A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 07:15 UTC
[Performance] Remove unnecessary zero-fill of MLA decode output tensor in Aiter backend (#37539)
Signed-off-by: xaguilar-amd <xaguilar@amd.com>
xaguilar-amd committed
Apr 10, 2026 at 13:27 UTC
f976e3b98ba45677a2213673a442c6cbff141e8e
1 file changed
+1
-1
vllm/v1/attention/backends/mla/rocm_aiter_mla.py
+1
-1
@@ -416,7 +416,7 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
416
else:
417
kernel_num_heads = self.num_heads
418
419
- o = torch.zeros(
419
+ o = torch.empty(
420
B,
421
kernel_num_heads,
422
self.kv_lora_rank,