config-file infer-config JSON Members - config-file infer-config JSON Members - 6.3 English

config-file infer-config JSON Members - config-file infer-config JSON Members - 6.3 English - XD056

Vitis Video Analytics SDK (VVAS) 6.3 User Guide (XD056)

Document_ID

XD056

Release_Date

2026-09-24

Version

6.3 English

Table 7 infer-config json members
Parameter	Type	Expected Values	Default	Description
attach-empty-metadata	Boolean	true/false	true	Adds empty inference metadata to GstBuffer, when there are no predictions available for that frame.
model-format	String	“RGB”/”BGR”	N/A	Model/network’s format.
inference-level	Integer	1 to 2	1	Specifies the inference level in cascaded pipeline; for example, object detection ML (level=1) followed by object classification ML (level=2) on detected objects.
attach-ppe-outbuf	Boolean	true/false	false	Specifies whether to attach the pre-processed ROI buffer inside the metadata tree so that subsequent stages can reuse the same crop (improving performance by avoiding repeated crop/resize operations).
inference-backend-runtime	String	“onnxrt”/”vart”	N/A	Specifies the inference backend runtime, accordingly config json object needs to be provided.
onnxrt-config	JSON Object	See Table onnxrt-config json members`	N/A	Configuration object for ONNX Runtime inference engine. Contains ONNX model path and optional execution provider settings.
vart-config	JSON Object	See Table vart-config json members`	N/A	Configuration object for vart inference engine. Contains vart model path and optional execution provider settings.

Table 8 onnxrt-config JSON Members
Parameter	Type	Expected Values	Default	Description
onnx-model-path	String	Path to `.onnx` file	N/A	Absolute or relative path to the ONNX model file to be used for inference.
input-tensor-layout	String	“NCHW” / “NHWC”	N/A	Specifies the layout of the input tensor expected by the ONNX model. Common values are “NCHW” (batch, channel, height, width) or “NHWC” (batch, height, width, channel). The selected layout and tensor data type determine the GStreamer tensor format name shown in caps and debug logs; see Custom AI Tensor Formats.
onnxrt-ep	String	“vitisai”`` / “cpu”	N/A	Specifies the Vitis AI Execution Provider (EP).
enable-profiling	Boolean	true / false	false	Enable or disable profiling.
profiling-file-path	String	File path to write profiling output	N/A	Specify the path to write profiling output.
vitisai-ep-config	JSON Object	See vitisai-ep json members	N/A	(Optional) Configuration for the Vitis AI Execution Provider (EP) in ONNX Runtime. If this node is set, the VitisAI EP will be used for inference; otherwise, the CPU Execution Provider (CPU-EP) will be used by default.

Table 9 vart-config JSON Members
Parameter	Type	Expected Values	Default	Description
vart-model-path	String	Path to VART model file	N/A	Absolute or relative path to the VART model file to be used for inference.
config-file-path	String	Path to config file	N/A	(Optional) Path to the Vitis AI EP configuration file (Example: `vitisai_config.json`).
ai-analyzer-profiling	Boolean	true / false	false	Parameter to enable or disable AI Analyzer profiling.
out-mem-bank	Integer	0 to 65535	0	VVAS output memory bank to allocate memory.
aie-columns-sharing	Boolean	true / false	true	Whether to use AIE columns sharing for the NPU overlay execution.
start-column	Integer	0 to 32	0	Start column for the NPU overlay execution. Value must be a multiple of 4.

Note

The valid range for start-column depends on the number of NPU columns available on the device. For devices with 24 NPU columns, the valid range is 0 to 20.

- input-tensor-type
- String
- “hw” / “cpu”
- “hw”
- Memory type for the model input tensor. Set according to the input subgraph boundaries of your compiled model.
- output-tensor-type
- String
- “hw” / “cpu”
- “hw”
- Memory type for the model output tensor. Set according to the output subgraph boundaries of your compiled model.
- use-async
- Boolean
- true / false
- true
- Controls whether VART inference runs in parallel with CPU processing. When true, inference is submitted to the NPU and the CPU immediately continues pre/post-processing the next frame. When false, the CPU waits for each NPU inference to complete before processing the next frame. Has no effect on the ONNX-RT backend.

Table 10 vitisai-ep JSON Members
Parameter	Type	Expected Values	Default	Description
config-file-path	String	Path to config file	N/A	Path to the Vitis AI EP configuration file (Example: `vitisai_config.json`).
ai-analyzer-profiling	Boolean	true / false	false	Parameter to enable or disable AI Analyzer profiling.
cache-dir	String	Path to cache directory or compiled model directory	`/tmp/$USER/vaip`	Directory path where cache files for the VAIML compiled model will be stored.
cache-key	String	Directory name	Directory named with a hash value under `/tmp/$USER/vaip`	Unique identifier (directory name) for the cache entry corresponding to the VAIML compiled model.