onnxruntime/docs/execution-providers/ROCm-ExecutionProvider.md
PeixuanZuo 4b9a7e270e
[Doc] Update latest ROCm version (#19665)
Update latest ROCm version
2024-02-29 15:07:22 +08:00

159 lines
4.6 KiB
Markdown

---
title: AMD - ROCm
description: Instructions to execute ONNX Runtime with the AMD ROCm execution provider
parent: Execution Providers
nav_order: 10
redirect_from: /docs/reference/execution-providers/ROCm-ExecutionProvider
---
# ROCm Execution Provider
{: .no_toc }
The ROCm Execution Provider enables hardware accelerated computation on AMD ROCm-enabled GPUs.
## Contents
{: .no_toc }
* TOC placeholder
{:toc}
## Install
**NOTE** Please make sure to install the proper version of Pytorch specified here [PyTorch Version](../install/#training-install-table-for-all-languages).
For Nightly PyTorch builds please see [Pytorch home](https://pytorch.org/) 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](../install).
## Requirements
|ONNX Runtime|ROCm |
|------------|-------------------------|
| main | 6.0 |
| 1.17 | 6.0<br/>5.7 |
| 1.16 | 5.6<br/>5.5<br/>5.4.2 |
| 1.15 | 5.4.2<br/>5.4<br/>5.3.2 |
| 1.14 | 5.4<br/>5.3.2 |
| 1.13 | 5.4<br/>5.3.2 |
| 1.12 | 5.2.3<br/>5.2 |
## Build
For build instructions, please see the [BUILD page](../build/eps.md#amd-rocm).
## 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:
```python
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](../api/python/api_summary.html) 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:
```python
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++
```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](../get-started/with-c.md).
### Python
Python APIs details are [here](https://onnxruntime.ai/docs/api/python/api_summary.html).
## Samples
### Python
```python
import onnxruntime as ort
model_path = '<path to model>'
providers = [
'ROCMExecutionProvider',
'CPUExecutionProvider',
]
session = ort.InferenceSession(model_path, providers=providers)
```