Allow multiple process per device #2916
Merged
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
|
Thanks for the PR. Do you have an example code snippet that produces the error you mentioned? |
|
@BenjaminBossan from accelerate import PartialState
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
distributed_state = PartialState()
pipe.to(distributed_state.device)
with distributed_state.split_between_processes(["a dog", "a cat", "a snail", "a hedgehog"]) as prompt:
print(f"Running {prompt} on process {distributed_state.process_index}, device {distributed_state.device}")
result = pipe(prompt).images
result[0].save(f"{prompt[0]}.png")Command: CUDA_VISIBLE_DEVICES=0,1 accelerate launch --num-processes 4 distributed_inference.pyError:With this patch, it outputs: |
|
Thanks for providing the example. I wonder if it could be adapted as a unit test on a multi GPU runner, assuming we want to proceed with the PR as is. I'll let @muellerzr decide this.
I think this line can safely be removed, as it's covered by the |
muellerzr
reviewed
Jul 8, 2024
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Co-authored-by: Zach Mueller <[email protected]>
|
@muellerzr Would you have any more guidance on where and how to add the test? I did try adding a test to @require_multi_device
def test_multiple_processes_per_device_ops(self):
num_processes = 2 * device_count
print(f"Found {device_count} devices, testing {num_processes} processes.")
cmd = get_launch_command(num_processes=num_processes) + [self.operation_file_path]
with patch_environment(omp_num_threads=1):
execute_subprocess_async(cmd)However, this fails on |
|
That makes me a bit hesitant to push this in, since generally users will likely want to do |
|
I just save the files to a directory as in the example in the docs. I agree it would be better to have support for Related discussions I found: |
muellerzr
approved these changes
Jul 10, 2024
|
I get the same set of failed tests for |
|
Great sounds good! Thanks for the improvement! |
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.
Running multiple replicas of a model on each device can speed up inference. However,
acceleratecurrently does not seem to support this. Settingnum_processesto a value larger than the number of available devices results in an error (for CUDA: "invalid device ordinal").What does this PR do?
Makes sure that the device index is always valid by taking the remainder upon division by
device_count(), hence allowing multiple processes to run on each device by settingnum_processesto a multiple ofdevice_count().Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
@muellerzr @BenjaminBossan @SunMarc