         extra_kwargs=strategy_kwargs,
     )

tests/data/conftest.py

@@ -0,0 +1,202 @@
+import tempfile
+
+import pytest
+from tokenizers import Tokenizer, models, pre_tokenizers, trainers
+
+from astrai.config.preprocess_config import (
+    InputConfig,
+    PipelineConfig,
+    ProcessingConfig,
+)
+from astrai.tokenize import AutoTokenizer
+
+_SPECIAL_TOKENS_CONFIG = {
+    "bos_token": "<|begin_of_sentence|>",
+    "eos_token": "<|end_of_sentence|>",
+    "pad_token": "<|_pad_|>",
+    "unk_token": "<|_unk_|>",
+    "im_start": "<|im_start|>",
+    "im_end": "<|im_end|>",
+}
+
+_SPECIAL_TOKENS = list(_SPECIAL_TOKENS_CONFIG.values())
+
+_CHAT_TEMPLATE = (
+    "{% for message in messages %}"
+    "{% if message['role'] == 'system' %}"
+    "<|im_start|>system\n{{ message['content'] }}<|im_end|>\n"
+    "{% elif message['role'] == 'user' %}"
+    "<|im_start|>user\n{{ message['content'] }}<|im_end|>\n"
+    "{% elif message['role'] == 'assistant' %}"
+    "<|im_start|>assistant\n{{ message['content'] }}<|im_end|>\n"
+    "{% endif %}"
+    "{% endfor %}"
+    "{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+)
+
+_CHAT_SECTIONS = [{"field": "messages", "action": "$role", "template": True}]
+
+_INSTRUCTION_SECTIONS = [
+    {"field": "prompt", "action": "mask", "add_special_tokens": True},
+    {"field": "response", "action": "train"},
+]
+
+_TEXT_SECTIONS = [{"field": "text", "action": "train"}]
+
+_GRPO_RESPONSE_SECTIONS = [{"field": "responses", "action": "train"}]
+
+
+def _build_chat_tokenizer():
+    tok = Tokenizer(models.BPE())
+    tok.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
+    tr = trainers.BpeTrainer(
+        vocab_size=512,
+        min_frequency=1,
+        special_tokens=_SPECIAL_TOKENS,
+    )
+    train_data = [
+        "hello world",
+        "Hi there!",
+        "You are helpful.",
+        "What is 2+2?",
+        "Tell me a story about dragons and knights.",
+        "Sure, here is a tale.",
+        "Translate to French: Hello",
+        "Bonjour",
+        "Artificial Intelligence is a field of computer science.",
+        "system",
+        "user",
+        "assistant",
+        "<|im_start|>",
+        "<|im_end|>",
+        *[chr(i) for i in range(32, 127)],
+    ]
+    tok.train_from_iterator(train_data, tr)
+
+    auto_tok = AutoTokenizer()
+    auto_tok._tokenizer = tok
+    auto_tok._special_token_map = {
+        "bos_token": "<|begin_of_sentence|>",
+        "eos_token": "<|end_of_sentence|>",
+        "pad_token": "<|_pad_|>",
+        "unk_token": "<|_unk_|>",
+    }
+    auto_tok.set_chat_template(_CHAT_TEMPLATE)
+    return auto_tok
+
+
+@pytest.fixture(scope="session")
+def chat_tokenizer():
+    return _build_chat_tokenizer()
+
+
+@pytest.fixture
+def temp_dir():
+    d = tempfile.mkdtemp()
+    yield d
+    import shutil
+
+    shutil.rmtree(d, ignore_errors=True)
+
+
+def make_chat_config():
+    return PipelineConfig(
+        input=InputConfig(sections=_CHAT_SECTIONS),
+        mask={"system": "mask", "user": "mask", "assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+
+
+def make_instruction_config():
+    return PipelineConfig(
+        input=InputConfig(sections=_INSTRUCTION_SECTIONS),
+        mask={"prompt": "mask", "response": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+
+
+def make_text_config():
+    return PipelineConfig(
+        input=InputConfig(sections=_TEXT_SECTIONS),
+        preprocessing=ProcessingConfig(
+            max_seq_len=2048, min_chars=1, max_chars=2_000_000
+        ),
+    )
+
+
+def make_dpo_chat_config():
+    return PipelineConfig(
+        input=InputConfig(
+            sources={
+                "chosen": {
+                    "sections": [
+                        {"field": "chosen", "action": "$role", "template": True}
+                    ]
+                },
+                "rejected": {
+                    "sections": [
+                        {"field": "rejected", "action": "$role", "template": True}
+                    ]
+                },
+            }
+        ),
+        mask={"user": "mask", "assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+
+
+def make_grpo_config():
+    return PipelineConfig(
+        input=InputConfig(
+            sources={
+                "prompts": {
+                    "sections": [
+                        {"field": "prompt", "action": "mask", "template": True}
+                    ]
+                },
+                "responses": {
+                    "sections": _GRPO_RESPONSE_SECTIONS,
+                    "list_field": True,
+                    "mask_key": "masks",
+                },
+                "rewards": {
+                    "sections": [{"field": "rewards", "action": "value"}],
+                },
+            }
+        ),
+        mask={"user": "mask", "assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+
+
+def make_grpo_no_template_config():
+    return PipelineConfig(
+        input=InputConfig(
+            sources={
+                "prompts": {
+                    "sections": [
+                        {
+                            "field": "prompt",
+                            "action": "mask",
+                            "add_special_tokens": True,
+                        }
+                    ]
+                },
+                "responses": {
+                    "sections": _GRPO_RESPONSE_SECTIONS,
+                    "list_field": True,
+                    "mask_key": "masks",
+                },
+                "rewards": {
+                    "sections": [{"field": "rewards", "action": "value"}],
+                },
+            }
+        ),
+        mask={"user": "mask", "assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )

