Repository navigation
Standardize samples on cupy-cuda13x[ctk] for consistent CUDA 13 dependencies #2961
Description
Description
Standardize samples' CUDA 13 dependency declarations (PEP 723 dependencies header and requirements.txt) on cupy-cuda13x[ctk], instead of mixing in separate CUDA-major-version-specific packages (e.g. nvidia-cufft, nvidia-curand`), for a consistent environment across samples.
See discussion: #2266 (comment)
Note: sample_runner.py's dependency gate (missing_dependencies()) currently checks extras via static metadata, which doesn't verify the extra's transitive packages are actually installed. This should be fixed alongside the standardization so waived-vs-run behavior stays correct.
Activity
[-]sample_runner: dependency gate falsely treats unfulfilled extras as satisfied[/-][+]Standardize samples on cupy-cuda13x[ctk] for consistent CUDA 13 dependencies[/+]fix(samples): use CUDA 13-native cufft/cublas instead of -cu12 packages
mnkj0021 commentedon Sep 29, 2026
I would like to work on this. I verified there is no existing PR covering the CUDA 13 sample dependency standardization or the extras-aware dependency gate. I will keep the change scoped to sample dependency declarations plus regression coverage for missing transitive packages behind requested extras.
mnkj0021 commentedon Sep 29, 2026
After checking current main more deeply, the sample migration that motivates this (#2266) is still open and conflicting, and the CUDA-library dependency declarations described here are not on main yet. I am going to hold off rather than build a follow-up against code that has not landed.
Migrate samples (#2959)