llm.inference(
    input_tokens: list[int],  # N tokens
    previous_kv_cache: list[Tensor],  # N tokens' kv cache ∪ <N
) -> output_tokens, new_kv_cache

output_tokens:  # N' new tokens
new_kv_cache:  # kv cache of N + N' tokens

class KVCacheStore:
    def store(tokens, kv_cache_tensors):
        pass

    def retrieve(tokens) -> kv_cache_tensors:
        pass

kv_cache_store.store("ABCDE", [KV1, KV2, KV3, KV4, KV5])
kv_cache_store.retrieve("ABCD") -> [KV1, KV2, KV3, KV4]

prefix_hash = ""
for chunk in chunked_tokens:  # ["AB", "CD", "EF"]
    chunk_hash = hash(prefix_hash + chunk)
    prefix_hash = chunk_hash

# Given chunked prefix hashes, chunked kv cache
# store
for chunk_hash, chunk_kv in zip(...):
    redis.put(chunk_hash, chunk_kv)

# retrieve
for chunk_hash in ...:
    kv_chunk = redis.get(chunk_hash)
    if kv_chunk is None:
        break