tests/data/test_dataset.py

@@ -1,4 +1,3 @@
-import json
 import os
 
 import numpy as np
@@ -8,7 +7,6 @@ import torch
 from astrai.dataset.dataset import DatasetFactory, SEQDataset
 from astrai.dataset.storage import (
     H5Store,
-    MmapStore,
     StoreFactory,
     detect_format,
     load_bin,
@@ -100,6 +98,7 @@ def test_sft_dataset_with_random_data(base_test_env):
     dummy_data = {
         "sequence": [torch.randint(0, 1000, (seq_length,), dtype=torch.int64)],
         "loss_mask": [torch.ones(seq_length, dtype=torch.bool)],
+        "position_ids": [torch.arange(seq_length, dtype=torch.int32)],
     }
 
     save_h5(test_dir, "sft_data", dummy_data)

tests/data/test_preprocess_builder.py

@@ -0,0 +1,396 @@
+from astrai.config.preprocess_config import (
+    InputConfig,
+    OutputConfig,
+    PipelineConfig,
+    ProcessingConfig,
+)
+from astrai.preprocessing.builder import (
+    MaskBuilderFactory,
+    SectionedMaskBuilder,
+)
+from tests.data.conftest import (
+    _CHAT_SECTIONS,
+    _INSTRUCTION_SECTIONS,
+    _TEXT_SECTIONS,
+    make_chat_config,
+    make_dpo_chat_config,
+    make_grpo_config,
+    make_instruction_config,
+    make_text_config,
+)
+
+
+def test_chat_simple(chat_tokenizer):
+    config = make_chat_config()
+    builder = SectionedMaskBuilder()
+    item = {
+        "messages": [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "Hello."},
+            {"role": "assistant", "content": "Hi there!"},
+        ]
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert result is not None
+    assert "sequence" in result
+    assert "loss_mask" in result
+    assert len(result["sequence"]) == len(result["loss_mask"])
+
+    ids = chat_tokenizer.decode(result["sequence"], skip_special_tokens=False)
+    assert "system" in ids.lower() or "<|im_start|>system" in ids
+    assert "assistant" in ids.lower() or "<|im_start|>assistant" in ids
+
+    total = len(result["sequence"])
+    trained = sum(result["loss_mask"])
+    assert trained > 0
+    assert trained < total
+
+
+def test_chat_mask_only_assistant(chat_tokenizer):
+    config = make_chat_config()
+    builder = SectionedMaskBuilder()
+    item = {
+        "messages": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ]
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    mask = result["loss_mask"]
+    ids = result["sequence"]
+    assert len(ids) == len(mask)
+
+    trained = [i for i, m in enumerate(mask) if m == 1]
+    masked = [i for i, m in enumerate(mask) if m == 0]
+    assert len(trained) > 0
+    assert len(masked) > 0
+
+
+def test_chat_all_masked(chat_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_CHAT_SECTIONS),
+        mask={"system": "mask", "user": "mask", "assistant": "mask"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+    builder = SectionedMaskBuilder()
+    item = {
+        "messages": [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "assistant", "content": "Hi there!"},
+        ]
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert sum(result["loss_mask"]) == 0
+
+
+def test_chat_all_trained(chat_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_CHAT_SECTIONS),
+        mask={},
+        mask_default="train",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+    builder = SectionedMaskBuilder()
+    item = {
+        "messages": [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "assistant", "content": "Hi there!"},
+        ]
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert sum(result["loss_mask"]) == len(result["sequence"]) - 1
+
+
+def test_chat_empty_messages(chat_tokenizer):
+    config = make_chat_config()
+    builder = SectionedMaskBuilder()
+    assert builder.build({"messages": []}, config, chat_tokenizer) is None
+    assert builder.build({}, config, chat_tokenizer) is None
+
+
+def test_chat_domain_extraction(chat_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_CHAT_SECTIONS),
+        mask={"assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+        output=OutputConfig(domain_key="source"),
+    )
+    builder = SectionedMaskBuilder()
+    item = {
+        "messages": [
+            {"role": "user", "content": "Hi"},
+            {"role": "assistant", "content": "Hello"},
+        ],
+        "source": "wiki",
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert result["domain"] == "wiki"
