- load_json/load_safetensors/load_state_dict 新增 broadcast 参数 - broadcast=True 时 rank-0 读取后 broadcast_object_list 分发到所有 rank - load_state_dict 改为逐张量 broadcast,避免大模型 pickle 内存瓶颈 - 删除 _get_meta/_get_config wrapper,Checkpoint.load 直接调用 load_json - 参数注解 str | Path 统一为 Union[str, Path] |
||
|---|---|---|
| .. | ||
| config | ||
| dataset | ||
| inference | ||
| model | ||
| parallel | ||
| tokenize | ||
| trainer | ||
| __init__.py | ||
| factory.py | ||
| protocols.py | ||
| serialization.py | ||