ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[breaking] bump transformers to 4.45.0 & improve ci (#7746)

Browse Source 
* update ci

* fix

* fix

* fix

* fix

* fix
This commit is contained in:
hoshi-hiyouga
2025-04-17 02:36:48 +08:00
committed by GitHub
parent d222f63cb7
commit 86ebb219d6
23 changed files with 211 additions and 140 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/llamafactory/model/model_utils/attention.py
View File

@@ -18,7 +18,6 @@ from transformers.utils import is_flash_attn_2_available, is_torch_sdpa_availabl
from ...extras import logging
from ...extras.constants import AttentionFunction
from ...extras.misc import check_version
if TYPE_CHECKING:
@@ -36,8 +35,6 @@ def configure_attn_implementation(
if getattr(config, "model_type", None) == "gemma2" and is_trainable:
if model_args.flash_attn == AttentionFunction.AUTO or model_args.flash_attn == AttentionFunction.FA2:
if is_flash_attn_2_available():
check_version("transformers>=4.42.4")
check_version("flash_attn>=2.6.3")
if model_args.flash_attn != AttentionFunction.FA2:
logger.warning_rank0("Gemma 2 should use flash attention 2, change `flash_attn` to fa2.")
model_args.flash_attn = AttentionFunction.FA2