This repository provides information about the OpenStack instance Quansight and MetroStar are providing to conda-forge and other communities.
The main intent of this service is to provide GPU CI to those conda-forge feedstocks that require it. To do so:
- Feedstocks must have access to the service. Refer to the
conda-forge/admin-requestsREADME. - Maintainers you must read and agree to the Terms of Service. Your username must be listed in
access/conda-forge-users.json. Please open a PR to add yourself to the list.
If you suspect the server is not operating as expected, please check:
If you think there should be an open incident report but there's none, please open a new issue and tag @Quansight/open-gpu-server so the team can take a look. Thanks!
- Model: AMD EPYC 7352 24-Core Processor
- Architecture:
x86_64, 32-bit, 64-bit - 48 Cores
- ~500 GB Memory
- 6x NVIDIA Tesla V100
The server can spin up VMs with the following configurations:
| Name | vCPUs | RAM | Disk | GPUs |
|---|---|---|---|---|
gpu_tiny |
4 | 2GB | 20GB | 1x NVIDIA Tesla V100 |
gpu_medium |
4 | 8GB | 50GB | 1x NVIDIA Tesla V100 |
gpu_large |
4 | 12GB | 60GB | 1x NVIDIA Tesla V100 |
gpu_xlarge |
8 | 16GB | 60GB | 1x NVIDIA Tesla V100 |
gpu_2xlarge |
8 | 32GB | 80GB | 1x NVIDIA Tesla V100 |
gpu_4xlarge |
8 | 64GB | 100GB | 1x NVIDIA Tesla V100 |
These runners use the ubuntu-2204-nvidia-20230914 image.
| Name | vCPUs | RAM | Disk |
|---|---|---|---|
ci_medium |
4 | 8GB | 60GB |
ci_large |
4 | 12GB | 60GB |
ci_xlarge |
4 | 32GB | 60GB |
ci_2xlarge |
8 | 32GB | 80GB |
ci_4xlarge |
8 | 64GB | 100GB |
These runners use the ubuntu-2204-20231018 image.
These runners run ISOs derived from Ubuntu 22.04. Images are built with the instructions provided in the images folder.
- Concurrency depends on available resources. Only 4 GPUs can be exposed to the VMs at a time; expect queues. This is not per repository, but a server-wide limitation. See docs/openstack.md for details.
- We have not yet implemented a time limit per job. Please be mindful of this and try to keep your jobs as short as possible. This might change in the future.
This service is provided as is, with no guarantees of uptime or support. If you have any questions, please open an issue in this repository and we'll try our best.