vllm - 💡(How to fix) Fix [Bug]: FLASHINFER_CUTLASS_MXFP4_MXFP8 produces wrong output under expert parallelism

Your current environment

Your output of `python collect_env.py` here

🐛 Describe the bug

The post-all-to-all swizzle in the EP/DP-EP prepare-finalize wrappers is gated on quant_dtype == "nvfp4", which means mxfp8 activation scales arrive at the CUTLASS kernel unswizzled. The CUTLASS MoE kernel requires swizzled scales, so outputs are garbage.

This is a pre-existing limitation surfaced by review of #42089, not a regression from that PR.

Affected lines (gate currently quant_dtype == "nvfp4" only):

• vllm/model_executor/layers/fused_moe/prepare_finalize/naive_dp_ep.py:62
• vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_one_sided.py:132-135
• vllm/model_executor/layers/fused_moe/prepare_finalize/flashinfer_nvlink_two_sided.py:197

Repro (GPQA Eval times out at 1800 s; ~6/1584 items in 11 minutes):

VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS=1 \
  vllm serve openai/gpt-oss-20b \
    --data-parallel-size 2 --enable-expert-parallel --gpu-memory-utilization 0.4
# then run gpt_oss.evals gpqa against it

PF in use: MoEPrepareAndFinalizeNaiveDPEPModular. Backend confirmed: FLASHINFER_CUTLASS_MXFP4_MXFP8.

Why the obvious fix isn't quite right

Just expanding the gate to include "mxfp8" and reusing nvfp4_block_scale_interleave would be wrong — that helper produces nvfp4-specific layout. mxfp8 needs swizzle_mxfp8_scale from vllm/model_executor/layers/quantization/utils/mxfp8_utils.py:14, which produces F8_128x4 layout, and needs (M, K) dims at the call site.

Proposed direction (also tracked separately): replace the orthogonal quant_dtype + is_scale_swizzled flag with merged dtype variants ("nvfp4_swizzled" / "mxfp8_swizzled") so the dispatcher can route to the right swizzle helper from a single field, removing this bug class entirely.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.