[Bugfix] Do not crash V0 engine on input errors #13101
Conversation
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>
joerunde
requested review from
alexm-redhat,
comaniac,
njhill,
youkaichao and
zhuohan123
as code owners
|
👋 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 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 🚀 |
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>
vllm/worker/model_runner.py
Outdated
| except Exception as e: | ||
| # Raise an exception that tracks the ID of the bad request | ||
| raise InputProcessingError(seq_group_metadata.request_id, | ||
| str(e)) from e |
There was a problem hiding this comment.
How about moving this try/except to the place where this method is called, here, so that it can contain that one line?
There was a problem hiding this comment.
Sure, that does sound like a better idea
|
Probably worth merging in latest main now anyhow, which will kick off all the tests again. |
Hah, reading that as: "This code may be rough but it's not worse than the surroundings" |
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: Louis Ulmer <ulmerlouis@gmail.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
4 participants
This PR adds exception handling to the V0 engine that catches errors processing the input for a sequence group, and removes that sequence group from the current batch instead of crashing the entire engine. We see this a lot when processing invalid image data for multimodal models, the mllama models are particularly bad offenders.
With this change, users will receive a 400 instead of a 500 when their request fails input processing in the engine, and other running requests will continue to process instead of also receiving 500s.