Skip to content

vllm_omni.model_executor.models.whisper_utils

Conv1d

Bases: Conv1d

Conv1d with automatic dtype casting for mixed precision inference.

ConvTranspose1d

Bases: ConvTranspose1d

Linear

Bases: Linear

Linear layer with automatic dtype casting for mixed precision inference.

forward

forward(x: Tensor) -> Tensor

sinusoids

sinusoids(length, channels, max_timescale=10000)

Returns sinusoids for positional embedding.