Convert SmolVLM to GGUF and push to HF hub

convert_smolvlm.sh

#!/bin/sh

convert_and_upload() {
    local src_user=$1
    local src_model=$2
    local has_q4=$3
    local base_name=$src_model
    target_model="$src_model-GGUF"
    target_repo="ggml-org/$target_model"

    mkdir -p $target_model

    huggingface-cli download "$src_user/$src_model" --local-dir $src_model
    python ../convert_hf_to_gguf.py \
        --outfile $target_model/$base_name-f16.gguf \
        --outtype f16 \
        $src_model

    ../build/bin/llama-quantize \
        $target_model/$base_name-f16.gguf \
        $target_model/$base_name-Q8_0.gguf \
        Q8_0

    if [ "$has_q4" = "true" ]; then
        ../build/bin/llama-quantize \
            $target_model/$base_name-f16.gguf \
            $target_model/$base_name-Q4_K_M.gguf \
            Q4_K_M
    fi

    python ../convert_hf_to_gguf.py \
        --outfile $target_model/mmproj-$base_name-f16.gguf \
        --outtype f16 --mmproj \
        $src_model

    echo "---" > $target_model/README.md
    echo "license: apache-2.0" >> $target_model/README.md
    echo "base_model: $src_user/$src_model" >> $target_model/README.md
    echo "---" >> $target_model/README.md
    echo "" >> $target_model/README.md
    echo "# $src_model" >> $target_model/README.md
    echo "" >> $target_model/README.md
    echo "Original model: https://huggingface.co/$src_user/$src_model" >> $target_model/README.md
    echo "" >> $target_model/README.md
    echo "For more info, please refer to this PR: https://github.com/ggml-org/llama.cpp/pull/13050" >> $target_model/README.md
    echo "" >> $target_model/README.md

    huggingface-cli upload $target_repo $target_model .

    rm -rf $target_model
    rm -rf $src_model
}

convert_and_upload HuggingFaceTB SmolVLM-Instruct              true
convert_and_upload HuggingFaceTB SmolVLM-256M-Instruct         false # no q4, quality degrades too much
convert_and_upload HuggingFaceTB SmolVLM-500M-Instruct         false # no q4, quality degrades too much
convert_and_upload HuggingFaceTB SmolVLM2-2.2B-Instruct        true
convert_and_upload HuggingFaceTB SmolVLM2-256M-Video-Instruct  false # no q4, quality degrades too much
convert_and_upload HuggingFaceTB SmolVLM2-500M-Video-Instruct  false # no q4, quality degrades too much