1#!/usr/bin/env bash
 2
 3set -e
 4
 5CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
 6QUANTIZED_TYPE="${2:-"$QUANTIZED_TYPE"}"
 7TOKEN_EMBD_TYPE="${3:-"${TOKEN_EMBD_TYPE}"}"
 8OUTPUT_TYPE="${4:-"${OUTPUT_TYPE}"}"
 9BUILD_DIR="${5:-"$BUILD_DIR"}"
10QUANTIZED_MODEL=$CONVERTED_MODEL
11
12# Final check if we have a model path
13if [ -z "$CONVERTED_MODEL" ]; then
14    echo "Error: Model path must be provided either as:" >&2
15    echo "  1. Command line argument" >&2
16    echo "  2. CONVERTED_MODEL environment variable" >&2
17    exit 1
18fi
19
20if [ -z "$QUANTIZED_TYPE" ]; then
21    echo "Error: QUANTIZED_TYPE is required" >&2
22    exit 1
23fi
24
25echo $CONVERTED_MODEL
26
27# Process the quantized model filename
28if [[ "$QUANTIZED_MODEL" == *.gguf ]]; then
29    # Remove .gguf suffix, add quantized type, then add .gguf back
30    BASE_NAME="${QUANTIZED_MODEL%.gguf}"
31    QUANTIZED_MODEL="${BASE_NAME}-${QUANTIZED_TYPE}.gguf"
32else
33    echo "Error: QUANTIZED_MODEL must end with .gguf extension" >&2
34    exit 1
35fi
36
37if [ -z "$BUILD_DIR" ]; then
38    BUILD_DIR="../../build"
39fi
40
41cmake --build $BUILD_DIR --target llama-quantize -j8
42
43echo $TOKEN_EMBD_TYPE
44echo $OUTPUT_TYPE
45
46CMD_ARGS=("${BUILD_DIR}/bin/llama-quantize")
47[[ -n "$TOKEN_EMBD_TYPE" ]] && CMD_ARGS+=("--token-embedding-type" "$TOKEN_EMBD_TYPE")
48[[ -n "$OUTPUT_TYPE" ]]     && CMD_ARGS+=("--output-tensor-type" "$OUTPUT_TYPE")
49CMD_ARGS+=("$CONVERTED_MODEL" "$QUANTIZED_MODEL" "$QUANTIZED_TYPE")
50
51"${CMD_ARGS[@]}"
52
53echo "Quantized model saved to: $QUANTIZED_MODEL"