[Feature]: Add support for the vLLM V1 connector

[flesher0813]

🚀 The feature, motivation and pitch

Motivation

The vLLM v1 supports using the KVConnectorBase_V1 class to support the use of external storage for storing and retrieving kv cache, which can achieve a trade-off between storage and computation. This feature has not yet been integrated into vllm-ascend. Considering the general support for multiple storage backends, the addition of this feature is worth considering.

Proposed Change

We can refer to gpu_model_runner.py and layer.py in vLLM v1 to add calls to the connector functions in vLLM-ascend, in order to achieve the purpose of implementing this feature. Additionally, since we aim to return each request's successfully dumped blocks after the worker side' saving is completed, we have added a return value to the wait_for_save function, which is a part not increased in vLLM.

Code

flesher0813@d5c47a5

Alternatives

No response

Additional context

No response

[wangxiyuan]

see #950 @ganyi1996ppo please double check this as well. Thanks

[ganyi1996ppo]

External storage is undergoing now, we are working tightly on the mooncake