{ "

Cache for Intermediate Activations

\n

During inference the model outputs token by token. We use this simple cache to store key's and value's attention layers, so that we don't have to recompute them for previous tokens.

\n": "

\u4e2d\u9593\u30a2\u30af\u30c6\u30a3\u30d9\u30fc\u30b7\u30e7\u30f3\u7528\u30ad\u30e3\u30c3\u30b7\u30e5

\n

\u63a8\u8ad6\u4e2d\u3001\u30e2\u30c7\u30eb\u306f\u30c8\u30fc\u30af\u30f3\u3054\u3068\u306b\u51fa\u529b\u3057\u307e\u3059\u3002\u3053\u306e\u30b7\u30f3\u30d7\u30eb\u306a\u30ad\u30e3\u30c3\u30b7\u30e5\u3092\u4f7f\u3063\u3066\u30ad\u30fc\u3068\u5024\u306e\u30a2\u30c6\u30f3\u30b7\u30e7\u30f3\u30ec\u30a4\u30e4\u30fc\u3092\u683c\u7d0d\u3059\u308b\u306e\u3067\u3001\u4ee5\u524d\u306e\u30c8\u30fc\u30af\u30f3\u3067\u305d\u308c\u3089\u3092\u518d\u8a08\u7b97\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u305b\u3093

\u3002\n", "

Cache

\n

This maintains a key-value cache and queues push values and pop them in the same order. The queues are useful since we have multiple attention layers.

\n": "

\u30ad\u30e3\u30c3\u30b7\u30e5

\n

\u3053\u308c\u306b\u3088\u308a\u3001\u30ad\u30fc\u3068\u5024\u306e\u30ad\u30e3\u30c3\u30b7\u30e5\u304c\u7dad\u6301\u3055\u308c\u3001\u5024\u306e\u30d7\u30c3\u30b7\u30e5\u3068\u30dd\u30c3\u30d7\u304c\u540c\u3058\u9806\u5e8f\u3067\u30ad\u30e5\u30fc\u306b\u5165\u308c\u3089\u308c\u307e\u3059\u3002\u30ad\u30e5\u30fc\u306f\u8907\u6570\u306e\u30a2\u30c6\u30f3\u30b7\u30e7\u30f3\u30ec\u30a4\u30e4\u30fc\u304c\u3042\u308b\u306e\u3067\u4fbf\u5229\u3067\u3059

\u3002\n", "

Cache a value

\n\n": "

\u5024\u3092\u30ad\u30e3\u30c3\u30b7\u30e5\u3059\u308b

\n\n", "

Clear a cache value

\n\n": "

\u30ad\u30e3\u30c3\u30b7\u30e5\u5024\u3092\u30af\u30ea\u30a2\u3059\u308b

\n\n", "

Clear cache

\n": "

\u30ad\u30e3\u30c3\u30b7\u30e5\u3092\u30af\u30ea\u30a2

\n", "

Get the cache instance

\n\n": "

\u30ad\u30e3\u30c3\u30b7\u30e5\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u3092\u53d6\u5f97

\n\n", "

Pop from a queue

\n\n": "

\u30ad\u30e5\u30fc\u304b\u3089\u30dd\u30c3\u30d7\u3059\u308b

\n\n", "

Push a value to a queue

\n\n": "

\u5024\u3092\u30ad\u30e5\u30fc\u306b\u30d7\u30c3\u30b7\u30e5

\n\n", "

Retrieve a value from cache

\n\n": "

\u30ad\u30e3\u30c3\u30b7\u30e5\u304b\u3089\u5024\u3092\u53d6\u5f97

\n\n", "

Return the size of the queue

\n\n": "

\u30ad\u30e5\u30fc\u306e\u30b5\u30a4\u30ba\u3092\u8fd4\u3059

\n\n", "

Create an empty queue if it's not present

\n": "

\u5b58\u5728\u3057\u306a\u3044\u5834\u5408\u306f\u7a7a\u306e\u30ad\u30e5\u30fc\u3092\u4f5c\u6210

\n", "

Push to the queue

\n": "

\u30ad\u30e5\u30fc\u306b\u30d7\u30c3\u30b7\u30e5

\n", "

Singleton for cache

\n": "

\u30ad\u30e3\u30c3\u30b7\u30e5\u7528\u30b7\u30f3\u30b0\u30eb\u30c8\u30f3

\n", "Cache for Intermediate Activations": "\u4e2d\u9593\u30a2\u30af\u30c6\u30a3\u30d9\u30fc\u30b7\u30e7\u30f3\u7528\u30ad\u30e3\u30c3\u30b7\u30e5", "Cache for intermediate activations for faster inference.": "\u63a8\u8ad6\u3092\u9ad8\u901f\u5316\u3059\u308b\u305f\u3081\u306e\u4e2d\u9593\u30a2\u30af\u30c6\u30a3\u30d9\u30fc\u30b7\u30e7\u30f3\u7528\u306e\u30ad\u30e3\u30c3\u30b7\u30e5\u3002" }