[V1] Support multiple kv connectors

[mgoin]

Co-authored with @njhill.

We take advantage of the kv_connector_extra_config: dict[str, Any] already present in KVTransferConfig to stash all the connectors we want in an ordered list of kwargs. The new MultiConnector then creates a connector from each item in its list of KVTransferConfig.kv_connector_extra_config["connectors"]:

class MultiConnector(KVConnectorBase_V1):

    def __init__(self, vllm_config: "VllmConfig", role: KVConnectorRole):
        super().__init__(vllm_config=vllm_config, role=role)
        self._connectors = []
        ktcs = vllm_config.kv_transfer_config.kv_connector_extra_config.get("connectors")
        assert ktcs is not None
        for ktc in ktcs:
            temp_config = copy.copy(vllm_config)
            temp_config.kv_transfer_config = KVTransferConfig(**ktc)
            self._connectors.append(KVConnectorFactory.create_connector_v1(temp_config, role))

Example usage:

CUDA_VISIBLE_DEVICES=0 NIXL_ROLE="SENDER" vllm serve meta-llama/Llama-3.2-1B-Instruct \
    --port 8100 \
    --enforce-eager \
    --disable-log-requests \
    --kv-transfer-config '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both"},{"kv_connector":"SharedStorageConnector","kv_role":"kv_both","kv_connector_extra_config":{"shared_storage_path":"local_storage"}}]}}'

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[mgoin]

Dependent on #17686 landing first for KVConnectorBase_V1 API changes/additions made in that PR.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @mgoin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[tlrmchlsmth]

LGTM, I think my main comment is that we should make sure to document the behavior of the loads. I.E. make sure the precedence order stuff is written down somewhere.

[maobaolong]

@mgoin @njhill Thanks for this feature! It looks useful. And I want to make connectorA as P/D transfer connector, make connectorB as offload-reuse purpose connector, how to config to make the connectorA and connectorB co-exist?

[njhill]

@mgoin @njhill Thanks for this feature! It looks useful. And I want to make connectorA as P/D transfer connector, make connectorB as offload-reuse purpose connector, how to config to make the connectorA and connectorB co-exist?

@maobaolong We'll add better doc/comments for this but the logic initially is to:

• Load from the first connector that advertises available tokens (in the order specified in the config)
• Save to all connectors

[maobaolong]

@mgoin @njhill Thanks for this feature! It looks useful. And I want to make connectorA as P/D transfer connector, make connectorB as offload-reuse purpose connector, how to config to make the connectorA and connectorB co-exist?

@maobaolong We'll add better doc/comments for this but the logic initially is to:

• Load from the first connector that advertises available tokens (in the order specified in the config)
• Save to all connectors

@njhill Thanks for your explain. And I'd like to know more about the motivation and the scenario, if you can add some more description, it will help. So I guess this PR can not solve the co-exist issue for P/D transfer connector and offload-reuse purpose connector ?

[njhill]

@mgoin @njhill Thanks for this feature! It looks useful. And I want to make connectorA as P/D transfer connector, make connectorB as offload-reuse purpose connector, how to config to make the connectorA and connectorB co-exist?

@maobaolong We'll add better doc/comments for this but the logic initially is to:

• Load from the first connector that advertises available tokens (in the order specified in the config)
• Save to all connectors

@njhill Thanks for your explain. And I'd like to know more about the motivation and the scenario, if you can add some more description, it will help. So I guess this PR can not solve the co-exist issue for P/D transfer connector and offload-reuse purpose connector ?

@maobaolong this is intended to be a starting point, it can work with P/D + offload-reuse if the P/D connector uses kv transfer params passed in the request to determine whether it should offer tokens to load. Which is how the current NixlConnector works - it will only return nonzero token count from get_matched_tokens if the kv transfer params in the request contain do_remote_prefill=True, otherwise it will fall back to the second connector which can be the one to reuse offloaded kvcache.

We can extend it with other loading selection logic if you have additional ideas...