Perplexity Evaluations

Below details how to run perplexity evaluations. Perplexity evaluations utilize the wikitext2 dataset. Supported devices currently are CPU and GPU.

Summary of support:

Model Types	Quark Quantized	Pretrained	Perplexity	ROUGE	METEOR	LM Eval Harness Tasks
LLMs
Torch	✓	✓	✓	✓	✓	✓
ONNX	✓	✓	X	✓	✓	✓
VLMs
Torch	✓	✓	✓	✓	✓	✓
ONNX	X	X	X	X	X	X

Recipes

• The --ppl argument specifies the perplexity task.

PPL on Torch Models

1. PPL on a pretrained LLM. Example with Llama2-7b-hf:

python llm_eval.py \
    --model_args pretrained=meta-llama/Llama-2-7b-hf \
    --ppl \
    --trust_remote_code \
    --batch_size 1 \
    --device cuda

Alternatively, to load a local checkpoint:

python llm_eval.py \
    --model_args pretrained=[local checkpoint path] \
    --ppl \
    --trust_remote_code \
    --batch_size 1 \
    --device cuda

2. PPL on a Quark Quantized model. Example with Llama-2-7b-chat-hf-awq-uint4-asym-g128-bf16-lmhead

python llm_eval.py \
    --model_args pretrained=meta-llama/Llama-2-7b-hf \
    --model_reload \
    --import_file_format hf_format \
    --import_model_dir [path to Llama-2-7b-chat-hf-awq-uint4-asym-g128-bf16-lmhead model] \
    --ppl \
    --trust_remote_code \
    --batch_size 1 \
    --device cuda

Other Arguments

1. Set --multi_gpu for multi-gpu support

2. Set --save_metrics_to_csv and metrics_output_dir to save PPL score to CSV

3. Set dtype by model_args dtype=float32 to change model dtype.