[Bugfix][Nixl] Fix full prefix cache hit bug

[robertgshaw2-redhat]

SUMMARY:

• in case of full prefix cache hit locally on D worker, we are leaking memory on the P worker side since we are not currently calling send_notif since we skip calling update_state_after_alloc
• also fixes the path where we do get a cache hit, which was passing the wrong thing

[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.

🚀

[robertgshaw2-redhat]

@njhill - can you let me know if this works okay with multi-connector?

[njhill]

@robertgshaw2-redhat I'm still not sure that this part or the change to always call update_state_after_alloc is needed. I'd already added logic for this case in get_num_new_matched_tokens above:

vllm/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

Lines 215 to 222 in f203673

	# NOTE: if count is 0 here, we have less than block_size
	# tokens to pull after subtracting the local prefix cache hit.
	# The remote only sends fully computed blocks, so there is
	# nothing to transfer but we still need to notify the
	# prefill worker so that the remote blocks are freed.
	if all(p in params for p in ("remote_engine_id", "remote_host",
	"remote_port")):
	self._reqs_need_recv[request.request_id] = (request, [])

I can see that the other two fixes below in build_connector_meta and _read_blocks are of course needed though.

If you think it's better to have this logic in this method then we can remove it from the other one. But again I feel it's logically clearer to not call update_state_after_alloc if 0 was returned from get_num_new_matched_tokens.

[robertgshaw2-redhat]

I think that get_num_new_matched_tokens should be a pure function. Adding a side effect to it is surprising given the name of the method and the fact that we will have different behavior depending on what happens if the request is or is not able to be scheduled. This issue is actually causing a bug right now.

• If allocate_slots returns None, the request will remain in the waiting queue. this will cause us to add the requests to reqs_need_recv more than one and as a result we will call read_blocks twice which will do a double free on the P worker side. Similarly this will happen if the request is preempted (it will get re-added to waiting). This is because we are not properly updating the request to have do_remote_prefill=False when it is added to reqs_need_recv from the get_num_new_matched_tokens function.

This is all just evidence that putting a side effect into this function is not a good idea. The update_state_after_alloc is where we should handle everything related to reqs_need_recv so we have a single place where all the logic is handled.

[robertgshaw2-redhat]

I removed those lines from get_num_new_matched_tokens

[njhill]

@robertgshaw2-redhat that makes sense, I agree about the pure function thing. I did also notice the fact that this could result in a double free on the P worker side in the case that it can't be scheduled, which isn't ideal (though I think would probably be harmless).

But to me, thinking from the pov of a generic connector interface, it still feels a bit odd given the connector isn't offering any tokens. I guess we should very clearly document the semantics and expectations for the interface.

A related quirk is that in the async load case, I think currently update_state_after_alloc will be called twice for a request (a second time once the request moves out of WAITING_FOR_REMOTE_KVS).

[robertgshaw2-redhat]

this is now handled in update_state_after_alloc

[njhill]

@robertgshaw2-redhat changes will be needed to multi-connector too, I've pushed them to a branch, feel free to pull into this PR: njhill@4150a41

[njhill]

LGTM, with the multi-connector changes