Advanced Runtime Techniques

These guides tie compiler and runtime features to concrete workflows: model preparation, quantization and compilation for accuracy and throughput, and application patterns—data flow, scheduling, and NPU column use—for effective execution on Versal AI Edge Gen2.

For overall application structure, initialization, and pipeline concepts, see also docs/appendix/vart_app_guide.

Feature-focused guides

Feature guide	Topics covered	Companion sample
Multi-Tenancy: Spatial and Temporal Sharing	Spatial and temporal sharing, runtime JSON configuration (start_column, aie_columns_sharing), column footprint rules, VART runner options.	Overview of Data and Tensor Parallelism (tutorial); vart_multi_tenancy (sample app)
Zero-copy preprocess, NPU, and postprocess	Preprocess, NPU inference, and postprocess with shared hardware buffers: IFM filled by preprocess and bound to the runner input without an extra copy; OFM read in place by postprocess.	vart_zerocopy
Compiling and Executing Mixed Precision Models	AMD Quark VINT8 with excluded tail subgraph, vitisai_config.json for BF16 or FP16 tail, build/run on target with the vision pipeline app.	x_plus_ml_vart