#!/bin/sh
# Downloads the configured GGML model on first start, then runs whisper-server.
# Uses the GPU (Vulkan) automatically when one is present; set NO_GPU=1
# to force CPU.
#
# Both the freshly downloaded file and any existing file are validated
# (minimum size + ggml magic bytes) so a truncated download or a saved
# error/redirect page can never be mistaken for a model — a poisoned
# file is deleted and re-downloaded instead.
set -eu

MODEL="${MODEL:-small}"
THREADS="${THREADS:-4}"
HOST="${HOST:-0.0.0.0}"
PORT="${PORT:-8085}"
MODEL_DIR="${MODEL_DIR:-/models}"
FILE="$MODEL_DIR/ggml-$MODEL.bin"
URL="https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-$MODEL.bin"

# smallest supported model (ggml-tiny.bin) is ~75 MB
MIN_SIZE=50000000
GGML_MAGIC="$(printf 'lmgg')"   # GGML_FILE_MAGIC 0x67676d6c, little-endian

is_valid_model() {
    [ -s "$1" ] || return 1
    [ "$(stat -c%s "$1")" -ge "$MIN_SIZE" ] || return 1
    [ "$(head -c 4 "$1")" = "$GGML_MAGIC" ] || return 1
    return 0
}

mkdir -p "$MODEL_DIR"

if ! is_valid_model "$FILE" && [ -e "$FILE" ]; then
    echo "existing $FILE is invalid ($(stat -c%s "$FILE") bytes) — " \
        "deleting and re-downloading"
fi

if ! is_valid_model "$FILE"; then
    rm -f "$FILE" "$FILE.part"
    echo "downloading $URL to $FILE ..."
    curl -fLS --retry 3 --retry-delay 5 -C - -o "$FILE.part" "$URL"

    if ! is_valid_model "$FILE.part"; then
        size="$(stat -c%s "$FILE.part" 2>/dev/null || echo 0)"
        echo "ERROR: download is not a valid model ($size bytes)."
        echo "------- first 400 bytes of the response (probably an error/redirect page):"
        head -c 400 "$FILE.part"
        echo ""
        echo "------- end of response"
        rm -f "$FILE.part"
        exit 1
    fi
    mv "$FILE.part" "$FILE"
    echo "downloaded $(stat -c%s "$FILE") bytes, magic ok."
fi

GPU_FLAGS=""
if [ "${NO_GPU:-0}" = "1" ]; then
    GPU_FLAGS="-ng"
fi

echo "starting whisper-server: model=$MODEL threads=$THREADS host=$HOST port=$PORT gpu=auto"
exec whisper-server \
    -m "$FILE" \
    -l auto \
    -t "$THREADS" \
    $GPU_FLAGS \
    --host "$HOST" \
    --port "$PORT"