cuda: add Kronecker FWHT support

[Lonny154]

Overview

I added CUDA support for MUL_MAT_HADAMARD / FWHT dimensions (384, 640, 768, and 1280). They are not powers of two.

This was done using Kronecker products with the fixed H12/H20 transforms.

The existing power-of-two CUDA path is untouched.

Additional information

16/16 CUDA tests passed
GPU used RTX 4060 Ti (Compute capability 8.9, CUDA 13.3):

./build/bin/test-backend-ops test -b CUDA0 -o MUL_MAT_HADAMARD

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES — I used ChatGPT to help with learning the existing FWHT code and Kronecker algorithm, and ChatGPT helped to generate the CUDA code. I reviewed the code and tested the resulting code.

[ggml-gh-bot]

Hi @Lonny154, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

• AI-generated content: While code is allowed to be generated by AI, please write the PR description and commit messages on your own without the help of AI.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[pwilkin]

/bot review

[ggml-gh-bot]

❌ Code review failed.

Error: failed to fetch diff (503) from https://github.com/ggml-org/llama.cpp/pull/29051.diff

[am17an]

where is this operation required?

[Lonny154]

@am17an I did some further research and I have not found a specific runtime path where these dimensions are required. When tracing the current KV-cache rotation path, the rotation size gets reduced to a power of two dimension.

I originally saw the SYCL implementation adding support for non power of two dimensions and saw an opportunity to do something similar for CUDA support.