README.md

BladeDISC Introduction

• Overview
  • Features and Roadmap
    • Frontend Framework Support Matrix
    • Backend Support Matrix
    • Deployment Solutions
  • Numbers of Typical Workloads
    • Advantage in Dynamic Shape Workloads
• API QuickView
  • For TensorFlow Users
  • For PyTorch Users
• Setup and Examples
• Publications
• Tutorials and Documents for Developers
• Presentations and Talks
• How to Contribute
• Building Status
• FAQ
  • Roadmap with mlir-hlo Project
• Contact Us

Overview

BladeDISC is an end-to-end DynamIc Shape Compiler project for machine learning workloads, which is one of the key components of Alibaba's PAI-Blade. BladeDISC provides general, transparent, and ease of use performance optimization for TensorFlow/PyTorch workloads on GPGPU and CPU backends. The architecture natively supports dynamic shape workloads, with many considerations in the performance of both static and dynamic shape scenarios. It also supports multiple and flexible deployment solutions, including both Plugin Mode inside TensorFlow/PyTorch runtime, and Standalone Mode for AOT standalone execution. The project is based on MLIR and highly related with mlir-hlo project.

Refer to our website for more information, including the setup tutorial, developer guide, demo examples and documents for developers.

Features and Roadmap

Frontend Framework Support Matrix

	TensorFlow [1]	PyTorch [2]
Inference	Yes	Yes
Training	Yes [3]	Ongoing

[1] TensorFlow 1.12, 1.15, 2.4 & 2.5 are supported and fully verified. For other versions some slight works on adaptation might be needed.

[2] 1.6.0 <= PyTorch version < 1.9.0 has been fully verified.

[3] Although supported, there's much room for improvement on Op coverage for training workloads.

Backend Support Matrix

	Status
Nvidia GPU	Yes
AMD GPU	Yes
Hygon DCU	Yes
X86	Yes
AArch64	Yes

Deployment Solutions

• Plugin Mode - BladeDISC works as a plugin of TensorFlow or PyTorch. Only the supported Ops are clustered and compiled, and the unsupported ones will be executed by the original TensorFlow or PyTorch runtime. We recommend this mode to most of the users for its transparency and ease of use.

• Standalone Mode - In Standalone mode, the input workload will be compiled into a binary that can be executed by it self, aka, does not rely on a TensorFlow or PyTorch runtime. In this mode all ops must be supported.

Numbers of Typical Workloads

By evaluating BladeDISC using a set of typical machine learning workloads for production purpose, BladeDISC shows up to 8.66x speedup compared with TensorFlow/PyTorch. Moreover, compared to static optimizing compilers (i.e., XLA and TensorRT), DISC shows comparable or even better performance.

Fig.1 Performance speedup over framework. Framework means either TensorFlow or PyTorch. FastSpeech2 is TensorFlow model and others are PyTorch models. The static compiler for TensorFlow is XLA and that for PyTorch is TensorRT. Note that S2T and T5 have no TensorRT performance due to wrong result.

Advantage in Dynamic Shape Workloads

Specifically, for the BERT large inference on T4 we provide in the examples, static compiler optimization (XLA) shows severe performance degradation due to its compilation overhead, while DISC shows a 1.75x speedup.

TensorFlow	XLA	DISC
1.78 s	41.69s	1.02s
1X		1.75X

API QuickView

For TensorFlow Users

Only two lines of code are needed on native Tensorflow program as the following:

import numpy as np
import tensorflow as tf

## enable BladeDISC on TensorFlow program
import blade_disc_tf as disc
disc.enable()

## construct TensorFlow Graph and run it
g = tf.Graph()
with g.as_default():
    ...
    with tf.session as sess:
        sess.run(...)

For more information, please refer to QuickStart for TensorFlow Users

For PyTorch Users

PyTorch users only need the following few lines of code to enable BladeDISC:

import torch_blade
# construct PyTorch Module
class MyModule(nn.Module):
    ...

module = MyModule()

with torch.no_grad():
    # blade_module is the optimized module by BladeDISC
    blade_module = torch_blade.optimize(module, allow_tracing=True, model_inputs=(x, y))

# run the optimized module
blade_module(x, y)

torch_blade.optimize accepts an nn.Module object and outputs the optimized module. For more information, please refer to Quickstart for PyTorch Users.

Setup and Examples

• How to Setup and Build from Source
• Use Case of TensorFlow Inference and Training
• Use Case of PyTorch Inference

Publications

• DISC: A Dynamic Shape Compiler for Machine Learning Workloads

Tutorials and Documents for Developers

• Tutorial: A Walkthough of the BladeDISC Pass Pipeline
• Introduction on Runtime Abstraction Layer
• TorchBlade Overview
• Tutorial: How to Add a New Torch Operator Converter

Presentations and Talks

• Performance optimization practice for dynamic shape AI workloads via a compiler-based approach
• 2022/07/31 BladeDISC: A Practice of Dynamic Shape Deep Learning Compiler(Chinese)
• 2022/07/07 BladeDISC and Torch-MLIR Roadmap Talk on Torch-MLIR Community
• GTC22-S41073, Generalized and Transparent AI Optimization Solutions with AI Compilers from Cloud Service
• GTC22-S41395, Easier-to-use and More Robust TensorRT via PAI-Blade

How to Contribute

• Contribute to BladeDISC

Building Status

Framework	Device	Status
PyTorch1.6.0	CPU
PyTorch1.8.1	CPU
PyTorch1.7.1	GPU
PyTorch1.12.0	GPU
PyTorch1.10.0	AArch64
TensorFlow1.15	CPU
TensorFlow2.4	GPU
TensorFlow2.8	AArch64

FAQ

Roadmap with mlir-hlo Project

BladeDISC is in a close relationship with mlir-hlo project. Part of the building blocks, including the MHLO Op definitions, TF to MHLO conversions, and some general purpose passes have been upstreamed to mlir-hlo repository. We'll continue to work in a close cooperative relationship with mlir-hlo project in the longer term.

Contact Us

• Mailgroup: bladedisc-dev@list.alibaba-inc.com

• DingTalk group for support and discussion: