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 9, 20:45 UTC
[ROCm] fix import for on_gfx9 (#32783)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Divakar Verma committed
Jan 21, 2026 at 12:41 UTC
180fba653ead2274bfcd3951a275f4d6cf9ade04
2 files changed
+16
-2
vllm/model_executor/layers/fused_moe/fused_batched_moe.py
+8
-1
@@ -913,7 +913,14 @@ class BatchedTritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
913
activation_key: QuantKey | None,
914
) -> bool:
915
p = current_platform
916
- device_supports_fp8 = (p.is_rocm() and p.rocm.on_gfx9()) or (
916
+ if p.is_rocm():
917
+ from vllm.platforms.rocm import on_gfx9
918
+
919
+ is_rocm_on_gfx9 = on_gfx9()
920
+ else:
921
+ is_rocm_on_gfx9 = False
922
+
923
+ device_supports_fp8 = is_rocm_on_gfx9 or (
924
p.is_cuda() and p.has_device_capability((8, 9))
925
)
926
vllm/model_executor/layers/fused_moe/fused_moe.py
+8
-1
@@ -1922,7 +1922,14 @@ class TritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
1922
activation_key: QuantKey | None,
1923
) -> bool:
1924
p = current_platform
1925
- device_supports_fp8 = (p.is_rocm() and p.rocm.on_gfx9()) or (
1925
+ if p.is_rocm():
1926
+ from vllm.platforms.rocm import on_gfx9
1927
+
1928
+ is_rocm_on_gfx9 = on_gfx9()
1929
+ else:
1930
+ is_rocm_on_gfx9 = False
1931
+
1932
+ device_supports_fp8 = is_rocm_on_gfx9 or (
1933
p.is_cuda() and p.has_device_capability((8, 9))
1934
)
1935