refactor: 重构打包模块,新增 BFD/FFD/Greedy 三种 bin-packing 算法,默认 BFD

- 将 pipeline/packing.py 拆分为 packing/ 子包 (base/stream/binpack)
- 新增 BfdPacker(默认)/FfDPacker/GreedyPacker,移除 StreamingPacker
- 超长序列直接截断至 pack_size
- group_size 语义改为"每 N 个 chunk 合并为一块",默认 1000
- 新增 AutoTokenizer.token_to_id(),修复 ChatML 中 hacky 的 nl_id 获取
- pad_value 默认改为 2(pad_token_id),position_ids pad=0, loss_mask pad=False
- 新增 position_ids 打包后归零一致性测试
- scripts/cache_h5.py 新增 --pack-algo 参数
This commit is contained in:
2026-07-03 16:17:27 +08:00
parent 2f919e9243
commit 598e1ce4ae
14 changed files with 715 additions and 295 deletions
+1 -1
View File
@@ -20,7 +20,7 @@ class ChatMLStrategy(PromptStrategy):
assistant_end: str = "<im▁end>",
):
super().__init__(tokenizer)
nl_id = tokenizer.encode("a\nb", add_special_tokens=False)[1]
nl_id = tokenizer.token_to_id("\n")
self._user_start_ids = self._encode_format(user_start) + [nl_id]
self._user_end_ids = self._encode_format(user_end) + [nl_id]