PushKvBlocks

Applicability

Product	Supported (Yes/No)
Atlas 350 Accelerator Card	No
Atlas A3 training product/Atlas A3 inference product	Yes
Atlas A2 training product/Atlas A2 inference product	Yes
Atlas 200I/500 A2 inference product	No
Atlas inference product	No
Atlas training product	No

Note: For the Atlas A2 training product/Atlas A2 inference product, only the Atlas 800I A2 inference server and A200I A2 Box heterogeneous subrack are supported.

Function Description

Pushes the cache from the local node to the remote node by configuring a block list. This API can be called only when the role is Prompt.

Prototype

Status PushKvBlocks(const Cache &src_cache,
                    const CacheIndex &dst_cache_index,
                    const std::vector<uint64_t> &src_blocks,
                    const std::vector<uint64_t> &dst_blocks,
                    const KvCacheExtParam &ext_param = {});

Parameters

Parameter	Input/Output	Description
src_cache	Input	Local source cache.
dst_cache_index	Input	Index of the remote destination cache.
src_blocks	Input	Block index list of the source cache.
dst_blocks	Input	Block index list of the destination cache.
ext_param	Input	The difference between second and first in src_layer_range must equal that in dst_layer_range. The default values of first and second in both src_layer_range and dst_layer_range are -1, indicating all layers. The valid range for each index is [0, maximum available layer index], and first must be less than or equal to second. The maximum available layer index is calculated as follows: (CacheDesc::num_tensors / KvCacheExtParam::tensor_num_per_layer) - 1 tensor_num_per_layer can range from 1 to the total number of tensors in the cache, with a default value of 2. When src_layer_range or dst_layer_range uses non-default values, tensor_num_per_layer can either remain at its default value or be set to another value, which must be divisible by the total number of tensors in the cache.

Example

CacheDesc kv_desc{};
kv_desc.data_type = llm_datadist::DT_INT32;
kv_desc.shape = {4, 16};
kv_desc.num_tensors = 4;
Cache cache{};
llm_datadist.AllocateCache(kv_desc, cache);
CacheIndex dst_cache_key{0, 1};
KvCacheExtParam ext_param{};
ext_param.src_layer_range =  std::pair<int32_t, int32_t>(3, 3);
ext_param.dst_layer_range =  std::pair<int32_t, int32_t>(3, 3);
ext_param.tensor_num_per_layer = 1;
std::vector<uint64_t> prompt_blocks = {0, 1, 2, 3};
std::vector<uint64_t> decoder_blocks = {3, 2, 1, 0};
Status ret = llm_datadist.PushKvBlocks(cache, dst_cache_key, prompt_blocks, decoder_blocks , ext_param);

Returns

LLM_SUCCESS: Success.
LLM_PARAM_INVALID: Incorrect parameter.
LLM_NOT_YET_LINK: No link is established with the remote cluster.
LLM_TIMEOUT: Push operation timeout.
LLM_KV_CACHE_NOT_EXIST: The local or remote KV cache does not exist.
Other values: Failure.

Constraints

Before calling this API, call the Initialize API to complete initialization. src_cache must be the cache allocated via the AllocateCache API.

Parent topic: LlmDataDist APIs