Recursive AI

Documentation

All onemodel docs

Normative, language-neutral notes live here, one file per topic, so a port can follow them without the C# code.

TopicFile
Tensors, memory, layouts, fp16/bf16 rounding, gradient recordstensors.md
Exact mode: definition and accumulation orderexact-mode.md
Exact mode golden hashes: rows, results, cost of exactnessgolden-hashes.md
Matrix product and convolution kernelskernels-gemm.md
Attention and softmax kernelskernels-attention.md
exp, log, tanh, erf, sigmoid, GELU: polynomials and errorexact-transcendentals.md
Reference oracles, pinned versions, comparison metricsoracles.md
Benchmark methodbench-method.md
Depthwise convolution: phase planes, lane layout, bitskernels-depthwise.md
Half-precision storage: conversions, the widening adapter, compact weightsfp16.md
ONNX import: protobuf reading, tensors, graph resolution, external data, errorsonnx-import.md
Operator contract, plans, sessions, memory reusegraph-runtime.md
Reference models: method, operators used, resultsreference-models.md
Graph passes: what each rewrites, levels, bit preservationgraph-passes.md
Normalisation, softmax, reduction, arg-reduction and cumulative-sum operatorsops-norms.md
Convolution, pooling, normalisation and resize operators: supported cases, edge rules, bitsops-conv.md
Shape and indexing operators: semantics, ONNX Runtime findingsops-shape-indexing.md
Recurrent operators (RNN, GRU, LSTM): order of operations, open choicesops-recurrent.md
CTC decoding: greedy and prefix beam search, confidencesctc.md
Element-wise operators, activations, comparisons, Where, Castops-elementwise.md
int8: operators, product kernel, requantization, QDQ rewriteint8.md
The OneModel model file: header, JSON manifest, aligned tensors, reading and writing rules (version 1, unstable until 1.0)format.md
Conformance: onnx/backend/test results per operator, method, gapsconformance.md
Operators added for coverage: settled semantics, exactnessops-coverage.md