MNN VS oneflow

I found a solution on GitHub check your app's build.gradle, defaultConfig section - you need to add x86 to your ndk abiFilters ndk.abiFilters 'armeabi-v7a','arm64-v8a', 'x86' GitHub Hope it will help. You have to find that file and edit it as given here

This article focuses on what is happening behind the execution of a Tensor in the deep learning framework OneFlow. It takes the operator oneflow.relu as an example to introduce the Interpreter and VM mechanisms that need to be relied on to execute this operator.

This article describes how OneFlow works with MLIR, how to add a graph-level Pass to OneFlow IR, how OneFlow Operations automatically become MLIR Operations, and why OneFlow IR can use MLIR to accelerate computations.

Backpressure mechanism, also known as credit-based flow control, is a classic scheme for network communication flow control problems. Its predecessor is the TCP sliding window. This idea is particularly simple and effective. As we will see in this article, based on the same principles, this idea is applicable to any flow control scheme and is found in the design of many hardware and software systems. In this article, the engineer of OneFlow will tell the chequered history of this simple idea.

Elementwise operation refers to applying a function transformation to every element of a tensor. In deep learning, many operators can be regraded as elementwise operators, such as common activation functions (like ReLU and GELU) and ScalarMultiply (multiplying each element of a tensor by a scalar). For this elementwise operation, OneFlow(https://github.com/Oneflow-Inc/oneflow/) abstracts a CUDA template. this article will introduce the design thoughts and optimization techniques of CUDA template.

You can download Oneflow on GitHub, and read the technical documents or blog on Medium to know more about OneFlow. If you have any problem with OneFlow, please write issues on github. (Sorry for the late reply)

Code:https://github.com/Oneflow-Inc/oneflow/

This series article introduced what a point-to-point communication library is, and discussed some of the general characteristics of the optimal P2P communication library. Furthermore, it dive into the details about how to design an optimal P2P library and introduce the design of CommNet in OneFlow.

This article lists papers on GShard, presents background information and inspiration from the papers, and finally evaluates what else can be done to improve GShard from similar work that has been done in OneFlow.

OneFlow Paper:https://arxiv.org/abs/2110.15032; Code:https://github.com/Oneflow-Inc/oneflow/

The paper of Gshard contains two main parts of work, one on parallel APIs and one on Mixture of experts. The former part is more interesting and I will only discuss this part. The contribution on parallel APIs is outlined clearly in the abstract of the paper:

GShard is a module composed of a set of lightweight annotation APIs and an extension to the XLA compiler.

Gshard Paper: https://arxiv.org/pdf/2006.16668.pdf

All ops computed in deep learning frameworks are translated into CUDA kernel functions on the GPU, and Softmax operations are no exception. Softmax is a widely used op in most networks, and the efficiency of its CUDA kernel implementation can affect the final training speed of many networks. So how can an efficient Softmax CUDA Kernel be implemented?

Code:https://github.com/Oneflow-Inc/oneflow

Deep learning frameworks such as TensorFlow and PyTorch provide a productive interface for expressing and training a DNN model on a single device or using data parallelism. Still, they may not be flexible or efficient enough in training emerging large models on distributed devices, which require more sophisticated parallelism beyond data parallelism. Plugins or wrappers have been developed to strengthen these frameworks for model or pipeline parallelism, but they complicate the usage and implementation of distributed deep learning. Paper:https://arxiv.org/pdf/2110.15032.pdf; Code: https://github.com/Oneflow-Inc/oneflow