CYBER WATCH / EQUINOXE INTELLIGENCE

← All articles

b12433: cuda: fix MUL_MAT_ID abort with odd src0 columns on AMD (#32385)

llama.cpp Releases · 2026-11-17 10:01:12

On AMD devices, the MUL_MAT_ID operation with src0 matrices of type F32, F16 or BF16 and up to 8 tokens always used MMVF. MMVF reads src0 in pairs and asserts an even ne[0], so a rank‑1 LoRA on MoE expert weights (lora_b with ne[0] = 1) caused an abort. The fix uses MMVF only for even ne[0] shapes; other shapes fall back to the existing implementation. This resolves issue #30576.

Read original source ↗