Struct RollingHashIdMapper 

pub struct RollingHashIdMapper { /* private fields */ }

Maps sequence-aware block hashes to compact, stable u64 ids.

The mapper is intentionally stateful and reusable across requests/turns: a block of tokens that appears at the same prefix position in two different requests will be assigned the same id. Equality of leading hash_ids between rows therefore signals shared prompt prefixes for replay purposes.

hash_ids here are workload identity labels, not literal Dynamo runtime KV-cache hashes. Producers should not try to reconcile them with a production cache.

Implementations

impl RollingHashIdMapper

pub fn new(block_size: usize) -> Self

Create a new mapper for the given block size.

pub fn block_size(&self) -> usize

Block size that this mapper was constructed with.

pub fn hash_token_blocks(&mut self, tokens: &[u32]) -> Vec<u64>

Hash a sequence of tokens into Mooncake hash_ids.

Tokens are chunked by block_size; each complete block contributes one compact id derived from Dynamo’s shared KV-hashing contract. A trailing partial block also contributes one compact id so replay capacity still covers the full prompt length. Identical prefixes across requests resolve to identical leading hash_ids once the mapper has seen them.

pub fn try_hash_token_blocks(&mut self, tokens: &[u32]) -> Result<Vec<u64>>

Fallible variant of Self::hash_token_blocks.

pub fn ids_for_sequence_hashes(&mut self, sequence_hashes: &[u64]) -> Vec<u64>

Map precomputed sequence-aware block hashes into compact Mooncake IDs.

This is useful for producers that record stable block hashes in the serving path and only compact them during offline trace conversion.