From 65b9808960d7f25a018c525d2aca3dba7c411cfb Mon Sep 17 00:00:00 2001 From: Michael Goin Date: Tue, 14 Apr 2026 19:03:57 -0400 Subject: [PATCH] [Bugfix] Disable FlashInfer CUTLASS MoE on SM121 (DGX Spark) (#39825) Signed-off-by: mgoin Co-authored-by: Claude --- .../layers/fused_moe/flashinfer_cutlass_moe.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py b/vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py index 26409804c4..0d47b0f317 100644 --- a/vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py +++ b/vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py @@ -130,7 +130,14 @@ class FlashInferExperts(mk.FusedMoEExpertsModular): p.is_device_capability(90) or p.is_device_capability_family(100) or p.is_device_capability_family(110) - or p.is_device_capability_family(120) + or p.is_device_capability(120) + # NOTE: SM121 (DGX Spark) is excluded because the bf16 + # unquantized CUTLASS MoE GEMM in flashinfer <= 0.6.7 has no + # Relu2 template instantiation and throws "Invalid activation + # type" on Nemotron-H. Fixed upstream by + # https://github.com/flashinfer-ai/flashinfer/pull/2926 + # (merged 2026-04-01, not yet in a stable release); lift this + # restriction once flashinfer >= 0.6.8 is the minimum. ) and has_flashinfer_cutlass_fused_moe() )