Skip to content

joyalbin/tvmdbg

 
 

Repository files navigation

NNVM: Open Compiler for AI Frameworks

Build Status GitHub license

Installation | Documentation | Tutorials | Release Notes

NNVM compiler offers reusable computation graph optimization and compilation for deep learning systems. It is backed by the TVM stack and provides modules to:

  • Represent deep learning workloads from front-end frameworks via a graph IR.
  • Optimize computation graphs to improve performance.
  • Compile into executable modules and deploy to different hardware backends with minimum dependency.

NNVM is designed to add new frontend, operators and graph optimizations in a decentralized fashion without changing the core interface. The compiled module can be deployed to server, mobile, embedded devices and browsers with minimum dependency, in languages including c++, python, javascript, java, objective-c. Checkout our release announcement

The following code snippet demonstrates the general workflow of nnvm compiler.

import tvm
from tvm.contrib import graph_runtime, rpc
import nnvm.frontend
import nnvm.compiler

# GET model from frameworks
# change xyz to supported framework name.
graph, params = nnvm.frontend.from_xyz(...)

# OPTIMIZE and COMPILE the graph to get a deployable module
# target can be "opencl", "llvm", "metal" or any target supported by tvm
target = "cuda"
graph, lib, params = nnvm.compiler.build(graph, target, {"data", data_shape}, params=params)

# DEPLOY and run on gpu(0)
module = graph_runtime.create(graph, lib, tvm.gpu(0))
module.set_input(**params)
module.run(data=data_array)
output = tvm.nd.empty(out_shape, ctx=tvm.gpu(0))
module.get_output(0, output)

# DEPLOY to REMOTE mobile/rasp/browser with minimum tvm rpc runtime
# useful for quick experiments on mobile devices
remote = rpc.connect(remote_host, remote_port)
lib.export_library("mylib.so")
remote.upload("mylib.so")
rlib = rpc.load_module("mylib.so")
# run on remote device
rmodule = graph_runtime.create(graph, rlib, remote.gpu(0))
rmodule.set_input(**params)
rmodule.run()

License

Licensed under an Apache-2.0 license.

Links

  • TinyFlow on how you can use NNVM to build a TensorFlow like API.
  • Apache MXNet uses NNVM as a backend.

About

Debug module for DMLC TVM project

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 61.0%
  • Python 31.1%
  • Java 3.0%
  • CMake 1.1%
  • Makefile 1.0%
  • JavaScript 0.6%
  • Other 2.2%