pub fn q12_batched_scratch_bytes_varlen(
n: usize,
total_tokens: usize,
max_chunk_len: usize,
top_k: usize,
mrope: bool,
) -> usizeExpand description
Exact scratch footprint for a ragged Q12 batch.
total_tokens is the packed cu_seqlens total while max_chunk_len
sizes the per-stream metadata slots. Uniform callers should use
q12_batched_scratch_bytes. Keeping this separate prevents VARLEN
admission from pessimistically charging every stream at the longest
request’s length.