Shark Runner

The Shark Runner provides inference and training APIs to run deep learning models on Shark Runtime.

How to configure.

Check out the code

git clone https://github.com/NodLabs/dSHARK.git 

Setup your Python VirtualEnvironment and Dependencies

# Setup venv and install necessary packages (torch-mlir, nodLabs/Shark, ...).
./setup_venv.sh
# Please activate the venv after installation.

Run a demo script

python -m  shark.examples.resnet50_script --device="cpu/gpu/vulkan"

Shark Inference API

from shark_runner import SharkInference

shark_module = SharkInference(
        module = torch.nn.module class.
        (input,)  = inputs to model (must be a torch-tensor)
        dynamic (boolean) = Pass the input shapes as static or dynamic.
        device = `cpu`, `gpu` or `vulkan` is supported.
        tracing_required = (boolean) = Jit trace the module with the given input, useful in the case where jit.script doesn't work. )

result = shark_module.forward(inputs)

Shark Trainer API

Work in Progress

Model Tracking (Shark Inference)

Hugging Face Models	Torch-MLIR lowerable	SHARK-CPU	SHARK-CUDA	SHARK-METAL
BERT	✔️ (JIT)	✔️
Albert	✔️ (JIT)	✔️
BigBird	✔️ (AOT)
DistilBERT	✔️ (AOT)
GPT2	❌ (AOT)

TORCHVISION Models	Torch-MLIR lowerable	SHARK-CPU	SHARK-CUDA	SHARK-METAL
AlexNet	✔️ (Script)
DenseNet121	✔️ (Script)
MNasNet1_0	✔️ (Script)
MobileNetV2	✔️ (Script)
MobileNetV3	✔️ (Script)
Unet	❌ (Script)
Resnet18	✔️ (Script)	✔️
Resnet50	✔️ (Script)	✔️
Resnext50_32x4d	✔️ (Script)
ShuffleNet_v2	❌ (Script)
SqueezeNet	❌ (Script)
EfficientNet	✔️ (Script)
Regnet	✔️ (Script)
Resnest	❌ (Script)
Vision Transformer	✔️ (Script)
VGG 16	✔️ (Script)
Wide Resnet	✔️ (Script)	✔️
RAFT	❌ (JIT)

For more information refer to MODEL TRACKING SHEET