feat(minimax-h3): invocations, denoise loop, and audio-video decode (PR 3/5)

[lstein]

Summary

PR 3 of the MiniMax H3 stack (on top of #139): the six generation nodes, the packed-row denoise loop, and audio+video decode. With this PR a hand-built graph in the Workflow Editor produces a playable MP4 with an AAC stereo soundtrack — no linear UI yet (that's PR 4).

Backend modules (invokeai/backend/minimax_h3/)

• packing.py — vendored from diffusers PR #14355 @ abc5e9bf71 (packed-sequence geometry + checkpoint constants; byte-identical apart from import rewrites).
• sampling.py — denoise-state construction: packed layout (condition rows lead), two MiniMaxH3Schedulers (video shift 12.0 / audio shift 3.0), one generator with the upstream draw order (keyframe → video → audio), frame-grid validation (num_frames % 17 == 5, 5–15 s window, plus exactly 5 frames as the still-image mode).
• denoise.py — the joint audio-video loop: one forward per step (guidance-distilled, no CFG), per-row timestep plans, keyframe rows pinned at t=0.999, condition rows never stepped.
• text_conditioning.py — Qwen3-VL-32B prompt encoding (unnormalized layer-50 hidden states; keyframes enter the vision context via the processor).
• keyframe_conditioning.py — keyframe VAE encoding (seed-42 posterior sample, fp16 rounding contract, stretch/cover canvas placement in packed order).

Nodes (all Prototype)

minimax_h3_model_loader, minimax_h3_text_encoder (idle-GPU-offloadable), minimax_h3_frame_conditioning, minimax_h3_denoise, minimax_h3_latents_to_video (video VAE decode + audio VAE decode → WAV → AAC mux via PR 1's make_mp4_writer), minimax_h3_latents_to_image (single-frame extraction = txt2img path).

Plus plumbing: MiniMaxH3ConditioningInfo (+ serializer safe-globals), fields/primitives/metadata generation modes, grayscale 24-channel latent preview fallback, H3 VAE working-memory estimator.

Deliberate divergences from upstream

• Non-grid num_frames values are rejected, not rounded (InvokeAI validator convention). Effective max is 345 frames = 14.375 s.
• Exactly 5 frames is allowed as a still-image mode (upstream enforces 5–15 s always).
• fp16-autocast decode (upstream's verified recipe) is used on CUDA only; ROCm/MPS/CPU decode in the weights' dtype.
• Latent previews are grayscale until RGB factors are calibrated against real weights (TODO noted in step_callback.py).

Adversarial review fixes (already in)

• The decode node no longer toggles enable_tiling()/disable_tiling() on the shared cached VAE — the H3 VAE tiles by default and the released frames are the blended-tile ones; toggling would have silently changed every later encode/decode of the cached instance.
• The VAE working-memory estimator now models the always-tiled decode (per-tile working set + on-device output clip) instead of a full-frame decode that never happens; the decoded clip moves to the CPU before the full-clip pixel math.
• The denoise node cross-checks keyframe coupling: the prompt's vision-context keyframes (anchors + canvas, now carried on MiniMaxH3ConditioningInfo) must match the frame-conditioning input.
• steps is ge=2 (the scheduler's contract; N grid points = N−1 model evaluations).
• Audio is zero-padded as well as trimmed to the exact video duration; temp-WAV cleanup on write failure.

Testing

• tests/backend/minimax_h3/: grid math, draw-order determinism, full denoise loop on a micro-config CPU transformer (anchor preservation, cancellation, callback counts).
• Full sweep tests/app + tests/backend: 2871 passed / 0 failures; post-review-fix sweep of tests/backend + tests/app/util: 1712 passed.
• ruff clean, pnpm lint:tsc green, openapi.json/schema.ts regenerated.
• Real-weights end-to-end (playable MP4 with audible audio) still needs a manual run — that's this PR's test gate.

🤖 Generated with Claude Code

[lstein]

Amended in 727f736a2f: the denoise node now estimates transformer activation memory from the packed sequence length (0.25 MiB/row + 5·seq² when padding rows force the boolean SDPA mask + 2 GiB fixed) and passes it as working_mem_bytes to model_on_device() — without it, partial loading packed the device with weights and the first forward OOMed (reproduced at 124 frames / 768×1344 on a 45 GiB card: 41.6 GiB weights + ~3.4 GiB free vs ~7.5 GiB needed). Same mechanism as the Krea-2 denoise / VAE nodes. Formula pinned by a unit test; adversarially reviewed (the seq² mask term budgets SDPA's convert_boolean_attn_mask additive copy + alignment transient, not just the bool mask). Known limitation shared with every estimator in the codebase: a user-set max_cache_vram_gb bypasses working-memory reservations entirely (model_cache.py:_get_vram_available returns before consulting it) — worth a separate upstream discussion.