[Doc] diffusion: Document reference resizing, LoRA auto mode and image profiling

[mickqian]

Motivation

Several recently merged diffusion controls are missing from, or incompletely described in, the existing public documentation.

Modifications

• Document MiniMax-H3 Ref2VA reference resizing policies, their scope and rounding behavior, and the ModelTC Turbo adapter's recommended policy ([Diffusion] Add configurable reference image resizing for MiniMax H3 Ref2VA #35839).
• Clarify when automatic LoRA merging keeps an adapter dynamic because of estimated rounding loss ([diffusion] fix: keep a LoRA unmerged under merge_mode auto when merging would round its update away #43385).
• Add per-request profiling examples for image generation and multipart image editing, including server-side trace placement and measurement caveats ([diffusion] Accept request-level profiling on the image API #43573).

The changes stay in the existing cookbook, CLI reference, and profiling guide. No README model section, navigation restructuring, or deprecated documentation is added.

Accuracy Tests

Documentation only; runtime behavior is unchanged.

Speed Tests and Profiling

No performance change is claimed.

---

CI States

Latest PR Test (Base): ✅ Run #38672452060
Latest PR Test (Extra): ❌ Run #38672452087
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
sglang-doc	🟢 Ready	View Preview	Oct 15, 2026, 2:47 AM

💡 Tip: Enable Automations to automatically generate PRs for you.