+
+
+def test_chat_truncation(chat_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_CHAT_SECTIONS),
+        mask={"assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=10),
+    )
+    builder = SectionedMaskBuilder()
+    item = {
+        "messages": [
+            {
+                "role": "user",
+                "content": "Tell me a very long story about dragons and knights and magic.",
+            },
+            {"role": "assistant", "content": "Sure! Here is a tale..."},
+        ]
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert len(result["sequence"]) <= 10
+    assert len(result["loss_mask"]) == len(result["sequence"])
+
+
+def test_instruction_basic(test_tokenizer):
+    config = make_instruction_config()
+    builder = SectionedMaskBuilder()
+    item = {"prompt": "Translate to French: Hello", "response": "Bonjour"}
+    result = builder.build(item, config, test_tokenizer)
+    assert result is not None
+    assert len(result["sequence"]) == len(result["loss_mask"])
+
+
+def test_instruction_prompt_masked(test_tokenizer):
+    config = make_instruction_config()
+    builder = SectionedMaskBuilder()
+    item = {"prompt": "hello", "response": "world"}
+    result = builder.build(item, config, test_tokenizer)
+    mask = result["loss_mask"]
+    ids = result["sequence"]
+
+    prompt_ids = test_tokenizer.encode("hello", add_special_tokens=True)
+    p_len = min(len(prompt_ids), len(ids))
+    assert all(m == 0 for m in mask[:p_len])
+    if p_len < len(ids):
+        assert all(m == 1 for m in mask[p_len:])
+
+
+def test_instruction_train_on_prompt(test_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(
+            sections=[
+                {"field": "prompt", "action": "train", "add_special_tokens": True},
+                {"field": "response", "action": "mask"},
+            ]
+        ),
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+    builder = SectionedMaskBuilder()
+    item = {"prompt": "hello", "response": "world"}
+    result = builder.build(item, config, test_tokenizer)
+    mask = result["loss_mask"]
+    ids = result["sequence"]
+
+    prompt_ids = test_tokenizer.encode("hello", add_special_tokens=True)
+    p_len = min(len(prompt_ids), len(ids))
+    assert all(m == 1 for m in mask[:p_len])
+
+
+def test_text_basic(test_tokenizer):
+    config = make_text_config()
+    builder = SectionedMaskBuilder()
+    item = {"text": "Hello world. This is a test document."}
+    result = builder.build(item, config, test_tokenizer)
+    assert result is not None
+    assert "sequence" in result
+    assert len(result["sequence"]) > 0
+    assert "loss_mask" not in result
+
+
+def test_text_empty(test_tokenizer):
+    config = make_text_config()
+    builder = SectionedMaskBuilder()
+    assert builder.build({"text": ""}, config, test_tokenizer) is None
+    assert builder.build({"text": "   "}, config, test_tokenizer) is None
+
+
+def test_text_too_short(test_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_TEXT_SECTIONS),
+        preprocessing=ProcessingConfig(min_chars=100),
+    )
+    builder = SectionedMaskBuilder()
+    assert builder.build({"text": "short"}, config, test_tokenizer) is None
+
+
+def test_text_truncation(test_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_TEXT_SECTIONS),
+        preprocessing=ProcessingConfig(max_seq_len=3, min_chars=1),
+    )
+    builder = SectionedMaskBuilder()
+    item = {"text": "This is a very long text that should be truncated"}
+    result = builder.build(item, config, test_tokenizer)
+    assert len(result["sequence"]) <= 3
+
+
+def test_sectioned_chat(chat_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_CHAT_SECTIONS),
+        mask={"system": "mask", "user": "mask", "assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+    )
+    builder = SectionedMaskBuilder()
+    item = {
+        "messages": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ]
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert result is not None
+    assert len(result["sequence"]) == len(result["loss_mask"])
+    assert sum(result["loss_mask"]) > 0
+    assert 0 in result["loss_mask"]
+
+
+def test_sectioned_instruction(test_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_INSTRUCTION_SECTIONS),
+        preprocessing=ProcessingConfig(max_seq_len=2048, min_chars=0),
