- ComposedPipelineBase owns a ComponentState: it records each component's
path after loading and attaches the state to every stage it registers.
from_pretrained(model_path, *, resolved_config, ...) takes a resolved
config instead of building FastVideoArgs from keywords or argparse; the
training_args alias and TrainingArgs checks are gone, and the training
DiT offload is a recorded override. Compile settings come from
engine.compile and torch_compile_kwargs(); LoRA settings from
pipeline.components.lora_* and training.lora.*.
- The worker and executors read engine.num_gpus, engine.parallelism.*,
engine.execution_backend, and pipeline.output_type; the Ray executor owns
its placement group and runtime env.
- VideoGenerator builds from a resolved config only (_from_resolved_config);
the prompt_txt fallback is gone (requests carry inputs.prompt_path). The
OpenAI server state holds the resolved config (get_resolved_config()).
- Tests use resolved configs; the SSIM and performance helpers map their
settings to dotted typed paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7