Current operator support

This document is the authoritative capability boundary for the current NCET implementation. It describes canonical GraphIR operators and their accepted semantics, not future roadmap targets and not every operation that PyTorch FX can trace.

The accepted PyTorch spellings that produce these canonical operators are listed separately in PyTorch-to-GraphIR operator mapping. The semantic guarantee for a graph inside this boundary is defined by the exactness contract.

Meaning of supported

A canonical operator is supported only when all of the following exist for its documented parameter range:

1. frontend normalization into GraphIR;
2. graph-based interval bound propagation;
3. an exact CVXPY formulation;
4. essential operator and network-integration tests.

An FX operation is not supported merely because torch.fx.symbolic_trace() can capture it. Any canonical operator or parameter case outside the table below is outside the current exact boundary.

Current canonical operator boundary

Category	Canonical operator	Exact formulation	Current restrictions
Graph	Input	Continuous tensor variable with elementwise box bounds	One tensor per FX placeholder; multiple model inputs are supported
Graph	Output	Reference to existing graph tensor variables	One or more tensor outputs; creates no new tensor or variable
Affine	Linear	Linear equality	nn.Linear or F.linear; fixed weight and optional fixed bias; acts on the last tensor dimension
Affine	Conv2d	Sparse affine equality	nn.Conv2d or F.conv2d; per-sample shape (C,H,W); fixed weight and optional fixed bias; groups=1; dilation=(1,1); numeric zero padding
Affine	BatchNorm	Per-channel affine equality	nn.BatchNorm1d on (C,) or (C,L) and nn.BatchNorm2d on (C,H,W); evaluation mode; fixed running statistics; affine or non-affine modules
Affine	ElementwiseAffine	Elementwise affine equality	Exactly one graph tensor and one finite real scalar/tensor constant in Add/Sub/Mul/Div; constant broadcasting must preserve the graph tensor shape; division requires the graph tensor as numerator and a nonzero constant denominator
Pooling	AdaptiveAvgPool2d	Sparse linear equality	Per-sample shape (C,H,W); module or functional form; static scalar or length-2 output size; each entry is a positive integer or None, where None preserves that input dimension
Pooling	AvgPool2d	Sparse linear equality	Per-sample shape (C,H,W); scalar or 2-D kernel, stride, and padding; stride=None uses the kernel size; ceil_mode=False; divisor_override=None; either value of count_include_pad
Pooling	MaxPool2d	Full exact one-hot formulation	Per-sample shape (C,H,W); scalar or 2-D kernel, stride, and padding; stride=None uses the kernel size; dilation=(1,1); ceil_mode=False; return_indices=False
Structural	Identity	Elementwise equality	nn.Identity; nn.Dropout, nn.Dropout1d/2d/3d, or corresponding functional calls only with training=False; in-place forms unsupported
Activation	ReLU	Exact big-M or stable equality	Elementwise ReLU; relu_binary_mode may be "full" or "reduced"; in-place forms are unsupported
Activation	LeakyReLU	Exact big-M or stable equality	nn.LeakyReLU or F.leaky_relu; finite negative_slope in [0,1]; slopes 0 and 1 normalize to ReLU and Identity; in-place forms unsupported
Arithmetic	Add, Sub	Linear equality	Exactly two graph tensor operands; finite real scalar alpha
Composition	Concat	Exact output-slice equalities	Static tensor inputs and dimension; tracing batch axis cannot be concatenated; out must be absent or None
Reduction	ReduceMean	Linear equality	torch.mean or Tensor.mean; one or more explicit static dimensions; keepdim is static; tracing batch axis cannot be reduced; dtype and out must be absent or None
Shape	Flatten	C-order element-preserving equality	Static start_dim and end_dim; flattened range cannot include the tracing batch axis
Shape	Reshape	C-order element-preserving equality	Reshape, View, and batch-preserving Squeeze/Unsqueeze spellings; statically resolved output shape; Squeeze requires explicit dimensions; tracing batch axis must remain first
Shape	Permute	Native N-D axis permutation equality	Complete static permutation; tracing batch axis remains first
Shape	Transpose	Native N-D axis permutation equality	Static dimensions; tracing batch axis cannot be exchanged
Indexing	GetItem, Slice	Exact static index selection	Static integer/slice/ellipsis/new-axis syntax; positive slice steps; tracing batch axis remains unchanged

The graph may contain branches, residual connections, concatenation, shared modules, and multiple graph inputs or outputs. Each non-boundary canonical operation currently produces one tensor.

Outside the current boundary

Standalone Constant operators and other canonical operators not shown above are not currently supported. Fixed state used by Linear, Conv2d, BatchNorm, and ElementwiseAffine is lifted into GraphIR.constants; this does not create standalone Constant operators or CVXPY decision variables.

Planned operators and formulations remain outside this published capability boundary; appearing in a development roadmap does not imply current support.