Quick Start Tutorial for Compiling Deep Learning Models

Author: Yao Wang, Truman Tian

This example shows how to build a neural network with Relay python frontend and generates a runtime library for Nvidia GPU with TVM. Notice that you need to build TVM with cuda and llvm enabled.

Overview for Supported Hardware Backend of TVM

The image below shows hardware backend currently supported by TVM:

https://github.com/dmlc/web-data/raw/main/tvm/tutorial/tvm_support_list.png

In this tutorial, we’ll choose cuda and llvm as target backends. To begin with, let’s import Relay and TVM.

import numpy as np

from tvm import relay
from tvm.relay import testing
import tvm
from tvm import te
from tvm.contrib import graph_runtime
import tvm.testing

Define Neural Network in Relay

First, let’s define a neural network with relay python frontend. For simplicity, we’ll use pre-defined resnet-18 network in Relay. Parameters are initialized with Xavier initializer. Relay also supports other model formats such as MXNet, CoreML, ONNX and Tensorflow.

In this tutorial, we assume we will do inference on our device and the batch size is set to be 1. Input images are RGB color images of size 224 * 224. We can call the tvm.relay.TupleWrapper.astext() to show the network structure.

batch_size = 1
num_class = 1000
image_shape = (3, 224, 224)
data_shape = (batch_size,) + image_shape
out_shape = (batch_size, num_class)

mod, params = relay.testing.resnet.get_workload(
    num_layers=18, batch_size=batch_size, image_shape=image_shape
)

# set show_meta_data=True if you want to show meta data
print(mod.astext(show_meta_data=False))

Out:

