ggml: support non-contiguous SUM (CPU, CUDA)

[rssndr]

Overview

test-backend-ops includes SUM(type=f32,ne=[33,256,1,1],permute=[1,0,2,3]), which CUDA rejects and which crashes the CPU reference if enabled. Two small changes:

• CUDA: supports_op required ggml_is_contiguous_rows, but the kernel only requires ggml_is_contiguously_allocated (asserted in ggml_cuda_op_sum since llama/ggml: add LLM training support #10544). The flat CUB reduction is valid for any gap-free view, since summation is order-independent. The contiguous_rows check was added in metal: optimise GGML_OP_SUM #16559 to keep CI green after that PR introduced permuted SUM test cases, it wasn't based on what the CUDA kernel can actually handle. This change makes supports_op match the condition the kernel itself asserts. The kernel is unchanged.
• CPU: ggml_compute_forward_sum_f32 asserted nb[0] == sizeof(float), so it aborted on non-contiguous rows, reached as soon as the CPU acts as reference (or fallback) for this case. Added a strided path keeping ggml_float accumulation; contiguous fast path unchanged. f16/bf16 variants left as-is.

Why CPU and CUDA together : the CPU fix is required for the CUDA case to be testable, since the reference crashes once the case is enabled.

Additional information

Testing (RTX 3070 Ti Laptop, CUDA 13.3):

• test-backend-ops -o SUM: 8/8, permuted case now passing
• Full test-backend-ops: 16094/16094
• perf -o SUM: unchanged

Ref #14909, #16559.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Used Claude for analysis help, code drafting for the CPU change, and the initial description draft. Reviewed, built, and tested locally.

[ggml-gh-bot]

Hi @4ndrearossetti, 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.

• Multiple backend changes in one PR: When adding support for a new model or feature, focus on CPU support only in the initial PR. Add support for other backends like CUDA in follow-up PRs. If you have a good reason to modify multiple backends in one PR, please explain it.

• 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.

[rssndr]

Hello,
I checked the three failures, here a summary:

• build-cmake-pkg / linux: it's a runner config error ("jobs without a container are forbidden"), so it failed at setup, the code was never checked.
• gpu-cuda: self-hosted runner lost communication with GitHub ~1h53m in; all tests up to that point, SUM included, were passing.
• gpu-openvino-low-perf: OpenVINO crashed with CL_OUT_OF_RESOURCES on the weak Intel-GPU runner. My code doesn't touch the OpenVINO backend.

So, it seems it's three infrastructure-related faillures, rather than anything from my code.
I remain available if there's anything more to change or check.

Thank you very much!

[JohannesGaessler]

I just realized: please add a corresponding test case to test-backend-ops.cpp as well.

[rssndr]

Hello,
just added two tests, as asked. I also rebased on master in the meantime, since quite some time passed since my last commits.

[JohannesGaessler]

A permutation is not enough for the tests. Please add the option to calculate a sum over a non-contiguous view.

[rssndr]

Hello Johannes,

That's right. I added the slice option following the test_sum_rows pattern. The sliced cases run on CPU and report unsupported on CUDA since such views aren't allocated contiguously. And, the slice + row-breaking-permute case skips grad because the backward pass goes through ACC, whose CPU implementation requires contiguous rows.

That should be all. Let me know if I missed anything. Thank you!