[Core] Fix abrupt request abort

[NickLucche]

There's currently an issue when a KV Connector is in use and requests are aborted before they have been scheduled (and so before any blocks have been allocated in the kv cache manager). The scheduler _connector_finished method is invoked and it calls kv_cache_manager.get_block_ids(request_id) which raises an exception because the kv cache manager doesn't recognize the request.

We do still want to invoke the connector in this case, so that it can perform external cleanup if needed, but it makes sense to pass it empty blocks.

It's not always 100% guaranteed due to timings, but this setup makes it pretty reproducible:

# start P
VLLM_NIXL_SIDE_CHANNEL_PORT=$5557 CUDA_VISIBLE_DEVICES=0 vllm serve Qwen/Qwen3-0.6B --port 8100 --enforce-eager --disable-log-requests --tensor-parallel-size 1 --gpu-memory-utilization 0.5 --trust-remote-code --max-model-len 128 --max-num-seqs 1 --max-num-batched-tokens 128 --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'

# start D
VLLM_NIXL_SIDE_CHANNEL_PORT=5558 CUDA_VISIBLE_DEVICES=1 vllm serve Qwen/Qwen3-0.6B --port 8200 --enforce-eager --disable-log-requests --tensor-parallel-size 1 --gpu-memory-utilization 0.5 --trust-remote-code --max-model-len 128 --max-num-seqs 1  --max-num-batched-tokens 128 --kv-transfer-config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'

# start proxy server
python tests/v1/kv_connector/nixl_integration/toy_proxy_server.py --port 8192 \
      --prefiller-port 8100 \
      --decoder-port 8200
...
# send some requests and abort some of those
python simulate_abort.py

simulate_abort.py

import asyncio
import httpx


API_URL = "http://localhost:40876/v1/completions" 
# API_URL = "http://localhost:8000/v1/completions" 
NUM_REQUESTS = 10
ABORT_INDICES = {1, 9} 

async def send_request(i, abort=False):
    async with httpx.AsyncClient(timeout=None) as client:
        task = asyncio.create_task(client.post(
            API_URL,
            json={
                "model": "Qwen/Qwen3-0.6B",
                "prompt": f"This is test request {i}. what do you say",
                "max_tokens": 100,
            },
            timeout=None
        ))

        if abort:
            # await asyncio.sleep(0.2 + random.random() * 0.3)  
            await asyncio.sleep(0.5)
            task.cancel()
            try:
                await task
            except asyncio.CancelledError:
                print(f"Request {i} aborted!!\n")
        else:
            try:
                response = await task
                print(f"Request {i} completed: {response.status_code}")
            except Exception as e:
                print(f"Request {i} failed: {e}")

async def main():
    tasks = [
        send_request(i, abort=(i in ABORT_INDICES))
        for i in range(NUM_REQUESTS)
    ]
    await asyncio.gather(*tasks, return_exceptions=True)

asyncio.run(main())

Client gets 500, D crashes with logs:

