Does this require VSA support?
No, this is the dense version that doesn't use VSA. Basic conversion from the officially released lora https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA/blob/main/dense-datafree/adapter_model.safetensors
What is the recommended setting?
also got these warning
[WARNING] lora key not loaded: diffusion_model.time_embedder.proj_in.diff
[WARNING] lora key not loaded: diffusion_model.time_embedder.proj_in.diff_b
[WARNING] lora key not loaded: diffusion_model.time_embedder.proj_out.diff
[WARNING] lora key not loaded: diffusion_model.time_embedder.proj_out.diff_b
Also it doesn't work with 4 steps; 6 is the minimum.
tested with fl2va_pruned_int8_convrot
you can ignore the warnings, it's residual layers for non-pruned versions I forgot to remove.
regarding the settings, the official parameters from FastH3 card are: 4 steps, CFG 1, euler/simple, 12 video shift, 3 audio shift.
Tho in my testings, 4 steps is okay-ish for video but terrible for audio, I recommend 8 steps for decent audio but then others 8 steps lora (lightx2v and alibaba ones) become more attractive.