Modelpublic

ENOT-AutoDL/gpt-j-6B-tensorrt-int8

sourceHugging Faceapache-2.0updated 3y agoView on Hugging Face

7likes

Model Card

INT8 GPT-J 6B

GPT-J 6B is a transformer model trained using Ben Wang's Mesh Transformer JAX. "GPT-J" refers to the class of model, while "6B" represents the number of trainable parameters.

This repository contains GPT-J 6B onnx model suitable for building TensorRT int8+fp32 engines. Quantization of model was performed by the ENOT-AutoDL framework. Code for building of TensorRT engines and examples published on github.

Metrics:

	TensorRT INT8+FP32	torch FP16	torch FP32
Lambada Acc	78.46%	79.53%	-
Model size (GB)	8.5	12.1	24.2

Test environment

—GPU RTX 4090
—CPU 11th Gen Intel(R) Core(TM) i7-11700K
—TensorRT 8.5.3.1
—pytorch 1.13.1+cu116

Latency:

Input sequance length	Number of generated tokens	TensorRT INT8+FP32 ms	torch FP16 ms	Acceleration
64	64	1040	1610	1.55
64	128	2089	3224	1.54
64	256	4236	6479	1.53
128	64	1060	1619	1.53
128	128	2120	3241	1.53
128	256	4296	6510	1.52
256	64	1109	1640	1.49
256	128	2204	3276	1.49
256	256	4443	6571	1.49

Test environment

—GPU RTX 4090
—CPU 11th Gen Intel(R) Core(TM) i7-11700K
—TensorRT 8.5.3.1
—pytorch 1.13.1+cu116

How to use

Example of inference and accuracy test published on github:

shell

git clone https://github.com/ENOT-AutoDL/ENOT-transformers