biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

use ceil_div in cutlass block scaling shape check (#17918)

Browse Source
This commit is contained in:
Lain
2025-05-16 03:02:58 -07:00
committed by GitHub
parent 390ec88905
commit e23564cb70
3 changed files with 62 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/model_executor/layers/quantization/utils/fp8_utils.py
View File

@@ -115,8 +115,19 @@ def apply_w8a8_block_fp8_linear(
output_shape = [*input.shape[:-1], weight.shape[0]]
if current_platform.is_cuda():
use_cutlass = cutlass_block_fp8_supported and (
weight.shape[0] % 128 == 0 and weight.shape[1] % 128 == 0)
if current_platform.has_device_capability(100):
def ceil_div(x: int, y: int) -> int:
return (x + y - 1) // y
use_cutlass = cutlass_block_fp8_supported and (
ceil_div(weight.shape[0], 128) == weight_scale.shape[0]
and ceil_div(weight.shape[1], 128) == weight_scale.shape[1])
else:
# TODO: update this after switching to public sm90 block scale gemm
# as it also supports weight.shape % 128 != 0
use_cutlass = cutlass_block_fp8_supported and (
weight.shape[0] % 128 == 0 and weight.shape[1] % 128 == 0)
else:
use_cutlass = False