CoolFace
Modelpublic

Felipe97/llama-cpp-compiled

sourceHugging Faceupdated 3d agoView on Hugging Face
0likes1.1kdownloads
compare-embeddings-logits.sh85 linesDownload Raw Back to embedding
1#!/usr/bin/env bash2 3set -e4 5# Parse command line arguments6MODEL_PATH=""7MODEL_NAME=""8PROMPTS_FILE=""9 10# First argument is always model path11if [ $# -gt 0 ] && [[ "$1" != --* ]]; then12    MODEL_PATH="$1"13    shift14fi15 16# Parse remaining arguments17while [[ $# -gt 0 ]]; do18    case $1 in19        --prompts-file|-pf)20            PROMPTS_FILE="$2"21            shift 222            ;;23        *)24            # If MODEL_NAME not set and this isn't a flag, use as model name25            if [ -z "$MODEL_NAME" ] && [[ "$1" != --* ]]; then26                MODEL_NAME="$1"27            fi28            shift29            ;;30    esac31done32 33# Set defaults34MODEL_PATH="${MODEL_PATH:-"$EMBEDDING_MODEL_PATH"}"35MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"36 37CONVERTED_MODEL_PATH="${CONVERTED_EMBEDDING_PATH:-"$CONVERTED_EMBEDDING_MODEL"}"38CONVERTED_MODEL_NAME="${CONVERTED_MODEL_NAME:-$(basename "$CONVERTED_MODEL_PATH" .gguf)}"39 40if [ -t 0 ]; then41    CPP_EMBEDDINGS="data/llamacpp-${CONVERTED_MODEL_NAME}-embeddings.bin"42else43    # Process piped JSON data and convert to binary (matching logits.cpp format)44    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)45    python3 -c "46import json47import sys48import struct49 50data = json.load(sys.stdin)51 52# Flatten all embeddings completely53flattened = []54for item in data:55    embedding = item['embedding']56    for token_embedding in embedding:57        flattened.extend(token_embedding)58 59print(f'Total embedding values: {len(flattened)}', file=sys.stderr)60 61# Write as binary floats - matches logitc.cpp fwrite format62with open('$TEMP_FILE', 'wb') as f:63    for value in flattened:64        f.write(struct.pack('f', value))65"66    CPP_EMBEDDINGS="$TEMP_FILE"67    trap "rm -f $TEMP_FILE" EXIT68fi69 70# Build the semantic_check.py command71SEMANTIC_CMD="python scripts/utils/semantic_check.py --model-path $MODEL_PATH \72    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \73    --cpp-embeddings $CPP_EMBEDDINGS"74 75# Add prompts file if specified, otherwise use default prompt76if [ -n "$PROMPTS_FILE" ]; then77    SEMANTIC_CMD="$SEMANTIC_CMD --prompts-file \"$PROMPTS_FILE\""78else79    SEMANTIC_CMD="$SEMANTIC_CMD --prompt \"Hello world today\""80fi81 82# Execute the command83eval $SEMANTIC_CMD84 85