[feat] v2 VideoGenerator: typed fastvideo.api entrypoint over the v2 engine
Mirrors fastvideo.entrypoints.VideoGenerator (from_config(GeneratorConfig) -> generate(
GenerationRequest) -> GenerationResult.video_path), reusing the OFFICIAL fastvideo.api config classes
so a basic_dmd_new_api.py-style script differs only by importing VideoGenerator from v2.
- model_path -> v2 card registry (Wan2.1 / FastWan -> wan21; SFWan -> wan_causal; LTX2 -> ltx2);
snapshot_download + stamp_wan21_checkpoints + Engine(cuda) + program; generate maps SamplingConfig
-> DiffusionParams -> make_request -> eng.run, saves the [C,T,H,W] decode as an mp4.
- Lazy v2.__getattr__ keeps 'import v2' torch-free (verified) so the CPU mini stays green (202+2).
- examples/inference/basic/v2_basic_new_api.py runs all three GPU models through this API.
- Single-GPU, resident, TORCH_SDPA (EngineConfig offload/num_gpus>1/VSA accepted for parity, not applied).
verified: wan21 from_config->generate->mp4 (frames (5,256,256,3) uint8, video_path written).
SolitaryThinker
f51497ee6d