+    )
+    builder = SectionedMaskBuilder()
+    item = {"prompt": "Q: Why?", "response": "A: Because."}
+    result = builder.build(item, config, test_tokenizer)
+    assert result is not None
+    mask = result["loss_mask"]
+    assert mask[0] == 0
+    assert mask[-1] == 1
+
+
+def test_sectioned_text(test_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_TEXT_SECTIONS),
+        preprocessing=ProcessingConfig(max_seq_len=2048, min_chars=1),
+    )
+    builder = SectionedMaskBuilder()
+    item = {"text": "Hello world, this is a test."}
+    result = builder.build(item, config, test_tokenizer)
+    assert result is not None
+    assert "loss_mask" not in result
+
+
+def test_sectioned_text_too_short(test_tokenizer):
+    config = PipelineConfig(
+        input=InputConfig(sections=_TEXT_SECTIONS),
+        preprocessing=ProcessingConfig(max_seq_len=2048, min_chars=100),
+    )
+    builder = SectionedMaskBuilder()
+    assert builder.build({"text": "short"}, config, test_tokenizer) is None
+
+
+def test_factory_registered():
+    names = MaskBuilderFactory._registry.list_names()
+    assert "sectioned" in names
+
+
+def test_factory_create():
+    builder = MaskBuilderFactory.create("sectioned")
+    assert isinstance(builder, SectionedMaskBuilder)
+
+
+def test_dpo_chat_basic(chat_tokenizer):
+    config = make_dpo_chat_config()
+    builder = SectionedMaskBuilder()
+    item = {
+        "chosen": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ],
+        "rejected": [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "5"},
+        ],
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert result is not None
+    assert "chosen" in result
+    assert "rejected" in result
+    assert "chosen_mask" in result
+    assert "rejected_mask" in result
+    assert "domain" in result
+    assert len(result["chosen"]) == len(result["chosen_mask"])
+    assert len(result["rejected"]) == len(result["rejected_mask"])
+    assert sum(result["chosen_mask"]) > 0
+    assert sum(result["rejected_mask"]) > 0
+
+
+def test_dpo_chosen_only_trained(chat_tokenizer):
+    config = make_dpo_chat_config()
+    builder = SectionedMaskBuilder()
+    item = {
+        "chosen": [
+            {"role": "user", "content": "Hi"},
+            {"role": "assistant", "content": "Hello"},
+        ],
+        "rejected": [
+            {"role": "user", "content": "Hi"},
+            {"role": "assistant", "content": "Go away"},
+        ],
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert 0 in result["chosen_mask"]
+    assert 1 in result["chosen_mask"]
+    assert 0 in result["rejected_mask"]
+    assert 1 in result["rejected_mask"]
+
+
+def test_dpo_missing_field_is_none(chat_tokenizer):
+    config = make_dpo_chat_config()
+    builder = SectionedMaskBuilder()
+    assert builder.build({"chosen": [], "rejected": []}, config, chat_tokenizer) is None
+
+
+def test_grpo_basic(chat_tokenizer):
+    config = make_grpo_config()
+    builder = SectionedMaskBuilder()
+    item = {
+        "prompt": [{"role": "user", "content": "What is 2+2?"}],
+        "responses": ["4", "The answer is four", "Four", "2+2=4"],
+        "rewards": [1.0, 0.5, 0.8, 0.2],
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert result is not None
+    assert "prompts" in result
+    assert "responses" in result
+    assert "masks" in result
+    assert "rewards" in result
+    assert len(result["responses"]) == len(result["masks"])
+    assert result["rewards"] == [1.0, 0.5, 0.8, 0.2]
+
+
+def test_grpo_response_tokens_all_trained(chat_tokenizer):
+    config = make_grpo_config()
+    builder = SectionedMaskBuilder()
+    item = {
+        "prompt": [{"role": "user", "content": "Q"}],
+        "responses": ["A", "B"],
+        "rewards": [0.8, 0.2],
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    masks = result["masks"]
+    assert all(m == 1 for m in masks)
+    assert len(masks) == len(result["responses"])
+
+
+def test_grpo_single_reward(chat_tokenizer):
+    config = make_grpo_config()
+    builder = SectionedMaskBuilder()
+    item = {
+        "prompt": [{"role": "user", "content": "Q"}],
+        "responses": ["A"],
+        "rewards": 0.9,
+    }
+    result = builder.build(item, config, chat_tokenizer)
+    assert result["rewards"] == [0.9]

