vllm icon indicating copy to clipboard operation
vllm copied to clipboard

[MISC] Dump model runner inputs when crashing

Open comaniac opened this issue 1 year ago • 8 comments

To better reproduce the model runner crashing due to illegal memory access and possibly other errors, this PR introduces a utility that dumps model runner inputs when crashing. Since the model runner inputs may be long, I dumped them using pickle for now. Any suggestions or better ideas are welcome.

cc @robertgshaw2-neuralmagic @simon-mo @DarkLight1337

comaniac avatar Sep 09 '24 19:09 comaniac

👋 Hi! Thank you for contributing to the vLLM project. 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 do one of these:

  • Add ready label to the PR
  • Enable auto-merge.

🚀

github-actions[bot] avatar Sep 09 '24 19:09 github-actions[bot]

Can we add instructions in the GitHub issues template so users can share their logs upon encountering such errors?

DarkLight1337 avatar Sep 10 '24 01:09 DarkLight1337

Can we add instructions in the GitHub issues template so users can share their logs upon encountering such errors?

Good point. Will do

comaniac avatar Sep 10 '24 04:09 comaniac

do we need to add a flag for this? looks like some debugging feature that can also be added in https://docs.vllm.ai/en/latest/getting_started/debugging.html

youkaichao avatar Sep 10 '24 07:09 youkaichao

do we need to add a flag for this? looks like some debugging feature that can also be added in https://docs.vllm.ai/en/latest/getting_started/debugging.html

The goal is to be able to get logs from production usage to help track down hard to replicate bugs (like illegal mem access in prefix caching). So having a flag defeats the purpose

robertgshaw2-redhat avatar Sep 10 '24 07:09 robertgshaw2-redhat

makes sense then. please ignore my comment.

youkaichao avatar Sep 10 '24 07:09 youkaichao

@DarkLight1337 added to issue template. PTAL.

comaniac avatar Sep 10 '24 18:09 comaniac

It should be fine as we never load it automatically? But yeah you may get virus if someone post a malicious pickle file to an issue...

comaniac avatar Sep 11 '24 01:09 comaniac