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

..

__init__.py

chore : 升级 torch 2.11.0+cu128，移除自定义 Muon，修复 gloo device_id

2026-06-27 16:10:37 +08:00

metric_util.py

refactor : grad_norm 指标简化，clip_grad_norm 移至 executor

2026-06-30 14:59:43 +08:00

schedule.py

fix: scheduler division-by-zero, loss_mask bool

2026-07-03 22:04:55 +08:00

strategy.py

fix: use sum reduction for DPO sequence logprob

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

train_callback.py

fix: init last_ckpt_step and last_log_flush_step from context.optimizer_step

2026-07-15 22:15:55 +08:00

train_context.py

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

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

trainer.py

feat: add --resume flag to decouple weight loading from training resumption

2026-07-16 14:23:23 +08:00