GPU deployment using CUDA

IREE can accelerate model execution on Nvidia GPUs using CUDA.

:octicons-download-16: Prerequisites

In order to use CUDA to drive the GPU, you need to have a functional CUDA environment. It can be verified by the following steps:

nvidia-smi | grep CUDA

If nvidia-smi does not exist, you will need to install the latest CUDA Toolkit SDK.

Get the IREE compiler

:octicons-package-16: Download the compiler from a release

Python packages are regularly published to PyPI. See the Python Bindings page for more details. The core iree-compiler package includes the CUDA compiler:

=== “Stable releases”

Stable release packages are
[published to PyPI](https://pypi.org/user/google-iree-pypi-deploy/).

``` shell
python -m pip install iree-compiler
```

=== “:material-alert: Nightly releases”

Nightly releases are published on
[GitHub releases](https://github.com/iree-org/iree/releases).

``` shell
python -m pip install \
  --find-links https://iree.dev/pip-release-links.html \
  --upgrade iree-compiler
```

!!! tip iree-compile is installed to your python module installation path. If you pip install with the user mode, it is under ${HOME}/.local/bin, or %APPDATA%Python on Windows. You may want to include the path in your system's PATH environment variable:

```shell
export PATH=${HOME}/.local/bin:${PATH}
```

:material-hammer-wrench: Build the compiler from source

Please make sure you have followed the Getting started page to build the IREE compiler, then enable the CUDA compiler target with the IREE_TARGET_BACKEND_CUDA option.

!!! tip iree-compile will be built under the iree-build/tools/ directory. You may want to include this path in your system's PATH environment variable.

Get the IREE runtime

Next you will need to get an IREE runtime that includes the CUDA HAL driver.

:material-hammer-wrench: Build the runtime from source

Please make sure you have followed the Getting started page to build IREE from source, then enable the CUDA HAL driver with the IREE_HAL_DRIVER_CUDA option.

Compile and run a program model

With the compiler and runtime ready, we can now compile programs and run them on GPUs.

:octicons-file-code-16: Compile a program

The IREE compiler transforms a model into its final deployable format in many sequential steps. A model authored with Python in an ML framework should use the corresponding framework's import tool to convert into a format (i.e., MLIR) expected by the IREE compiler first.

Using MobileNet v2 as an example, you can download the SavedModel with trained weights from TensorFlow Hub and convert it using IREE's TensorFlow importer. Then run one of the following commands to compile:

iree-compile \
    --iree-hal-target-backends=cuda \
    --iree-cuda-target=<...> \
    mobilenet_iree_input.mlir -o mobilenet_cuda.vmfb

Canonically a CUDA target (iree-cuda-target) matching the LLVM NVPTX backend of the form sm_<arch_number> is needed to compile towards each GPU architecture. If no architecture is specified then we will default to sm_60.

Here is a table of commonly used architectures:

CUDA GPU	Target Architecture	Architecture Code Name
NVIDIA P100	`sm_60`	`pascal`
NVIDIA V100	`sm_70`	`volta`
NVIDIA A100	`sm_80`	`ampere`
NVIDIA H100	`sm_90`	`hopper`
NVIDIA RTX20 series	`sm_75`	`turing`
NVIDIA RTX30 series	`sm_86`	`ampere`
NVIDIA RTX40 series	`sm_89`	`ada`

In addition to the canonical sm_<arch_number> scheme, iree-cuda-target also supports two additonal schemes to make a better developer experience:

Architecture code names like volta or ampere
GPU product names like a100 or rtx3090

These two schemes are translated into the canonical form under the hood. We add support for common code/product names without aiming to be exhaustive. If the ones you want are missing, please use the canonical form.

:octicons-terminal-16: Run a compiled program

Run the following command:

iree-run-module \
    --device=cuda \
    --module=mobilenet_cuda.vmfb \
    --function=predict \
    --input="1x224x224x3xf32=0"

The above assumes the exported function in the model is named as predict and it expects one 224x224 RGB image. We are feeding in an image with all 0 values here for brevity, see iree-run-module --help for the format to specify concrete values.