ViperEkura f433672140 fix: use sum reduction for DPO sequence logprob ...

- DPO requires sequence-level sum of token logprobs, not per-token mean
- mean reduction made beta*ratio_diff ~0.03 (near-zero gradient)
- loss stalled at 0.6931 because logsigmoid(0.03) has vanishing grad
- sum gives beta*ratio_diff ~10 with meaningful gradients

2026-07-18 23:48:35 +08:00

..

config

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

2026-07-18 21:02:29 +08:00

dataset

refactor: DatasetFactory.load accepts pre-built store instance

2026-07-18 23:23:51 +08:00

extension

refactor: stride-based attn interface with layout and causal mask

2026-07-14 21:34:42 +08:00

inference

fix: incremental decode to avoid U+FFFD in streaming

2026-07-18 13:05:36 +08:00

model

perf: replace paged KV cache with contiguous ContiguousCache, decode all groups

2026-07-05 11:34:36 +08:00

parallel

feat : add distributed checkpoint via executor checkpoint_context

2026-07-13 12:27:09 +08:00

preprocessing

refactor: decouple tokenizer from Store into Transform layer

2026-07-18 21:37:31 +08:00

serialization

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

2026-07-18 21:02:29 +08:00

tokenize

refactor : Pipeline 去除去重，ids 重命名为 sequence，泛型透传

2026-05-31 15:14:27 +08:00

trainer

fix: use sum reduction for DPO sequence logprob

2026-07-18 23:48:35 +08:00

__init__.py

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

2026-07-18 21:02:29 +08:00

factory.py

fix: ForwardRef._evaluate Python 3.12 compatibility

2026-07-03 18:41:19 +08:00

protocols.py

refactor: 重构训练后端为 Executor 模式

2026-05-24 20:35:44 +08:00