CUDA: fix crash with partial offloading of MoE

[JohannesGaessler]

Fixes #13437 .

There is an edge case for clearing the padding that I did not consider. I don't think this edge case can be sensibly fixed other than by falling back to cuBLAS where no padding is needed.

[slaren]

What is the view? I am not able to reproduce the linked issue, and in the qwen3moe graph I don't see any views (of weights).

[slaren]

I was able to reproduce it by building for a lower cc. There is no view in the graph, this is caused by the slice created in ggml_cuda_mul_mat_id. This also fixes it:

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 7643c4b8b..f8a4e7216 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -2071,8 +2071,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         ggml_tensor src0_slice = *src0;
         src0_slice.ne[2]    = 1;
         src0_slice.nb[3]    = src0_slice.nb[2];
-        src0_slice.op       = GGML_OP_VIEW;
-        src0_slice.view_src = dst->src[0]; // non-const pointer to src0
+        src0_slice.op       = GGML_OP_NONE;
+        src0_slice.view_src = nullptr;
         src0_slice.data     = (char *) src0->data + i02*nb02;

         ggml_tensor src1_slice;

[JohannesGaessler]

It's correct that an assert is being triggered because clearing the supposed padding would actually be unsafe in that situation. It's just that I did not correctly consider that the effective number of tokens for the slice of src0 could fall into a range where the code in ggml_cuda_mul_mat could decide to use MMVQ or MMQ after all.

[JohannesGaessler]

I relaxed the condition a bit because for src0->ne[0] % MATRIX_ROW_PADDING == 0 it would still be fine to use MMVQ or MMQ.

[slaren]

This doesn't fix the issue because the assert is still triggered. IMO the correct solution here would be to clear the padding of the view_src, but only if actually necessary.

[JohannesGaessler]

I think the asserts were also in the wrong place. There were cases where they were being triggered even if cudaMemsetAsync is not called.