wschin / onnxruntime

ONNX Runtime: cross-platform, high performance scoring engine for ML models

ONNX Runtime is a cross-platform inference and training machine-learning accelerator.

ONNX Runtime inference can enable faster customer experiences and lower costs, supporting models from deep learning frameworks such as PyTorch and TensorFlow/Keras as well as classical machine learning libraries such as scikit-learn, LightGBM, XGBoost, etc. ONNX Runtime is compatible with different hardware, drivers, and operating systems, and provides optimal performance by leveraging hardware accelerators where applicable alongside graph optimizations and transforms. Learn more →

ONNX Runtime training can accelerate the model training time on multi-node NVIDIA GPUs for transformer models with a one-line addition for existing PyTorch training scripts. Learn more →

Get Started

General Information: onnxruntime.ai

Usage documention and tutorials: onnxruntime.ai/docs

Companion sample repositories:

• ONNX Runtime Inferencing: microsoft/onnxruntime-inference-examples
• ONNX Runtime Training: microsoft/onnxruntime-training-examples

Build Pipeline Status

System	CPU	GPU	EPs
Windows
Linux
Mac
Android
iOS
WebAssembly

Data/Telemetry

Windows distributions of this project may collect usage data and send it to Microsoft to help improve our products and services. See the privacy statement for more details.

Contributions and Feedback

We welcome contributions! Please see the contribution guidelines.

For feature requests or bug reports, please file a GitHub Issue.

For general discussion or questions, please use GitHub Discussions.

Code of Conduct

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

License

This project is licensed under the MIT License.

onnxruntime's Issues

Unify Scalar Representation in Pytorch JIT, LazyTensor, and ORT

Debug Mode (Assert of ORT and Pytorch Generating the Same Result per Subgraph)

Ideally, outputs generated by ORT should match whatever Pytorch (torch::jit::GraphExecutor) would produces. Under debug mode, we want to run torch::jit::GraphExecutor::run in parallel to InferenceSession::Run and compare their outcomes. This helps identifying error at earlier stage.

Unregister LazyTensor backend

As title. We call

import lazy_tensor_core as lt
import lazy_tensor_core.core.lazy_model as ltm
import torch
import onnxruntime as ort
from onnxruntime.training.ortmodule.torch_cpp_extensions import aten_op_executor
from onnxruntime.capi import _pybind_state as C

# Set up ORT as Pytorch backend
lt._LAZYC._ltc_init_ts_backend()
C.register_aten_op_executor(str(aten_op_executor.is_tensor_argument_address()),
                            str(aten_op_executor.execute_aten_operator_address()))
ost.register_ort_as_torch_jit_executor()

to register ORT as Pytorch backend (via LazyTensor). We should also have a function to reverse this action.

Return byte arrays as exporting result

Current exporting flow is

1. Pytorch creates ONNX graph in memory.
2. Pytorch saves ONNX graph to disk.
3. ORT loads ONNX graph from disk.

Ideally, step (2) is not necessary and we should remove it.

Print helper functions

We need helper to print types and shapes for at::IValue and OrtValue. It's helpful for debugging.

Add environment flag to turn on/off session caching

Storing too many ORT sessions can cause memory exhaustion. We probably want to have an environment variable, say, ORTLTMAXSESSCOUNT=int. If its value is 0 or unset, no caching.

Support `aten::add`

aten::add is not supported because

• ONNX exporter doesn't handle scalar correctly.
• After applying Bowen's patch, LazyTensor generates incorrect sub-graph. The last input to aten::add should be at::Scalar but got at::Tensor.

Add C++ interface to Pytorch Exporter

Pytorch doesn't have C/C++ interface for calling exporter. We made a porotype here and that function is called this way. We should add it back to Pytorch.

Improve Pytorch GraphFuser

GraphFuser in Pytorch doesn't support all aten ops. To increase ORT's coverage, we need to relax these constraints.

Only Add GPU EP when Input Tensors Are on GPU

Currently, we use GPU EP for both CPU and GPU tensors.

[Update] The problem is deeper than I thought. With ORT built with USE_CUDA=1, some CUDA-specific graph fusion may be applied and ORT may throw if GPU EP is not registered. We move some graph transformations to under their EPs and then come back to fix this problem.

Support `aten::nll_loss_forward`

Got an error

RuntimeError: outputs_.size() == 1INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/ir.h":500, please report a bug to PyTorch. 

when running LazyTensor-ORT with the following script.

from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import lazy_tensor_core
from onnxruntime.capi import _pybind_state as ost
import lazy_tensor_core.core.lazy_model as ltm
torch.manual_seed(0)

ost.register_ort_as_torch_jit_executor()
lazy_tensor_core._LAZYC._ltc_init_ts_backend()

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 32, 3, 1, bias=False)
        self.conv2 = nn.Conv2d(32, 64, 3, 1, bias=False)
        self.dropout1 = nn.Dropout(0.25)
        self.dropout2 = nn.Dropout(0.5)
        self.fc1 = nn.Linear(9216, 128, bias=False)
        self.fc2 = nn.Linear(128, 10, bias=False)

    def forward(self, x):
        x = self.conv1(x)
        x = F.relu(x)
        x = self.conv2(x)
        x = F.relu(x)
        x = F.max_pool2d(x, 2)
        x = self.dropout1(x)
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = F.relu(x)
        x = self.dropout2(x)
        x = self.fc2(x)
        output = F.log_softmax(x, dim=1)
        return output

device = 'lazy'
x = torch.rand((64, 1, 28, 28), dtype=torch.float32).to(device)
y = torch.randint(0, 9, (64,), dtype=torch.int64).to(device)
model = Net().to(device)
optimizer = optim.Adagrad(model.parameters(), lr=0.001)
for i in range(5):
  print(i)
  optimizer.zero_grad()
  print(f'tr::forward::{i}')
  output = model(x)
  ltm.mark_step()
  print('tr:loss')
  loss = F.nll_loss(output, y)
  print(loss)
  print(f'tr::backward::{i}')
  loss.backward() #2.1838
  ltm.mark_step()
  print(f'tr::optim::{i}')
  optimizer.step()
  ltm.mark_step()