xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-20 21:07:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

Disable the broken gemm_a8w8_blockscale

Browse Source 
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
This commit is contained in:
Gregory Shtrasberg 2025-12-23 17:28:08 +00:00
parent 23d8239cdf
commit 981e1d8343
1 changed files with 1 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -362,23 +362,10 @@ class W8A8BlockFp8LinearOp:
n, k = weight.shape n, k = weight.shape
use_triton = ( gemm_a8w8_blockscale_op = rocm_aiter_ops.gemm_a8w8_blockscale
not current_platform.is_fp8_fnuz()
and rocm_aiter_ops.is_triton_gemm_w8a8_tuned(n, k)
)
if use_triton:
gemm_a8w8_blockscale_op = rocm_aiter_ops.triton_gemm_a8w8_blockscale
else:
gemm_a8w8_blockscale_op = rocm_aiter_ops.gemm_a8w8_blockscale
if input_scale is not None: if input_scale is not None:
q_input = input_2d q_input = input_2d
elif use_triton:
q_input, input_scale = torch.ops.vllm.triton_per_token_group_quant_fp8(
input_2d,
self.act_quant_group_shape.col,
)
else: else:
q_input, input_scale = rocm_aiter_ops.group_fp8_quant( q_input, input_scale = rocm_aiter_ops.group_fp8_quant(
input_2d, self.act_quant_group_shape.col input_2d, self.act_quant_group_shape.col