EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

deepseek torchair adapt for torch_npu version (#3876)

Browse Source 
### What this PR does / why we need it?
To adapt the torch_npu version to avoid the precision problem of
torchair deepseek. The torch_npu version may result in the different
branches in the ops register, the rms_norm ops has two branches
according to the verson_check, this pr unify the rms_norm in torchair by
patch method. #3862

Signed-off-by: hust17yixuan <303660421@qq.com>
This commit is contained in:
Wang Yixuan
2025-10-29 22:44:44 +08:00
committed by GitHub
parent 29bd9235ed
commit b323be9fe4
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm_ascend/torchair/utils.py
View File

@@ -210,7 +210,7 @@ def torchair_quant_method_register():
def torchair_ops_patch(): def torchair_ops_patch():
from vllm_ascend.ops.activation import AscendSiluAndMul from vllm_ascend.ops.activation import AscendSiluAndMul
from vllm_ascend.ops.layernorm import AscendRMSNorm from vllm_ascend.ops.layernorm import AscendQuantRMSNorm, AscendRMSNorm
from vllm_ascend.ops.rotary_embedding import ( from vllm_ascend.ops.rotary_embedding import (
AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding) AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding)
from vllm_ascend.ops.vocab_parallel_embedding import \ from vllm_ascend.ops.vocab_parallel_embedding import \
@@ -232,6 +232,9 @@ def torchair_ops_patch():
AscendRMSNorm.__init__ = torchair_layernorm.torchair_rmsnorm_init_ # type: ignore[method-assign] AscendRMSNorm.__init__ = torchair_layernorm.torchair_rmsnorm_init_ # type: ignore[method-assign]
AscendRMSNorm.forward_oot = torchair_layernorm.torchair_rmsnorm_forward_oot # type: ignore[method-assign] AscendRMSNorm.forward_oot = torchair_layernorm.torchair_rmsnorm_forward_oot # type: ignore[method-assign]
AscendQuantRMSNorm.__init__ = torchair_layernorm.torchair_rmsnorm_init_ # type: ignore[method-assign]
AscendQuantRMSNorm.forward_oot = torchair_layernorm.torchair_rmsnorm_forward_oot # type: ignore[method-assign]
AscendSiluAndMul.forward_oot = torchair_activation.torchair_silu_and_mul_forward_oot # type: ignore[method-assign] AscendSiluAndMul.forward_oot = torchair_activation.torchair_silu_and_mul_forward_oot # type: ignore[method-assign]
AscendVocabParallelEmbedding.forward = vocab_embedding_forward # type: ignore[method-assign] AscendVocabParallelEmbedding.forward = vocab_embedding_forward # type: ignore[method-assign]