Compressed Meta Llama-3-8B-Instruct with Palu
Overview
This repository contains a compressed version of the Meta Llama-3-8B-Instruct model, utilizing the Palu framework for KV-Cache compression. Palu reduces the hidden dimensions of the KV-Cache through low-rank decomposition, significantly reducing the model's memory footprint while maintaining or enhancing performance.
Meta Llama-3-8B-Instruct: Palu Compression Results
Perplexity (PPL)
| Model | PPL |
|---|
| meta-llama-3-8b-instruct-palu | 8.8309 |
| meta-llama-3-8b-instruct (Base) | 8.2845 |
Zero-shot Evaluation
meta-llama-3-8b-instruct-palu
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr |
|---|
| winogrande | 1 | none | 0 | acc | 0.7277 | ±0.0125 |
| arc_challenge | 1 | none | 0 | acc | 0.4949 | ±0.0146 |
| | | 0 | acc_norm | 0.5427 | ±0.0146 |
| arc_easy | 1 | none | 0 | acc | 0.7942 | ±0.0083 |
| | | 0 | acc_norm | 0.7551 | ±0.0088 |
| piqa | 1 | none | 0 | acc | 0.7655 | ±0.0099 |
| | | 0 | acc_norm | 0.7644 | ±0.0099 |
| hellaswag | 1 | none | 0 | acc | 0.5664 | ±0.0049 |
| | | 0 | acc_norm | 0.7511 | ±0.0043 |
| openbookqa | 1 | none | 0 | acc | 0.3360 | ±0.0211 |
| | | 0 | acc_norm | 0.4380 | ±0.0222 |
meta-llama-3-8b-instruct (Base)
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr |
|---|
| winogrande | 1 | none | 0 | acc | 0.7206 | ±0.0126 |
| arc_challenge | 1 | none | 0 | acc | 0.5299 | ±0.0146 |
| | | 0 | acc_norm | 0.5683 | ±0.0145 |
| arc_easy | 1 | none | 0 | acc | 0.8161 | ±0.0079 |
| | | 0 | acc_norm | 0.7976 | ±0.0082 |
| piqa | 1 | none | 0 | acc | 0.7867 | ±0.0096 |
| | | 0 | acc_norm | 0.7856 | ±0.0096 |
| hellaswag | 1 | none | 0 | acc | 0.5769 | ±0.0049 |
| | | 0 | acc_norm | 0.7581 | ±0.0043 |
| openbookqa | 1 | none | 0 | acc | 0.3420 | ±0.0212 |
| | | 0 | acc_norm | 0.4320 | ±0.0222 |
Long-Bench Evaluation
triviaqa
| Model | Score |
|---|
| meta-llama-3-8b-instruct-palu | 89.45 |
| meta-llama-3-8b-instruct (Base) | 90.56 |
qasper
| Model | Score |
|---|
| meta-llama-3-8b-instruct-palu | 34.92 |
| meta-llama-3-8b-instruct (Base) | 31.74 |
Key Features
- Model: Meta Llama-3-8B-Instruct
- Compression Framework: Palu
- Compression Rate: Up to 91.25% memory reduction
- Accuracy: Maintained or improved perplexity compared to the base model
Installation
Clone the Repository
Ensure you have Git and Conda installed on your system.
1git clone --recurse-submodules https://github.com/shadowpa0327/Palu.git
2cd Palu
Set Up the Environment
Create and activate a Conda environment.
1conda create -n Palu python=3.10
2conda activate Palu
3pip install -r requirements.txt
Install Third-Party Libraries
1pip install -e 3rdparty/lm-evaluation-harness
2pip install -e 3rdparty/fast-hadamard-transform
Usage
Compress the Model
To compress Meta Llama-3-8B-Instruct using Palu's low-rank decomposition, use the following command:
1python compress.py \
2--model_id="meta-llama/Llama-3-8b-instruct" \
3--calib_dataset wikitext2 \
4--param_ratio_target 0.7 \
5--search_method fisher_uniform \
6--head_group_size 4 \
7--dump_huggingface_model \
8--use_cache
The compressed model will be saved in the Meta-Llama-3-8b-instruct_ratio-0.7_gs-4-fisher_uniform directory in Hugging Face format.
Evaluate the Compressed Model
Perplexity
To evaluate the perplexity on the wikitext2 dataset with sequence length 2048, run:
1python run_ppl_eval.py \
2--model_name_or_path /Path/To/Palu/Model \
3--datasets wikitext2 \
4--seqlen 2048
To evaluate with 3-bit low-rank aware quantization, use:
1python run_ppl_eval.py \
2--model_name_or_path /Path/To/Palu/Model \
3--datasets wikitext2 \
4--seqlen 4096 \
5--lt_bits 3 \
6--lt_hadamard
Zero-shot Evaluation
For zero-shot evaluations, use the following command:
1CUDA_VISIBLE_DEVICES=0 python run_lm_eval.py \
2--model_name_or_path "/Path/To/Palu/Model" \
3--tasks "openbookqa,hellaswag,piqa,arc_easy,arc_challenge,winogrande"
Long-Bench Evaluation
Evaluate the compressed model on long-bench tasks:
1CUDA_VISIBLE_DEVICES=0 python run_long_bench.py \
2--model_name_or_path /Path/To/Palu/Model
Latency Evaluation
Attention Module
Evaluate the latency of the Palu-compressed attention module:
1CUDA_VISIBLE_DEVICES=0 python run_latency_attention.py \
2--rank_k 1024 --rank_v 3072 --group_size 4 \
3--prompt_len 65536 --palu
Reconstruction Kernel
Evaluate the latency of the reconstruction kernel:
1CUDA_VISIBLE_DEVICES=0 python run_latency_kernel.py \
2--total_rank 1024 --group_size 4
Conclusion
This compressed version of Meta Llama-3-8B-Instruct, powered by Palu, is optimized for memory efficiency without compromising performance. Whether you're working with large datasets or deploying models in memory-constrained environments, this setup is designed to provide robust results.
