ViperEkura 9bcd696580 fix: token-level ratio and prompt masking in GRPO strategy ... - Mask prompt tokens to 0 so their logprobs excluded from ratio/KL - Switch to token-level ratio + PPO clipping via reduction='none' - Slice response token logprobs from full sequence output - Replace k3 KL estimator with non-negative k1 estimator - Fix epsilon from finifo.eps (~1e-38) to 1e-8 - Remove unused 'reduction' param from GRPOStrategy.__init__ - Clarify offline batch semantics in docstring - Add 11 unit tests for masking, advantage, KL, sync, clipping - Sync training.md and architecture.md docs		2026-07-12 21:24:09 +08:00
..
data	refactor: split mask builder by single/multi output	2026-07-08 21:18:34 +08:00
inference	perf: replace paged KV cache with contiguous ContiguousCache, decode all groups	2026-07-05 11:34:36 +08:00
module	refactor : 压缩测试代码，消除重复	2026-06-19 14:54:39 +08:00
parallel	feat: 增加server, 并且修改测试单元	2026-04-02 15:05:07 +08:00
trainer	fix: token-level ratio and prompt masking in GRPO strategy	2026-07-12 21:24:09 +08:00
conftest.py	refactor : replace iteration with consumed_samples	2026-06-30 18:42:42 +08:00