llama-3-8b-pruned-30-percent

Llama-3.1 8B Instruct (Pruned 30%)

This model was structurally pruned using the L2-Norm method from LLM-Pruner.

Model Details

• Base Model: Meta Llama 3.1 8B Instruct
• Pruning Ratio: 30% (MLP layers only)
• Attention Layers: Intact (Protected to preserve GQA)
• Pruning Method: L2 Norm
• Parameters: ~6.76B (reduced from 8.03B)
• Model Size: ~13.5 GB (FP16)

Architecture Changes

• Hidden Size: 4096 (unchanged)
• Intermediate Size: ~10035 (reduced from 14336)
• Attention Heads: 32 (unchanged)
• Key-Value Heads: 8 (unchanged, GQA preserved)

How to Load

Standard Transformers (Recommended)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "naveedashfaq/llama-3-8b-pruned-30-percent",
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("naveedashfaq/llama-3-8b-pruned-30-percent")

# Generate text
inputs = tokenizer("Hello, how are you?", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

llama.cpp / GGUF Conversion

This model is compatible with llama.cpp. Convert to GGUF format:

# Clone llama.cpp
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make

# Download and convert the model
huggingface-cli download naveedashfaq/llama-3-8b-pruned-30-percent --local-dir ./llama-3-8b-pruned
python convert.py ./llama-3-8b-pruned/

# Optional: Quantize to reduce size further
./quantize ./llama-3-8b-pruned/ggml-model-f16.gguf ./llama-3-8b-pruned/ggml-model-q4_0.gguf q4_0

Performance

• Size Reduction: 30% fewer parameters
• Memory Savings: ~3 GB less memory usage
• Speed: Slightly faster inference due to reduced computations
• Quality: Minimal quality degradation on most tasks

Limitations

• This is a structurally pruned model with modified dimensions
• Performance may vary depending on the task
• Further evaluation needed for domain-specific applications

Citation

If you use this model, please cite the original LLM-Pruner work:

@article{ma2023llmpruner,
  title={LLM-Pruner: On the Structural Pruning of Large Language Models},
  author={Ma, Xinyin and Fang, Gongfan and Wang, Xinchao},
  journal={arXiv preprint arXiv:2305.11627},
  year={2023}
}

License

This model inherits the license from the base Llama 3.1 model. Please refer to Meta's Llama 3.1 license for usage terms.