Skip to content

tp=2,pp=2时,kvcache存储时是只存储 rank0 节点的kvcache嘛?还是说rank0-ran3 的kv都进行存储 #1660

@tianlang-wq

Description

@tianlang-wq

🚀 Feature Description and Motivation

构建了一个ray集群 一个head一个work,连接器使用的是AIBrixOffloadingConnectorV1Type1,kvcache后端使用的infinistore。
vllm配置 tp=2 pp=2 两个节点4张卡。

根据python/aibrix_kvcache/aibrix_kvcache/l2/l2_cache.py:88
cat_head_ids = "".join(
[
str(self.block_spec.tensor_spec.heads[0]),
str(self.block_spec.tensor_spec.heads[-1]),
]
)
cat_layer_ids = "
".join(
[
str(self.block_spec.tensor_spec.layers[0]),
str(self.block_spec.tensor_spec.layers[-1]),
]
)
partition_id = f"h{cat_head_ids}_l{cat_layer_ids}"
猜测 rank0-3的kvcache都会存储到infinistore中,是这样子嘛?由于环境问题目前还无法自测验证

Use Case

Proposed Solution

Metadata

Metadata

Assignees

Labels

area/kv-cachetriage/acceptedIndicates an issue or PR is ready to be actively worked on.

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions