Struggling to get Vulkan to work for 285hx iGPU

[StudioNirin]

Basically tried every flavour and type of llama.cpp / lm-studio / wrapper / etc to try and get this to work. But my best efforts either end up with the attempt force-closing due to some incompatibility error, or it'll run but only at about 1tok/s at best.

CPU-only I'm getting 18t/s and with SYCL I'm getting about 14t/s.

It might just be that vulkan is no improvement, but I thought it would be at least better than SYCL, while also using less power than CPU only.

I am running this on an unraid box using the llama.cpp docker container.

warn: LLAMA_ARG_HOST environment variable is set, but will be overwritten by command line argument --host
0.00.159.529 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
0.00.162.022 I srv    load_model: loading model '/models/Qwen3.6-35B-A3B-IQ4_XS-4.19bpw-BS-MTP.gguf'
0.18.476.913 I srv    load_model: initializing, n_slots = 4, n_ctx_slot = 262144, kv_unified = 'true'
0.18.500.124 I srv          init: chat template supports preserving reasoning, consider enabling it via --reasoning-preserve
0.18.500.193 I srv  llama_server: model loaded
0.18.500.206 I srv  llama_server: listening on http://0.0.0.0:8000
0.34.322.178 I slot get_availabl: id  3 | task -1 | selected slot by LRU, t_last = -1
0.34.322.323 I slot launch_slot_: id  3 | task 0 | processing task, is_child = 0
1.08.275.858 I slot print_timing: id  3 | task 0 | prompt processing, n_tokens =     16, progress = 0.80, t =  33.95 s / 0.47 tokens per second
1.53.710.325 I srv    operator(): operator(): cleaning up before exit...
warn: LLAMA_ARG_HOST environment variable is set, but will be overwritten by command line argument --host
0.00.140.453 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
0.00.142.428 I srv    load_model: loading model '/models/Qwen3.6-35B-A3B-IQ4_XS-4.19bpw-BS-MTP.gguf'

This is an attempt where it 'runs' but the output is extremely slow and is just a line of gibberish characters.

docker run
  -d
  --name='llama.cpp-vulkan'
  --net='bridge'
  --pids-limit 2048
  -e TZ="Europe/London"
  -e HOST_OS="Unraid"
  -e HOST_HOSTNAME="Hex"
  -e HOST_CONTAINERNAME="llama.cpp-vulkan"
  -e 'GGML_VK_DISABLE_INTEGER_DOT_PRODUCT'='1'
  -l net.unraid.docker.managed=dockerman
  -l net.unraid.docker.webui='http://[IP]:[PORT:8000]/'
  -l net.unraid.docker.icon='https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png'
  -p '8009:8000/tcp'
  -v '/mnt/user/appdata/llama_cpp/model':'/models':'rw'
  --device='/dev/dri' 'ghcr.io/ggml-org/llama.cpp:server-vulkan'
  -m /models/Qwen3.6-35B-A3B-IQ4_XS-4.19bpw-BS-MTP.gguf
  --port 8000
  --host 0.0.0.0
  --n-gpu-layers 99

Already tried the GGML_VK_DISABLE thing I read elsewhere but it seemed to do nothing.

[jruhe-adesso]

Give llama-bench a try to see what kind of performance you can expect from Intel's older 4-core Xe-LPG Meteor Lake architecture

[StudioNirin]

wouldn't this require having vulkan working? cos... otherwise surely the bench will just give the same result as the generation. Being very little.

[jruhe-adesso]

Lama-bench is using other defaults than llama-server which makes results comparable (see #10879). You are trying Qwen3.6-35B-A3B-IQ4_XS-4.19bpw-BS-MTP.gguf full context, four slots on your tiny iGPU. What do you expect?

[StudioNirin]

Well, going from what people keep telling me, I expected to get similar performance to the cpu but at lower power draw.

So pp of about 150 and tg of around 25tok/s.

At the very least id have expected to get the same results as SYCL which is 40pp and 15tg.

Instead its 1pp and 0.5tg (or just constant crashing) so 🤷

Ill probably just stick with cpu on ik-llama as that is at least reliable. If I need anything elaborate ill do it on my main pc or get a LP gpu for the server.