GPU deployment using Vulkan

IREE can accelerate model execution on GPUs via Vulkan, a low-overhead graphics and compute API. Vulkan is cross-platform: it is available on many operating systems, including Android, Linux, and Windows. Vulkan is also cross-vendor: it is supported by most GPU vendors, including AMD, ARM, Intel, NVIDIA, and Qualcomm.

Support matrix

As IREE and the compiler ecosystem it operates within matures, more target specific optimizations will be implemented. At this stage, expect reasonable performance across all GPUs and for improvements to be made over time for specific vendors and architectures.

GPU Vendor	Category	Performance	Focus Architecture
ARM Mali GPU	Mobile	Good	Valhall+
Qualcomm Adreno GPU	Mobile	Reasonable	640+
AMD GPU	Desktop/server	Good	RDNA+
NVIDIA GPU	Desktop/server	Reasonable	Turing+

Device compatibility

IREE's Vulkan HAL targets modern compute-capable Vulkan implementations. The runtime baseline is Vulkan 1.3 with a compute queue and the timelineSemaphore, scalarBlockLayout, and synchronization2 device features. Devices missing that baseline still appear in device dumps, but are reported below the supported feature tier and are not selected for execution.

Descriptor-set dispatch is the portable baseline executable ABI. Buffer device address (BDA) dispatch is an optional faster ABI used when the device exposes bufferDeviceAddress and the program was compiled with BDA support. The runtime default accepts all supported executable ABI variants and prefers BDA when available; descriptor executables remain the fallback for devices without BDA.

Other Vulkan features are enabled opportunistically when the implementation reports them:

• VK_KHR_push_descriptor reduces descriptor allocation and update overhead.
• VK_EXT_calibrated_timestamps enables device-side queue and dispatch timing in IREE HAL profile captures.
• External memory extensions enable native interop and file/staging strategies when the platform and driver expose compatible handle types.
• Sparse binding and residency features are used only when explicitly requested and supported by an appropriate queue family.
• Validation layers, debug utils, robust buffer access, object names, and command labels are debug/profiling features rather than execution requirements.

Use iree-run-module --dump_devices to inspect the exact Vulkan API version, feature tier, device features, queues, and enabled extension opportunities for a machine. A supported device reports feature_tier[vulkan-1.3]: supported.

Prerequisites

In order to use Vulkan to drive the GPU, you need to have a functional Vulkan environment. IREE requires a device satisfying the Vulkan 1.3 baseline above. It can be verified by the following steps:

Android Linux Windows

Android version alone is not enough to determine IREE compatibility. Run vulkaninfo or iree-run-module --dump_devices on the device and verify that the selected physical device reports Vulkan 1.3 and the required baseline features.

Run the following command in a shell:

vulkaninfo | grep apiVersion

If vulkaninfo does not exist, you will need to install the latest Vulkan SDK. Installing via LunarG's package repository is recommended, as it places Vulkan libraries and tools under system paths so it's easy to discover.

If the listed version is lower than Vulkan 1.3, you will need to update the driver for your GPU.

Run the following command in a shell:

vulkaninfo | grep apiVersion

If vulkaninfo does not exist, you will need to install the latest Vulkan SDK.

If the listed version is lower than Vulkan 1.3, you will need to update the driver for your GPU.

Get the IREE compiler

Vulkan expects the program running on GPU to be expressed by the SPIR-V binary exchange format, which the model must be compiled into.

Download the compiler from a release

Python packages are distributed through multiple channels. See the Python Bindings page for more details. The core iree-base-compiler package includes the SPIR-V compiler:

Stable releases Nightly releases Development packages

Stable release packages are published to PyPI.

python -m pip install iree-base-compiler

Nightly pre-releases are published on GitHub releases.

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

Development packages are built at every commit and on pull requests, for limited configurations.

On Linux with Python 3.11, development packages can be installed into a Python venv using the build_tools/pkgci/setup_venv.py script:

# Install packages from a specific commit ref.
# See also the `--fetch-latest-main` and `--fetch-gh-workflow` options.
python ./build_tools/pkgci/setup_venv.py /tmp/.venv --fetch-git-ref=8230f41d
source /tmp/.venv/bin/activate

Tip

iree-compile and other tools are 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:

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

Build the compiler from source

Please make sure you have followed the Getting started page to build IREE for your host platform. The SPIR-V compiler backend is compiled in by default on all platforms, though you should ensure that the IREE_TARGET_BACKEND_VULKAN_SPIRV CMake option is ON when configuring.

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 supports the Vulkan HAL driver.

Download the runtime from a release

Python packages are distributed through multiple channels. See the Python Bindings page for more details. The core iree-base-runtime package includes the Vulkan HAL driver:

Stable releases Nightly releases Development packages

Stable release packages are published to PyPI.

python -m pip install iree-base-runtime

Nightly pre-releases are published on GitHub releases.

python -m pip install \
  --find-links https://iree.dev/pip-release-links.html \
  --upgrade --pre iree-base-runtime

Development packages are built at every commit and on pull requests, for limited configurations.

