UPSTREAM PR #18724: Corrected: In ggml_cuda_mul_mat_q(), s13 is declared incorrectly to nb[2] instead of nb[3]

[loci-dev]

Mirrored from ggml-org/llama.cpp#18724

In ggml_cuda_mul_mat_q(), when in the MoE path and when quantizing src1, s13 is incorrectly declared to nb[2] instead of nb[3].

Earlier in the function:
const int64_t s03 = src0->nb[3] / ts_src0;
const int64_t s3 = dst->nb[3] / ts_dst;
...

if (!ids) {
...
{
const int64_t s11 = src1->nb[1] / ts_src1;
const int64_t s12 = src1->nb[2] / ts_src1;
const int64_t s13 = src1->nb[3] / ts_src1;
}
Later, in the ids path:
const int64_t s11 = src1->nb[1] / ts_src1;
const int64_t s12 = src1->nb[2] / ts_src1;
const int64_t s13 = src1->nb[2] / ts_src1; <--- [2] here, repeated, should be [3].

ne13 is currently asserted to 1 , so this very likely doesn't affect anything yet, but it should still be fixed for correctness.

Copilot summary:
This pull request fixes an indexing bug in the ggml_cuda_mul_mat_q function in mmq.cu. The change corrects the calculation of the s13 variable to use the correct dimension of the src1 tensor, which improves the accuracy of subsequent operations.

• Bug fix:
  • Corrected the calculation of s13 to use src1->nb[3] instead of src1->nb[2], ensuring the correct tensor dimension is used in matrix multiplication operations.

[loci-review]

Explore the complete analysis inside the Version Insights

I've successfully generated a summary report for your project. Here are the key findings:

Summary Report for llama.cpp PR #871

Good News: ✅ No significant performance regressions were detected in this pull request.

Key Points:

• The analysis compared two versions of the code (base vs. target)
• No modified functions showed performance changes greater than 2% threshold
• Both Response Time (execution time per function) and Throughput Time (time including callees) remained stable

Recommendation: This pull request appears safe to merge from a performance perspective, as all functions maintained stable performance within acceptable thresholds.

The analysis indicates that the changes in PR #871 for the auroralabs-loci/llama.cpp repository have not introduced any measurable performance degradation.