Struct PrefillInput 

pub struct PrefillInput {
    pub request_id: Option<RequestId>,
    pub maximum_sequence_tokens: Option<usize>,
    pub chunk: Option<PrefillChunk>,
    pub input_ids: TensorRef,
    pub attention_mask: Option<TensorRef>,
    pub position_ids: Option<TensorRef>,
    pub kv_cache: Option<Arc<dyn KvCacheHandle>>,
    pub recurrent_state: Option<Arc<dyn RecurrentStateHandle>>,
    pub metadata: HashMap<String, Value>,
}

Input for prefill phase (processing the initial prompt)

Fields

request_id: Option<RequestId>

Stable product request identity for plan-runtime resources.

maximum_sequence_tokens: Option<usize>

Maximum sequence extent this request may reach, including the prompt. Executors use this for fit validation without allocating future pages.

chunk: Option<PrefillChunk>

Exact scheduler-owned prompt chunk for this invocation.

The input tensor still contains the full prompt so token identity and global offsets remain stable. Plan runtimes execute only this range.

input_ids: TensorRef

Input token IDs [batch_size, sequence_length]

attention_mask: Option<TensorRef>

Attention mask [batch_size, sequence_length] (optional)

position_ids: Option<TensorRef>

Position IDs [batch_size, sequence_length] (optional, for RoPE)

kv_cache: Option<Arc<dyn KvCacheHandle>>

Pre-allocated KV cache handle (optional, for paged attention)

recurrent_state: Option<Arc<dyn RecurrentStateHandle>>

Pre-allocated recurrent-state handle (optional, for state-space layers)

metadata: HashMap<String, Value>

Request metadata that can affect model execution.

Implementations

impl PrefillInput

pub fn new(input_ids: TensorRef) -> Self

Create new prefill input

pub fn with_request_context( self, request_id: RequestId, maximum_sequence_tokens: usize, ) -> Self

Attach the typed request boundary consumed by plan runtimes.

pub fn with_chunk(self, chunk: PrefillChunk) -> Self

Attach the exact scheduler-published prompt chunk.

pub fn with_kv_cache(self, kv_cache: Arc<dyn KvCacheHandle>) -> Self

Create prefill input with a pre-allocated KV cache handle.

pub fn with_recurrent_state( self, recurrent_state: Arc<dyn RecurrentStateHandle>, ) -> Self

Create prefill input with a pre-allocated recurrent-state handle.

pub fn with_metadata(self, metadata: HashMap<String, Value>) -> Self

Attach request metadata.

pub fn with_attention_mask(self, mask: TensorRef) -> Self

Add attention mask

pub fn with_position_ids(self, positions: TensorRef) -> Self

Add position IDs

pub fn batch_size(&self) -> usize

Get batch size

pub fn sequence_length(&self) -> usize

Get sequence length

Trait Implementations

impl Clone for PrefillInput

fn clone(&self) -> PrefillInput

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PrefillInput

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.