dbrx-instruct-FP8-KV

dbrx-instruct-FP8-KV

• Introduction

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

• Quantization Stragegy

  • Quantized Layers: All linear layers excluding "lm_head" and "router.layer"
  • Weight: FP8 symmetric per-tensor
  • Activation: FP8 symmetric per-tensor
  • KV Cache: FP8 symmetric per-tensor

• 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 databricks/dbrx-instruct
# single GPU
python3 quantize_quark.py \ 
        --model_dir $MODEL_DIR \
        --output_dir dbrx-instruct-FP8-KV \                           
        --quant_scheme w_fp8_a_fp8 \
        --kv_cache_dtype fp8 \
        --num_calib_data 128 \
        --model_export quark_safetensors \
        --no_weight_matrix_merge \
        --custom_mode fp8
# If model size is too large for single GPU, please use multi GPU instead.
python3 quantize_quark.py
        --model_dir $MODEL_DIR \
        --output_dir dbrx-instruct-FP8-KV\
        --quant_scheme w_fp8_a_fp8 \
        --kv_cache_dtype fp8 \
        --num_calib_data 128 \
        --multi_gpu \
        --model_export quark_safetensors \
        --no_weight_matrix_merge \
        --custom_mode fp8

Deployment

Quark has its own export format and allows FP8 quantized models to be efficiently deployed using the vLLM backend(vLLM-compatible). In the dbrx-instruct model, "transformer.blocks.*.ffn.experts" modules can be divided into experts-num mlps, and if the shape of the weight of w1 in one of the mlps is [dim1, dim2], then the shape of “transformer.blocks.*.ffn.experts.mlp.w1.weight“ in the exported safetensors file is [dim1*experts-num, dim2]. The shapes of "transformer.blocks.*.ffn.experts.mlp.w1.weight_scale" and "transformer.blocks.*.ffn.experts.mlp.w1.input_scale" are [dim1]. Similarly, this also applies to the w2 and v1 of "transformer.blocks.*.ffn.experts.mlp".

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

<table> <tr> <td><strong>Benchmark</strong> </td> <td><strong>dbrx-instruct </strong> </td> <td><strong>dbrx-instruct-FP8-KV(this model)</strong> </td> </tr> <tr> <td>Perplexity-wikitext2 </td> <td>4.2275 </td> <td>4.3033 </td> </tr> </table>

License

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