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

[Bugfix][Hardware][AMD] Gate FP4 ops on gfx950 to prevent MI300X crash (#35250)

Browse Source 
Signed-off-by: c0de128 <kevin.mckay@outlook.com>
This commit is contained in:
Kevin McKay
2026-02-26 02:11:07 -06:00
committed by GitHub
parent 6042e66cd5
commit ade81f17fe
1 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/_aiter_ops.py
View File

@@ -1052,12 +1052,16 @@ class rocm_aiter_ops:
@classmethod
@if_aiter_supported
def is_fp4bmm_enabled(cls) -> bool:
return cls._AITER_ENABLED and cls._FP4BMM_ENABLED
from vllm.platforms.rocm import on_gfx950
return cls._AITER_ENABLED and cls._FP4BMM_ENABLED and on_gfx950()
@classmethod
@if_aiter_supported
def is_asm_fp4_gemm_dynamic_quant_enabled(cls) -> bool:
return cls._AITER_ENABLED and cls._FP4_GEMM_DYNAMIC_QUANT_ASM
from vllm.platforms.rocm import on_gfx950
return cls._AITER_ENABLED and cls._FP4_GEMM_DYNAMIC_QUANT_ASM and on_gfx950()
@classmethod
@if_aiter_supported