Not working on dgx spark with vllm
Not working on dgx spark with vllm
FLASHINFER_TRTLLM (trtllm-gen) is SM100-only (datacenter B100/B200). On your SM120 RTX PRO 6000, is_supported_config() returns False and it gets eliminated.
No backends remain → NotImplementedError.
Note that FLASHINFER_CUTLASS does support both the SwiGLU clamp and SM120 — but it's simply missing from the NVFP4_BACKENDS_WITH_CLAMP set.
Fix Status
Artifact Status
PR #41834(opens in new tab) — adds FLASHINFER_CUTLASS to the clamp set + full SM12x enablement for DeepSeek V4 Flash Still open (needs rebase, last updated June 1, 2026)
PR #42209(opens in new tab) — NVFP4 MoE support for DeepSeek V4 (merged May 22, 2026) Merged, but didn't fix the clamp-set issue
Same issue. Have 2x DGX Spark clustered
It is better to use dspark version
Your diagnosis matches the fix in vllm#41834 (FLASHINFER_CUTLASS missing from NVFP4_BACKENDS_WITH_CLAMP) — that PR is still open but active as of today. Two things that may save you time meanwhile: on DGX Spark (sm_121) the flashinfer side also needs the topk instantiation fix in flashinfer#4309, and if you're open to sglang instead of vLLM, the Blackwell-workstation path works today — our 4x RTX PRO 6000 recipe with the full patch list is at https://github.com/ombori/deepseek-v4-flash-0731-sglang-4x-rtx-pro-6000 (sm_120, so verify the pins on sm_121 before trusting them). If DSpark comes into play anywhere: set the draft depth explicitly, the checkpoint-default 5 is the one value we measured corrupting (sglang#33800).