Update trainer.py

This commit is contained in:
Daniel Han 2025-12-10 06:11:03 -08:00
commit 1e08afbca1

View file

@ -60,7 +60,7 @@ _AUTO_PADDING_FREE_ENV_DISABLED = os.environ.get(
# [TODO]
# Below cannot work with padding-free
_PADDING_FREE_BLOCK_LIST = {
PADDING_FREE_BLOCKLIST = {
"gemma2", # - gemma2: Uses slow_attention_softcapping which has torch.compile issues
"gpt_oss", # - gpt_oss: Uses Flex Attention which doesn't handle padding_free correctly
"mistral", # - mistral: Unfortunately I think sliding window attention doesn't work correctly?