tests/data/test_preprocess_config.py

@@ -0,0 +1,77 @@
+import os
+
+from astrai.config.preprocess_config import (
+    InputConfig,
+    PipelineConfig,
+)
+from tests.data.conftest import (
+    _INSTRUCTION_SECTIONS,
+    _TEXT_SECTIONS,
+    make_dpo_chat_config,
+)
+
+
+def test_default_values():
+    config = PipelineConfig()
+    assert config.version == 1
+    assert config.mask == {}
+    assert config.mask_default == "mask"
+    assert config.preprocessing.max_seq_len == 2048
+    assert config.output.storage_format == "bin"
+    assert config.input.sections is None
+
+
+def test_from_dict_flat():
+    data = {
+        "version": 1,
+        "input": {
+            "sections": [{"field": "messages", "action": "$role", "template": True}]
+        },
+        "mask": {"system": "mask", "assistant": "train"},
+        "mask_default": "mask",
+        "preprocessing": {"max_seq_len": 1024},
+        "output": {"storage_format": "h5"},
+    }
+    config = PipelineConfig.from_dict(data)
+    assert config.input.sections == [
+        {"field": "messages", "action": "$role", "template": True}
+    ]
+    assert config.mask == {"system": "mask", "assistant": "train"}
+    assert config.preprocessing.max_seq_len == 1024
+    assert config.output.storage_format == "h5"
+
+
+def test_to_dict_roundtrip():
+    config = PipelineConfig(
+        input=InputConfig(sections=_INSTRUCTION_SECTIONS),
+        mask={"prompt": "mask", "response": "train"},
+        mask_default="mask",
+    )
+    d = config.to_dict()
+    config2 = PipelineConfig.from_dict(d)
+    assert config2.input.sections == _INSTRUCTION_SECTIONS
+    assert config2.mask == {"prompt": "mask", "response": "train"}
+
+
+def test_to_json_from_json(temp_dir):
+    config = PipelineConfig(
+        input=InputConfig(sections=_TEXT_SECTIONS),
+        mask={"text": "train"},
+        mask_default="mask",
+    )
+    path = os.path.join(temp_dir, "config.json")
+    config.to_json(path)
+    loaded = PipelineConfig.from_json(path)
+    assert loaded.input.sections == _TEXT_SECTIONS
+    assert loaded.mask == {"text": "train"}
+
+
+def test_dpo_config_roundtrip(temp_dir):
+    config = make_dpo_chat_config()
+    path = os.path.join(temp_dir, "config.json")
+    config.to_json(path)
+    loaded = PipelineConfig.from_json(path)
+    assert loaded.input.sources is not None
+    assert "chosen" in loaded.input.sources
+    assert "rejected" in loaded.input.sources
+    assert loaded.input.sections is None

