README.md

metadata

license: apache-2.0
base_model: mistralai/Mistral-7B-v0.1
language:
  - en
tags:
  - mistral
  - onnxruntime
  - onnx
  - llm

Mistral-7b for ONNX Runtime

Introduction

This repository hosts the optimized versions of Mistral-7B-v0.1 to accelerate inference with ONNX Runtime CUDA execution provider.

See the usage instructions for how to inference this model with the ONNX files hosted in this repository.

Model Description

• Developed by: MistralAI
• Model type: Pretrained generative text model
• License: Apache 2.0 License
• Model Description: This is a conversion of the Mistral-7B-v0.1 for ONNX Runtime inference with CUDA execution provider.

Performance Comparison

Latency for token generation

Below is average latency of generating a token using a prompt of varying size using NVIDIA A100-SXM4-80GB GPU:

Prompt Length	Batch Size	PyTorch 2.1 torch.compile	ONNX Runtime CUDA
32	1	32.58ms	12.08ms
256	1	54.54ms	23.20ms
1024	1	100.6ms	77.49ms
2048	1	236.8ms	144.99ms
32	4	63.71ms	15.32ms
256	4	86.74ms	75.94ms
1024	4	380.2ms	273.9ms
2048	4	N/A	554.5ms

Usage Example

Following the benchmarking instructions. Example steps:

1. Clone onnxruntime repository.

git clone https://github.com/microsoft/onnxruntime
cd onnxruntime

2. Install required dependencies

python3 -m pip install -r onnxruntime/python/tools/transformers/models/llama/requirements-cuda.txt

5. Inference using manual model API, or use Hugging Face's ORTModelForCausalLM

from optimum.onnxruntime import ORTModelForCausalLM
from onnxruntime import InferenceSession
from transformers import AutoConfig, AutoTokenizer

sess = InferenceSession("Mistral-7B-v0.1.onnx", providers = ["CUDAExecutionProvider"])
config = AutoConfig.from_pretrained("mistralai/Mistral-7B-v0.1")

model = ORTModelForCausalLM(sess, config, use_cache = True, use_io_binding = True)

tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")

inputs = tokenizer("Instruct: What is a fermi paradox?\nOutput:", return_tensors="pt")

outputs = model.generate(**inputs)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))