For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo module

types

This module contains the types for the key-value cache APIs.

The module includes structs implementing several different types of KV caches.

This module defines two traits that define the roles of the different structs

• KVCacheT: Defines the interface for a single (key or value) cache.
• KVCollectionT: Defines the interface for a pair of caches (keys and values).

Structs

• ​ContinuousBatchingKVCache: Wrapper for the ContinuousKVCache of a given layer in the transformer model.
• ​ContinuousBatchingKVCacheCollection: This is a "view" of the cache for the given sequences in the batch.
• ​KVCacheStaticParams: Compile-time shape parameters shared across all layers of a KV cache.
• ​PagedKVCache: The PagedKVCache is a wrapper around the KVCache blocks for a given layer. It is used to access the KVCache blocks for PagedAttention.
• ​PagedKVCacheCollection: Paged pair of key and value caches backed by a block-allocated tensor.
• ​PagedRowIndices: Pre-computed physical row indices for a BN-row range of paged KV cache.

Traits

• ​KVCacheT: Trait for different KVCache types and implementations.
• ​KVCollectionT: Trait for a pair of caches (keys and values).

Functions

• ​kv_num_sub_tiles: Number of sub-tile TMA copies needed for tile_BN rows.
• ​kv_sub_tile_rows: Sub-tile row count for a TMA load of tile_BN rows.
• ​kv_tma_fold_chunks: Single source of truth for the SM100 depth-chunk TMA-fold predicate.
• ​padded_depth: Aligns depth up to the nearest multiple of the swizzle granularity.
• ​swizzle_granularity: Returns the TMA swizzle granularity measured in elements of dtype.