| #!/usr/bin/env bash |
| set -euo pipefail |
|
|
| ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" |
| BUILD=b11177 |
| RELEASE="https://github.com/ggml-org/llama.cpp/releases/download/${BUILD}" |
| mkdir -p "$ROOT/runtime" "$ROOT/models" "$ROOT/logs" "$ROOT/secrets" "$ROOT/results" |
|
|
| download() { |
| local url="$1" dest="$2" expected="$3" |
| if [[ -f "$dest" ]] && [[ "$(stat -c %s "$dest")" == "$expected" ]]; then |
| printf 'Already complete: %s\n' "$dest" |
| return |
| fi |
| curl --http1.1 -fL --retry 8 --retry-all-errors -C - "$url" -o "$dest" |
| [[ "$(stat -c %s "$dest")" == "$expected" ]] || { echo "Size mismatch: $dest" >&2; exit 1; } |
| } |
|
|
| download "$RELEASE/llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" \ |
| "$ROOT/runtime/llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" 169512356 |
| download "$RELEASE/cudart-llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" \ |
| "$ROOT/runtime/cudart-llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" 594377693 |
| printf '%s %s\n' \ |
| 7f584b4911be091788468412022dde72c8337af8db880d5f748e87563b25a82a "$ROOT/runtime/llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" \ |
| cea160366caea83923d76a676fc8591033d92138f35250cd3dc8b11f2edd57d8 "$ROOT/runtime/cudart-llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" \ |
| | sha256sum --check --status |
|
|
| mkdir -p "$ROOT/runtime/${BUILD}" |
| tar -xzf "$ROOT/runtime/llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" -C "$ROOT/runtime/${BUILD}" --strip-components=1 |
| tar -xzf "$ROOT/runtime/cudart-llama-${BUILD}-bin-ubuntu-cuda-12.8-x64.tar.gz" -C "$ROOT/runtime/${BUILD}" --strip-components=1 |
|
|
| download_model() { |
| local repo="$1" file="$2" revision="$3" expected="$4" |
| if [[ -f "$ROOT/models/$file" ]] && [[ "$(stat -c %s "$ROOT/models/$file")" == "$expected" ]]; then |
| printf 'Already complete: %s\n' "$ROOT/models/$file" |
| return |
| fi |
| if [[ ! -x "$ROOT/.venv/bin/python" ]]; then |
| python -m venv "$ROOT/.venv" |
| fi |
| "$ROOT/.venv/bin/python" -m pip install -q 'huggingface_hub[hf_xet]==1.33.0' |
| HF_XET_HIGH_PERFORMANCE=1 "$ROOT/.venv/bin/python" - "$repo" "$file" "$revision" "$ROOT/models" <<'PY' |
| import sys |
| from huggingface_hub import hf_hub_download |
| repo, filename, revision, local_dir = sys.argv[1:] |
| print(hf_hub_download(repo_id=repo, filename=filename, revision=revision, local_dir=local_dir)) |
| PY |
| [[ "$(stat -c %s "$ROOT/models/$file")" == "$expected" ]] || { echo "Size mismatch: $file" >&2; exit 1; } |
| } |
|
|
| download_model '6block/Qwen3.8-27B-GGUF' 'Qwen3.8-27B-Q6_K.gguf' \ |
| 5312b0f1c3705b45a48a18f051b4dc482627c17f 22563871392 |
| download_model 'unsloth/Qwen3.6-35B-A3B-GGUF' 'Qwen3.6-35B-A3B-UD-Q5_K_M.gguf' \ |
| a483e9e6cbd595906af30beda3187c2663a1118c 26456194016 |
| if [[ "${WITH_MTP:-0}" == 1 ]]; then |
| download_model '6block/Qwen3.8-27B-GGUF' 'mtp-Qwen3.8-27B-Q8_0.gguf' \ |
| 5312b0f1c3705b45a48a18f051b4dc482627c17f 3164006816 |
| printf '%s %s\n' e871427e49333b6eb05bef1745c811f4259455af9d67e5b577e0ef5fe9a8c1e5 \ |
| "$ROOT/models/mtp-Qwen3.8-27B-Q8_0.gguf" | sha256sum --check --status |
| fi |
|
|
| printf '%s %s\n' \ |
| 87a379aa1ec5d06b740c73ab5610d3925de0e9d7bc99c933447f27eb6a8f2bcc "$ROOT/models/Qwen3.8-27B-Q6_K.gguf" \ |
| c13ce26253ea334df472bd8fbd2d6da66d8a41195c17f6fcbf44c4d20ece0932 "$ROOT/models/Qwen3.6-35B-A3B-UD-Q5_K_M.gguf" \ |
| | sha256sum --check --status |
|
|
| printf 'Installed llama.cpp %s and both models.\n' "$BUILD" |
|
|
