onnx-std

A pure-Rust ONNX standard library — model I/O, textual format, and an extensible validator — built on the shared runtime IR. This is the first wave of the design captured in docs/ONNX_RS.md.

Relationship to the rest of the workspace

onnx-std does not define its own IR. Per ONNX_RS §4.1 ("Shared Crate"), the Graph / Node / Value / Tensor / WeightRef types are reused verbatim from [onnx_runtime_ir], and the protobuf parse/encode stack is reused from [onnx_runtime_loader]. onnx-std is the standard-library layer on top: ergonomic model I/O, a human-readable text format, and a checker — the concerns the design assigns to the ONNX library rather than the inference runtime.

onnx-std  ──(standard lib: load/save/print/check)
   │
   └─ depends on ─▶ onnx-runtime-ir      (shared Graph/Node/Value IR)
   └─ depends on ─▶ onnx-runtime-loader  (protobuf serde + weight mmap)

What this wave covers

JSON and protobuf TextFormat are descriptor-driven from the exact vendored ONNX proto, so every bound message, field, oneof, and enum is covered. The readable text DSL appends a protobuf-TextFormat extension containing only fields outside the graph syntax. The parsed DSL is authoritative for every field it represents, while the explicit extension preserves the remainder.

Deferred to later waves (see // FOLLOW-UP markers): the complete operator catalog, remaining function/IR-gate checker rules, custom-op registration (§11), and Python bindings (§12). Training semantics are out of scope.

Example

let model = onnx_std::load_model("model.onnx")?;
println!("{}", model.to_text());
let report = model.validate();
assert!(report.is_valid());
onnx_std::save_model(&model, "roundtrip.onnx")?;
# Ok::<(), onnx_std::Error>(())