Name and Version

❯ build/bin/llama-cli --version
ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
Device 0: AMD Radeon RX 6800 XT, gfx1030 (0x1030), VMM: no, Wave Size: 32, VRAM: 16368 MiB
version: 139 (a9883db)
built with Clang 22.1.3 for Linux x86_64

Operating systems

Linux

GGML backends

Vulkan

Hardware

Intel Core i7-10700K 3.8 GHz 8-Core Processor
RX 6800 XT 16GB VRAM

ROCm: 7.2.2
Vulkan: RADV NAVI21 (Mesa 26.1.0)

Models

Unsloth Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf
IQ3_XXS

Problem description & steps to reproduce

Running llama-bench with Vulkan backend on AMD RX 6800 XT crashes immediately with a segfault starting with build b139. The same build works fine with ROCm, and llama-cli with Vulkan works fine too.

I tracked it down to the commit that broke it. 706fbd8

RDNA2 doesn't support cooperative matrix operations (coopmat), so mul_mat_l is set to false. The switch statement sets it properly on lines 5651 and 5685.

Changing lines 5724 - 5729 to use the values from the switch statement fixes the issue.

  device->mul_mat_l_int[i]    = device->mul_mat_l[i];
  device->mul_mat_m_int[i]    = device->mul_mat_m[i];
  device->mul_mat_s_int[i]    = device->mul_mat_s[i];
  device->mul_mat_id_l_int[i] = device->mul_mat_id_l[i];
  device->mul_mat_id_m_int[i] = device->mul_mat_id_m[i];
  device->mul_mat_id_s_int[i] = device->mul_mat_id_s[i];

First Bad Commit

706fbd8

Relevant log output

build-vulkan/bin/llama-bench -m ~/.cache/huggingface/hub/models--unsloth--Qwen3.6-35B-A3B-GGUF/snapshots/*/Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf -p 512,1024 -n 128
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 6800 XT (RADV NAVI21) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 65536 | int dot: 1 | matrix cores: none
| model                          |       size |     params | backend    | ngl |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
fish: Job 1, 'build-vulkan/bin/llama-bench -m…' terminated by signal SIGSEGV (Address boundary error)