ROCm Execution Provider

The ROCm Execution Provider enables hardware accelerated computation on AMD ROCm-enabled GPUs.

Contents

Install

NOTE Please make sure to install the proper version of Pytorch specified here PyTorch Version.

For Nightly PyTorch builds please see Pytorch home and select ROCm as the Compute Platform.

Pre-built binaries of ONNX Runtime with ROCm EP are published for most language bindings. Please reference Install ORT.

Requirements

ONNX Runtime ROCm
main 6.0
1.17 6.0
5.7
1.16 5.6
5.5
5.4.2
1.15 5.4.2
5.4
5.3.2
1.14 5.4
5.3.2
1.13 5.4
5.3.2
1.12 5.2.3
5.2

Build

For build instructions, please see the BUILD page.

Configuration Options

The ROCm Execution Provider supports the following configuration options.

device_id

The device ID.

Default value: 0

tunable_op_enable

Set to use TunableOp.

Default value: false

tunable_op_tuning_enable

Set the TunableOp try to do online tuning.

Default value: false

user_compute_stream

Defines the compute stream for the inference to run on. It implicitly sets the has_user_compute_stream option. It cannot be set through UpdateROCMProviderOptions. This cannot be used in combination with an external allocator.

Example python usage:

providers = [("ROCMExecutionProvider", {"device_id": torch.cuda.current_device(),
                                        "user_compute_stream": str(torch.cuda.current_stream().cuda_stream)})]
sess_options = ort.SessionOptions()
sess = ort.InferenceSession("my_model.onnx", sess_options=sess_options, providers=providers)

To take advantage of user compute stream, it is recommended to use I/O Binding to bind inputs and outputs to tensors in device.

do_copy_in_default_stream

Whether to do copies in the default stream or use separate streams. The recommended setting is true. If false, there are race conditions and possibly better performance.

Default value: true

gpu_mem_limit

The size limit of the device memory arena in bytes. This size limit is only for the execution provider’s arena. The total device memory usage may be higher. s: max value of C++ size_t type (effectively unlimited)

Note: Will be over-ridden by contents of default_memory_arena_cfg (if specified)

arena_extend_strategy

The strategy for extending the device memory arena.

Value Description
kNextPowerOfTwo (0) subsequent extensions extend by larger amounts (multiplied by powers of two)
kSameAsRequested (1) extend by the requested amount

Default value: kNextPowerOfTwo

Note: Will be over-ridden by contents of default_memory_arena_cfg (if specified)

gpu_external_[alloc|free|empty_cache]

gpu_external_* is used to pass external allocators. Example python usage:

from onnxruntime.training.ortmodule.torch_cpp_extensions import torch_gpu_allocator

provider_option_map["gpu_external_alloc"] = str(torch_gpu_allocator.gpu_caching_allocator_raw_alloc_address())
provider_option_map["gpu_external_free"] = str(torch_gpu_allocator.gpu_caching_allocator_raw_delete_address())
provider_option_map["gpu_external_empty_cache"] = str(torch_gpu_allocator.gpu_caching_allocator_empty_cache_address())

Default value: 0

Usage

C/C++ 

Ort::Env env = Ort::Env{ORT_LOGGING_LEVEL_ERROR, "Default"};
Ort::SessionOptions so;
int device_id = 0;
Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_ROCm(so, device_id));

The C API details are here.

Python

Python APIs details are here.

Samples

Python 

import onnxruntime as ort

model_path = '<path to model>'

providers = [
    'ROCMExecutionProvider',
    'CPUExecutionProvider',
]

session = ort.InferenceSession(model_path, providers=providers)