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

Merge pull request #4446 from stceum/bug-fix

Browse Source 
Bug Fix: `off` is parsed as `False` in yaml file

Former-commit-id: 243478a3d6c08f5677ee57871862694561617f64
This commit is contained in:
hoshi-hiyouga
2024-06-24 21:41:28 +08:00
committed by GitHub
parent 826d7808b4 d475dd3809
commit f4c518b370
3 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llamafactory/hparams/model_args.py
View File

@@ -97,7 +97,7 @@ class ModelArguments:
default=None,
metadata={"help": "Which scaling strategy should be adopted for the RoPE embeddings."},
)
flash_attn: Literal["off", "sdpa", "fa2", "auto"] = field(
flash_attn: Literal["disabled", "sdpa", "fa2", "auto"] = field(
default="auto",
metadata={"help": "Enable FlashAttention for faster training and inference."},
)

2
src/llamafactory/model/model_utils/attention.py
View File

@@ -32,7 +32,7 @@ def configure_attn_implementation(config: "PretrainedConfig", model_args: "Model
if model_args.flash_attn == "auto":
return
elif model_args.flash_attn == "off":
elif model_args.flash_attn == "disabled":
requested_attn_implementation = "eager"
elif model_args.flash_attn == "sdpa":