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 4, 22:30 UTC
Properly enable wvSplitK fp8 path for RDNA (#37712)
Signed-off-by: Hashem Hashemi <hashem.hashemi@amd.com>
Hashem Hashemi committed
Apr 20, 2026 at 08:09 UTC
3a30eaa1d7b6497ce70c15558d410e37d8399b87
1 file changed
+3
-3
vllm/model_executor/kernels/linear/scaled_mm/rocm.py
+3
-3
@@ -79,10 +79,10 @@ class ROCmFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
79
if not current_platform.is_rocm():
80
return False, "requires ROCm."
81
82
- from vllm.platforms.rocm import on_mi3xx
82
+ from vllm.platforms.rocm import on_gfx12x, on_mi3xx
83
84
- if not on_mi3xx():
85
- return False, "requires MI3xx."
84
+ if not (on_mi3xx() or on_gfx12x()):
85
+ return False, "requires MI3xx or gfx12x"
86
87
if not envs.VLLM_ROCM_USE_SKINNY_GEMM:
88
return False, "requires VLLM_ROCM_USE_SKINNY_GEMM to be enabled."