Compiler
Tools and APIs for compiling custom ONNX models to run on Metis hardware. Compilation converts a floating-point model to an int8 binary optimized for the AIPU.
| Page | What it covers |
|---|---|
| Compiler CLI | The compile command-line tool. Covers basic usage, quantize-only mode, real-image calibration, output artifacts, and error status codes. |
| Compiler Python API | The compiler.quantize() and compiler.compile() Python API. Covers CompilerConfig, the two-step quantize → compile workflow, and usage examples. |
| Compiler Configuration | Multi-core compilation modes: Batch-1 (independent cores, lower latency) vs Batch-4 (shared memory, higher throughput). Includes resource allocation for multi-model pipelines. |
| ONNX Operator Support | Which ONNX operators are accelerated on the AIPU (opsets 14–17), which are constrained, and what falls back to CPU. |
| CompilerConfig Reference | Full property listing for CompilerConfig: all quantization, scheduling, memory, and hardware parameters with types, defaults, and enum values. |
See the Deploy Custom Weights tutorial for the end-to-end workflow.