ArslanRobo/llama-3.1-8b-instruct-smoothquant-Pruned30Tailor-fp16

Model Card for LLaMA-3.1-8B-Instruct (SmoothQuant FP16)

This repository contains a SmoothQuant-processed FP16 version of Meta LLaMA-3.1-8B-Instruct.
The model is not quantized. SmoothQuant is applied only as an outlier-suppression preprocessing step to improve downstream low-bit quantization.

---

Model Details

Model Description

This model applies SmoothQuant (activation-aware weight smoothing) to the FP16 weights of LLaMA-3.1-8B-Instruct.
No architectural changes or precision reduction are performed.

The purpose of this model is to serve as an intermediate checkpoint in a deployment pipeline targeting CPU and edge devices using llama.cpp and mixed-precision quantization.

• Developed by: Muhammad Arslan Rafiq
• Funded by: Academic coursework (AI on Edge Devices)
• Shared by: ArslanRobo
• Model type: Decoder-only Transformer (Causal Language Model)
• Language(s) (NLP): English
• License: LLaMA 3.1 Community License
• Finetuned from model: meta-llama/Llama-3.1-8B-Instruct

Model Sources

• Repository: https://huggingface.co/ArslanRobo
• Paper: SmoothQuant – https://arxiv.org/abs/2211.10438
• Base Model: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct

---

Uses

Direct Use

• Research on outlier suppression techniques
• Preprocessing before post-training quantization
• Analysis of quantization robustness
• Academic experiments on edge deployment

Downstream Use

• Conversion to GGUF format
• Mixed-precision or low-bit quantization using llama.cpp
• CPU-only and embedded device inference (e.g., Raspberry Pi)

Out-of-Scope Use

• Production deployment without quantization
• Safety-critical or regulated applications
• Fine-tuning without further evaluation

---

Bias, Risks, and Limitations

• Inherits all biases and limitations of the original LLaMA-3.1 model
• SmoothQuant does not reduce model size or latency by itself
• Quality improvements appear only after quantization
• No safety fine-tuning applied beyond the base model

Recommendations

Users should:

• Apply quantization after SmoothQuant
• Evaluate perplexity and downstream task performance
• Use appropriate safety filters for real-world deployment

---

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ArslanRobo/llama-3.1-8b-instruct-smoothquant-fp16",
    torch_dtype=torch.float16,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained(
    "ArslanRobo/llama-3.1-8b-instruct-smoothquant-fp16"
)
Training Details
Training Data
No additional training data was used.
The model weights are derived directly from the base LLaMA-3.1-8B-Instruct model.

Training Procedure
Preprocessing
Activation statistics collected on WikiText-2

SmoothQuant applied with α = 0.5

FP16 precision preserved

Training Hyperparameters
Training regime: FP16 (no training, preprocessing only)

Speeds, Sizes, Times
Model size: ~15 GB (FP16)

SmoothQuant calibration: ~5–10 minutes on NVIDIA T4

Evaluation
Testing Data, Factors & Metrics
Testing Data
WikiText-2 (used for calibration and analysis)

Factors
Activation outliers

Weight distribution

Quantization sensitivity

Metrics
Perplexity (evaluated post-quantization)

Quantization MSE and SNR (simulated)

Results
SmoothQuant reduces quantization error

Improves robustness for 4-bit and mixed-precision quantization

Minimal impact on FP16 perplexity

Model Examination
Weight distribution smoothing

Reduced activation outliers

Improved numerical stability for low-bit quantization

Environmental Impact
Hardware Type: NVIDIA T4 GPU

Hours used: ~1 hour

Cloud Provider: Google Colab / Kaggle

Compute Region: Not specified

Carbon Emitted: Not measured

Technical Specifications
Model Architecture and Objective
Decoder-only Transformer

Causal language modeling objective

Compute Infrastructure
Hardware
NVIDIA T4 GPU (16 GB VRAM)

Software
PyTorch

Hugging Face Transformers

Accelerate

Datasets

Citation
BibTeX
bibtex
Copy code
@article{xiao2023smoothquant,
  title={SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models},
  author={Xiao, Guangxuan and others},
  journal={arXiv preprint arXiv:2211.10438},
  year={2023}
}
APA
Xiao, G., et al. (2023). SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models. arXiv:2211.10438.

Model Card Authors
Muhammad Arslan Rafiq