omniback 0.2.24

A flexible and efficient inference framework

TorchPipe

TorchPipe is an alternative choice for Triton Inference Server, mainly featuring similar functionalities such as Shared-momory, Ensemble, and BLS mechanism.

For serving scenarios, TorchPipe is designed to support multi-instance deployment, pipeline parallelism, adaptive batching, GPU-accelerated operators, and reduced head-of-line (HOL) blocking.It acts as a bridge between lower-level acceleration libraries (e.g., TensorRT, OpenCV, CVCUDA) and RPC frameworks (e.g., Thrift). At its core, it is an engine that enables programmable scheduling.

News

• [2026-01-23] 📦 Available on PyPI: pip install torchpipe
• [2026-01-04] 🔧 We switched to tvm_ffi to provide clearer C++-Python interaction.

Usage

Below are some usage examples, for more check out the examples.

Initialize and Prepare Pipeline

from torchpipe import pipe
import torch

from torchvision.models.resnet import resnet101

# create some regular pytorch model...
model = resnet101(pretrained=True).eval().cuda()

# create example model
model_path = f"./resnet101.onnx"
x = torch.ones((1, 3, 224, 224)).cuda()
torch.onnx.export(model, x, model_path, opset_version=17,
                    input_names=['input'], output_names=['output'], 
                    dynamic_axes={'input': {0: 'batch_size'},
                                'output': {0: 'batch_size'}})

thread_safe_pipe = pipe({
    "preprocessor": {
        "backend": "S[DecodeTensor,ResizeTensor,CvtColorTensor,SyncTensor]",
        # "backend": "S[DecodeMat,ResizeMat,CvtColorMat,Mat2Tensor,SyncTensor]",
        'instance_num': 2,
        'color': 'rgb',
        'resize_h': '224',
        'resize_w': '224',
        'next': 'model',
    },
    "model": {
        "backend": "SyncTensor[TensorrtTensor]",
        "model": model_path,
        "model::cache": model_path.replace(".onnx", ".trt"),
        "max": '4',
        'batching_timeout': 4,  # ms, timeout for batching
        'instance_num': 2,
        'mean': "123.675, 116.28, 103.53",
        'std': "58.395, 57.120, 57.375",  # merged into trt
    }}
)

Execute

We can execute the returned thread_safe_pipe just like the original PyTorch model, but in a thread-safe manner.

data = {'data': open('/path/to/img.jpg', 'rb').read()}
thread_safe_pipe(data) # <-- this is thread-safe
result = data['result']

Installation

• NGC Docker containers (recommended):

test on 25.05, 25.06, 24.05, 23.05

img_name=nvcr.io/nvidia/pytorch:25.05-py3

docker run --rm --gpus all -it --network host \
    -v $(pwd):/workspace/ --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 \
    -w /workspace/ \
    $img_name \
    bash

pip install torchpipe
python -c "import torchpipe"

The backends it introduces will be JIT-compiled and cached.

There are one core backend group(torchpipe_core) and three optional groups (torchpipe_opencv, torchpipe_nvjpeg, and torchpipe_tensorrt) with different dependencies. For details, see here.

Dependencies such as OpenCV and TensorRT can also be provided in the following ways:

• providing environment variables:
  Users can specify paths via the following environment variables:
  OPENCV_INCLUDE, OPENCV_LIB, TENSORRT_INCLUDE, TENSORRT_LIB.

Other installation options

How does it work?

See Basic Usage.

How to add (or override) a backend

WIP