DEBUG 05-21 12:59:23 [loggers.py:116] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%
ERROR 05-21 12:59:23 [core.py:495] EngineCore encountered a fatal error.
ERROR 05-21 12:59:23 [core.py:495] Traceback (most recent call last):
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/engine/core.py", line 486, in run_engine_core
ERROR 05-21 12:59:23 [core.py:495]     engine_core.run_busy_loop()
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/engine/core.py", line 511, in run_busy_loop
ERROR 05-21 12:59:23 [core.py:495]     self._process_input_queue()
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/engine/core.py", line 532, in _process_input_queue
ERROR 05-21 12:59:23 [core.py:495]     self._handle_client_request(*req)
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/engine/core.py", line 550, in _handle_client_request
ERROR 05-21 12:59:23 [core.py:495]     self.abort_requests(request)
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/engine/core.py", line 202, in abort_requests
ERROR 05-21 12:59:23 [core.py:495]     self.scheduler.finish_requests(request_ids,
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/core/sched/scheduler.py", line 866, in finish_requests
ERROR 05-21 12:59:23 [core.py:495]     self._free_request(request)
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/core/sched/scheduler.py", line 872, in _free_request
ERROR 05-21 12:59:23 [core.py:495]     delay_free_blocks, kv_xfer_params = self._connector_finished(request)
ERROR 05-21 12:59:23 [core.py:495]                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/core/sched/scheduler.py", line 952, in _connector_finished
ERROR 05-21 12:59:23 [core.py:495]     block_ids = self.kv_cache_manager.get_block_ids(request.request_id)[0]
ERROR 05-21 12:59:23 [core.py:495]                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 05-21 12:59:23 [core.py:495]   File "/home/nicolo/vllmd/vllm/vllm/v1/core/kv_cache_manager.py", line 369, in get_block_ids
ERROR 05-21 12:59:23 [core.py:495]     assert request_id in self.single_type_manager.req_to_blocks
ERROR 05-21 12:59:23 [core.py:495]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 05-21 12:59:23 [core.py:495] AssertionError

The following is a straightforward fix to the issue, let me know if you see a better solution to this.

[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.

🚀

[mergify]

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

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

[mergify]

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

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

[heheda12345]

@NickLucche Does #18829 fix this bug?

[NickLucche]

Quickly tried to run it with my setup but it's crashing.
Is it working for you with the snippet I posted?

[heheda12345]

I didn't try. I'm asking because #18829 (comment) by @Abatom says that PR can fix this problem.

[NickLucche]

I will address asap to try and cover your PD separation scheme

[Abatom]

I will address asap to try and cover your PD separation scheme

Thanks

[NickLucche]

I've addressed the PR so that it should be compatible with your connector as well as latest changes to get_block_ids logic. Thanks for reviewing @Abatom!

[njhill]

@NickLucche I don't think this has anything to do with the connector really or P/D lifecycle, it's a general problem related to aborting a request before it's been scheduled. It's just we only hit that path when a connector is in use.

I actually agree with @Abatom and I think it would make sense to change the KVCacheManager get_block_ids() method to return empty blocks if the request isn't recognized. I think this is reasonable because this just means there are no blocks allocated for this request, so returning zero blocks is logically "correct".

I have pushed my own version of this here: main...njhill:vllm:fix-pd-abort (also moved the single cache group assertion to the constructor)

[NickLucche]

I think this is reasonable because this just means there are no blocks allocated for this request, so returning zero blocks is logically "correct"

I am still not 100% sure that changing the interface from a .get(k) to a .get_with_default(k) is always going to be safe, given that we won't raise on possible buggy states where a request slips in the loop.
But I am ok with just getting this fixed after all.

And yeah this has nothing to do with pd my bad, it just pops up more easily given the higher latency.

[njhill]

Thanks @NickLucche, I pushed my update as you agreed to offline :)

I've also now opened a related PR #19223 which I think helps shrink our P/D abort race condition window, which relies on the behaviour we've kept here to always notify the connector when requests are finished (it sounds similar to what @Abatom needs it for).

[njhill]

I will add @Abatom as co-author too since this is similar to his other PR #18829.

[njhill]

Thanks @NickLucche!

[NickLucche]

I will add @Abatom as co-author too since this is similar to his other PR #18829.

Absolutely, thanks for your work @njhill !

[NickLucche]

I've rebased in light of the recent hybrid allocator changes, ptal when you get the chance @njhill

[NickLucche]

This test tests/v1/kv_connector/unit/test_multi_connector.py::test_multi_shared_storage_connector_consistency got broken on main so CI will report a failure here too, looking into it.

[njhill]

@NickLucche I think we'll want to actually return a list of empty lists in this case. I didn't update it yet since we might as well do it after rebasing on the other fix.

Suggested change

	manager.req_to_blocks[request_id]
	for manager in self.single_type_managers
	if request_id in manager.req_to_blocks
	manager.req_to_blocks.get(request_id) or []
	for manager in self.single_type_managers