Shears Adapter Card: shears-llama-7b-50-cs-heuristic-adapter

The heuristic adapter discovered from the super-adapter fine-tuned on sparsified LLaMA-7B with some commonsense reasoning datasets using Shears.

Paper Abstract

Recently, several approaches successfully demonstrated that weight-sharing Neural Architecture Search (NAS) can effectively explore a search space of elastic low-rank adapters (LoRA), allowing the parameter-efficient fine-tuning (PEFT) and compression of large language models. In this paper, we introduce a novel approach called Shears, demonstrating how the integration of cost-effective sparsity and a proposed Neural Low-rank adapter Search (NLS) algorithm can further improve the efficiency of PEFT approaches. Results demonstrate the benefits of Shears compared to other methods, reaching high sparsity levels while improving or with little drop in accuracy, utilizing a single GPU for a pair of hours.

Model Details

Note

Please note, we only provide the model adapter and do not provide a copy of the base yahma/llama-7b-hf model or its sparsified one. Any use of this adapter requires a separate download of the base model and follow this instruction to sparse the base model.

Information

• Adapter name: shears-llama-7b-50-cs-heuristic-adapter
• Base model: Sparsified LLaMA-7B
• Sparsity: 50%
• Domain: Commonsense
• Subnetwork version: Heuristic
• NNCF Configuration: nncf_shears_llama.json

Sparsified Base Model

Shears employs a simple but effective pruning approach Wanda to sparsify the language model, serving as the base model. Clone the Wanda repo:

git clone https://github.com/locuslab/wanda.git && cd wanda && git checkout 8e8fc87 && cd ..

The command for unstructured sparsifying LLaMA-7B with Wanda, to achieve unstructured 50% sparsity:

python wanda/main.py \
    --model yahma/llama-7b-hf \
    --prune_method wanda \
    --sparsity_ratio 0.5 \
    --sparsity_type unstructured \
    --save wanda_out \
    --save_model shears-llama-7b-50-base

• --model: The identifier for the model on the Hugging Face model hub or local path.
• --sparsity_ratio: Specifies the percentage of weights to be pruned.
• --save_model: Specifies the directory where the pruned language model will be stored.

Refer to our repo for the environment information to run this command.

Adapter Configuration

• LoRA rank: 32
• LoRA alpha: 64
• LoRA target modules: q_proj, k_proj, v_proj, up_proj, down_proj
• LoRA rank search space: [32, 24, 16] (for each LoRA module)

Training Hyperparameters

• Batch size: 16
• Learning rate: 3e-4
• Epoch: 5

Training Data

Unified commonsense reasoning dataset: commonsense_170k.json.

Evaluation Data

BoolQ, PIQA, SIQA, HellaSwag, WinoGrande, ARC-e, ARC-c, OBQA.

How to use

Use our modified PEFT library (apply patch):

git clone https://github.com/huggingface/peft.git
cd peft && git checkout v0.5.0 && git apply --ignore-space-change --ignore-whitespace peft-modifications-for-shears-inference-usage.patch && pip install -e . && cd ..

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM
from transformers import AutoTokenizer

def generate_prompt(instruction):
    return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request. 

                    ### Instruction:
                    {instruction}

                    ### Response:
                    """

base_model = AutoModelForCausalLM.from_pretrained("shears-llama-7b-50-base")
model = PeftModel.from_pretrained(base_model, "IntelLabs/shears-llama-7b-50-cs-heuristic-adapter")
model.eval()

non_zero_params = sum([(param.data != 0).sum().item() for _, param in model.named_parameters()])
print(f"Number of all non-zero parameters: {non_zero_params}")

tokenizer = AutoTokenizer.from_pretrained("shears-llama-7b-50-base")

instruction = "Please choose the correct answer to the question: A cactus stem is used to store\n\nAnswer1: fruit "
        "Answer2: liquid Answer3: food Answer4: spines\n\nAnswer format: answer1/answer2/answer3/answer4"
prompt = generate_prompt(instruction)
inputs = tokenizer(prompt, return_tensors="pt")
input_ids = inputs["input_ids"].to(model.device)
with torch.no_grad():
    generation_output = model.generate(
        input_ids=input_ids,
        return_dict_in_generate=True,
        output_scores=True,
        max_new_tokens=256,
        use_cache=True,
        num_beams=4,
    )
    s = generation_output.sequences[0]
    output = tokenizer.decode(s)
print(output)

Evaluation Results

Model	Sparsity	BoolQ	PIQA	SIQA	HellaSwag	WinoG	ARC-e	ARC-c	OBQA	Average
ChatGPT	-	73.1	85.4	68.5	78.5	66.1	89.8	79.9	74.8	77.0
LLaMA-7B-LoRA	-	68.9	80.7	77.4	78.1	78.8	77.8	61.3	74.8	74.7
LLaMA-7B-Shears	50%	67.3	79.1	77.5	73.3	77.7	74.4	57.9	72.8	72.5

Model Sources

• Repository: https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning/tree/main/Shears
• Paper: Shears: Unstructured Sparsity with Neural Low-rank Adapter Search

Ethical Considerations

Intel is committed to respecting human rights and avoiding causing or contributing to adverse impacts on human rights. See Intel’s Global Human Rights Principles. Intel’s products and software are intended only to be used in applications that do not cause or contribute to adverse impacts on human rights.

Ethical Considerations	Description
Data	The adapter was trained using the commonsense_170k.json data mixture as described above.
Human life	The model is not intended to inform decisions central to human life or flourishing.
Mitigations	No additional risk mitigation strategies were considered during model development.
Risks and harms	This model has not been assessed for harm or biases, and should not be used for sensitive applications where it may cause harm.
Use cases	-

Citation

@inproceedings{munoz2024shears,
  title = {Shears: Unstructured Sparsity with Neural Low-rank Adapter Search},
  author={J. Pablo Munoz and Jinjie Yuan and Nilesh Jain},
  booktitle={The 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-2024)},
  year={2024}
}

License

Apache-2.0