Compilation errors due to incorrect GPU architecture

I got these errors:
src/quantized.cu(1962): error: identifier "__dp4a" is undefined
sumi = __dp4a(v[i], u[i], sumi);
^

18 errors detected in the compilation of "src/quantized.cu".
thread 'main' panicked at /home/remote/.cargo/registry/src/index.crates.io-6f17d22bba15001f/bindgen_cuda-0.1.5/src/lib.rs:391:13:
nvcc error while compiling "src/quantized.cu":

# CLI "nvcc" "--gpu-architecture=sm_60" "--ptx" "--default-stream" "per-thread" "--output-directory" "/projects/ai/candle/candle/candle-kernels/target/debug/build/candle-kernels-2b82da83d6e51215/out" "-Isrc" "-I/usr/include" "src/quantized.cu"

They were solved after some ChatGPT and putting this:

export CUDA_COMPUTE_CAP=61

However it would be better to solve this in a more generic way.

Just as an FYI. This happens with Ubuntu 24.04, CUDA 12.6 and 3 Tesla P100.

Start with candle-kernels/src/quantized.cu and the build invocation showing nvcc --gpu-architecture=sm_60. Check how CUDA_COMPUTE_CAP is read and how the GPU architecture is selected, using the reported Tesla P100 failure as the reproduction case. Done means the kernels compile with the appropriate architecture without requiring a manually exported environment variable.

Written by the indexing model from the issue text.