Commit Graph

5 Commits (295d681b9ccd0811b0cd97a0c9b3ed86431bfe9a)

Author SHA1 Message Date
jingyaogong 295d681b9c [fix] fp16 attention overflow 2026-09-20 20:39:51 +08:00
jingyaogong 02ebfb33d2 [fix] top-1 router gradient 2026-09-18 16:34:53 +08:00
jingyaogong 0b85f814ed [fix] existing bugs 2026-05-07 19:07:26 +08:00
zhuxiaoxuhit c06144907e Fix 4 bugs in training and inference
1. Fix EOS label leak in multi-turn masking (omni_dataset.py)
2. Fix repetition penalty for negative logits (model_minimind.py, model_omni.py)
3. Fix stream_generate stop condition (model_omni.py)
4. Fix O(n^2) tolist in generation loop (model_omni.py)
2026-05-07 17:50:44 +00:00
jingyaogong c009567d06 [update] minimind-3o 2026-05-06 13:53:28 +08:00