Reduce cuda complexity by dropping 11.8 #6917

hmaarrfk · 2025-01-12T20:29:44Z

For years we have struggled in maintainning the multiple cuda versions.

Personally, I feel that Nvidia has done tremendous work in ameliorating the situation here at Conda-forge.

However, I'm hoping to shed a little of bit of mental weight from cuda 11.8.

IMO the main projects that utilize it:

have all dropped CUDA 11.8 for lack of maintainer interest.

I think it is time for us to drop it from the global pinnings

Checklist

Used a personal fork of the feedstock to propose changes
Bumped the build number (if the version is unchanged)
Reset the build number to 0 (if the version changed)
Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
Ensured the license file is being packaged.

For years we have struggled in maintainning the multiple cuda versions. Personally, I feel that Nvidia has done tremendous work in ameliorating the situation here at Conda-forge. However, I'm hoping to shed a little of bit of mental weight from cuda 11.8. IMO the main projects that utilize it: * pytorch - https://github.com/conda-forge/pytorch-cpu-feedstock/blob/main/recipe/meta.yaml#L58 * tensorflow - https://github.com/conda-forge/tensorflow-feedstock/blob/main/recipe/meta.yaml#L60 * jax - https://github.com/conda-forge/jaxlib-feedstock/blob/main/recipe/meta.yaml#L27 * onnxruntime https://github.com/conda-forge/onnxruntime-feedstock/blob/main/recipe/meta.yaml#L48 have all dropped CUDA 11.8 for lack of maintainer interest. I think it is time for us to drop it from the global pinnings

hmaarrfk · 2025-01-12T20:30:59Z

recipe/conda_build_config.yaml

-cuda_compiler:
-  - None
-  - nvcc                       # [(linux or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
-  - cuda-nvcc                  # [(linux or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]


do we really need to use this key? could we have users simply depend on cuda-nvcc? I likely don't understand the downstream implications.

Users shouldn't depend on cuda-nvcc but continue to depend on {{ compiler("cuda") }}. For that, we need to set cuda_compiler in the pinning, also in a 12+ world

conda-forge-admin · 2025-01-12T20:31:05Z

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/meta.yaml) and found it was in an excellent condition.

ngam

Shedding nice-to-haves is always good 👍

h-vetinari

Neither for nor against the 11.8 removal, but this PR is not yet correct for its stated goal.

h-vetinari · 2025-01-14T09:28:11Z

recipe/conda_build_config.yaml

-cuda_compiler:
-  - None
-  - nvcc                       # [(linux or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
-  - cuda-nvcc                  # [(linux or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]


Users shouldn't depend on cuda-nvcc but continue to depend on {{ compiler("cuda") }}. For that, we need to set cuda_compiler in the pinning, also in a 12+ world

h-vetinari · 2025-01-14T09:29:56Z

recipe/conda_build_config.yaml

-    - cuda_compiler             # [linux and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
    - cuda_compiler_version     # [linux and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
    - docker_image              # [linux and os.environ.get("CF_CUDA_ENABLED", "False") == "True" and os.environ.get("BUILD_PLATFORM", "").startswith("linux-")]
-  -                             # [win64 and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
-    - cuda_compiler             # [win64 and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
-    - cuda_compiler_version     # [win64 and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]


This will thus continue to be necessary

h-vetinari · 2025-01-14T09:30:00Z

recipe/conda_build_config.yaml

 cuda_compiler_version:
  - None
-  - 11.8                       # [(linux or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
-  - 12.4                       # [linux and ppc64le and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
-  - 12.6                       # [((linux and not ppc64le) or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]
+  - 12                         # [(linux or win64) and os.environ.get("CF_CUDA_ENABLED", "False") == "True"]


I don't want to switch to free-floating minor for 12.x in an unrelated PR. If you think that's a good idea, please open an issue and we can discuss it there.

