[diffusion] Add MiniMax-H3 to ComfyUI integrated mode

[niehen6174]

Motivation

sgl-project#35960 turns ComfyUI integrated mode into a DiT-only profile on the native pipeline (adapter + checkpoint spec). Flux / Z-Image / Qwen already sit on that path. MiniMax-H3 does not: it is a packed [video, audio] DiT with T2V, I2VA, FL2VA, and R2V layouts that ComfyUI already ships as community graphs.

This PR adds H3 to that same contract so ComfyUI keeps CLIP / VAE / the sampler loop, and SGLang runs one packed DiT forward per apply_model.

24GB skip-step speed-ups are not in this stack. They land in #35858 (Cache-DiT + layerwise) and #35684 (Spectrum). After those merge, a follow-up will expose the recommend recipes on SGLDOptions.

Modifications

• H3 adapter + worker step — pack/unpack [video, audio], translate ComfyUI layout / RoPE / denoise mask, match the official audio sigma-shift ODE.
• Checkpoint spec — register MiniMax-H3; GGUF overlay; read AdaLN curve from pruned BF16 so the DiT is not built as the unpruned MLP.
• Session / generator — cache scoped to one sampler.sample(); one SGLD generator per process (two-pass upscaler and back-to-back H3 graphs).
• TP — shard fused MergedColumn per logical matrix.
• Shipped graphs — T2V, R2V, two-pass latent upscaler.

Workflows

ComfyUI R2V UI：

Shipped under apps/ComfyUI_SGLDiffusion/workflows/:

Graph	Task
minimax_h3_t2v_sgld.json	T2V / fl2va DiT
minimax_h3_r2v_sgld.json	Reference-to-video
minimax_h3_t2v_sgld_upscaler.json	Low-res then 3D ×2 refine

Also tested on this branch (H3 only):

Task	Result
I2VA (first frame) / FL2VA (first+last)	OK
R2V image / R2V video+audio	OK (needs the ref2va checkpoint)
kitchen_int8 + layerwise	OK
pruned BF16 + layerwise	OK
GGUF Q4 overlay	OK
sol_attn / sage_attn	OK
Cache-DiT	OK (24GB + layerwise coexist waits on sgl-project#35858)
lightx2v turbo LoRA on BF16	OK
Same UNET sampled twice	OK
GGUF × LoRA	Rejected (expected)

GGUF × LoRA is fail-closed. kitchen-int8 × LoRA is not. H3 SP needs ulysses_degree=2.

Tests

Unit: test_comfyui_h3.py (layout vs ComfyUI, adapter, step kwargs, profile).

4090 D 24GB · ComfyUI 0.30.0 · CLIP on CPU. Wall is prompt e2e.

Smoke (432×240 · 25f · 4 step) unless noted:

H3 path	Wall
kitchen_int8 + sol + layerwise	96 s
pruned BF16 + layerwise	104 s
R2V ref2va + kitchen	145 s
Two-pass upscaler 432→864	217 s

768P quality (1344×768 · 5s · 20 step · kitchen_int8 + FA + layerwise):

Task	Wall
I2VA	418 s
FL2VA	426 s
R2V image	609 s
R2V video+audio	1433 s

Test Video：

i2va_ramen_first.mp4

ref2va_girl.mp4

ref2va_video.mp4

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.