[niehen6174]

Motivation

ComfyUI integrated mode could load a MiniMax H3 single file, but several H3 workflows either did not run or ran with the wrong settings:

• Distilled models did not run.
  • FastH3 8-Step V2 and VDN-H3 need their native runtime directory, their trained sigma grid and shifts, and the matching step count.
  • Since [Fix] MiniMax H3: honor requested inference step count #38671, num_inference_steps counts DiT forwards, and FastH3 requires exactly 8.
• Sparse attention had no step metadata in the ComfyUI step (VSA-H3, cube, VDN hybrid).
• Audio velocity was scaled twice (time_shift_slope on top of ComfyUI's carried-audio transform), and denoise masks were not applied before that transform as ComfyUI does.
• Some loader settings were silently wrong:
  • A FastH3 file loaded as base H3 failed deep in weight loading.
  • Runtime quantization of a BF16 H3 file was ignored.
  • A renamed local VDN copy loaded as a generic diffusers pipeline.
• Spawned workers re-ran ComfyUI's main.py, whose import utils can resolve to comfy/utils.py and kill the worker before it reports ready.

Changes

Output correctness (executors/minimax_h3.py)

• Drop the extra time_shift_slope on the audio velocity and apply the video / audio denoise masks before the carried-audio transform, matching comfy/ldm/minimax/model.py.

Distilled H3 (executors/minimax_h3.py, core/generator.py, nodes.py)

• fast_h3 / vdn_h3 model types take a runtime_model_path to the native model directory. The plugin checks it, and FastH3's release metadata, before the worker starts.
• One shared executor base checks each step against the trained grid and sends num_inference_steps = len(grid) - 1 (8). FastH3 uses its trained 10/3 shifts; VDN takes its steps and shift from its sampling params. Reference and keyframe tasks the models were not trained for are rejected.
• SGLDH3DistilledSigmas builds the trained grid. A FastH3 single file selected as minimax_h3 is rejected, with a pointer to fast_h3.

Sparse attention in the ComfyUI step (stages/comfyui_step.py, stages/denoising.py)

• The step builds cube, VSA-H3 or VDN-hybrid metadata for its packed rows and passes the same per-step metadata to the BCG capture and the forward.
• _maybe_prepare_vsa_h3_step_metadata takes is_ref2va and the latent shape instead of the native loop context, so both paths share it. The native path is unchanged.
• BCG with layerwise offload is rejected in verify_input.
• SGLDOptions.attention_backend_config (a plain string, parsed like --attention-backend-config) carries the cube sparse-attention settings.

Loader and worker start (core/generator.py)

• For a single file the plugin passes warmup_mode=off and model_id=MiniMax-H3; for VDN, model_id=vdn-minimax-h3, so a renamed local copy still resolves the VDN pipeline.
• Runtime quantization of a floating-point H3 file and native AdaLN caches are rejected before the worker starts. Pre-quantized files load as before.
• Spawned workers no longer re-execute ComfyUI's main.py (_spawn_without_launcher_main).

Two core files change, both in the H3 ComfyUI stage: comfyui_step.py and denoising.py.

Test plan

Unit: test_comfyui_*.py, test_minimax_h3_dit_contract.py, test_conditioning_cache.py: 190 passed. pre-commit is clean.

End-to-end on H200: ComfyUI loads only this plugin tree and the worker imports the same tree; 864x480, 33 frames, euler, fixed seed. Final latents are compared with stock ComfyUI (UNETLoader) where it supports the model.

Workflow	Result	vs stock ComfyUI
MiniMax H3 t2v, INT8 single file, 20 steps	completes	video latent cos 0.982, audio 0.998; same scene
FastH3 8-Step V2 (native runtime dir, trained grid)	completes	no stock equivalent
VDN-H3 8 NFE (renamed local dir, hybrid window attention)	completes	no stock equivalent

The trimmed branch produces bit-identical latents (sha256) to the previous head of this PR on all three workflows.

Not in this PR

• Per-request options (per-request attention backend with a worker restart, cache-dit / Spectrum / skip-softmax request options, arbitrary ServerArgs JSON) were dropped from this PR; if needed they will come back with an allowlist.
• BCG is wired in the ComfyUI step, but SGLDOptions has no switch for enable_breakable_cuda_graph yet.
• Performance (Sage QKV stride reuse, allocator release in ComfyUI mode, per-run RoPE cache) will follow separately with measurements.
• SageAttention3 packed varlen is Blackwell only; we could not verify it on H200.
• Runtime quantization of BF16 H3 files will get its own PR.
• Odd-size H3 layouts are fixed in [diffusion] comfyui: plugin fixes #43493.

---

CI States

Latest PR Test (Base): ✅ Run #38223136143
Latest PR Test (Extra): ❌ Run #38223135877
Latest PR Test (AMD ROCm 10): ❌ Run #38223136172

[niehen6174]

/tag-and-rerun-ci