#[version = "0.0.5"]
def @main(%data: Tensor[(1, 3, 224, 224), float32], %bn_data_gamma: Tensor[(3), float32], %bn_data_beta: Tensor[(3), float32], %bn_data_moving_mean: Tensor[(3), float32], %bn_data_moving_var: Tensor[(3), float32], %conv0_weight: Tensor[(64, 3, 7, 7), float32], %bn0_gamma: Tensor[(64), float32], %bn0_beta: Tensor[(64), float32], %bn0_moving_mean: Tensor[(64), float32], %bn0_moving_var: Tensor[(64), float32], %stage1_unit1_bn1_gamma: Tensor[(64), float32], %stage1_unit1_bn1_beta: Tensor[(64), float32], %stage1_unit1_bn1_moving_mean: Tensor[(64), float32], %stage1_unit1_bn1_moving_var: Tensor[(64), float32], %stage1_unit1_conv1_weight: Tensor[(64, 64, 3, 3), float32], %stage1_unit1_bn2_gamma: Tensor[(64), float32], %stage1_unit1_bn2_beta: Tensor[(64), float32], %stage1_unit1_bn2_moving_mean: Tensor[(64), float32], %stage1_unit1_bn2_moving_var: Tensor[(64), float32], %stage1_unit1_conv2_weight: Tensor[(64, 64, 3, 3), float32], %stage1_unit1_sc_weight: Tensor[(64, 64, 1, 1), float32], %stage1_unit2_bn1_gamma: Tensor[(64), float32], %stage1_unit2_bn1_beta: Tensor[(64), float32], %stage1_unit2_bn1_moving_mean: Tensor[(64), float32], %stage1_unit2_bn1_moving_var: Tensor[(64), float32], %stage1_unit2_conv1_weight: Tensor[(64, 64, 3, 3), float32], %stage1_unit2_bn2_gamma: Tensor[(64), float32], %stage1_unit2_bn2_beta: Tensor[(64), float32], %stage1_unit2_bn2_moving_mean: Tensor[(64), float32], %stage1_unit2_bn2_moving_var: Tensor[(64), float32], %stage1_unit2_conv2_weight: Tensor[(64, 64, 3, 3), float32], %stage2_unit1_bn1_gamma: Tensor[(64), float32], %stage2_unit1_bn1_beta: Tensor[(64), float32], %stage2_unit1_bn1_moving_mean: Tensor[(64), float32], %stage2_unit1_bn1_moving_var: Tensor[(64), float32], %stage2_unit1_conv1_weight: Tensor[(128, 64, 3, 3), float32], %stage2_unit1_bn2_gamma: Tensor[(128), float32], %stage2_unit1_bn2_beta: Tensor[(128), float32], %stage2_unit1_bn2_moving_mean: Tensor[(128), float32], %stage2_unit1_bn2_moving_var: Tensor[(128), float32], %stage2_unit1_conv2_weight: Tensor[(128, 128, 3, 3), float32], %stage2_unit1_sc_weight: Tensor[(128, 64, 1, 1), float32], %stage2_unit2_bn1_gamma: Tensor[(128), float32], %stage2_unit2_bn1_beta: Tensor[(128), float32], %stage2_unit2_bn1_moving_mean: Tensor[(128), float32], %stage2_unit2_bn1_moving_var: Tensor[(128), float32], %stage2_unit2_conv1_weight: Tensor[(128, 128, 3, 3), float32], %stage2_unit2_bn2_gamma: Tensor[(128), float32], %stage2_unit2_bn2_beta: Tensor[(128), float32], %stage2_unit2_bn2_moving_mean: Tensor[(128), float32], %stage2_unit2_bn2_moving_var: Tensor[(128), float32], %stage2_unit2_conv2_weight: Tensor[(128, 128, 3, 3), float32], %stage3_unit1_bn1_gamma: Tensor[(128), float32], %stage3_unit1_bn1_beta: Tensor[(128), float32], %stage3_unit1_bn1_moving_mean: Tensor[(128), float32], %stage3_unit1_bn1_moving_var: Tensor[(128), float32], %stage3_unit1_conv1_weight: Tensor[(256, 128, 3, 3), float32], %stage3_unit1_bn2_gamma: Tensor[(256), float32], %stage3_unit1_bn2_beta: Tensor[(256), float32], %stage3_unit1_bn2_moving_mean: Tensor[(256), float32], %stage3_unit1_bn2_moving_var: Tensor[(256), float32], %stage3_unit1_conv2_weight: Tensor[(256, 256, 3, 3), float32], %stage3_unit1_sc_weight: Tensor[(256, 128, 1, 1), float32], %stage3_unit2_bn1_gamma: Tensor[(256), float32], %stage3_unit2_bn1_beta: Tensor[(256), float32], %stage3_unit2_bn1_moving_mean: Tensor[(256), float32], %stage3_unit2_bn1_moving_var: