[core][distributed] fix zmq hang #6759
Conversation
|
👋 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:
🚀 |
| @@ -40,7 +40,7 @@ def _validate_http_url(self, url: str): | |||
| raise ValueError("Invalid HTTP URL: A valid HTTP URL " | |||
| "must have scheme 'http' or 'https'.") | |||
|
|
|||
| def _headers(self, **extras: str) -> Mapping[str, str]: | |||
There was a problem hiding this comment.
this is a lint error I fix by the way
| @@ -9,7 +9,7 @@ | |||
| import torch | |||
| import torch.distributed as dist | |||
| from torch.distributed import ProcessGroup | |||
| from zmq import PUB, REP, REQ, SUB, SUBSCRIBE, Context # type: ignore | |||
| from zmq import SUB, SUBSCRIBE, XPUB, XPUB_VERBOSE, Context # type: ignore | |||
There was a problem hiding this comment.
No, you can check https://netmq.readthedocs.io/en/latest/xpub-xsub/ .
XPUB connects to SUB.
XSUB is used to connect many PUB, which is not our usecase.
(cherry picked from commit 740374d)
[core][distributed] fix zmq hang (vllm-project#6759)
Signed-off-by: Alvant <alvasian@yandex.ru>
fixes #6700
this is caused by incorrect usage of zmq, or a bug of zmq, reported at zeromq/libzmq#4713 .
By using XPUB channel, we can make sure all subscribers already subscribed, and we are ready to publish (broadcast).
Locally tested, previously it hangs once in 20 runs.
Now it runs without any problem in 1000 runs.