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]

Confidence Score: 4/5

The PR is not yet safe to merge because the existing maximum-duration failure remains unresolved.

Summary

• Adds MiniMax H3 conditioning, transformer, scheduler, codec, checkpoint, LoRA, and pipeline implementations.
• Adds shared cacheless attention, synchronized sampling, selective checkpoint loading, and T2V request-input support.
• Migrates the integration to the v2 application/session runtime and adds CPU-focused tests.
• Updates licensing and third-party attribution for adapted MiniMax and Hugging Face code.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  CLI["MiniMax H3 T2V adapter"] --> Request["Validated prompt, duration, keyframes or references"]
  Request --> Condition["Text, image, video, and audio conditioning"]
  Condition --> Joint["Joint video/audio denoising"]
  Joint --> Decode["Video VAE decode"]
  Decode --> Runtime["T2V v2 session output"]

Loading

[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

[yamada-sexta]

I tried to use the built in RoPE, but H3 precomputes cos and sin once, shaped [S, 96], then reuses them across its transformer blocks.

H3 rotates only the first 96 channels of each 128-channel attention head; the final 32 channels pass through unchanged.

Therefore, I cannot use the current RoPE effectively. I wonder if you guys are open to changes in the RoPE kernel?

[ArielG-NV]

Making changes like adding to the core of flashdreams so that future development of world models is easier for future devleopers is very much encouraged.

This would include changes like adding functionality (knobs for performance) or making alternative implementations of RoPE.

[yamada-sexta]

I tested my integration and it seems that the RoPE changes were not really needed.