Tensor[(256), float32], %stage3_unit2_conv1_weight: Tensor[(256, 256, 3, 3), float32], %stage3_unit2_bn2_gamma: Tensor[(256), float32], %stage3_unit2_bn2_beta: Tensor[(256), float32], %stage3_unit2_bn2_moving_mean: Tensor[(256), float32], %stage3_unit2_bn2_moving_var: Tensor[(256), float32], %stage3_unit2_conv2_weight: Tensor[(256, 256, 3, 3), float32], %stage4_unit1_bn1_gamma: Tensor[(256), float32], %stage4_unit1_bn1_beta: Tensor[(256), float32], %stage4_unit1_bn1_moving_mean: Tensor[(256), float32], %stage4_unit1_bn1_moving_var: Tensor[(256), float32], %stage4_unit1_conv1_weight: Tensor[(512, 256, 3, 3), float32], %stage4_unit1_bn2_gamma: Tensor[(512), float32], %stage4_unit1_bn2_beta: Tensor[(512), float32], %stage4_unit1_bn2_moving_mean: Tensor[(512), float32], %stage4_unit1_bn2_moving_var: Tensor[(512), float32], %stage4_unit1_conv2_weight: Tensor[(512, 512, 3, 3), float32], %stage4_unit1_sc_weight: Tensor[(512, 256, 1, 1), float32], %stage4_unit2_bn1_gamma: Tensor[(512), float32], %stage4_unit2_bn1_beta: Tensor[(512), float32], %stage4_unit2_bn1_moving_mean: Tensor[(512), float32], %stage4_unit2_bn1_moving_var: Tensor[(512), float32], %stage4_unit2_conv1_weight: Tensor[(512, 512, 3, 3), float32], %stage4_unit2_bn2_gamma: Tensor[(512), float32], %stage4_unit2_bn2_beta: Tensor[(512), float32], %stage4_unit2_bn2_moving_mean: Tensor[(512), float32], %stage4_unit2_bn2_moving_var: Tensor[(512), float32], %stage4_unit2_conv2_weight: Tensor[(512, 512, 3, 3), float32], %bn1_gamma: Tensor[(512), float32], %bn1_beta: Tensor[(512), float32], %bn1_moving_mean: Tensor[(512), float32], %bn1_moving_var: Tensor[(512), float32], %fc1_weight: Tensor[(1000, 512), float32], %fc1_bias: Tensor[(1000), float32]) -> Tensor[(1, 1000), float32] {
  %0 = nn.batch_norm(%data, %bn_data_gamma, %bn_data_beta, %bn_data_moving_mean, %bn_data_moving_var, epsilon=2e-05f, scale=False) /* ty=(Tensor[(1, 3, 224, 224), float32], Tensor[(3), float32], Tensor[(3), float32]) */;
  %1 = %0.0;
  %2 = nn.conv2d(%1, %conv0_weight, strides=[2, 2], padding=[3, 3, 3, 3], channels=64, kernel_size=[7, 7]) /* ty=Tensor[(1, 64, 112, 112), float32] */;
  %3 = nn.batch_norm(%2, %bn0_gamma, %bn0_beta, %bn0_moving_mean, %bn0_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 112, 112), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %4 = %3.0;
  %5 = nn.relu(%4) /* ty=Tensor[(1, 64, 112, 112), float32] */;
  %6 = nn.max_pool2d(%5, pool_size=[3, 3], strides=[2, 2], padding=[1, 1, 1, 1]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %7 = nn.batch_norm(%6, %stage1_unit1_bn1_gamma, %stage1_unit1_bn1_beta, %stage1_unit1_bn1_moving_mean, %stage1_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %8 = %7.0;
  %9 = nn.relu(%8) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %10 = nn.conv2d(%9, %stage1_unit1_conv1_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %11 = nn.batch_norm(%10, %stage1_unit1_bn2_gamma, %stage1_unit1_bn2_beta, %stage1_unit1_bn2_moving_mean, %stage1_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %12 = %11.0;
  %13 = nn.relu(%12) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %14 = nn.conv2d(%13, %stage1_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %15 = nn.conv2d(%9, %stage1_unit1_sc_weight, padding=[0, 0, 0, 0], channels=64, kernel_size=[1, 1]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %16 = add(%14, %15) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %17 = nn.batch_norm(%16, %stage1_unit2_bn1_gamma, %stage1_unit2_bn1_beta, %stage1_unit2_bn1_moving_mean, %stage1_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %18 = %17.0;
