AMD Vitis™ AI EP Options

The provider_options parameter configures the AMD Vitis™ AI Execution Provider when creating an ONNX Runtime InferenceSession. Pass these options as a dictionary to the provider_options parameter, as shown in the Model Compilation examples.

The Vitis AI EP supports the following options, organized by category:

Required Options

These options must be provided when creating the InferenceSession.

config_file

Path to the JSON configuration file containing compilation options. This file specifies device configuration, optimization settings, and other compiler parameters. For complete details on configuration file structure and available options, see AMD Vitis™ AI EP Configuration File.

• Type: String

• Required: Yes

target

Target platform for the Vitis AI Execution Provider.

• Type: String

• Required: Yes

• Supported values: “VAIML”

Minimal example with required options:

provider_options = {
    "config_file": "vitisai_config.json",
    "target": "VAIML"
}

Caching Options (Recommended)

These options control where compiled model binaries are stored. Setting these options is recommended to ensure compiled models are cached in a predictable location and can be reused across sessions, avoiding recompilation.

cache_dir

Path to the directory where compiled model binaries are cached. When specified, the Vitis AI EP stores the compiled model in this directory. On subsequent runs with the same model, the cached binary is loaded instead of recompiling, significantly reducing startup time.

• Type: String

• Default: Current working directory

cache_key

Subfolder name within the cache directory for storing this specific compiled model. This allows multiple compiled models to coexist in the same cache directory without conflicts. Useful when compiling different models or the same model with different configurations.

• Type: String

• Default: Input ONNX model filename without extension

Example with caching options:

provider_options = {
    "config_file": "vitisai_config.json",
    "target": "VAIML",
    "cache_dir": "/path/to/cache",
    "cache_key": "my_model_v1"
}

Analysis and Debugging Options

These options generate analysis data for debugging and performance optimization. They are intended for short development and analysis sessions only. Profiling adds memory and performance overhead and should be avoided for long-running sessions. Do not enable these options in production deployment.

ai_analyzer_visualization

Enables generation of compile-time analysis data for visualizing model partitioning, operator placement decisions, and compilation flow. The output helps understand how the compiler distributes operations between NPU and CPU.

• Type: Boolean

• Default: False

ai_analyzer_profiling

Enables generation of runtime profiling data for analyzing inference performance metrics, including execution time per operator and memory usage patterns.

• Type: Boolean

• Default: False

Example with analysis options:

provider_options = {
    "config_file": "vitisai_config.json",
    "target": "VAIML",
    "cache_dir": "/path/to/cache",
    "ai_analyzer_visualization": True,
    "ai_analyzer_profiling": True
}

Note

Analysis and profiling options increase compilation time and add memory and performance overhead at runtime. Use them only for short development and analysis runs—not for long-running sessions or production deployment.