This model has been pushed to the Hub using the PytorchModelHubMixin integration.


Qwen2.5-7B-Instruct INT8 (Optimum Quanto)

This repository contains an INT8 quantized version of Qwen2.5-7B-Instruct created using Optimum Quanto.

The objective of this project is to reduce model size and memory usage while maintaining inference quality for research and local deployment.

Model Details

  • Base Model: Qwen/Qwen2.5-7B-Instruct
  • Quantization Method: Optimum Quanto
  • Weight Precision: INT8
  • Original Model Size: ~15.2 GB
  • Quantized Model Size: ~8.18 GB
  • Storage Reduction: ~46%

Usage

from optimum.quanto import QuantizedModelForCausalLM
from transformers import AutoTokenizer

model = QuantizedModelForCausalLM.from_pretrained(
    "YOUR_USERNAME/Qwen2.5-7B-Instruct-INT8"
)

tokenizer = AutoTokenizer.from_pretrained(
    "YOUR_USERNAME/Qwen2.5-7B-Instruct-INT8"
)

Benchmark

Benchmarking is currently in progress.

Planned evaluations include:

  • HellaSwag
  • ARC Challenge
  • Winogrande
  • GSM8K

Benchmark results will be added after evaluation.

Hardware

Quantization performed on:

  • Windows
  • Python 3.11
  • PyTorch
  • NVIDIA GeForce RTX 3050 Laptop GPU (4 GB VRAM)

Limitations

  • Quantization may slightly reduce model accuracy depending on the evaluation task.
  • Performance should be validated for production workloads.
  • GPU memory requirements depend on sequence length and runtime configuration.

License

This repository contains a quantized derivative of Qwen/Qwen2.5-7B-Instruct.

Please follow the original model's license and usage terms when using or redistributing this model.


This model has been pushed to the Hugging Face Hub using:

Downloads last month
68
Safetensors
Model size
8B params
Tensor type
BF16
·
I8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Helllbos/Qwen2.5-7B-Instruct-INT8-Quanto

Base model

Qwen/Qwen2.5-7B
Finetuned
(2935)
this model