Right we also discussed this previously. The reason we pin this here is to avoid picking up incomplete CUDA Toolkit's during upgrades

We do plan to bump this version whenever the next CUDA Toolkit is fully available on conda-forge

jakirkham · 2025-01-14T10:17:23Z

Thanks for the kind words Mark! 🙏

Am glad to hear the CUDA 12 packages are serving the community well 😊

Can understand individual feedstock maintainers taking the step of shedding maintenance burden. And they certainly should feel free to skip versions they no longer can or want to support

However when it comes to the global matrix in conda-forge, we have been taking the approach of keeping a broad range of versions. For example with Python we have been keeping all versions that have not reached EOL

By the same measure we have kept around CUDA 11 (specifically 11.8). We certainly have tried to cutdown on the burden by dropping old CUDAs ( #1708 ), moving to CUDA 11.8 ( #4834 ), and dropping anything previous ( #5799 ). Not to mention the significant collective effort put into migrating feedstocks. All this to ensure only one version of CUDA 11 (the latest one) is around

We do this in part because CUDA 11.8 is the last version that still supports Kepler. In CUDA 12.0, Kepler support was removed. Once we drop CUDA 11.8, those users will no longer have support

NVIDIA continues to publish its own libraries (like CUDA Python and RAPIDS with CUDA 11 support). Also a number of packages in conda-forge either from NVIDIA or maintained in part by NVIDIA, ship with CUDA 11 support, which NVIDIA helps maintain

Looking through conda-forge, there are a number of feedstocks that continue to support CUDA 11. So it seems probable that others still value providing support for CUDA 11 too

While I understand why individual maintainers would like to drop CUDA 11 from their feedstocks (and they certainly should feel free to using skips). In terms of many folks still maintaining CUDA 11, it is much easier to do so with it in the global matrix

Would appreciate hearing a bit more how this issue is coming up for you. Perhaps we can find another acceptable solution

hmaarrfk · 2025-01-14T11:55:58Z

Thanks for the thorough explanation

hmaarrfk requested a review from a team as a code owner January 12, 2025 20:29

hmaarrfk mentioned this pull request Jan 12, 2025

distinguish github_actions_labels for CPU vs. CUDA builds #6910

Open

hmaarrfk commented Jan 12, 2025

View reviewed changes

ngam approved these changes Jan 13, 2025

View reviewed changes

hmaarrfk changed the title ~~Reduce cuda complexity dry dropping 11.8~~ Reduce cuda complexity by dropping 11.8 Jan 13, 2025

h-vetinari reviewed Jan 14, 2025

View reviewed changes

hmaarrfk closed this Jan 14, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reduce cuda complexity by dropping 11.8 #6917

Reduce cuda complexity by dropping 11.8 #6917

hmaarrfk commented Jan 12, 2025

hmaarrfk Jan 12, 2025

h-vetinari Jan 14, 2025

conda-forge-admin commented Jan 12, 2025

ngam left a comment

h-vetinari left a comment

h-vetinari Jan 14, 2025

h-vetinari Jan 14, 2025

h-vetinari Jan 14, 2025

jakirkham Jan 14, 2025 •

edited

Loading

jakirkham commented Jan 14, 2025 •

edited

Loading

hmaarrfk commented Jan 14, 2025

Reduce cuda complexity by dropping 11.8 #6917

Reduce cuda complexity by dropping 11.8 #6917

Conversation

hmaarrfk commented Jan 12, 2025

hmaarrfk Jan 12, 2025

Choose a reason for hiding this comment

h-vetinari Jan 14, 2025

Choose a reason for hiding this comment

conda-forge-admin commented Jan 12, 2025

ngam left a comment

Choose a reason for hiding this comment

h-vetinari left a comment

Choose a reason for hiding this comment

h-vetinari Jan 14, 2025

Choose a reason for hiding this comment

h-vetinari Jan 14, 2025

Choose a reason for hiding this comment

h-vetinari Jan 14, 2025

Choose a reason for hiding this comment

jakirkham Jan 14, 2025 • edited Loading

Choose a reason for hiding this comment

jakirkham commented Jan 14, 2025 • edited Loading

hmaarrfk commented Jan 14, 2025

jakirkham Jan 14, 2025 •

edited

Loading

jakirkham commented Jan 14, 2025 •

edited

Loading