scripts/fetch_vision.sh

#!/usr/bin/env bash
# Janus-35B — fetch the vision projector (mmproj) for image input.
#
# Why this is separate from build.sh:
#   build.sh is for the Ollama text path. The mmproj is only known-good on
#   llama.cpp / llama-cpp-python: attaching it to Ollama was observed failing
#   with `unknown model architecture` and has not been re-tested since
#   ollama/ollama#14575 closed (see README Vision section). Note there is no
#   "vendored llama.cpp fork" -- Ollama serves every GGML model through the
#   bundled llama-server (ollama/ollama#16031); the old wording here was wrong.
#
# Usage:
#   ./scripts/fetch_vision.sh                    # default: BF16 (902 MB)
#   MMPROJ_PATH=/path/to/proj.gguf ./scripts/fetch_vision.sh   # destination override
#
# The Heretic GGUF repo publishes the projector as BF16 only (its Files
# listing, HF API, 2026-09-18); for another precision, point REPO_ID /
# FILE_NAME at a projector from any other Qwen3.6-35B-A3B GGUF repo:
#   REPO_ID=<other-qwen3.6-35b-a3b-gguf-repo> FILE_NAME=mmproj-F16.gguf ./scripts/fetch_vision.sh
# (vision tokens are projected the same way across Qwen 3.6 35B-A3B
# finetunes, so a family projector is functionally interchangeable.)
#
# Requires: huggingface-cli (or hf).
set -euo pipefail

# Resolve a Hugging Face CLI. The bare name `hf` is ambiguous: Arch's `hf`
# package ships an unrelated hidden-file utility under that name, so probing
# with `command -v hf` alone can select the wrong binary and turn a download
# into a baffling argument error. Prefer the unambiguous `huggingface-cli`,
# and accept `hf` only when its help text identifies it as Hugging Face's.
detect_hf() {
    local candidate
    for candidate in huggingface-cli hf; do
        command -v "${candidate}" >/dev/null 2>&1 || continue
        if "${candidate}" --help 2>&1 | grep -qiE 'huggingface|hf\.co'; then
            printf '%s\n' "${candidate}"
            return 0
        fi
    done
    return 1
}

PRECISION="${1:-${PRECISION:-BF16}}"
REPO_ID="${REPO_ID:-llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-GGUF}"
FILE_NAME="${FILE_NAME:-Qwen3.6-35B-A3B-uncensored-heretic-mmproj-${PRECISION}.gguf}"
ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
DEST="${MMPROJ_PATH:-${ROOT}/${FILE_NAME}}"

echo "[*] repo:      ${REPO_ID}"
echo "[*] precision: ${PRECISION}"
echo "[*] file:      ${FILE_NAME}"
echo "[*] dest:      ${DEST}"

if [[ -f "${DEST}" ]]; then
    echo "[=] already present at ${DEST}, skipping."
    exit 0
fi

if ! HF="$(detect_hf)"; then
    echo "[!] No Hugging Face CLI found (an unrelated 'hf' on PATH does not count)." >&2
    echo "    pip install -U huggingface_hub" >&2
    exit 1
fi

DEST_DIR="$(dirname "${DEST}")"
mkdir -p "${DEST_DIR}"

"${HF}" download "${REPO_ID}" "${FILE_NAME}" --local-dir "${DEST_DIR}"

# `hf download REPO FILE --local-dir DIR` writes to DIR/FILE: only the directory
# part of MMPROJ_PATH is honoured and the basename is discarded. Without this
# move, MMPROJ_PATH=~/models/mmproj.gguf downloaded ~900 MB successfully and then
# reported "download failed" and exited 1, and the re-run guard above probed a
# path that was never created.
FETCHED="${DEST_DIR}/${FILE_NAME}"
if [[ "${FETCHED}" != "${DEST}" && -f "${FETCHED}" ]]; then
    mv -- "${FETCHED}" "${DEST}"
fi

if [[ ! -f "${DEST}" ]]; then
    echo "[!] download failed: ${DEST} not present." >&2
    exit 1
fi

echo
echo "[+] Done. Use it via:"
echo "    python ${ROOT}/examples/llama_cpp_vision.py \\"
echo "        --gguf   ${ROOT}/Janus-35B-A3B.Q4_K_M.gguf \\"
echo "        --mmproj ${DEST} \\"
echo "        --image  /path/to/photo.jpg \\"
echo "        --prompt 'Describe this image.'"