Name and Version

$ ./llama-cli --version
version: 9159 (5c0e94683)
built with GNU 15.2.1 for Linux x86_64

The large matmul warptile configuration (mul_mat_l) was previously
enabled for AMD GPUs with coopmat support, but the MUL_MAT_ID path
(mul_mat_id_l), used by Mixture-of-Experts models, was hardcoded
to false regardless of coopmat availability.

Enable mul_mat_id_l when coopmat is supported and the driver is not
proprietary AMD, matching the same condition used for mul_mat_l.

0001-vulkan-enable-large-warptile-for-MUL_MAT_ID-on-AMD-w.patch

Operating systems

Linux

GGML backends

Vulkan

Hardware

AMD RYZEN AI MAX+ 395 w/ Radeon 8060S

Models

Problem description & steps to reproduce

First Bad Commit

No response

Relevant log output