Conversation
| - cuda-nvcc # [linux or win] | ||
| cuda_compiler_version: # [linux or win] | ||
| - 12.9 # [linux or win] | ||
| - 13.1 # [linux or win] |
There was a problem hiding this comment.
Unpopular (but safer) option: What about using 13.0 instead of 13.1? The reason is simple: CUDA minor version compatibility.
I agree some libraries have to depend on 13.1, e.g. with CUDA Tile dependencies, and some of them, like cuda-bindings require specifying the exact CUDA version. Apart from these, for 13.x, majority of the libraries support the minor version compatibility (or advanced enhanced compatibility).
Use of 13.1 (or later) could be reconsidered in a couple of months, depending on the changes in CTK.
There was a problem hiding this comment.
After some research, I agree with you.
The GPU feature list did not change between 13.0 and 13.1:
https://docs.nvidia.com/cuda/archive/13.0.0/cuda-compiler-driver-nvcc/index.html#gpu-feature-list
https://docs.nvidia.com/cuda/archive/13.1.0/cuda-compiler-driver-nvcc/index.html#gpu-feature-list
13.1 requires a slightly higher driver than 13.0.
This doc: https://github.com/AnacondaRecipes/cuda-feedstock/blob/main/recipe/doc/recipe_guide.md#cuda-enhanced-compatibility indicates we could build with 13.1 while keeping maximum range (provided we take the appropriate steps). However, identifying appropriate steps seems error prone.
There is also cuda-compat, but that is linux only.
Considering all this, setting the default to 13.0 does seem to be the right call.
There was a problem hiding this comment.
@ELundby45 @xkong-anaconda you're good with 13.0 instead of 13.1?
There was a problem hiding this comment.
This doc: https://github.com/AnacondaRecipes/cuda-feedstock/blob/main/recipe/doc/recipe_guide.md#cuda-enhanced-compatibility indicates we could build with 13.1 while keeping maximum range (provided we take the appropriate steps). However, identifying appropriate steps seems error prone.
That documentation is a hidden gem :)
There was a problem hiding this comment.
I'm good with that. Thanks for flagging and investigating.
onurbingol
left a comment
There was a problem hiding this comment.
LGTM! Thank you very much @cbouss, @ELundby45 and @xkong-anaconda !
Following https://github.com/anaconda/aggregate-duct-tape/pull/38
13.113.0linux_varsfunction settingBUILDwrong whencdt_nameis in use conda/conda-build#5733)