tests/data/test_preprocess_pipeline.py

@@ -0,0 +1,349 @@
+import json
+import os
+
+from astrai.config.preprocess_config import (
+    InputConfig,
+    OutputConfig,
+    PipelineConfig,
+    ProcessingConfig,
+)
+from astrai.preprocessing.pipeline import Pipeline, filter_by_length
+from tests.data.conftest import (
+    _CHAT_SECTIONS,
+    _CHAT_TEMPLATE,
+    _INSTRUCTION_SECTIONS,
+    _SPECIAL_TOKENS_CONFIG,
+    _TEXT_SECTIONS,
+    make_dpo_chat_config,
+    make_grpo_no_template_config,
+)
+
+
+def test_filter_by_length():
+    assert filter_by_length("hello world", min_len=5)
+    assert not filter_by_length("hi", min_len=5)
+    assert not filter_by_length("x" * 100, max_len=50)
+    assert filter_by_length("just right", min_len=5, max_len=20)
+
+
+def test_full_chat_pipeline(temp_dir, chat_tokenizer):
+    tokenizer_dir = os.path.join(temp_dir, "tok")
+    os.makedirs(tokenizer_dir, exist_ok=True)
+    chat_tokenizer._tokenizer.save(os.path.join(tokenizer_dir, "tokenizer.json"))
+    with open(os.path.join(tokenizer_dir, "tokenizer_config.json"), "w") as f:
+        json.dump(
+            {
+                "special_tokens": _SPECIAL_TOKENS_CONFIG,
+                "chat_template": _CHAT_TEMPLATE,
+            },
+            f,
+        )
+
+    jsonl_path = os.path.join(temp_dir, "chat.jsonl")
+    with open(jsonl_path, "w", encoding="utf-8") as f:
+        f.write(
+            json.dumps(
+                {
+                    "messages": [
+                        {"role": "system", "content": "You are helpful."},
+                        {"role": "user", "content": "Hi."},
+                        {"role": "assistant", "content": "Hello!"},
+                    ]
+                }
+            )
+            + "\n"
+        )
+        f.write(
+            json.dumps(
+                {
+                    "messages": [
+                        {"role": "user", "content": "What is 2+2?"},
+                        {"role": "assistant", "content": "4"},
+                    ]
+                }
+            )
+            + "\n"
+        )
+
+    config = PipelineConfig(
+        input=InputConfig(sections=_CHAT_SECTIONS),
+        mask={"system": "mask", "user": "mask", "assistant": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+        output=OutputConfig(storage_format="bin", domain_key=None),
+    )
+
+    out_dir = os.path.join(temp_dir, "output")
+    Pipeline(
+        config=config,
+        input_paths=[jsonl_path],
+        output_dir=out_dir,
+        tokenizer_path=tokenizer_dir,
+    ).run()
+
+    meta_path = os.path.join(out_dir, "__default__", "shard_0000", "meta.json")
+    assert os.path.exists(meta_path)
+    with open(meta_path, "r") as f:
+        meta = json.load(f)
+    assert "sequence" in meta
+    assert "loss_mask" in meta
+    assert meta["sequence"]["dtype"] == "int32"
+    assert meta["loss_mask"]["dtype"] == "int32"
+
+
+def test_full_text_pipeline(temp_dir, test_tokenizer):
+    tokenizer_dir = os.path.join(temp_dir, "tok")
+    os.makedirs(tokenizer_dir, exist_ok=True)
+    test_tokenizer._tokenizer.save(os.path.join(tokenizer_dir, "tokenizer.json"))
+    with open(os.path.join(tokenizer_dir, "tokenizer_config.json"), "w") as f:
+        json.dump(
+            {
+                "special_tokens": {
+                    "pad_token": "<|_pad_|>",
+                    "unk_token": "<|_unk_|>",
+                }
+            },
+            f,
+        )
+
+    jsonl_path = os.path.join(temp_dir, "text.jsonl")
+    with open(jsonl_path, "w", encoding="utf-8") as f:
+        f.write(
+            json.dumps(
+                {
+                    "text": "Hello world this is a test document with enough characters to pass the minimum length filter."
+                }
+            )
+            + "\n"
+        )
+        f.write(
+            json.dumps(
+                {
+                    "text": "Another document for testing purposes with sufficient length to be processed."
+                }
+            )
+            + "\n"
+        )
+
+    config = PipelineConfig(
+        input=InputConfig(sections=_TEXT_SECTIONS),
+        preprocessing=ProcessingConfig(max_seq_len=2048, min_chars=10),
+        output=OutputConfig(storage_format="bin"),
+    )
+
+    out_dir = os.path.join(temp_dir, "output")
+    Pipeline(
+        config=config,
+        input_paths=[jsonl_path],
+        output_dir=out_dir,
+        tokenizer_path=tokenizer_dir,
+    ).run()
+
+    meta_path = os.path.join(out_dir, "__default__", "shard_0000", "meta.json")
+    assert os.path.exists(meta_path)
+    with open(meta_path, "r") as f:
+        meta = json.load(f)
+    assert "sequence" in meta
+    assert "loss_mask" not in meta
+    assert meta["sequence"]["dtype"] == "int32"
+
+
+def test_full_instruction_pipeline(temp_dir, test_tokenizer):
+    tokenizer_dir = os.path.join(temp_dir, "tok")
+    os.makedirs(tokenizer_dir, exist_ok=True)
+    test_tokenizer._tokenizer.save(os.path.join(tokenizer_dir, "tokenizer.json"))
+    with open(os.path.join(tokenizer_dir, "tokenizer_config.json"), "w") as f:
+        json.dump(
+            {
+                "special_tokens": {
+                    "pad_token": "<|_pad_|>",
+                    "unk_token": "<|_unk_|>",
+                }
+            },
+            f,
+        )
+
+    jsonl_path = os.path.join(temp_dir, "instruct.jsonl")
+    with open(jsonl_path, "w", encoding="utf-8") as f:
+        f.write(
+            json.dumps(
+                {
+                    "prompt": "Tell me a joke",
+                    "response": "Why did the chicken cross the road?",
+                }
+            )
+            + "\n"
+        )
+        f.write(
