[Bug] Vulkan ErrorOutOfDeviceMemory on Windows with AMD Radeon HD 8570 despite --gpu none

[PORTRAITART1]

Environment:

• Windows 10 (26200)
• Docker Desktop 4.92.0
• GPU: AMD Radeon HD 8570 (1 GB VRAM)
• Docker Model Runner: llama.cpp b9879-cpu

Problem:
Docker Model Runner tries to use Vulkan even when installed with --gpu none.
Models >= 1B parameters crash with ggml_vulkan: ErrorOutOfDeviceMemory.
Small model (smollm2:360M) runs but produces garbage output.

Steps to reproduce:

1. docker model install-runner --backend llama.cpp --gpu none
2. docker model run ai/gemma3:1B-Q4_K_M "Bonjour"
3. Observe crash

Expected behavior:
Model should run on CPU when --gpu none is specified.

Actual behavior:
llama.cpp still attempts Vulkan allocation, fails, and aborts.

Error logs:
_layers already set by user to 999, abort
ggml_vulkan: Device memory allocation of size 57671680 failed.
ggml_vulkan: vk::Device::allocateMemory: ErrorOutOfDeviceMemory
llama_init_from_model: failed to initialize the context: failed to allocate buffer for kv cache

Note:
The environment variable DOCKER_MODEL_RUNNER_LLAMA_CPP_ARGS="--n-gpu-layers 0 --no-vulkan" did not resolve the issue.
Docker Desktop requirements list only NVIDIA GPUs for Windows (amd64), but the error suggests Vulkan is still being attempted on AMD hardware.

[ericcurtin]

Likely cause: --gpu none isn't disabling Vulkan device selection, so llama.cpp still tries GPU offload. Worth checking the GPU/backend selection logic to force CPU-only when --gpu none is set.

Note: model-runner is being deprecated in favor of https://github.com/llmmanorg/llmman. Please open this issue there instead. A PR there would be very welcome if you want to fix it.