On Linux with Python 3.11, development packages can be installed into a Python venv using the build_tools/pkgci/setup_venv.py script:

# Install packages from a specific commit ref.
# See also the `--fetch-latest-main` and `--fetch-gh-workflow` options.
python ./build_tools/pkgci/setup_venv.py /tmp/.venv --fetch-git-ref=8230f41d
source /tmp/.venv/bin/activate

Build the runtime from source

Please make sure you have followed one of the Building from source pages to build IREE for your target platform. The Vulkan HAL driver is compiled in by default on supported platforms, though you should ensure that the IREE_HAL_DRIVER_VULKAN CMake option is ON when configuring.

Check for Vulkan devices

You can check for Vulkan support by looking for a matching driver and device:

$ iree-run-module --list_drivers

# ============================================================================
# Available HAL drivers
# ============================================================================
# Use --list_devices={driver name} to enumerate available devices.

        cuda: NVIDIA CUDA HAL driver (via dylib)
         hip: HIP HAL driver (via dylib)
  local-sync: Local execution using a lightweight inline synchronous queue
  local-task: Local execution using the IREE multithreading task system
      vulkan: Vulkan 1.x (dynamic)

$ iree-run-module --list_devices

hip://GPU-00000000-1111-2222-3333-444444444444
local-sync://
local-task://
vulkan://00000000-1111-2222-3333-444444444444

To see device details, including hints about what to use as a Vulkan target when compiling a program:

$ iree-run-module --dump_devices

...
# ============================================================================
# Enumerated devices for driver 'vulkan'
# ============================================================================

# ===----------------------------------------------------------------------===
# --device=vulkan://00000000-1111-2222-3333-444444444444
#   AMD Radeon PRO W7900 Dual Slot  (RADV GFX1100)
# ===----------------------------------------------------------------------===

Compile and run a program

With the requirements out of the way, we can now compile a model and run it.

Compile a program

The IREE compiler transforms a model into its final deployable format in several 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 a MobileNet model as an example, import using IREE's ONNX importer:

# Download the model you want to compile and run.
wget https://github.com/onnx/models/raw/refs/heads/main/validated/vision/classification/mobilenet/model/mobilenetv2-10.onnx

# Import to MLIR using IREE's ONNX importer.
pip install iree-base-compiler[onnx]
iree-import-onnx mobilenetv2-10.onnx --opset-version 17 -o mobilenetv2.mlir

Then run the following command to compile with the vulkan target device:

iree-compile \
    --iree-hal-target-device=vulkan \
    --iree-vulkan-dispatch-abi=all \
    --iree-vulkan-target=<...> \
    mobilenetv2.mlir -o mobilenet_vulkan.vmfb

Tip - Compiler Optimizations

Use --iree-opt-level=[O0,O1,O2,O3] to enable additional compiler optimizations. The default value of O0 enables only minimal optimizations while higher levels enable progressively more aggressive optimizations. See Optimization Options for more details.

--iree-vulkan-dispatch-abi controls which Vulkan executable ABI variants are emitted:

• descriptors emits the portable descriptor-set ABI.
• bda emits only the BDA root binding table ABI.
• all emits all supported ABI variants in runtime preference order.

Use all when producing a program intended to run across multiple Vulkan devices. Use bda only when the deployment fleet is known to support bufferDeviceAddress.

Choosing Vulkan targets

The --iree-vulkan-target specifies the GPU architecture to target. It accepts a few schemes:

• LLVM CodeGen backend style: this is using LLVM AMDGPU/NVPTX CodeGen targets like gfx1100 for AMD RX 7900XTX and sm_86 for NVIDIA RTX 3090 GPUs.
• Architecture code name style like rdna3/valhall4/ampere/adreno for AMD/ARM/NVIDIA/Qualcomm GPUs.
• Product name style: e.g., using rx7900xtx/a100 for corresponding GPUs.

Here are a few examples showing how you can target various recent common GPUs:

GPU	Product Name	Target Architecture	Architecture Code Name
AMD RX 5000 series			rdna1
AMD RX 6000 series			rdna2
AMD RX 7700XT	rx7700xt	gfx1101	rdna3
AMD RX 7800XT	rx7800xt	gfx1101	rdna3
AMD RX 7900XT	rx7900xt	gfx1100	rdna3
AMD RX 7900XTX	rx7900xtx	gfx1100	rdna3
AMD RX 9060XT	rx9060xt	gfx1200	rdna4
AMD RX 9070	rx9070	gfx1201	rdna4
AMD RX 9070XT	rx9070xt	gfx1201	rdna4
ARM GPUs			valhall
ARM Mali G510	mali-g510		valhall3
ARM Mali G715	mali-g715		valhall4
NVIDIA RTX20 series	rtx2070super	sm_75	turing
NVIDIA RTX30 series	rtx3080ti	sm_86	ampere
NVIDIA RTX40 series	rtx4090	sm_89	ada
Qualcomm GPUs			adreno

If no target is specified, then a safe but more limited default will be used.

Note

We don't support the full spectrum of GPUs here and it is impossible to capture all details of a Vulkan implementation with a target triple, given the allowed variances on extensions, properties, limits, etc. So the target triple is just an approximation for usage. This is more of a mechanism to help us develop IREE itself. In the long term we want to perform multi-targeting to generate code for multiple architectures if no explicit target is given.

Run a compiled program

To run the compiled program:

iree-run-module \
    --device=vulkan \
    --module=mobilenet_vulkan.vmfb \
    --function=torch-jit-export \
    --input="1x3x224x224xf32=0"

The above assumes the exported function in the model is named torch-jit-export 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.