From f976e3b98ba45677a2213673a442c6cbff141e8e Mon Sep 17 00:00:00 2001 From: xaguilar-amd Date: Fri, 10 Apr 2026 13:27:35 +0200 Subject: [PATCH] [Performance] Remove unnecessary zero-fill of MLA decode output tensor in Aiter backend (#37539) Signed-off-by: xaguilar-amd --- vllm/v1/attention/backends/mla/rocm_aiter_mla.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla.py index 8b764cd627..351fdba085 100644 --- a/vllm/v1/attention/backends/mla/rocm_aiter_mla.py +++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla.py @@ -416,7 +416,7 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]): else: kernel_num_heads = self.num_heads - o = torch.zeros( + o = torch.empty( B, kernel_num_heads, self.kv_lora_rank,