Add MiniMax H3 integration

[yamada-sexta]

As mentioned here: #443

This commit adds MiniMax H3 integration to flashdreams.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[greptile-apps]

Greptile Summary

The PR adds a separately packaged MiniMax H3 integration supporting text-to-video-with-audio, keyframe-conditioned generation, and ordered media references.

• Registers three MiniMax H3 runner configurations.
• Implements staged conditioning, joint video/audio denoising, decoding, LoRA conversion, and resumable checkpoints.
• Adds package dependencies and smoke coverage for the integration.

Confidence Score: 4/5

The PR is not yet safe to merge because the supported 15-second boundary still aborts during request initialization.

The current alignment logic accepts 15 seconds as within range, advances its 360 frames to the next decodable count of 362, and then rejects the request before generation.

Files Needing Attention: integrations/minimax_h3/minimax_h3/constants.py

Important Files Changed

Filename	Overview
integrations/minimax_h3/minimax_h3/constants.py	Defines geometry and duration validation; the maximum-duration alignment defect remains outstanding.
integrations/minimax_h3/minimax_h3/pipeline.py	Implements staged conditioning, resumable checkpoints, joint denoising, and video decoding for all three workflows.
integrations/minimax_h3/minimax_h3/runner.py	Defines workflow-specific arguments and sequences generation, post-processing, MP4 output, and completion recording.
integrations/minimax_h3/minimax_h3/transformer.py	Implements the native H3 transformer configuration, checkpoint mapping, attention selection, and joint video/audio forward path.
integrations/minimax_h3/pyproject.toml	Packages the integration and registers its three runner entry points.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Request[MiniMax H3 request] --> Validate[Validate workflow, canvas, and duration]
  Validate --> Condition[Encode prompt and optional media]
  Condition --> Prepare[Prepare video and audio latent streams]
  Prepare --> Denoise[Joint transformer denoising]
  Denoise --> Checkpoint[Persist latent checkpoint]
  Checkpoint --> Decode[Decode video frames]
  Decode --> Output[Write MP4 and completion status]

Loading

_{Reviews (4): Last reviewed commit: "Optimize MiniMax H3 GPU handoff" | Re-trigger Greptile}

[greptile-apps]

Maximum duration fails alignment

When duration is 15 seconds, alignment advances 360 frames to the next decodable count of 362 and then rejects the request, causing the advertised maximum duration to fail before generation starts.

Suggested change

	if frames / FPS > MAX_DURATION:
	raise ValueError("duration aligns beyond MiniMax H3's 15-second maximum")

[greptile-apps]

Exceptional exits leak output target

When generation, finalization, processing, or writing raises after output_target.open(), the runner exits without closing the target, leaking encoder and file resources and leaving a partial MP4 that can interfere with retries.

Knowledge Base Used: Integrations

[yamada-sexta]

Tested with both the regular MiniMax H3 and the the Looping-Sketch LoRA on A100.

No audio support yet.

[ArielG-NV]

questions/suggestions

[ArielG-NV]

we could make this async since otherwise we are blocking generate just to generate a cache file we don't immediately need every time we have a new cache?