ros/LlamaFactory
1
0
Fork 0
mirror of https://github.com/hiyouga/LlamaFactory.git synced 2026-02-01 20:23:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

[ci] disable pip cache for ci (#9654)

Browse Source
This commit is contained in:
Yaowei Zheng
2025-12-23 18:37:40 +08:00
committed by GitHub
parent 1c8a42d2f8
commit 84485406b7
5 changed files with 19 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/llamafactory/model/model_utils/attention.py
View File

@@ -29,16 +29,19 @@ logger = logging.get_logger(__name__)
def configure_attn_implementation(config: "PretrainedConfig", model_args: "ModelArguments") -> None:
from transformers.utils import is_flash_attn_2_available
if getattr(config, "model_type", None) == "gpt_oss":
from transformers.integrations.hub_kernels import load_and_register_kernel
flash_attn3_kernel = "kernels-community/vllm-flash-attn3"
load_and_register_kernel(flash_attn3_kernel)
setattr(config, "_attn_implementation", flash_attn3_kernel)
setattr(config, "_attn_implementation_internal", flash_attn3_kernel)
model_args.flash_attn = flash_attn3_kernel
model_args.flash_attn = AttentionFunction.FA3
logger.info_rank0("Using FlashAttention-3 with attention sink for the gpt-oss model.")
return
from transformers.utils import is_flash_attn_2_available
if getattr(config, "model_type", None) == "gemma2":
if model_args.flash_attn == AttentionFunction.AUTO or model_args.flash_attn == AttentionFunction.FA2: