[tyronecai]

Overview

simd_gemm() has an incorrect A-matrix index in the scalar tail-column path for full row blocks.

Additional information

This appears to have been introduced by commit 08e6d914b8fef477d2a5aeeb89e08d42709981cf (ggml : avoid UB in gemm ukernel (llama/19642)), which changed the SIMD GEMM ukernel from using global ii / jj offsets to passing already-offset A, B, and C pointers.

Before that commit, the scalar tail-column path used the global row index:

a += A[(ii + i) * K + kk] * B[kk * N + jj];

After the commit, A is advanced after each full row block:

A += GEMM_RM * K;
C += GEMM_RM * N;

So inside the current block, the correct row-major index should be:

A[i * K + kk]

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:

[ggml-gh-bot]

Hi @tyronecai, 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.

---

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

[ggerganov]

Thanks for catching that.

[ggerganov]

For reference, this was the PR #19642