From d8eebe6d9750534ab309f684f0b04748f4195f8a Mon Sep 17 00:00:00 2001 From: Xin Yang <105740670+xyang16@users.noreply.github.com> Date: Tue, 26 May 2026 19:55:52 -0700 Subject: [PATCH] [Perf] Optimize Fp8BlockScaledMMLinearKernel input_scale tensor using new_empty() (#43677) Signed-off-by: Xin Yang --- .../kernels/linear/scaled_mm/BlockScaledMMLinearKernel.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/model_executor/kernels/linear/scaled_mm/BlockScaledMMLinearKernel.py b/vllm/model_executor/kernels/linear/scaled_mm/BlockScaledMMLinearKernel.py index d738796ce6..c83885ca51 100644 --- a/vllm/model_executor/kernels/linear/scaled_mm/BlockScaledMMLinearKernel.py +++ b/vllm/model_executor/kernels/linear/scaled_mm/BlockScaledMMLinearKernel.py @@ -126,7 +126,7 @@ class Fp8BlockScaledMMLinearKernel( # always Tensors. Subclasses with apply_input_quant=False must not # use As in apply_block_scaled_mm. input_scale = ( - input_scale if input_scale is not None else input_2d.new_ones(1) + input_scale if input_scale is not None else input_2d.new_empty(1) ) output = self.apply_block_scaled_mm(