README.md

metadata

license: apache-2.0
base_model: tiiuae/falcon-7b
language:
  - en
tags:
  - falcon-7b
  - falcon
  - onnxruntime
  - onnx
  - llm

falcon-7b for ONNX Runtime

Introduction

This repository hosts the optimized version of falcon-7b 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: TIIUAE
• Model type: Pretrained generative text model
• License: Apache 2.0 License
• Model Description: This is a conversion of the falcon-7b 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	53.64ms	15.68ms
256	1	59.55ms	26.05ms
1024	1	89.82ms	99.05ms
2048	1	208.0ms	227.0ms
32	4	70.8ms	19.62ms
256	4	78.6ms	81.29ms
1024	4	373.7ms	369.6ms
2048	4	N/A	879.2ms

Usage Example

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 custom model API, or use Hugging Face's ORTModelForCausalLM

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

sess = InferenceSession("falcon-7b.onnx", providers = ["CUDAExecutionProvider"])
config = AutoConfig.from_pretrained("tiiuae/falcon-7b")

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

tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-7b")

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))