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, first/last-frame conditioning, reference conditioning, native denoising, LoRA conversion, checkpoint recovery, and plugin registration.

• Adds MiniMax H3 pipeline, transformer, scheduler, model, runner, reference, and LoRA implementations.
• Registers three runner workflows and adds CPU/CUDA integration tests.
• Updates workspace dependencies and adds RoPE prefix-view coverage.

Confidence Score: 3/5

The PR is not yet safe to merge because the maximum advertised duration is rejected and exceptional runner exits still fail to close the MP4 output target.

A 15-second request deterministically aligns to 362 frames and fails validation, while exceptions during generation, finalization, or output processing propagate past the sole output-target close call; both previously reported blocking failures remain in the current code.

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

Important Files Changed

Filename	Overview
integrations/minimax_h3/minimax_h3/constants.py	Defines geometry and duration validation, but the advertised 15-second boundary remains unreachable because frame alignment advances it beyond the maximum.
integrations/minimax_h3/minimax_h3/runner.py	Implements the three workflow runners, but exceptional exits after opening the MP4 target still bypass target cleanup.
integrations/minimax_h3/minimax_h3/pipeline.py	Implements staged conditioning, checkpoint recovery, native denoising, and decoding for all MiniMax H3 workflows.
integrations/minimax_h3/minimax_h3/model.py	Adds paired video/audio scheduler execution and reconstruction of denoised video latents.
integrations/minimax_h3/minimax_h3/transformer.py	Adds the native MiniMax H3 transformer implementation and checkpoint configuration.
integrations/minimax_h3/minimax_h3/lora.py	Adds validation, conversion, caching, and merging of Musubi-format MiniMax H3 LoRA adapters.
integrations/minimax_h3/pyproject.toml	Packages and registers the three MiniMax H3 runner configurations as FlashDreams entry points.
integrations/minimax_h3/tests/test_smoke.py	Adds CPU smoke coverage for plugin registration, configuration, LoRA conversion, pipeline behavior, and recovery helpers.
flashdreams/tests/test_rope_kernel.py	Adds coverage proving in-place RoPE operation on a sliced rotary prefix without changing trailing channels.
uv.lock	Incorporates the new workspace package and its resolved dependency graph.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  CLI[flashdreams-run] --> Registry[Runner plugin registry]
  Registry --> Runner[MiniMax H3 runner]
  Runner --> Condition[Text, keyframe, or reference conditioning]
  Condition --> Checkpoint[Conditioning checkpoint]
  Checkpoint --> Denoise[Native joint video/audio denoising]
  Denoise --> Latents[Latent checkpoint]
  Latents --> Decode[Video decode]
  Decode --> Output[MP4 output target]
  Output --> Complete[Completion status]

Loading

_{Reviews (8): Last reviewed commit: "Merge remote-tracking branch 'upstream/m..." | 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

[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.