Pull a llamacpp Model

This machine uses kyuz0/amd-strix-halo-toolboxes:rocm-7.2 for llamacpp inference (AMD Strix Halo / gfx1151 — not supported by the official ROCm build). Harbor's pull mechanism starts an ephemeral container with --n-gpu-layers 0; the custom image fails in that context without ROCm device access. Use the standard CPU image just for pulling, then restore.

Steps

1. Switch to the standard CPU image

harbor config set llamacpp.image.rocm ghcr.io/ggml-org/llama.cpp:server

2. Pull the model

harbor pull <hf-owner/model-repo:quantization>
# Examples:
harbor pull bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M
harbor pull unsloth/Mistral-Small-3.1-24B-Instruct-2503-GGUF:UD-Q4_K_XL

Harbor detects the HuggingFace model spec and routes it through llamacpp automatically.

3. Restore the custom image

harbor config set llamacpp.image.rocm kyuz0/amd-strix-halo-toolboxes:rocm-7.2

Notes

• The restore step is mandatory — without it, llamacpp will not use the GPU on the next harbor up llamacpp
• Extra args (-fa 1 -dio --no-mmap --ctx-size 64000 --fit off) are stored separately in config and are not affected
• To verify the model after pulling: harbor llamacpp model
• To verify the image was restored: harbor config get llamacpp.image.rocm