Use CUDA_VISIBLE_DEVICES instead of gpu_id variables everywhere. #2824
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This is recommended by PyTorch: > In most cases it’s better to use CUDA_VISIBLE_DEVICES environmental variable. (https://pytorch.org/docs/stable/generated/torch.cuda.set_device.html) It also helps avoid certain classes or errors (e.g., initializing CUDA on the wrong device and using extra memory there).
heiner
requested review from
merrymercy,
Ying1123,
hnyls2002,
zhyncs,
ispobock and
ByronHsu
as code owners
rkooo567
reviewed
Jan 10, 2025
Something at process start seems to already initialize CUDA.
merrymercy
requested changes
Jan 10, 2025
| proc = mp.Process( | ||
| target=run_scheduler_process, | ||
| args=(server_args, port_args, gpu_id, tp_rank, dp_rank, writer), | ||
| ) | ||
| proc.start() | ||
| os.environ["CUDA_VISIBLE_DEVICES"] |
There was a problem hiding this comment.
del os.environ["CUDA_VISIBLE_DEVICES"]?
| @@ -450,13 +450,15 @@ def launch_engine( | |||
| for tp_rank in tp_rank_range: | |||
| reader, writer = mp.Pipe(duplex=False) | |||
| gpu_id = server_args.base_gpu_id + tp_rank % tp_size_per_node | |||
| os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id) | |||
There was a problem hiding this comment.
Does this work for AMD? We might need to find the correct env vars for AMD as well.
cc @HaiShaw
There was a problem hiding this comment.
Perhaps not; the tests pass for me locally but don't seem to pass in CI.
There was a problem hiding this comment.
According to these docs, this should work: https://rocm.docs.amd.com/en/latest/conceptual/gpu-isolation.html#cuda-visible-devices
|
The tests are still failing. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation
This is recommended by PyTorch:
(https://pytorch.org/docs/stable/generated/torch.cuda.set_device.html)
It also helps avoid certain classes or errors (e.g., initializing CUDA on the wrong device and using extra memory there).
Modifications
CUDA_VISIBLE_DEVICESat the start ofrun_scheduler_process(annoyingly, Python'smultiprocessingmodule gives no way of setting the env variables of the new child process).gpu_idvariables throughout the code.Checklist