Decommission open-gpu-server#174
Conversation
|
Based on the .cirun access list, the following feedstocks need CPU runners:
These need GPU runners:
Perhaps more interestingly is to see it together. Note I've removed test feedstocks for clarity. These need both types of runners:
These ONLY need CPU:
These ONLY need GPU:
|
|
IIRC the cutlass feedstock does not need this runner anymore conda-forge/cutlass-feedstock#22 |
|
@jaimergp Thanks for your time. We use this machine for the cuda support not for the long-running jobs. conda-forge/muscat-split-feedstock can live without cuda support. So what do I have to do to keep things working?. After this PR is merge can I continue building linux packages using the "classic" conda-forge runners? Thanks again, conda-forge rules |
|
It may vary from project to project, but it should be possible to compile things with GPU support even without a GPU present. Testing the outputs is the main problem. We are working on finding a replacement for both use cases of the GPU server, but it's likely that the long running CPU support will land before GPU runners. |
|
Ok, we were just informed that the server will be disconnected today at 4PM ET so I'm going to remove the integrations now to avoid any killed jobs. Sorry about the sudden change! We can continue the discussions in a follow-up issue. |
The Quansight Open GPU Server (OpenStack) was decommissioned on 2026-03-13 (conda-forge/.cirun#174), so cirun-openstack-cpu-4xlarge runners no longer exist. All Linux GHA jobs were timing out after 24h waiting for runners that will never come. Cirrus runners access was granted via conda-forge/admin-requests#1982. This switches Linux builds to ghcr.io/cirruslabs/ubuntu-runner-amd64:24.04-lg (16 CPUs, 48GB RAM). Windows builds remain on cirun-azure-windows-4xlarge. Re-rendered with conda-smithy 3.59.0.
3 participants
IMPORTANT:
We've been informed that the server hosted by Quansight (
open-gpu-server) will be decommissioned next week. In order to prevent mid-job cancelations that can affect package uploads, we'll remove access by merging this PR on Monday, March 16th, no later than 11AM UTC.We are working on getting access to alternative services for long-running jobs, but it might take a bit, so expect some disruption while we figure this out.
cc @conda-forge/core, @conda-forge/astra-toolbox, @conda-forge/cutlass, @conda-forge/flash-attn, @conda-forge/jaxlib, @conda-forge/libmagma, @conda-forge/mongodb, @conda-forge/muscat-split, @conda-forge/nodejs, @conda-forge/onnxruntime, @conda-forge/pytorch-cpu, @conda-forge/pytorch_scatter, @conda-forge/qt-webengine, @conda-forge/ray-packages, @conda-forge/tensorflow, @conda-forge/tinygrad, @conda-forge/torchao, @conda-forge/viskores, @conda-forge/vllm, @conda-forge/vtk-m, @conda-forge/webkit2gtk4-1 , @conda-forge/xformers.