+            json.dumps(
+                {
+                    "prompt": "What is AI?",
+                    "response": "Artificial Intelligence is a field of computer science.",
+                }
+            )
+            + "\n"
+        )
+
+    config = PipelineConfig(
+        input=InputConfig(sections=_INSTRUCTION_SECTIONS),
+        mask={"prompt": "mask", "response": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+        output=OutputConfig(storage_format="bin"),
+    )
+
+    out_dir = os.path.join(temp_dir, "output")
+    Pipeline(
+        config=config,
+        input_paths=[jsonl_path],
+        output_dir=out_dir,
+        tokenizer_path=tokenizer_dir,
+    ).run()
+
+    meta_path = os.path.join(out_dir, "__default__", "shard_0000", "meta.json")
+    assert os.path.exists(meta_path)
+    with open(meta_path, "r") as f:
+        meta = json.load(f)
+    assert "sequence" in meta
+    assert "loss_mask" in meta
+    assert meta["sequence"]["dtype"] == "int32"
+    assert meta["loss_mask"]["dtype"] == "int32"
+
+
+def test_dtype_override(temp_dir, test_tokenizer):
+    tokenizer_dir = os.path.join(temp_dir, "tok")
+    os.makedirs(tokenizer_dir, exist_ok=True)
+    test_tokenizer._tokenizer.save(os.path.join(tokenizer_dir, "tokenizer.json"))
+    with open(os.path.join(tokenizer_dir, "tokenizer_config.json"), "w") as f:
+        json.dump(
+            {
+                "special_tokens": {
+                    "pad_token": "<|_pad_|>",
+                    "unk_token": "<|_unk_|>",
+                }
+            },
+            f,
+        )
+
+    jsonl_path = os.path.join(temp_dir, "data.jsonl")
+    with open(jsonl_path, "w", encoding="utf-8") as f:
+        f.write(json.dumps({"prompt": "Q", "response": "A"}) + "\n")
+
+    config = PipelineConfig(
+        input=InputConfig(sections=_INSTRUCTION_SECTIONS),
+        mask={"prompt": "mask", "response": "train"},
+        mask_default="mask",
+        preprocessing=ProcessingConfig(max_seq_len=2048),
+        output=OutputConfig(storage_format="bin", dtype={"loss_mask": "bool"}),
+    )
+
+    out_dir = os.path.join(temp_dir, "output")
+    Pipeline(
+        config=config,
+        input_paths=[jsonl_path],
+        output_dir=out_dir,
+        tokenizer_path=tokenizer_dir,
+    ).run()
+
+    meta_path = os.path.join(out_dir, "__default__", "shard_0000", "meta.json")
+    with open(meta_path, "r") as f:
+        meta = json.load(f)
+    assert meta["sequence"]["dtype"] == "int32"
+    assert meta["loss_mask"]["dtype"] == "bool"
+
+
+def test_dpo_pipeline(temp_dir, chat_tokenizer):
+    tokenizer_dir = os.path.join(temp_dir, "tok")
+    os.makedirs(tokenizer_dir, exist_ok=True)
+    chat_tokenizer._tokenizer.save(os.path.join(tokenizer_dir, "tokenizer.json"))
+    with open(os.path.join(tokenizer_dir, "tokenizer_config.json"), "w") as f:
+        json.dump(
+            {
+                "special_tokens": _SPECIAL_TOKENS_CONFIG,
+                "chat_template": _CHAT_TEMPLATE,
+            },
+            f,
+        )
+
+    jsonl_path = os.path.join(temp_dir, "dpo.jsonl")
+    with open(jsonl_path, "w", encoding="utf-8") as f:
+        f.write(
+            json.dumps(
+                {
+                    "chosen": [
+                        {"role": "user", "content": "Hi."},
+                        {"role": "assistant", "content": "Hello!"},
+                    ],
+                    "rejected": [
+                        {"role": "user", "content": "Hi."},
+                        {"role": "assistant", "content": "Go away."},
+                    ],
+                }
+            )
+            + "\n"
+        )
+
+    out_dir = os.path.join(temp_dir, "output")
+    Pipeline(
+        config=make_dpo_chat_config(),
+        input_paths=[jsonl_path],
+        output_dir=out_dir,
+        tokenizer_path=tokenizer_dir,
+    ).run()
+
+    meta_path = os.path.join(out_dir, "__default__", "shard_0000", "meta.json")
+    assert os.path.exists(meta_path)
+    with open(meta_path, "r") as f:
+        meta = json.load(f)
+    assert "chosen" in meta
+    assert "rejected" in meta
+    assert "chosen_mask" in meta
+    assert "rejected_mask" in meta
+    assert "sequence" not in meta
+
+
+def test_grpo_pipeline(temp_dir, test_tokenizer):
+    tokenizer_dir = os.path.join(temp_dir, "tok")
+    os.makedirs(tokenizer_dir, exist_ok=True)
+    test_tokenizer._tokenizer.save(os.path.join(tokenizer_dir, "tokenizer.json"))
+    with open(os.path.join(tokenizer_dir, "tokenizer_config.json"), "w") as f:
+        json.dump(
+            {
+                "special_tokens": {
+                    "pad_token": "<|_pad_|>",
+                    "unk_token": "<|_unk_|>",
+                }
+            },
+            f,
+        )
+
+    jsonl_path = os.path.join(temp_dir, "grpo.jsonl")
+    with open(jsonl_path, "w", encoding="utf-8") as f:
+        f.write(
+            json.dumps(
+                {
+                    "prompt": "Question?",
+                    "responses": ["Answer A", "Answer B"],
+                    "rewards": [0.8, 0.3],
+                }
+            )
+            + "\n"
+        )
+
+    out_dir = os.path.join(temp_dir, "output")
+    Pipeline(
+        config=make_grpo_no_template_config(),
+        input_paths=[jsonl_path],
+        output_dir=out_dir,
+        tokenizer_path=tokenizer_dir,
+    ).run()
+
+    meta_path = os.path.join(out_dir, "__default__", "shard_0000", "meta.json")
+    assert os.path.exists(meta_path)
+    with open(meta_path, "r") as f:
+        meta = json.load(f)
+    assert "prompts" in meta
+    assert "responses" in meta
+    assert "masks" in meta
+    assert "rewards" in meta
+    assert "sequence" not in meta

