423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-23 06:12:50 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #5580 from amrear/main

Browse Source 
made a small change to a warning about fa2 for gemma2 models.

Former-commit-id: 5ec161d6bdbb3ecc0ca0b10ea24ac7270618263d
This commit is contained in:
hoshi-hiyouga 2024-09-29 12:45:03 +08:00 committed by GitHub
commit 786e013375
1 changed files with 4 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/llamafactory/model/model_utils/attention.py
View File

@ -37,10 +37,11 @@ def configure_attn_implementation(
if is_flash_attn_2_available():
require_version("transformers>=4.42.4", "To fix: pip install transformers>=4.42.4")
require_version("flash_attn>=2.6.3", "To fix: pip install flash_attn>=2.6.3")
logger.warning("Gemma-2 should use flash attention 2, change `flash_attn` to fa2.")
model_args.flash_attn = "fa2"
if model_args.flash_attn != "fa2":
logger.warning("Gemma-2 should use flash attention 2, change `flash_attn` to fa2.")
model_args.flash_attn = "fa2"
else:
logger.warning("Gemma-2 should use eager attention, change `flash_attn` to disabled.")
logger.warning("FlashAttention-2 is not installed, use eager attention.")
model_args.flash_attn = "disabled"
elif model_args.flash_attn == "sdpa":
logger.warning("Gemma-2 should use soft-capping attention, while the SDPA attention does not support it.")