flashinfer.comm._ulysses_lowp.zero_scale_and_padding

flashinfer.comm._ulysses_lowp.zero_scale_and_padding(send_u8: Tensor, spec: Dict[str, int | float]) → None

Deterministically zero every scale slot and the alignment tail.

Must run before the first V2-G pack launch: the pack kernels write only the touched slots, and the ABI requires every unused byte to be zero.

Parameters:

• send_u8 – Contiguous CUDA uint8 destination-major buffer [world_size, chunk_bytes] from the matching payload specification.

• spec – Payload geometry returned by the matching layout’s payload_spec.