amd/Llama-2-7b-chat-hf-awq-g128-int4-asym-fp16-onnx-dml

meta-llama/Llama-2-7b-chat-hf

• Introduction

  • Quantization Tool: Quark 0.6.0
  • OGA Model Builder: v0.5.1

• Quantization Strategy

  • AWQ / Group 128 / Asymmetric / UINT4 Weights / FP16 activations
  • Excluded Layers: None

  python3 quantize_quark.py \
        --model_dir "$model" \
        --output_dir "$output_dir" \
        --quant_scheme w_uint4_per_group_asym \
        --num_calib_data 128 \
        --quant_algo awq \
        --dataset pileval_for_awq_benchmark \
        --seq_len 512 \
        --model_export quark_safetensors \
        --data_type float16 \
        --exclude_layers [] \
        --custom_mode awq
  

• OGA Model Builder

  python builder.py \
    -i <quantized safetensor model dir> \
    -o <oga model output dir> \
    -p int4 \
    -e dml