  %19 = nn.relu(%18) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %20 = nn.conv2d(%19, %stage1_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %21 = nn.batch_norm(%20, %stage1_unit2_bn2_gamma, %stage1_unit2_bn2_beta, %stage1_unit2_bn2_moving_mean, %stage1_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %22 = %21.0;
  %23 = nn.relu(%22) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %24 = nn.conv2d(%23, %stage1_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %25 = add(%24, %16) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %26 = nn.batch_norm(%25, %stage2_unit1_bn1_gamma, %stage2_unit1_bn1_beta, %stage2_unit1_bn1_moving_mean, %stage2_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 64, 56, 56), float32], Tensor[(64), float32], Tensor[(64), float32]) */;
  %27 = %26.0;
  %28 = nn.relu(%27) /* ty=Tensor[(1, 64, 56, 56), float32] */;
  %29 = nn.conv2d(%28, %stage2_unit1_conv1_weight, strides=[2, 2], padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %30 = nn.batch_norm(%29, %stage2_unit1_bn2_gamma, %stage2_unit1_bn2_beta, %stage2_unit1_bn2_moving_mean, %stage2_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %31 = %30.0;
  %32 = nn.relu(%31) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %33 = nn.conv2d(%32, %stage2_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %34 = nn.conv2d(%28, %stage2_unit1_sc_weight, strides=[2, 2], padding=[0, 0, 0, 0], channels=128, kernel_size=[1, 1]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %35 = add(%33, %34) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %36 = nn.batch_norm(%35, %stage2_unit2_bn1_gamma, %stage2_unit2_bn1_beta, %stage2_unit2_bn1_moving_mean, %stage2_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %37 = %36.0;
  %38 = nn.relu(%37) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %39 = nn.conv2d(%38, %stage2_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %40 = nn.batch_norm(%39, %stage2_unit2_bn2_gamma, %stage2_unit2_bn2_beta, %stage2_unit2_bn2_moving_mean, %stage2_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %41 = %40.0;
  %42 = nn.relu(%41) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %43 = nn.conv2d(%42, %stage2_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=128, kernel_size=[3, 3]) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %44 = add(%43, %35) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %45 = nn.batch_norm(%44, %stage3_unit1_bn1_gamma, %stage3_unit1_bn1_beta, %stage3_unit1_bn1_moving_mean, %stage3_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 128, 28, 28), float32], Tensor[(128), float32], Tensor[(128), float32]) */;
  %46 = %45.0;
  %47 = nn.relu(%46) /* ty=Tensor[(1, 128, 28, 28), float32] */;
  %48 = nn.conv2d(%47, %stage3_unit1_conv1_weight, strides=[2, 2], padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %49 = nn.batch_norm(%48, %stage3_unit1_bn2_gamma, %stage3_unit1_bn2_beta, %stage3_unit1_bn2_moving_mean, %stage3_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %50 = %49.0;
  %51 = nn.relu(%50) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %52 = nn.conv2d(%51, %stage3_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %53 = nn.conv2d(%47, %stage3_unit1_sc_weight, strides=[2, 2], padding=[0, 0, 0, 0], channels=256, kernel_size=[1, 1]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %54 = add(%52, %53) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %55 = nn.batch_norm(%54, %stage3_unit2_bn1_gamma, %stage3_unit2_bn1_beta, %stage3_unit2_bn1_moving_mean, %stage3_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %56 = %55.0;
