amd/gemma-2-2b-awq-uint4-asym-g128-lmhead-g32-fp16-onnx-hybrid_v2

language:

• en pipeline_tag: text-generation base_model:
• google/gemma-2-2b license: gemma

gemma-2-2b-awq-uint4-asym-g128-lmhead-g32-fp16-onnx

• Introduction

  This model was created by applying Quark with calibration samples from Pile dataset.

• Quantization Strategy

  • Quantized Layers: All linear layers
  • Weight: uint4 asymmetric per-group. group_size=32 for lm_head, and group_size=128 for the rest.

• Quick Start

1. Download and install Quark
2. Run the quantization script in the example folder using the following command line:

   export MODEL_DIR = [local model checkpoint folder] or google/gemma-2-2b
   # single GPU
   python quantize_quark.py --model_dir $MODEL_DIR \
                           --output_dir output_dir $MODEL_NAME-awq-uint4-asym-g128-lmhead-g32-fp16 \
                           --quant_scheme w_uint4_per_group_asym \
                           --num_calib_data 128 \
                           --quant_algo awq \
                           --dataset pileval_for_awq_benchmark \
                           --model_export hf_format \
                           --group_size 128 \
                           --group_size_per_layer lm_head 32 \
                           --data_type float16 \
                           --exclude_layers
   # cpu
   python quantize_quark.py --model_dir $MODEL_DIR \
                           --output_dir output_dir $MODEL_NAME-awq-uint4-asym-g128-lmhead-g32-fp16 \
                           --quant_scheme w_uint4_per_group_asym \
                           --num_calib_data 128 \
                           --quant_algo awq \
                           --dataset pileval_for_awq_benchmark \
                           --model_export hf_format \
                           --group_size 128 \
                           --group_size_per_layer lm_head 32 \
                           --data_type float16 \
                           --exclude_layers \
                           --device cpu
   

Deployment

Quark has its own export format, quark_safetensors, which is compatible with autoAWQ exports.

Evaluation

Quark currently uses perplexity(PPL) as the evaluation metric for accuracy loss before and after quantization.The specific PPL algorithm can be referenced in the quantize_quark.py. The quantization evaluation results are conducted in pseudo-quantization mode, which may slightly differ from the actual quantized inference accuracy. These results are provided for reference only.

Evaluation scores

Benchmark | google/gemma-2-2b (float16) | amd/gemma-2-2b-awq-uint4-asym-g128-lmhead-g32-fp16-onnx (this model) |

Perplexity-wikitext2 | 64.41 | 71.43 (evalauted by CPU) |

License

Modifications copyright(c) 2024 Advanced Micro Devices,Inc. All rights reserved.