Skip to content

vllm_omni.diffusion.models.minimax_h3.sampling ¶

Request-local H3 solvers reconstructed from the surviving experiment records.

RES coefficients follow ComfyUI_RH_MinMaxH3 sampler_core.py at 6abdf97ae8cdce548c751ce850adc5010b583b7c. Euler delegates to the existing implementation to preserve its exact floating-point operation order.

H3SampleSolver ¶

Own the denoised history for one request and one video/audio stream.

coeffs `instance-attribute` ¶

coeffs = res_multistep_coeffs(self.sigmas)

name `instance-attribute` ¶

name = normalize_h3_sampler(name)

old_denoised `instance-attribute` ¶

old_denoised: Tensor | None = None

sigmas `instance-attribute` ¶

sigmas = tuple(float(s) for s in sigmas)

step_index `instance-attribute` ¶

step_index = 0

step ¶

step(state: Tensor, denoised: Tensor, step: int) -> Tensor

create_h3_sample_solver ¶

create_h3_sample_solver(
    name: str | None, sigmas: Sequence[float]
) -> H3SampleSolver

normalize_h3_sampler ¶

normalize_h3_sampler(name: str | None) -> str

res_multistep_coeffs ¶

res_multistep_coeffs(
    sigmas: Sequence[float],
) -> list[tuple[float, float] | None]