Fijit is a resarch prototype built in 2019 that consumes ONNX protobuf and execute the compute node while letting you configure:
- How operators are placed in different CUDA streams
- The exact implementation for the operator
- cuDNN
- cuBLAS
- TVM generated CUDA kernel
It can current runs a ResNet 50!
System Requirements:
- Install git-lfs
- CUDA 10 in the system
Steps:
git clone --recursive [email protected]:simon-mo/fijit-sys.git
cd fijit-sys
mkdir build
cd build
cmake ..
make -j8 fijit-sys
# testing
./fijit-sys --model ../data/sanity_check.onnx --input ../data/resnet50_input.onnx --max-block 20
Usage:
./fijit-sys -h
FIJIT Inference Engine
Usage:
fijit-sys [OPTION...]
-m, --model arg Path to the model ONNX file
-i, --input arg Path to the input ONNX file
--max-block arg Max block for TVM ops
--input-name arg Override input tensor name
-h, --help Print help
