[perf] Sequential MiniMax H3 start with GPU-direct DiT load

[aryan5v]

Purpose

MiniMax H3 CUDA cannot keep Qwen3-VL resident with the DiT and VAEs on GB10 unified memory. Host offload is already disabled there.

This PR encodes first, releases the encoder, then loads DiT and VAEs when sequential load is on. DiT safetensors go onto the accelerator when cpu_offload is off (no CPU-then-copy).

Changes

• Add h3_sequential_load (--h3-sequential-load / --no-h3-sequential-load). Default auto: on unified-memory devices, off on discrete GPUs.
• Defer vae, audio_vae, and transformer until after conditioning when sequential load is on.
• Read DiT weights with to_cpu=cpu_offload instead of always to_cpu=True.
• Document both in offloading and Spark GB10 docs.

A second generate() on a sequential-load worker still needs a new process. Training load order and LoRA wiring are unchanged.

Test Plan

python -m pytest fastvideo/tests/stages/test_minimax_h3_sequential_start.py \
  fastvideo/tests/inference/test_basic_fasth3_profile.py \
  fastvideo/tests/loader/test_fsdp_load_releases_checkpoint.py -q

Spark GB10, FastH3 4-step Preview VSA-DataFree, basic_fasth3.py --num-gpus 1 --vsa-kernel triton --no-fa4 --no-warmup --repeats 1, 768×1344×124, seed 2026, n=1.

Test Results

CPU: 17 passed for sequential-load + FastH3 profile tests. pre-commit on changed files: passed.

Same Spark box, sequential alpine. Before = CPU-then-copy DiT load. After = GPU-direct DiT load.

	Before	After
DiT weight load	445 s	39 s
Denoise	164.1 s	139.1 s
Video VAE decode	76.6 s	68.0 s
Generate	768.5 s	333.1 s
E2E	771.9 s	336.1 s

VAE load stays ~75–80 s (different loader). Denoise/decode are n=1 and include first-run Triton. Both runs wrote a 4.4 MiB MP4. SSIM not run.

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, also check:

• I verified SSIM regression tests pass
• I updated the support matrix if adding a new model

[mergify]

Merge Protections

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

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

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=full-suite-passed

This rule is failing.

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