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 5, 04:45 UTC
[Bugfix] Disable FlashInfer CUTLASS MoE on SM121 (DGX Spark) (#39825)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Claude <noreply@anthropic.com>
Michael Goin committed
Apr 14, 2026 at 19:03 UTC
65b9808960d7f25a018c525d2aca3dba7c411cfb
1 file changed
+8
-1
vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
+8
-1
@@ -130,7 +130,14 @@ class FlashInferExperts(mk.FusedMoEExpertsModular):
130
p.is_device_capability(90)
131
or p.is_device_capability_family(100)
132
or p.is_device_capability_family(110)
133
- or p.is_device_capability_family(120)
133
+ or p.is_device_capability(120)
134
+ # NOTE: SM121 (DGX Spark) is excluded because the bf16
135
+ # unquantized CUTLASS MoE GEMM in flashinfer <= 0.6.7 has no
136
+ # Relu2 template instantiation and throws "Invalid activation
137
+ # type" on Nemotron-H. Fixed upstream by
138
+ # https://github.com/flashinfer-ai/flashinfer/pull/2926
139
+ # (merged 2026-04-01, not yet in a stable release); lift this
140
+ # restriction once flashinfer >= 0.6.8 is the minimum.
141
)
142
and has_flashinfer_cutlass_fused_moe()
143
)