vllm_omni.diffusion.models.minimax_h3.temporal_chunks ¶

Small compatibility loop for the released MiniMax-H3 temporal decoder.

decode_temporal_chunks ¶

decode_temporal_chunks(
    model: Module,
    latent: Tensor,
    callback: Callable[[Tensor], None] | None,
) -> Tensor

Decode H3 temporal clips and synchronously publish committed frames.

The released VAE exposes the temporal clip primitives but materializes the final tensor. This is the same loop with the write point replaced by a frames-only callback. callback=None retains the complete-output path.