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 2, 19:30 UTC
[Perf] Optimize Fp8BlockScaledMMLinearKernel input_scale tensor using new_empty() (#43677)
Signed-off-by: Xin Yang <xyangx@amazon.com>
Xin Yang committed
May 26, 2026 at 19:55 UTC
d8eebe6d9750534ab309f684f0b04748f4195f8a
1 file changed
+1
-1
vllm/model_executor/kernels/linear/scaled_mm/BlockScaledMMLinearKernel.py
+1
-1
@@ -126,7 +126,7 @@ class Fp8BlockScaledMMLinearKernel(
126
# always Tensors. Subclasses with apply_input_quant=False must not
127
# use As in apply_block_scaled_mm.
128
input_scale = (
129
- input_scale if input_scale is not None else input_2d.new_ones(1)
129
+ input_scale if input_scale is not None else input_2d.new_empty(1)
130
)
131
132
output = self.apply_block_scaled_mm(