[renyongxing]

Stacked on #6711 — review only the 4 files below; the rest of the diff is #6711's own content, which is already under review there. This PR will be rebased onto main once #6711 lands.

Benchmark tooling for MixFusion, plus the profiling knob it needs. Contents:

• benchmarks/diffusion/mixfusion_kernel_benchmark.py — synthetic HunyuanImage-3.0-shaped DiT stack (RMSNorm, fused QKV, SwiGLU MLP). Times independent / padded / MixFusion layouts of the same sequences and reports correctness, latency, peak memory and relative work. Loads no checkpoint, so it isolates the layout effect and runs in seconds.
• benchmarks/diffusion/mixfusion_benefit.py — real-pipeline A/B: one-request-at-a-time vs the same prompts admitted concurrently, parameterized over --family {qwen,hunyuan}. Replaces the two near-duplicate per-family scripts from [Perf][Diffusion][Qwen-Image] Replace per-bucket varlen loop with flat varlen attention for mixfusion #6711.
• benchmarks/diffusion/README.md — a ## MixFusion benchmarks section covering all three entry points, including an online-serving recipe that reuses the repo's own diffusion_benchmark_serving.py via extra_args (no new serving code).
• vllm_omni/profiler/omni_torch_profiler.py (+3) — OMNI_PROFILER_ACTIVITIES lets the pipeline profiler's activity list be narrowed from the default CPU,CUDA. Nothing in the tree sets it; unset, behaviour is byte-identical to main.

Size exemption request

This diff is 1071 non-test added lines, 71 over the 1000-line budget, so stating a concrete exemption as the review comment invited:

File	Lines
mixfusion_benefit.py	597
mixfusion_kernel_benchmark.py	392
README.md	79
omni_torch_profiler.py	3

There is no production code path here. No model, scheduler, attention backend or request path is touched — the only non-benchmarks/ file is a 3-line profiler env override. The lines are measurement code and its documentation, which cannot be meaningfully shrunk further: the two per-family scripts from #6711 were already consolidated into one parameterized script (753 → 597 lines), and the synthetic benchmark was reviewed line by line and is dense by design. Removing documentation or the Qwen small-GCD pre-check to fit the budget would make the benchmark less trustworthy, not smaller in any useful sense.

Validation status

• ruff check and ruff format --check pass; both scripts carry the correct SPDX header.
• The scripts' pure helpers (argument parsing and per-family defaults, seed semantics, request construction, extra_args contents, engine config, output aggregation) are covered by 55 assertions.
• Not run end-to-end: no checkpoint was loaded and no real generation was measured on the hardware this targets. The scripts are validated as code, not as reported numbers.

🤖 Generated with Claude Code

[hsliuustc0106]

This PR touches vllm_omni/diffusion/, benchmarks/diffusion/, tests/diffusion/, .gitignore, vllm_omni/profiler/ (17 files). Based on CODEOWNERS coverage of the changed files, the most-related reviewers appear to be:

@Isotr0py @princepride @SamitHuang

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