Description
In the NLP tasks, there are lots of INT64 ID tensor as input, but currently tensorrt doesn’t support the INT64 datatype. Is there any workarounds or tips on it?
Thanks.
Environment
TensorRT Version : 8.0.1.6
GPU Type : T4
Nvidia Driver Version : 460.32.03
CUDA Version : 11.2
CUDNN Version : 8.0.5.39
Operating System + Version : Ubuntu 18.04.5 LTS
Hi,
TensorRT will attempt to cast down INT64 to INT32. For your reference,
opened 07:20PM - 10 May 18 UTC
closed 10:44PM - 20 Oct 20 UTC
question
triaged
I am trying to deploy onnx model using tensorrt, but cannot find tensorrt suppor… t for int64 data type.
nvinfer1::ITensor supprts the following types
kFLOAT FP32 format.
kHALF FP16 format.
kINT8 INT8 format.
kINT32 INT32 format.
In the onnx model exported from pytorch, I have several usages of int64
1. Shape operator generate int64 type tensor per definition
2. some constants are int64 type.
Any thoughts on how to address that? Should we modify pytorch export to avoid int64? Assume
we never use a number > 2G? Or should we add that into onnx-tensorrt ?
Thank you.