[MUSA] Use MATE Mutlass backend for MAGI-2 FA3 consumer

[yeahdongcn]

MUSA FA3 consumer for MAGI-2 Preview

Draft child PR for the MUSA consumer side of MAGI-2 attention. This is based on the shared FA3 varlen adapter in #7249. Please merge/review #7249 first, or treat this PR as a stacked Draft. The historical #7156 branch remains unchanged.

Scope

• On MUSA, use the MATE/torchada wrapper through its default auto backend. Normalize packed LSE to Omni's [H,T] contract before MAGI-2 sink correction.
• Normalize Omni's negative softcap sentinel to MATE's non-negative convention in the direct-provider compatibility path.
• Use the exact Torch reference only for query sequences shorter than 32 tokens, where MATE's fast-exp2 approximation accumulates error across repeated tiny layers. Production MAGI-2 sequences remain on FA3; the six operator cases force FA3 independently.
• Preserve CUDA/ROCm/CPU dispatch. No EP, MoE kernel, mHC, sampler, launch-tile, dependency or performance-tuning changes.

Validation

Published commit: a9921de (full SHA verified on the fork). The current source includes dynamic MATE LSE ABI detection, provider-default selection, FP32 sink correction and the explicit short-sequence policy. Earlier pre-ABI runs are retained only as diagnostic evidence.

• 17 CPU consumer tests passed; the generic adapter backend-forwarding change has 24 passing CPU tests.
• Final MUSA run: 7/7 tests passed, 0 skipped. Six sink/LSE operator oracles exercise the provider-default FA3 path; the tiny-transformer case verifies the exact short-sequence fallback and that the provider is not called. Test runtime was 287.62 s, dominated by first-run MATE compilation, and is not a benchmark.
• Hardware: one60-SM MTT S5000, driver3.3.8-server (not pinned), pinned image registry.mthreads.com/mcconline/inference/vllm-omni@sha256:7f4f2cd83a88979025ef92968334f6e075eb24d45d02e0c30dbb222f7780d58a.
• Stack: torch/torch_musa2.11.0.post1+musa5.2.0, torchada0.1.83, MATE0.2.6, flash_attn_3 0.2.6+musa, Triton3.2.0, vLLM0.28.0.
• Targeted pre-commit passed. Source/import and current-lease container checks passed; container exited0 and was removed. No GPU remains leased.

The generic adapter does not expose a mandatory backend parameter. This consumer only adapts installed MATE ABI variants (return_softmax_lse versus return_attn_probs) and LSE layout; backend policy stays with MATE's auto default. The tiny-transformer gate is not relaxed: it uses the documented exact fallback for this short-sequence regime. The six passing operator oracles do not imply full-model FA3 accuracy for long sequences.

Remaining gates

No full272p/10s checkpoint, multi-rank EP, graph/capture, memory-capacity or DiT step performance validation was performed. The historical consumer and #7156 branches remain unchanged. Backend numerical improvement is not a performance claim; full-model accuracy and <=0.7s/step remain open. This PR stays Draft.

[hsliuustc0106]

This PR touches tests/diffusion/, vllm_omni/diffusion/ (4 files). Based on CODEOWNERS coverage of the changed files, the most-related reviewers appear to be:

@wtomin @yenuo26 @NickCao

Could one of you take a look when you get a chance? Thanks!

[yeahdongcn]

Superseded by #8498. The MUSA FA3 consumer there keeps MATE's default backend selection.