Struct PrefillSlice 

pub struct PrefillSlice<'a> {
    pub prompt_tokens: &'a [u32],
    pub seq: &'a mut SequenceState,
    pub chunk_start: usize,
    pub chunk_len: usize,
    pub is_last_chunk: bool,
}

Per-stream input slice for batched prefill.

One of these per concurrent prefilling stream — prefill_batch_chunk and mixed_forward_batch accept a &mut [PrefillSlice<'_>] and process all streams’ chunks in a single forward pass. See Q12 in /workspace/atlas-internal/qwen-refactor/notes.md for the bug this fixes (concurrent prefills serialized through prefilling.first_mut() in the scheduler, causing 5× asymmetric TTFT).

Fields

prompt_tokens: &'a [u32]

Full prompt tokens for this stream.

seq: &'a mut SequenceState

Per-stream sequence state (KV blocks, SSM slot, etc.).

chunk_start: usize

Token offset into prompt_tokens where this chunk starts.

chunk_len: usize

Number of tokens in this chunk.

is_last_chunk: bool

Whether this is the final chunk for this stream (controls whether the model emits last-token logits for sampling).