[bugfix]: drop redundant h3_sequential_load from Spark FastH3 configs

[aryan5v]

Problem

The three Spark FastH3 YAML presets shipped with both lazy_module_load: true and h3_sequential_load: true. On GB10/unified memory, lazy_module_load owns component deferral and explicitly suppresses sequential load (see _defer_denoise_modules in minimax_h3_pipeline.py, fixed in the hao-ai-lab#1803 review). The h3_sequential_load flag in those configs was therefore dead and misleading.

Solution

• Remove h3_sequential_load: true from:
  • examples/inference/basic/basic_fasth3_spark.yaml
  • examples/inference/basic/basic_fasth3_spark_pair.yaml
  • examples/serving/openai_fasth3_spark.yaml
• Document in each file that lazy_module_load is the Spark deferral path; h3_sequential_load remains the H3-only fallback when lazy is off (--no-lazy-module-load).
• Add a regression test in tests/local_tests/test_cookbook_serving.py.

Context

• [perf] Sequential MiniMax H3 start with GPU-direct DiT load hao-ai-lab/FastVideo#1793 introduced h3_sequential_load (load Qwen3-VL → run conditioning → release → load DiT/VAE).
• [feat]: run FastH3 across two DGX Sparks with Ray sequence parallel hao-ai-lab/FastVideo#1803 introduced lazy_module_load (general per-component load-on-use / release-after-stage).
• When both would arm on GB10, lazy wins so VAE torch.compile can attach to the lazy proxy.

Sequential load is not dead code — it is still used on discrete GPUs without lazy load, or when --no-lazy-module-load is passed on Spark.

Testing

• Added test_spark_configs_use_lazy_load_not_sequential (requires dev env with torch/pytest to run locally).

  

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}