feat: add record-mode to Store for DPO/GRPO

- Store gains fetch_record/num_records alongside stream fetch/__len__
- save_bin/load_bin support per-record offsets via record_keys param
- H5Store/MmapStore/JsonlStore all support dual stream+record access
- DPODataset/GRPODataset use fetch_record, no cross-record concat
- dpo_collate_fn + collate_fn wired through TrainConfig
- fixes attention context leakage in DPO from windowed concatenation
This commit is contained in:
2026-07-18 21:02:29 +08:00
parent 28886e4241
commit a74e5b91a3
12 changed files with 334 additions and 92 deletions
+4
View File
@@ -87,6 +87,10 @@ class TrainConfig(BaseConfig):
pin_memory: bool = field(
default=False, metadata={"help": "Pin memory for dataloader."}
)
collate_fn: Optional[Callable[[List[Any]], Any]] = field(
default=None,
metadata={"help": "Collate function for dataloader (e.g. dpo_collate_fn)."},
)
# distributed training
nprocs: int = field(