王得利 8972dab6f5 fix sft resume with compile mode ... Load checkpoint state before torch.compile to avoid key mismatch on resume, and make optimizer/scaler restore tolerant to missing fields in older checkpoints. Made-with: Cursor		2026-03-19 12:44:53 +00:00
..
train_distillation.py	[fix] data skip	2026-01-18 16:56:29 +08:00
train_dpo.py	[fix] data skip	2026-01-18 16:56:29 +08:00
train_full_sft.py	fix sft resume with compile mode	2026-03-19 12:44:53 +00:00
train_grpo.py	[fix] data skip	2026-01-18 16:56:29 +08:00
train_lora.py	[update] save interval	2026-01-30 20:30:50 +08:00
train_ppo.py	[fix] data skip	2026-01-18 16:56:29 +08:00
train_pretrain.py	[fix] data skip	2026-01-18 16:56:29 +08:00
train_reason.py	[fix] data skip	2026-01-18 16:56:29 +08:00
train_spo.py	[fix] data skip	2026-01-18 16:56:29 +08:00
train_tokenizer.py	[update] rename train tokenizer	2026-01-06 01:17:33 +08:00
trainer_utils.py	[fix] compile unpack	2026-01-14 20:13:32 +08:00