423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-29 02:00:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

[model] remove npu sdpa patch (#9368)

Browse Source 
Co-authored-by: frozenleaves <frozen@Mac.local>
This commit is contained in:
魅影
2025-10-30 16:26:35 +08:00
committed by GitHub
parent 3057db15c3
commit 767b344fb4
2 changed files with 0 additions and 148 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/llamafactory/model/patcher.py
View File

@@ -193,23 +193,6 @@ def patch_model(
if not model_args.use_unsloth:
print_attn_implementation(model.config)
# ======== NPU fused attention redirect: SDPA -> torch_npu.npu_fusion_attention ========
# Place after all structural modifications and before DeepSpeed/Trainer initialization;
# does not modify any Module/_parameters, safe for ZeRO-3 + offload.
try:
import os
import torch
if hasattr(torch, "npu") and torch.npu.is_available() and os.environ.get("NPU_FA_DISABLE", "0") != "1":
from .model_utils.sdpa_npu_redirect import apply_sdpa_npu_redirect
apply_sdpa_npu_redirect(verbose=not model_args.use_unsloth)
logger.info_rank0("[sdpa_npu_redirect] Enabled: SDPA will use Ascend npu_fusion_attention when available.")
except Exception as e:
logger.warning_rank0(f"[sdpa_npu_redirect] Failed to enable redirect, will keep native SDPA. Reason: {e}")
# =====================================================================================
try:
model.add_model_tags(["llama-factory"])
except Exception: