Skip to content

api-reference.md

Latest commit

 

History

History
56 lines (44 loc) · 3.04 KB

api-reference.md

File metadata and controls

56 lines (44 loc) · 3.04 KB
Raw

Python API

The public classes are native wrappers: ML, MLContext, MLGraphBuilder, MLOperand, MLGraph, and MLTensor. The type stubs in python/webnn/_rustnn.pyi list the full operation signatures; the public stub re-exports those definitions.

Context and tensors

  • ML.create_context(power_preference="default", accelerated=True, backend="auto") adapts arguments to RustNN's context options and backend hint.
  • context.create_graph_builder() creates a runtime-bound native builder.
  • context.create_tensor(shape, data_type, readable=False, writable=False) explicitly allocates a native tensor.
  • context.write_tensor(tensor, data) and context.read_tensor(tensor) adapt NumPy buffers to native writes and reads without changing the native tensor dtype.
  • context.dispatch(graph, inputs, outputs) forwards named tensor bindings.
  • context.resize_tensor(tensor, shape) and set_tensor_capacity(tensor, max_shape) forward RustNN's tensor extensions.
  • context.accelerated queries the native context.

RustNN has not implemented capability reporting or tensor destruction; corresponding Python calls raise NotImplementedError rather than simulating those operations.

Builder

Operations forward to their RustNN counterparts. Numeric coefficients retain double precision; arg_min and arg_max default to int32; quantization zero points are optional. Each operation accepts a keyword-only label.

builder.input(name, shape, data_type) creates an input handle. builder.constant adapts NumPy values to native constant bytes, converting an explicit dtype override. Reshape/expand shapes and slice sizes accept integer dimensions or named dynamic dimensions such as {"name": "batch", "maxSize": 8}. RustNN determines their semantics.

Use builder.operand_shape(operand) and builder.operand_data_type(operand) to query native metadata before build. Python operands are opaque handles with no cached shape, dtype, or name properties.

builder.save(outputs, path) forwards RustNN's native save method before compilation. builder.build(outputs) calls RustNN's existing build method and consumes the builder. Subsequent native builder queries fail as they do in RustNN.

Compiled graph

MLGraph is a native compiled handle used by context.dispatch. It retains its runtime owner for lifetime safety. Python stores no GraphInfo or separate graph representation. Its input_descriptors and output_descriptors properties forward RustNN's public I/O fields as dictionaries with shape and data_type entries. Dynamic dimensions remain { "name": ..., "maxSize": ... } dictionaries.

context.load_graph(path) forwards RustNN's file loader and compiles the loaded graph for that context. RustNN resolves external weights relative to the graph file. Allocate tensors explicitly from the native I/O descriptors and reuse them across dispatches. Download Hub files with the independent webnn.Hub helper or huggingface_hub. Hub provides download_model, list_cached_models, and clear_cache; it performs no graph parsing or validation.