Roformer.cpp/.github/workflows/convert-model.yml

name: Convert Model to GGUF

on:
  workflow_dispatch:
    inputs:
      hf_repo:
        description: 'HuggingFace 仓库名称 (例如: GaboxR67/MelBandRoformers)'
        required: true
        type: string
      checkpoint_path:
        description: '权重文件路径 (相对于仓库, 例如: melbandroformers/vocals/voc_fv6.ckpt)'
        required: true
        type: string
      config_path:
        description: '配置文件路径 (相对于仓库, 例如: melbandroformers/vocals/voc_gabox.yaml)'
        required: true
        type: string
      model_name:
        description: '输出模型名称 (用于文件命名, 例如: voc_fv6)'
        required: false
        type: string
        default: 'MelBandRoformer-'
      quantization_types:
        description: '要转换的量化类型 (用逗号分隔, 留空则转换全部)'
        required: false
        type: string
        default: 'fp32,fp16,q8_0,q4_0,q4_1,q5_0,q5_1'
      gguf_name:
        description: 'GGUF 元数据中的模型名称 (可选, 默认: Mel-Band-Roformer Separator)'
        required: false
        type: string
      gguf_description:
        description: 'GGUF 元数据中的模型描述 (可选, 默认: Music source separation model)'
        required: false
        type: string

env:
  SUPPORTED_QUANT_TYPES: 'fp32,fp16,q8_0,q4_0,q4_1,q5_0,q5_1'

jobs:
  convert-to-all-quant-types:
    runs-on: ubuntu-latest
    
    steps:
      - name: Checkout
        uses: actions/checkout@v4
      
      - name: Setup Python
        uses: actions/setup-python@v5
        with:
          python-version: '3.11'
          
      - name: Install Dependencies
        run: |
          pip install torch --index-url https://download.pytorch.org/whl/cpu
          pip install huggingface_hub gguf librosa einops pyyaml numpy
          
      - name: Download Model from HuggingFace
        env:
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          python -c "
          from huggingface_hub import hf_hub_download
          import os
          
          token = os.environ.get('HF_TOKEN') or None
          repo = '${{ inputs.hf_repo }}'
          
          print(f'Downloading from {repo}...')
          
          # Download checkpoint
          ckpt_path = hf_hub_download(
              repo, 
              '${{ inputs.checkpoint_path }}', 
              local_dir='./model', 
              token=token
          )
          print(f'  Checkpoint: {ckpt_path}')
          
          # Download config
          config_path = hf_hub_download(
              repo, 
              '${{ inputs.config_path }}',
              local_dir='./model', 
              token=token
          )
          print(f'  Config: {config_path}')
          "
          
      - name: Show Model Info
        run: |
          echo "=== Model Configuration ==="
          cat model/${{ inputs.config_path }}
          echo ""
          echo "=== Checkpoint File Size ==="
          ls -lh model/${{ inputs.checkpoint_path }}
          
      - name: Create Output Directory
        run: mkdir -p output
        
      - name: Convert to All Quantization Types
        run: |
          #!/bin/bash
          set -e
          
          CHECKPOINT="model/${{ inputs.checkpoint_path }}"
          CONFIG="model/${{ inputs.config_path }}"
          MODEL_NAME="${{ inputs.model_name }}"
          GGUF_NAME="${{ inputs.gguf_name }}"
          GGUF_DESC="${{ inputs.gguf_description }}"
          QUANT_TYPES="${{ inputs.quantization_types }}"
          
          # If no types specified, use all supported types
          if [ -z "$QUANT_TYPES" ]; then
            QUANT_TYPES="${{ env.SUPPORTED_QUANT_TYPES }}"
          fi
          
          echo "=== Converting Model: $MODEL_NAME ==="
          echo "Checkpoint: $CHECKPOINT"
          echo "Config: $CONFIG"
          echo "Quantization Types: $QUANT_TYPES"
          echo ""
          
          # Convert comma-separated to array
          IFS=',' read -ra TYPES <<< "$QUANT_TYPES"
          
          for qtype in "${TYPES[@]}"; do
            # Trim whitespace
            qtype=$(echo "$qtype" | xargs)
            
            # Generate output filename following GGUF naming convention
            # Format: {model_name}-{quantization}.gguf
            OUTPUT_FILE="output/${MODEL_NAME}-${qtype^^}.gguf"
            
            echo ">>> Converting to $qtype -> $OUTPUT_FILE"
            
            # Build convert command with optional metadata args
            CMD="python scripts/convert_to_gguf.py --ckpt \"$CHECKPOINT\" --config \"$CONFIG\" --out \"$OUTPUT_FILE\" --dtype \"$qtype\""
            [ -n "$GGUF_NAME" ] && CMD="$CMD --name \"$GGUF_NAME\""
            [ -n "$GGUF_DESC" ] && CMD="$CMD --description \"$GGUF_DESC\""
            
            eval $CMD
              
            # Show file size
            SIZE=$(ls -lh "$OUTPUT_FILE" | awk '{print $5}')
            echo "    Output size: $SIZE"
            echo ""
          done
          
          echo "=== Conversion Complete ==="
          ls -lh output/
          
      - name: Generate Summary Report
        run: |
          python -c "
          import os
          import json
          
          output_dir = 'output'
          model_name = '${{ inputs.model_name }}'
          
          report = {
              'model_name': model_name,
              'hf_repo': '${{ inputs.hf_repo }}',
              'checkpoint': '${{ inputs.checkpoint_path }}',
              'config': '${{ inputs.config_path }}',
              'files': []
          }
          
          total_size = 0
          for f in sorted(os.listdir(output_dir)):
              if f.endswith('.gguf'):
                  path = os.path.join(output_dir, f)
                  size_bytes = os.path.getsize(path)
                  total_size += size_bytes
                  
                  # Extract quant type from filename
                  quant_type = f.replace(f'{model_name}-', '').replace('.gguf', '')
                  
                  report['files'].append({
                      'filename': f,
                      'quant_type': quant_type,
                      'size_bytes': size_bytes,
                      'size_mb': round(size_bytes / 1024 / 1024, 2)
                  })
          
          report['total_size_bytes'] = total_size
          report['total_size_mb'] = round(total_size / 1024 / 1024, 2)
          
          # Write JSON report
          with open('output/conversion_report.json', 'w') as f:
              json.dump(report, f, indent=2)
          
          # Print markdown table for GitHub Actions summary
          print('## Conversion Report')
          print('')
          print(f'**Model**: {model_name}')
          print(f'**Source**: [{report[\"hf_repo\"]}](https://huggingface.co/{report[\"hf_repo\"]})')
          print('')
          print('| Quantization | Filename | Size |')
          print('|--------------|----------|------|')
          for f in report['files']:
              print(f'| {f[\"quant_type\"]} | {f[\"filename\"]} | {f[\"size_mb\"]} MB |')
          print('')
          print(f'**Total Size**: {report[\"total_size_mb\"]} MB')
          " | tee -a $GITHUB_STEP_SUMMARY
          
      - name: Upload Converted Models
        uses: actions/upload-artifact@v4
        with:
          name: gguf-models-${{ inputs.model_name }}
          path: |
            output/*.gguf
            output/conversion_report.json
          retention-days: 30
          compression-level: 0  # GGUF is already compressed for quantized types