[Raymondlol]

Motivation

Closes #34954. validate_quality_deployment pins sixteen server args and the
device, but none of them observes which DiT weights are resident, so
quality: "high" keeps its audited SSIM/PSNR promise attached to weights it was
never measured on. Two CLI args substitute those weights:

• --lora-path — on the audited 4×H200 deployment, with the cookbook's own
  Turbo-LoRA recipe merged, quality: "high" completes at the audited shape in
  42.98 s against the 43.28 s the same tier takes on base weights. The LoRA
  recipes section says that combination "has not been quality-validated".
  Repro in [diffusion] MiniMax-H3 quality="high" runs on Turbo-LoRA-merged weights — the audit gate has no adapter field #34954.
• --transformer-weights-path — quantization: None does not cover this. That
  field carries only the explicit --quantization override, whose help states
  the method is otherwise auto-detected from the checkpoint, and nothing writes
  the detected method back onto server_args.
  resolve_transformer_safetensors_to_load substitutes the entire transformer
  weight set whenever the path is set; when no quant config resolves,
  TransformerLoader warns and proceeds.

Modifications

Add both args to the actual/expected pair, following the existing
one-line-per-field shape, so the gate fails closed on a substituted DiT.
Extend _quality_server_args() and add a parametrized test per arg.

This affects only quality: "high". MiniMaxH3PartitionAdmissionStage.forward
calls the validator under if high_quality and not batch.is_warmup, so lossless
serving with a LoRA and the warmup path are unchanged.

The docs warning names distilled adapters, while lora_path: None also rejects
style adapters like the Realism-People LoRA that keep the 50-step schedule. That
is deliberate. The audited figure was measured on base weights and the Cache-DiT
policy keys on residual drift, so any merged adapter moves what was measured, and
the server cannot tell the two kinds apart from the path. Allowing non-distilled
adapters would want an explicit allowlist rather than an unchecked field.

model_path is deliberately out of scope. It has no single expected value the
way the other fields do: the cookbook gives a root model ID per hub
(MiniMaxAI/MiniMax-H3 on Hugging Face, MiniMax/MiniMax-H3 on ModelScope),
and a repo ID on its own does not fix the content without a pinned revision.
Base-model identity needs that revision pin or a weight digest, which is a
separate decision. component_transformer_weights_paths is left alone as well:
it has no CLI flag and I did not find an H3 path that populates it.

Accuracy Tests

N/A — admission-time validation only; no model output or kernel path changes.

test_minimax_h3_admission.py passes 16/16 with this change. Reverting only the
config change and keeping the new test fails exactly the two new parametrizations
(DID NOT RAISE ValueError) with the other 14 still passing, so the test covers
the gap rather than restating the fix.

Speed Tests and Profiling

N/A — one dict comparison per quality: "high" batch admission.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Docs are left out on purpose: the cookbook wording that this now enforces sits in
the same .mdx that #34841 and #34909 are both editing. Happy to follow up
separately once those land.

---

This touches the same function as #34841, which moves the topology fields into
per-profile updates. Both fields here are topology-independent and stay in the
shared literal, so the two changes are disjoint; happy to rebase whichever merges
second.

I don't have permission to add the run-ci label, so pr-gate will be red until
someone applies it — grateful if a maintainer can.

---

CI States

Latest PR Test (Base): ✅ Run #37272691758
Latest PR Test (Extra): ❌ Run #37272691509
Latest PR Test (AMD ROCm 10): ❌ Run #37272691716

[mickqian]

/tag-and-rerun-ci