llama : pad KV cache size (#4280)

Commit

1 year ago

llama : pad KV cache size (#4280) * llama : pad KV cache size to 32 * metal : try to improve batched decoding