# Exact llama-server command used for the Vulkan numbers (one run per KV cache type: q8_0, then f16).
# Build:    llama-b11177-bin-ubuntu-vulkan-x64 (official ggml-org release b11177, commit 1ab7e5ad2), unmodified.
#           llama-server sha256   d8febc7c0d19b220dbeb51efc2b4d0b305e8a339317d4d29bf9c7b6693af23b8
#           libggml-vulkan.so     cc8e98040ca4579a926a69cba36b07b71402a01ebdb210ca29006134964dc3d2
# Driver:   Mesa 25.2.8 RADV (mesa-vulkan-drivers 25.2.8-0ubuntu0.24.04.4), libvulkan1 1.3.275.
# Model:    qwen3-14b-Q4_K_M.gguf (9 276 184 896 bytes, sha256 starts with a8cc1361f3145dc0).
# Devices:  `llama-server --list-devices` shows Vulkan0 = AMD Radeon RX 9070 XT (RADV GFX1201), Vulkan1 = Raphael iGPU.
# No HIP variables are set; the only environment variable is the library path of the build directory.

LD_LIBRARY_PATH=/path/to/llama-b11177 \
/path/to/llama-b11177/llama-server --model /path/to/qwen3-14b-Q4_K_M.gguf --alias qwen3:14b \
  --ctx-size 32768 --n-gpu-layers 99 --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0 \
  --device Vulkan0 --host 127.0.0.1 --port 8081