Build fails on Maxwell GPU due to __dp4a undefined in quantized.cu

I’m trying to build a Rust project locally that depends on candle-kernels on my laptop with an NVIDIA GeForce 940MX (Maxwell, compute capability 5.0). The build fails with errors like:

src/quantized.cu(1997): error: identifier "__dp4a" is undefined
...
18 errors detected in the compilation of "src/quantized.cu".

GPU: NVIDIA GeForce 940MX (GM107, compute capability 5.0)
OS: Kali Linux (rolling)
CUDA toolkit: 12.3
NVIDIA driver: 550.163.01
candle-kernels: v0.7.2

The error is caused by the use of the CUDA intrinsic __dp4a, which is only available on GPUs with compute capability 6.1+ (Pascal and newer).
My GPU is compute 5.0, so this intrinsic is not available.

Questions:
Is there a way to disable quantized kernels or the use of __dp4a for older GPUs?
If not, could a feature flag or build option be added to support older hardware, or at least skip building quantized kernels on unsupported GPUs?

Start with src/quantized.cu and inspect how the quantized kernels using __dp4a are selected and built for compute capability 5.0. Check whether the project exposes an existing build option for disabling them; done should be a documented, tested way to build candle-kernels on the reported Maxwell GPU without the unsupported intrinsic.