Is this page helpful?

`core.inference.sampling.request_seed_noise`#

Batched, request-local exponential noise for CUDA sampling.

Module Contents#

Functions#

`_request_seed_noise_kernel`
`fill_request_seed_noise`	Fill all seeded rows in one launch; leave shared-generator rows untouched.

API#

core.inference.sampling.request_seed_noise._request_seed_noise_kernel( Noise, Seeds, Positions, V: triton.language.constexpr, SINGLE_ROW: triton.language.constexpr, BLOCK: triton.language.constexpr, )#

core.inference.sampling.request_seed_noise.fill_request_seed_noise( noise: torch.Tensor, seeds: list[int | None], positions: list[int] | None = None, ) → None#

Fill all seeded rows in one launch; leave shared-generator rows untouched.

Every variate is indexed by (request seed, absolute position, vocabulary index), not batch slot, launch geometry, or shared-generator state. CUDA draws are not bit-identical to Torch’s generator-specific implementation.