[FEAT] Add MiniMax H3 joint audio-video generation

[lzx1413]

Description

This PR adds production-quality MiniMax H3 joint audio-video generation to TeleFuser.

It integrates the released MiniMax H3 DiT, Qwen3-VL text encoder, audio VAE, and video VAE components and supports
T2VA, FL2VA, and Ref2VA from the original local checkpoint partitions. The pipeline produces 24 FPS video with
synchronized 32 kHz stereo audio and supports local Python execution and standard telefuser serve entrypoints.

The PR also adds resident four-GPU inference using DiT SP2+TP2, text-encoder TP4, parallel video-VAE tiling,
optimized Triton/public-ops paths, and direct CUDA tensor handoff between independently spawned worker groups.

Motivation

TeleFuser did not previously support MiniMax H3 or its joint audio-video task contracts.

A faithful integration requires more than loading checkpoints. Processor behavior, ordered heterogeneous
conditioning, packed joint token layout, independent video/audio schedulers, mixed-precision boundaries, VAE
decoding, and final audio-video muxing must remain aligned with the released implementation.

Full 50-step validation also exposed precision-sensitive parameters that must remain FP32 during checkpoint loading.
The integration preserves those values and validates intermediate and final generation boundaries against pinned
SGLang references.

For production multi-GPU use, the initial replicated/offloaded path also left substantial performance on the table.
This PR therefore adds resident TP/SP execution, fused kernels, reusable request-static buffers, parallel VAE tiling,
and worker-to-worker CUDA tensor transport.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change
• Performance improvement
• Code refactoring
• Documentation update
• Other

Changes Made

• Added MiniMax H3 DiT, Qwen3-VL text encoder, audio VAE, and video VAE implementations with strict released-checkpoint conversion and explicit BF16/FP32 boundaries.
• Added canonical T2VA, first/last/first-last FL2VA, and ordered Ref2VA request handling with published material-count and duration validation.
• Added packed joint video/audio denoising, independent modality schedules, condition noise augmentation, synchronized decoding, and MP4 muxing.
• Added one-GPU sequential offload plus resident multi-GPU execution using Ulysses SP, DiT TP, text-encoder TP/FSDP2, and TP-based video-VAE tiling.
• Added the default four-GPU SP2+TP2 DiT, TP4 text encoder, TP4 video VAE, and resident audio VAE profile.
• Added fused BF16 Triton kernels for indexed modulation, Q/K normalization plus NeoX RoPE, SwiGLU, and Ulysses relayout.
• Extended public telefuser.ops dispatch with optional tf-kernel acceleration and native PyTorch fallbacks for unsupported devices, dtypes, and compile mode.
• Added reusable denoising buffers, request-static layouts, optimized scheduler updates, and raw phase, memory, and communication timing metrics.
• Connected text encoder → DiT, visual VAE encoder → DiT, and DiT → video VAE decoder with WorkerTensorChannel, avoiding parent-process and CPU staging for CUDA intermediates.
• Added standardized PPL_CONFIG, get_pipeline, run, and run_with_file entrypoints plus manifests for telefuser serve.
• Added bundled default FL2VA and Ref2VA inputs with provenance metadata.
• Added SGLang/TeleFuser reference runners, loading probes, trajectory capture, artifact comparison, and stable boundary diagnostics.
• Added model, pipeline, distributed, ops, tensor-channel, checkpoint-contract, and parity tests.
• Added support-table entries, example documentation, and a MiniMax H3 integration and validation guide.

Testing

• Unit tests pass (pytest tests/)
• Manual testing performed
• Benchmarks added/updated (if applicable)

Focused automated validation:

• MiniMax H3 model, pipeline, ops, and tensor-channel suite: 110 passed.
• T2VA, FL2VA, and Ref2VA request, packed-layout, scheduler, VAE, and pipeline contracts passed.
• TP/SP/FSDP configuration, checkpoint conversion, dtype preservation, and forward-contract tests passed.
• Artifact and trajectory comparison regression tests passed.
• Commit-time ruff and ruff-format hooks passed for the finalized commits.

