Add H3 support into Dreamverse

[Davids048]

Purpose

Add FastH3 Preview support to the DreamVerse dreamverse-server runtime while preserving the existing DreamVerse
WebSocket, session, GPU-pool, and AV-streaming flow.

Operators select the model with DREAMVERSE_MODEL_ID=fast-h3. DreamVerse then runs FastVideo's Python
VideoGenerator inside its GPU worker; this integration does not route requests through the fastvideo serve HTTP
server.

Changes

Model-specific generation ownership

• Add a shared GenerationBackend contract and StepResult media contract for DreamVerse GPU workers.
• Add VideoGenerationWorker as the model-selection boundary for LTX-2 and MiniMax H3.
• Rename the LTX-2 implementation to ltx2_generation.py so each backend has an explicit model-specific owner.

FastH3 integration

• Register the fast-h3 DreamVerse model profile for MiniMaxAI/MiniMax-H3.
• Load vsa-datafree/adapter_model.safetensors from
  FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA.
• Build the FastH3 generator through FastVideo's typed Python configuration with four-way sequence parallelism,
  MiniMax H3 Video Sparse Attention, inference fusions, and parallel VAE decode.
• Generate synchronized 124-frame, 768x1344 video and audio segments.
• Retain each segment's last decoded frame for MiniMax H3 first-frame continuation, and trim the repeated leading
  video frame and matching audio duration before streaming.
• Warm both text-to-video and first-frame-to-video paths before the DreamVerse readiness endpoint reports success.
• Keep the FastH3 startup adapter fixed and reject DreamVerse runtime LoRA mutation for this profile.

Runtime and loader behavior

• Default the DreamVerse GPU allocation to the active model's sequence-parallel size.
• Use the active model profile for worker initialization and model reloads, and clear model ownership after a failed
  reload.
• Move CPU-loaded dense adapter deltas to the base tensor's device before applying them. This supports FastH3
  checkpoints that load directly onto CUDA.
• Update DreamVerse architecture documentation, launch instructions, and GPU-pool diagrams.

Test Plan

pytest apps/dreamverse/dreamverse/tests -q
pytest fastvideo/tests/loader/test_lora_patch.py -q

Manual end-to-end validation:

CUDA_VISIBLE_DEVICES=0,1,2,3 \
  DREAMVERSE_MODEL_ID=fast-h3 \
  DREAMVERSE_SP_SIZE=4 \
  FASTVIDEO_GPU_COUNT=4 \
  dreamverse-server --host 127.0.0.1 --port 8009

The manual run loaded the VSA data-free adapter, completed text-to-video and first-frame-to-video warmup, and
streamed six consecutive AV segments through the DreamVerse UI.

Test Results

• DreamVerse backend suite: 99 passed.
• Dense LoRA loader suite: 27 passed.
• Applicable pre-commit hooks passed: YAPF, Ruff, codespell, PyMarkdown, filename validation, and suggestion checks.
• Local mypy was skipped because the hook rejects the hyphenated worktree directory name
  fastvideo-dreamverse-h3; CI remains the authoritative mypy result.
• SSIM was not run because this change integrates an existing FastVideo model with DreamVerse streaming and does not
  define a DreamVerse golden-video fixture.

Checklist

• I ran pre-commit run --all-files and fixed all issues.
• I added or updated tests for my changes.
• I updated documentation if needed.
• I considered GPU memory impact of my changes.

For model/pipeline changes:

• I verified SSIM regression tests pass. Not applicable: DreamVerse has no FastH3 SSIM fixture.
• I updated the relevant DreamVerse model support documentation.

[mergify]

⚠️ PR title format required

Your PR title must start with a type tag in brackets. Examples:

• [feat] Add new model support
• [bugfix] Fix VAE tiling corruption
• [refactor] Restructure training pipeline
• [perf] Optimize attention kernel
• [ci] Update test infrastructure
• [infra] Add activation trace hooks
• [docs] Add inference guide
• [misc] Clean up configs
• [new-model] Port Flux2 to FastVideo
• [skill] Add add-model agent skill

Valid tags: feat, feature, bugfix, fix, refactor, perf, ci, infra, doc, docs, misc, chore, kernel, new-model, skill, skills

Please update your PR title and the merge protection check will pass automatically.

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews, 🤖 CI and 🙋 you

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews, 🤖 CI and 🙋 you

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=full-suite-passed
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]

This rule is failing.

• #approved-reviews-by>=1
• check-success=full-suite-passed
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]
• check-success=fastcheck-passed
• check-success~=pre-commit

[shaoxiongduan]

Is there a reason to use the lora weights instead of the full finetuned weights for this?