From b323be9fe415280c12408202fdd1c6a08adbd20d Mon Sep 17 00:00:00 2001 From: Wang Yixuan <88923622+hust17yixuan@users.noreply.github.com> Date: Wed, 29 Oct 2025 22:44:44 +0800 Subject: [PATCH] deepseek torchair adapt for torch_npu version (#3876) ### 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> --- vllm_ascend/torchair/utils.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm_ascend/torchair/utils.py b/vllm_ascend/torchair/utils.py index af61c65..97fc3b1 100644 --- a/vllm_ascend/torchair/utils.py +++ b/vllm_ascend/torchair/utils.py @@ -210,7 +210,7 @@ def torchair_quant_method_register(): def torchair_ops_patch(): 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 ( AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding) 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.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] AscendVocabParallelEmbedding.forward = vocab_embedding_forward # type: ignore[method-assign]