Skip to content

Simplify GPU distribution#16

Open
mpoemsl wants to merge 3 commits intotorchserve-23mt-v0.8.0from
feature/simplify-gpu-distribution
Open

Simplify GPU distribution#16
mpoemsl wants to merge 3 commits intotorchserve-23mt-v0.8.0from
feature/simplify-gpu-distribution

Conversation

@mpoemsl
Copy link
Copy Markdown

@mpoemsl mpoemsl commented Mar 4, 2024

Simplify GPU distribution so it can be controlled better in a multi-model_server setup.

Changes:

  • Add env var override_gpu_id that makes GPUManager always assign a specified GPU to new workers. Default value is -1, which does not have any effect.
  • Remove GPU failure tracking logic since it is not practical
  • Stochastic GPU distribution weighted by current free memory distribution remains

@mpoemsl mpoemsl requested a review from pypae March 4, 2024 10:41
@mpoemsl
Copy link
Copy Markdown
Author

mpoemsl commented Mar 4, 2024

I'm building multi-platform images for this now as textshuttle/pytorch-serve:torchserve-23mt-v0.8.0-v3-{DEVICE}.

@mpoemsl
Copy link
Copy Markdown
Author

mpoemsl commented Mar 4, 2024

There exists now a multi-arch image textshuttle/pytorch-serve:torchserve-23mt-v0.8.0-v3-cpu and a single-arch amd image textshuttle/pytorch-serve:torchserve-23mt-v0.8.0-v3-gpu.

Copy link
Copy Markdown

@JeffWigger JeffWigger left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These changes look good.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants