From dbfd6a7391a2b4267e81cd917171a6ab3e2050ac Mon Sep 17 00:00:00 2001 From: rjg-lyh <1318825571@qq.com> Date: Tue, 31 Mar 2026 15:49:12 +0800 Subject: [PATCH] [v0.13.0][bugfix][accuracy] Fix ds indexer accuracy problem caused by k rope Signed-off-by: rjg-lyh <1318825571@qq.com> --- vllm_ascend/attention/sfa_v1.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm_ascend/attention/sfa_v1.py b/vllm_ascend/attention/sfa_v1.py index 52574bcc359..674c90a0333 100644 --- a/vllm_ascend/attention/sfa_v1.py +++ b/vllm_ascend/attention/sfa_v1.py @@ -953,7 +953,7 @@ def indexer_select_pre_process( sin = sin.view(-1, 1, 1, self.qk_rope_head_dim) k_pe = k_pe.unsqueeze(2) - k_pe = torch_npu.npu_interleave_rope(k_pe, cos, sin) + k_pe = torch_npu.npu_rotary_mul(k_pe, cos, sin) k_pe = k_pe.squeeze(2) k = torch.cat([k_pe, k_nope], dim=-1) # [b*s,128]