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 7, 16:30 UTC
[Bugfix][Hardware][AMD] Gate FP4 ops on gfx950 to prevent MI300X crash (#35250)
Signed-off-by: c0de128 <kevin.mckay@outlook.com>
Kevin McKay committed
Feb 26, 2026 at 02:11 UTC
ade81f17feeebef775e8cddf9a8f23848ec694a3
1 file changed
+6
-2
vllm/_aiter_ops.py
+6
-2
@@ -1052,12 +1052,16 @@ class rocm_aiter_ops:
1052
@classmethod
1053
@if_aiter_supported
1054
def is_fp4bmm_enabled(cls) -> bool:
1055
- return cls._AITER_ENABLED and cls._FP4BMM_ENABLED
1055
+ from vllm.platforms.rocm import on_gfx950
1056
+
1057
+ return cls._AITER_ENABLED and cls._FP4BMM_ENABLED and on_gfx950()
1058
1059
@classmethod
1060
@if_aiter_supported
1061
def is_asm_fp4_gemm_dynamic_quant_enabled(cls) -> bool:
1060
- return cls._AITER_ENABLED and cls._FP4_GEMM_DYNAMIC_QUANT_ASM
1062
+ from vllm.platforms.rocm import on_gfx950
1063
+
1064
+ return cls._AITER_ENABLED and cls._FP4_GEMM_DYNAMIC_QUANT_ASM and on_gfx950()
1065
1066
@classmethod
1067
@if_aiter_supported