Coelanox · Repository docsBrowse by category

COELANOX — ONNX support for customers (opset 13)

Purpose: Human-readable summary of what the in-tree ONNX translator can turn into Universal IR so coelanox package succeeds. It is not the internal CLF op_id registry (see sdk/clf.md for developers).

Scope: ONNX opset 13 import path. Other opsets are not documented here as supported.

1. How to read this document

1:1 mapping — One ONNX operator becomes one IR operator (direct support).
Decomposed — One ONNX operator becomes a small subgraph of IR primitives (still valid).
Rejected — Packaging fails with a clear message (you must change the model or app).
Custom / unsupported — Translator emits Custom("OpName"); package fails until resolved.

2. Directly supported ONNX ops (1:1 examples)

The following categories are representative; the canonical list is maintained in the repository:

Arithmetic: Add, Sub, Mul, Div
Activations: Relu, Sigmoid, Tanh, Softmax, LogSoftmax, Gelu
Matrix: MatMul, Gemm
Conv / pool / norm: Conv, MaxPool, AveragePool, GlobalAveragePool, GlobalMaxPool, BatchNormalization, LayerNormalization, Dropout
Tensor: Reshape, Transpose, Concat, Split, Slice, Gather, Scatter (variants), Expand, Broadcast
Reductions: ReduceSum, ReduceMean, ReduceMax, ReduceMin, ReduceProd
Math: Sqrt, Pow, Exp, Log, Cos, Sin, Min, Max
Comparisons / logical: Equal, NotEqual, Greater, GreaterEqual, Less, LessEqual, And, Or, Not
Other: Identity (rewired)

Full machine-readable list: ONNX decomposition tree §1.

3. Decomposed ONNX ops (still supported)

These ONNX ops are not missing—they are lowered to supported primitives:

ONNX	Lowering idea
Clip	Max / Min chain
Neg	Multiply by −1
Reciprocal	Divide 1 / x
Flatten, Unsqueeze, Squeeze	Reshape with attributes
Abs	Pow(2) + Sqrt (or equivalent chain)
Mean	Add chain + Divide
MeanVarianceNormalization	ReduceMean, Subtract, Pow, Sqrt, Divide
Constant	Float32 constant materialized as weight
LSTM	Forward-only subgraph (see limitations)

Details: ONNX decomposition tree §3–5.

4. Explicitly rejected at translation (not “unsupported silently”)

ONNX	Behaviour
If, Loop	Error: control flow must live in your application; export a model with a fixed execution path.

5. Unsupported ONNX ops (what happens)

Any ONNX operator not covered by §2–4 is typically translated as Custom("OnnxOpName"). coelanox package will fail with an error indicating unresolved Custom nodes.

Examples of ops that fall into “Custom” today (non-exhaustive):

Many Reduce* variants not mapped (e.g. ReduceL1, ReduceL2, …)
Resize, Pad, Cast, Shape, Erf, GridSample, RNN, GRU (unless specifically handled elsewhere)
QuantizeLinear / DequantizeLinear as first-class paths
Einsum, ArgMax, ArgMin, and many others

See §4 Custom in the full decomposition tree.

6. What to do if your op does not compile

Open the error — It should name Custom("OpName") or a translator error.
Check the full tree — ONNX_DECOMPOSITION_TREE.md for a known decomposition.
Change the export — Simplify graph (remove unsupported ops), or fuse ops in the framework before export.
Move control flow to the app — If If/Loop, restructure.
Extend the translator — For vendors, this means Rust changes in coelanox-packager (see frontend SDK); not a runtime configuration toggle.

7. LSTM caveat

LSTM is decomposed with constraints (e.g. forward direction, documented assumptions). Read the LSTM section in the full tree before relying on it for production.

← Non-technical hub