  %57 = nn.relu(%56) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %58 = nn.conv2d(%57, %stage3_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %59 = nn.batch_norm(%58, %stage3_unit2_bn2_gamma, %stage3_unit2_bn2_beta, %stage3_unit2_bn2_moving_mean, %stage3_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %60 = %59.0;
  %61 = nn.relu(%60) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %62 = nn.conv2d(%61, %stage3_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=256, kernel_size=[3, 3]) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %63 = add(%62, %54) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %64 = nn.batch_norm(%63, %stage4_unit1_bn1_gamma, %stage4_unit1_bn1_beta, %stage4_unit1_bn1_moving_mean, %stage4_unit1_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 256, 14, 14), float32], Tensor[(256), float32], Tensor[(256), float32]) */;
  %65 = %64.0;
  %66 = nn.relu(%65) /* ty=Tensor[(1, 256, 14, 14), float32] */;
  %67 = nn.conv2d(%66, %stage4_unit1_conv1_weight, strides=[2, 2], padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %68 = nn.batch_norm(%67, %stage4_unit1_bn2_gamma, %stage4_unit1_bn2_beta, %stage4_unit1_bn2_moving_mean, %stage4_unit1_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %69 = %68.0;
  %70 = nn.relu(%69) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %71 = nn.conv2d(%70, %stage4_unit1_conv2_weight, padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %72 = nn.conv2d(%66, %stage4_unit1_sc_weight, strides=[2, 2], padding=[0, 0, 0, 0], channels=512, kernel_size=[1, 1]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %73 = add(%71, %72) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %74 = nn.batch_norm(%73, %stage4_unit2_bn1_gamma, %stage4_unit2_bn1_beta, %stage4_unit2_bn1_moving_mean, %stage4_unit2_bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %75 = %74.0;
  %76 = nn.relu(%75) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %77 = nn.conv2d(%76, %stage4_unit2_conv1_weight, padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %78 = nn.batch_norm(%77, %stage4_unit2_bn2_gamma, %stage4_unit2_bn2_beta, %stage4_unit2_bn2_moving_mean, %stage4_unit2_bn2_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %79 = %78.0;
  %80 = nn.relu(%79) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %81 = nn.conv2d(%80, %stage4_unit2_conv2_weight, padding=[1, 1, 1, 1], channels=512, kernel_size=[3, 3]) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %82 = add(%81, %73) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %83 = nn.batch_norm(%82, %bn1_gamma, %bn1_beta, %bn1_moving_mean, %bn1_moving_var, epsilon=2e-05f) /* ty=(Tensor[(1, 512, 7, 7), float32], Tensor[(512), float32], Tensor[(512), float32]) */;
  %84 = %83.0;
  %85 = nn.relu(%84) /* ty=Tensor[(1, 512, 7, 7), float32] */;
  %86 = nn.global_avg_pool2d(%85) /* ty=Tensor[(1, 512, 1, 1), float32] */;
  %87 = nn.batch_flatten(%86) /* ty=Tensor[(1, 512), float32] */;
  %88 = nn.dense(%87, %fc1_weight, units=1000) /* ty=Tensor[(1, 1000), float32] */;
  %89 = nn.bias_add(%88, %fc1_bias, axis=-1) /* ty=Tensor[(1, 1000), float32] */;
  nn.softmax(%89) /* ty=Tensor[(1, 1000), float32] */
}

