[Bug][Frontend] Improve ZMQ client robustness #7443
Conversation
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
|
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge). To run full CI, you can do one of these:
🚀 |
|
Thanks @joerunde! Possibly related to this, I noticed a problem where we call abort in the cleanup of uncompleted generate requests, the context can sometimes already be closed so it raises a secondary exception. I'm guessing we should just not attempt to abort in this case. |
|
@njhill Yeah good call, let me see about catching that too and adding a test for it edit: aborts are now a no-op after the client closes, and I added a test for it |
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
| @@ -389,7 +414,6 @@ async def check_health(self, | |||
| socket: Optional[zmq.asyncio.Socket] = None | |||
| ) -> None: | |||
| """Raise if unhealthy""" | |||
|
|
|||
| expected_type: Any, | ||
| error_message: str) -> Any: | ||
| error_message: str, | ||
| timeout: Optional[int] = None) -> Any: |
There was a problem hiding this comment.
I think we can drop the timeout argument if we are using the self._data_timeout?
There was a problem hiding this comment.
We should have one or the other IMO
There was a problem hiding this comment.
Yeah that's true, looks like this arg is unused
| @@ -32,6 +33,10 @@ | |||
| INPROC_PROXY_PATH = f"inproc://{uuid4()}" | |||
|
|
|||
|
|
|||
| class ClientClosedError(Exception): | |||
There was a problem hiding this comment.
Can we call this RPCClientClosedError? It should make it clearer when used in the api_server.py
| @@ -32,6 +33,10 @@ | |||
| INPROC_PROXY_PATH = f"inproc://{uuid4()}" | |||
|
|
|||
|
|
|||
| class ClientClosedError(Exception): | |||
| """Exception class raised when the client is used post-shutdown""" | |||
There was a problem hiding this comment.
I think you should add a comment about why this is needed
Specifically:
- the client can be closed, which closes the zmq context
- certain methods like
abort,do_log_statswill then try to open a socket, which causes aZMQError - this creates a huge stack trace on exit
- so, we throw this error such that we can suppress it
There was a problem hiding this comment.
Added, mostly with your wording there ^^
| @@ -18,6 +18,7 @@ | |||
| VLLM_RPC_ZMQ_HWM, RPCAbortRequest, | |||
| RPCGenerateRequest, RPCUtilityRequest) | |||
| # yapf: enable | |||
| from vllm.envs import VLLM_RPC_GET_DATA_TIMEOUT_MS | |||
There was a problem hiding this comment.
Can we then delete the other variables:
VLLM_RPC_HEALTH_TIMEOUT_MSVLLM_RPC_SERVER_START_TIMEOUT_MS
There was a problem hiding this comment.
👍 , replaced with this one timeout
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
|
/ready |
| @@ -82,7 +82,8 @@ async def lifespan(app: FastAPI): | |||
| async def _force_log(): | |||
| while True: | |||
| await asyncio.sleep(10) | |||
| await async_engine_client.do_log_stats() | |||
| with suppress(Exception): | |||
There was a problem hiding this comment.
@joerunde - do we need this? Dont we suppress the warning inside client.py?
There was a problem hiding this comment.
Ah, yeah you're right. I can clean that up tomorrow
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com> Signed-off-by: Alvant <alvasian@yandex.ru>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Related to #7394
Followup for #6797
This PR adds some client-side tests for the zmq client that interacts with the async engine process, and adds timeouts for all of the connections. This ensures that we won't end up deadlocked waiting for a server response that never comes.
Also includes fixes to re-raise exceptions returned by the ZMQ server and to handle situations where the client is shutting down
Draft atm, took a while to figure out the correct test setups. Still need to get timeouts on .generate() calls that also correctly abort the request.PR Checklist (Click to Expand)
Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.
PR Title and Classification
Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:
[Bugfix]for bug fixes.[CI/Build]for build or continuous integration improvements.[Doc]for documentation fixes and improvements.[Model]for adding a new model or improving an existing model. Model name should appear in the title.[Frontend]For changes on the vLLM frontend (e.g., OpenAI API server,LLMclass, etc.)[Kernel]for changes affecting CUDA kernels or other compute kernels.[Core]for changes in the core vLLM logic (e.g.,LLMEngine,AsyncLLMEngine,Scheduler, etc.)[Hardware][Vendor]for hardware-specific changes. Vendor name should appear in the prefix (e.g.,[Hardware][AMD]).[Misc]for PRs that do not fit the above categories. Please use this sparingly.Note: If the PR spans more than one category, please include all relevant prefixes.
Code Quality
The PR need to meet the following code quality standards:
format.shto format your code.docs/source/if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.Notes for Large Changes
Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with
rfc-requiredand might not go through the PR.What to Expect for the Reviews
The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:
action-requiredlabel on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.Thank You
Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!