Manual H100 validation:

• Frozen official T2VA, FL2VA, and Ref2VA requests using original condition media, seed 0, and release-default 50-step schedules.
• Pinned SGLang and TeleFuser frames, waveforms, latents, and MP4 outputs.
• Processor/encoder, condition VAE, initial noise, packed layout, DiT predictions, scheduler outputs, final latents, and decode boundaries.
• T2VA passed all reference-derived topology checks.
• FL2VA and Ref2VA passed locked frame and waveform thresholds.
• Visual, audio, and conditioning results for all three tasks were manually approved.
• T2VA and FL2VA direct tensor-channel paths completed warmup and measured runs on four H100 GPUs.
• Standard local and serve entrypoints were manually exercised.

Test commands:

.venv/bin/python -m pytest \
  tests/unit/pipelines/minimax_h3 \
  tests/unit/models/test_minimax_h3_audio_vae.py \
  tests/unit/models/test_minimax_h3_dit.py \
  tests/unit/models/test_minimax_h3_encoder.py \
  tests/unit/models/test_minimax_h3_video_vae.py \
  tests/unit/worker/test_tensor_channel.py -q

ruff check \
  telefuser/models/minimax_h3* \
  telefuser/pipelines/minimax_h3 \
  telefuser/ops \
  telefuser/kernel/triton \
  tools/validation/*minimax_h3* \
  tests/unit/models/test_minimax_h3* \
  tests/unit/pipelines/minimax_h3

ruff format --check \
  telefuser/models/minimax_h3* \
  telefuser/pipelines/minimax_h3 \
  tools/validation/*minimax_h3* \
  tests/unit/models/test_minimax_h3* \
  tests/unit/pipelines/minimax_h3

git diff --check origin/main...HEAD

Checklist

• Code follows the project's coding standards (ruff)
• Pre-commit hooks pass (pre-commit run --all-files)
• All tests pass (pytest tests/)
• New tests added for new functionality
• Documentation updated
• Commit messages are clear and descriptive
• PR title follows the convention: [TYPE] Brief description

Commit-time hooks passed, but pre-commit run --all-files and the complete repository-wide pytest tests/ suite
were not rerun after the final changes.

Related Issues

N/A

Additional Notes

• The checkpoint root must contain the original FL2VA and Ref2VA partitions.
• T2VA uses the FL2VA partition without visual conditions.
• The integration targets MiniMax H3 Base and does not imply support for hosted Context-IR or Regenerate-2K services.
• Existing two-step outputs are smoke tests only; production acceptance used the complete 50-step schedule.
• Generated parity and benchmark artifacts under work_dirs/ are intentionally excluded from Git.
• No new environment variables, shared service-schema fields, or public framework configuration fields are introduced.
• Optional tf-kernel acceleration is used only when compatible; public ops retain native fallbacks.
• Bidirectional DiT/VAE CUDA IPC may emit a PyTorch shutdown warning while worker processes exit; generation completes and GPU memory is released.

GPU Architecture Support

• SM80 (Ampere, Ada Lovelace)
• SM90 (Hopper H100)
• SM100+ (Blackwell)

The new fused Triton and resident distributed paths were validated end to end on H100. Native fallbacks exist for
unsupported kernel conditions, but SM80 and SM100+ were not validated as part of this PR.

Performance Impact

This PR adds resident and fused performance paths while preserving correctness and SGLang alignment.

On the frozen 768p, five-second, 50-step T2VA request using four H100 GPUs and matched SP2+TP2 topology:

Backend	Measured wall time	Peak sampled GPU0 memory
SGLang	79.37 s	67.8 GiB
TeleFuser	79.34 s	62.7 GiB

TeleFuser reached practical runtime parity with the pinned local SGLang build while using approximately 7.6% less
sampled peak memory on GPU0 in this run. The TeleFuser DiT denoising phase took 76.41 seconds, including 4.12 seconds
of recorded SP/TP communication.

The final worker tensor-channel change removes parent/CPU staging but did not materially change wall time versus the
immediately preceding TeleFuser result (79.25 s versus 79.34 s); this difference is measurement noise. The
remaining runtime is dominated by DiT computation and per-layer TP/SP collectives.

These measurements are specific to the frozen request, local checkpoints, dependency versions, four-H100 topology,
and warm resident profile. They are not general performance guarantees.