Compilation

Next step is to compile the model using the Relay/TVM pipeline. Users can specify the optimization level of the compilation. Currently this value can be 0 to 3. The optimization passes include operator fusion, pre-computation, layout transformation and so on.

relay.build() returns three components: the execution graph in json format, the TVM module library of compiled functions specifically for this graph on the target hardware, and the parameter blobs of the model. During the compilation, Relay does the graph-level optimization while TVM does the tensor-level optimization, resulting in an optimized runtime module for model serving.

We’ll first compile for Nvidia GPU. Behind the scene, relay.build() first does a number of graph-level optimizations, e.g. pruning, fusing, etc., then registers the operators (i.e. the nodes of the optimized graphs) to TVM implementations to generate a tvm.module. To generate the module library, TVM will first transfer the high level IR into the lower intrinsic IR of the specified target backend, which is CUDA in this example. Then the machine code will be generated as the module library.

opt_level = 3
target = tvm.target.cuda()
with tvm.transform.PassContext(opt_level=opt_level):
    lib = relay.build(mod, target, params=params)

Out:

...1%, 0.01 MB, 227 KB/s, 0 seconds passed
...3%, 0.02 MB, 452 KB/s, 0 seconds passed
...5%, 0.02 MB, 671 KB/s, 0 seconds passed
...6%, 0.03 MB, 891 KB/s, 0 seconds passed
...8%, 0.04 MB, 978 KB/s, 0 seconds passed
...10%, 0.05 MB, 1169 KB/s, 0 seconds passed
...11%, 0.05 MB, 1361 KB/s, 0 seconds passed
...13%, 0.06 MB, 1551 KB/s, 0 seconds passed
...15%, 0.07 MB, 1741 KB/s, 0 seconds passed
...16%, 0.08 MB, 1907 KB/s, 0 seconds passed
...18%, 0.09 MB, 2092 KB/s, 0 seconds passed
...20%, 0.09 MB, 2244 KB/s, 0 seconds passed
...21%, 0.10 MB, 2403 KB/s, 0 seconds passed
...23%, 0.11 MB, 2578 KB/s, 0 seconds passed
...25%, 0.12 MB, 2555 KB/s, 0 seconds passed
...26%, 0.12 MB, 2716 KB/s, 0 seconds passed
...28%, 0.13 MB, 2879 KB/s, 0 seconds passed
...30%, 0.14 MB, 3042 KB/s, 0 seconds passed
...31%, 0.15 MB, 3152 KB/s, 0 seconds passed
...33%, 0.16 MB, 3310 KB/s, 0 seconds passed
...35%, 0.16 MB, 3469 KB/s, 0 seconds passed
...36%, 0.17 MB, 3627 KB/s, 0 seconds passed
...38%, 0.18 MB, 3785 KB/s, 0 seconds passed
...40%, 0.19 MB, 3914 KB/s, 0 seconds passed
...41%, 0.20 MB, 4058 KB/s, 0 seconds passed
...43%, 0.20 MB, 4212 KB/s, 0 seconds passed
...45%, 0.21 MB, 4365 KB/s, 0 seconds passed
...46%, 0.22 MB, 4518 KB/s, 0 seconds passed
...48%, 0.23 MB, 4671 KB/s, 0 seconds passed
...50%, 0.23 MB, 4775 KB/s, 0 seconds passed
...51%, 0.24 MB, 4921 KB/s, 0 seconds passed
...53%, 0.25 MB, 5051 KB/s, 0 seconds passed
...55%, 0.26 MB, 5198 KB/s, 0 seconds passed
...56%, 0.27 MB, 5345 KB/s, 0 seconds passed
...58%, 0.27 MB, 5275 KB/s, 0 seconds passed
...60%, 0.28 MB, 5397 KB/s, 0 seconds passed
...61%, 0.29 MB, 5536 KB/s, 0 seconds passed
...63%, 0.30 MB, 5676 KB/s, 0 seconds passed
...65%, 0.30 MB, 5815 KB/s, 0 seconds passed
...66%, 0.31 MB, 5883 KB/s, 0 seconds passed
...68%, 0.32 MB, 6018 KB/s, 0 seconds passed
...70%, 0.33 MB, 6154 KB/s, 0 seconds passed
...71%, 0.34 MB, 6289 KB/s, 0 seconds passed
...73%, 0.34 MB, 6425 KB/s, 0 seconds passed
...75%, 0.35 MB, 6556 KB/s, 0 seconds passed
...76%, 0.36 MB, 6691 KB/s, 0 seconds passed
...78%, 0.37 MB, 6824 KB/s, 0 seconds passed
...80%, 0.38 MB, 6924 KB/s, 0 seconds passed
...81%, 0.38 MB, 7054 KB/s, 0 seconds passed
...83%, 0.39 MB, 7185 KB/s, 0 seconds passed
...85%, 0.40 MB, 7309 KB/s, 0 seconds passed
...86%, 0.41 MB, 7438 KB/s, 0 seconds passed
...88%, 0.41 MB, 7567 KB/s, 0 seconds passed
...90%, 0.42 MB, 7697 KB/s, 0 seconds passed
...91%, 0.43 MB, 7827 KB/s, 0 seconds passed
...93%, 0.44 MB, 7956 KB/s, 0 seconds passed
...95%, 0.45 MB, 8084 KB/s, 0 seconds passed
...96%, 0.45 MB, 8213 KB/s, 0 seconds passed
...98%, 0.46 MB, 8341 KB/s, 0 seconds passed
...100%, 0.47 MB, 8451 KB/s, 0 seconds passed
Cannot find config for target=cuda -keys=cuda,gpu -max_num_threads=1024 -model=unknown -thread_warp_size=32, workload=('dense_small_batch.cuda', ('TENSOR', (1, 512), 'float32'), ('TENSOR', (1000, 512), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.

Run the generate library

Now we can create graph runtime and run the module on Nvidia GPU.

# create random input
ctx = tvm.gpu()
data = np.random.uniform(-1, 1, size=data_shape).astype("float32")
# create module
module = graph_runtime.GraphModule(lib["default"](ctx))
# set input and parameters
module.set_input("data", data)
# run
module.run()
# get output
out = module.get_output(0, tvm.nd.empty(out_shape)).asnumpy()

# Print first 10 elements of output
print(out.flatten()[0:10])

Out:

[0.00089283 0.00103331 0.0009094  0.00102275 0.00108751 0.00106737
 0.00106262 0.00095838 0.00110792 0.00113151]

Save and Load Compiled Module

We can also save the graph, lib and parameters into files and load them back in deploy environment.

# save the graph, lib and params into separate files
from tvm.contrib import utils

temp = utils.tempdir()
path_lib = temp.relpath("deploy_lib.tar")
lib.export_library(path_lib)
print(temp.listdir())

Out:

['deploy_lib.tar']
# load the module back.
loaded_lib = tvm.runtime.load_module(path_lib)
input_data = tvm.nd.array(np.random.uniform(size=data_shape).astype("float32"))

module = graph_runtime.GraphModule(loaded_lib["default"](ctx))
module.run(data=input_data)
out_deploy = module.get_output(0).asnumpy()

# Print first 10 elements of output
print(out_deploy.flatten()[0:10])

# check whether the output from deployed module is consistent with original one
tvm.testing.assert_allclose(out_deploy, out, atol=1e-3)

Out:

[0.00090713 0.00105705 0.00094459 0.00103146 0.00110017 0.00105846
 0.00104143 0.00095862 0.0010827  0.00111618]

Gallery generated by Sphinx-Gallery