tests/inference/conftest.py

@@ -5,21 +5,22 @@ from unittest.mock import MagicMock
 import pytest
 from fastapi.testclient import TestClient
 
-from astrai.inference import app
+from astrai.inference import get_app
 
 
 @pytest.fixture
 def client():
     """Provide a test client for the FastAPI app."""
-    app.state.server_config = {
+    _app = get_app()
+    _app.state.server_config = {
         "device": "cpu",
         "dtype": "bfloat16",
         "param_path": None,
         "max_batch_size": 1,
         "_test": True,
     }
-    app.state.engine = None
-    return TestClient(app)
+    _app.state.engine = None
+    return TestClient(_app)
 
 
 @pytest.fixture
@@ -49,5 +50,5 @@ def mock_engine():
 @pytest.fixture
 def loaded_model(client, mock_engine):
     """Simulate that the engine is loaded."""
-    app.state.engine = mock_engine
+    get_app().state.engine = mock_engine
     return mock_engine

tests/inference/test_server.py

@@ -2,12 +2,12 @@
 
 import pytest
 
-from astrai.inference import app
+from astrai.inference import get_app
 
 
 def test_health_no_model(client):
     """GET /health should return 200 even when engine not loaded."""
-    app.state.engine = None
+    get_app().state.engine = None
     response = client.get("/health")
     assert response.status_code == 200
     data = response.json()
@@ -30,7 +30,7 @@ def test_chat_completions_non_stream(client, loaded_model):
     async def async_gen():
         yield "Assistant reply"
 
-    app.state.engine = loaded_model
+    get_app().state.engine = loaded_model
     loaded_model.generate_async.return_value = async_gen()
     response = client.post(
         "/v1/chat/completions",
@@ -56,7 +56,7 @@ def test_chat_completions_stream(client, loaded_model):
         yield "cumulative1"
         yield "cumulative2"
 
-    app.state.engine = loaded_model
+    get_app().state.engine = loaded_model
     loaded_model.generate_async.return_value = async_gen()
     response = client.post(
         "/v1/chat/completions",
@@ -83,7 +83,7 @@ def test_messages_non_stream(client, loaded_model):
     async def async_gen():
         yield "Assistant reply"
 
-    app.state.engine = loaded_model
+    get_app().state.engine = loaded_model
     loaded_model.generate_async.return_value = async_gen()
     response = client.post(
         "/v1/messages",
@@ -111,7 +111,7 @@ def test_messages_stream(client, loaded_model):
         yield "cumulative1"
         yield "cumulative2"
 
-    app.state.engine = loaded_model
+    get_app().state.engine = loaded_model
     loaded_model.generate_async.return_value = async_gen()
     response = client.post(
         "/v1/messages",
@@ -141,7 +141,7 @@ def test_messages_with_system(client, loaded_model):
     async def async_gen():
         yield "Reply"
 
-    app.state.engine = loaded_model
+    get_app().state.engine = loaded_model
     loaded_model.generate_async.return_value = async_gen()
     response = client.post(
         "/v1/messages",
@@ -165,7 +165,7 @@ def test_chat_completions_stop_sequence(client, loaded_model):
         yield "X"
         yield "world"
 
-    app.state.engine = loaded_model
+    get_app().state.engine = loaded_model
     loaded_model.generate_async.return_value = async_gen()
     response = client.post(
         "/v1/chat/completions",
@@ -191,7 +191,7 @@ def test_chat_completions_stop_sequence_stream(client, loaded_model):
         yield "X"
         yield "world"
 
-    app.state.engine = loaded_model
+    get_app().state.engine = loaded_model
     loaded_model.generate_async.return_value = async_gen()
     response = client.post(
         "/v1/chat/completions",