[Core] add an option to log every function call to for debugging hang/crash in distributed inference #4079
Conversation
|
Currently this functionality is only enabled for We can enable it for all the executors, but would like to hear opinions on whether the effort is worthwhile. |
|
I'm going to set this as a release blocker for |
| (f"VLLM_TRACE_FUNCTION_for_process_{os.getpid()}" | ||
| f"_thread_{threading.get_ident()}_" | ||
| f"at_{datetime.datetime.now()}.log").replace(" ", "_")) | ||
| os.makedirs(os.path.dirname(log_path), exist_ok=True) |
There was a problem hiding this comment.
this should be moved into enable_trace_function_call
There was a problem hiding this comment.
I separate and simplify the logic in enable_trace_function_call so that it can be tested in a standalone way. The caller should be responsible for the logic of creating the log file path.
Co-authored-by: Simon Mo <simon.mo@hey.com>
|
@simon-mo thanks for the quick and detailed review! |
…/crash in distributed inference (vllm-project#4079) Co-authored-by: Simon Mo <simon.mo@hey.com>
…/crash in distributed inference (vllm-project#4079) Co-authored-by: Simon Mo <simon.mo@hey.com>
…/crash in distributed inference (vllm-project#4079) Co-authored-by: Simon Mo <simon.mo@hey.com>
…/crash in distributed inference (vllm-project#4079) Co-authored-by: Simon Mo <simon.mo@hey.com>
…/crash in distributed inference (vllm-project#4079) Co-authored-by: Simon Mo <simon.mo@hey.com>
…/crash in distributed inference (vllm-project#4079) Co-authored-by: Simon Mo <simon.mo@hey.com>
We often receive issue report from users that programs hang or crash in certain scenario. It is very difficult to debug. This PR adds an option to run with
export VLLM_TRACE_FUNCTION=1, which will log every function call invllm. This way, we can know the final function and call stack before program hangs or crashes, so we can easily tell what is the bug.Sometimes the bug might happen in unexpected places, e.g. #4027 finds the hang is because s3 bucket read is too slow; #3916 finds the core dump is due to corrupted libnccl.so .
Hopefully, this PR will help debugging in the future. Also might help #4019 .
TODO
launch_id, as described in [RFC]: Interface and Abstraction for Distributed Inference Environment #3587 and we should place these logs under/tmp/launch_id/. Then users can zip the whole directory so that we can easily debug it. (This isVLLM_INSTANCE_ID, and default isvllm-instance-{random_uuid()})