SM120 (RTX PRO 6000) MoE backend gaps: MoeConfig.backend schema missing MegaMoE entries, and no EP-capable path for this hardware

[ItamarTerem]

Hi TensorRT-LLM team,
Deploying a large MoE model (Kimi-K2.7-Code-NVFP4, 384 experts) on 8x RTX PRO 6000 (SM120) via the pytorch backend, trtllm-serve, version 1.3.0rc20. Two related findings we wanted to raise:

1. MoeConfig.backend's pydantic Literal doesn't include MEGAMOE_CUTEDSL/MEGAMOE_DEEPGEMM, even though create_moe.py's dispatch logic already handles both as real branches. Setting moe_config.backend: MEGAMOE_CUTEDSL in extra_llm_api_options.yaml fails pydantic validation before create_moe.py is ever reached, even though the underlying backend selection logic supports it. This looks like the config schema hasn't caught up to the dispatch logic yet.
2. More fundamentally: none of the EP-capable MoE backends currently work on SM120 in this build.

• MEGAMOE_CUTEDSL / MEGAMOE_DEEPGEMM: both explicitly require SM100/SM103 ("MegaMoECuteDsl requires SM100 family (SM100 or SM103); got SM120"), and fall back to CutlassFusedMoE.
• WIDEEP: not SM-gated, but its GEMM compute still routes through the same CUTLASS kernel that produces silently incorrect output on SM120 in our testing, and additionally hit a binding mismatch in this build (run_moe() is missing value for argument '_37') before we could evaluate EP itself.
• CUTEDSL (b12x): the only backend we've confirmed produces correct output on SM120, but has no EP support at the flashinfer level (num_local_experts != num_experts raises NotImplementedError in b12x_fused_moe).

Net result: there is currently no correct + EP-capable MoE path for SM120 in this stack, which limits per-GPU memory headroom for large-expert-count models on this hardware class.
Questions:

1. Is SM120 support planned for any of the MegaMoE/WideEP paths, or is SM100/SM103-first prioritization intentional for the near term?
2. Should we file the MoeConfig.backend schema gap as a separate, smaller issue, since that looks like a straightforward fix independent of the larger EP question?
3. Is there a recommended path for SM120 users running large MoE models today, or is CUTEDSL/no-EP the expected configuration for this hardware class in the current release?

Happy to share full repro steps/configs if helpful. Thanks for your work on this — appreciate the pointers.