deep-gemm: add

[drbh]

This PR adds the deep-gemm kernels and relies on an experimental feature added in this PR huggingface/kernels#298

The deep-gemm kernels heavily rely on JIT compilation and need access to nvcc, cutlass headers and internal deep-gemm headers at runtime. This pr includes the internal headers and minor changes to lazily load nvrtc at runtime, and the related PR in the kernels builder updates the build process to inject cutlass headers into the build artifacts so the kernel has all of the required dependencies at runtime.

example usage

nvidia-smi -L
# GPU 0: NVIDIA H100 80GB HBM3 

# navigate to example and run
cd kernels-community/deep-gemm
uv run scripts/readme_example.py

[cuBLASLt BF16] shape: 256x1024x512, cosine_sim: 1.000000, max_diff: 0.0000
[FP8 1D2D] shape: 256x1024x512, cosine_sim: 0.999325, max_diff: 3.9062

note

• if you are on a machine with cuda cap of >=9 you'll need cuda 12.9 and up for the JIT to build successfully dues to inlined asm that is not available on earlier version.
• if you are on a machine with more than one cuda driver you may have to specify the cuda home like CUDA_HOME=/usr/local/cuda-12.9 uv run scripts/readme_example.py

[danieldk]

Really cool work! Added a comment about the submodules.

[danieldk]

I think we should put this in the flake itself: https://discourse.nixos.org/t/nix-2-27-0-released/62003

If users build a kernel manually they'll miss the ?submodules=1.

[drbh]

oo thanks for the tip, this is much cleaner. I fully agree its better to declare this in the flake than depend on users to include ?submodules=1 when building. 🙏