ExecuTorch

import { parseExecutorch } from "@wetron/executorch";

const graph = parseExecutorch(bytes: Uint8Array): ModelGraph

Synchronous. Detected by ET12 at bytes 4-7.

What is parsed#

• Root Program table -> first ExecutionPlan
• Operator list - name + overload -> opType string (e.g. aten.conv2d.default)
• EValue list scanned for Tensor entries -> shape and dtype for tensorShapes
• Chain instructions -> KernelCall entries -> operator index resolved to op type
• ExecutionPlan.inputs and .outputs -> graph inputs and outputs

ScalarType mapping#

Graph construction#

KernelCall arguments are categorised by whether their EValue index was already produced by a prior instruction:

• First occurrence of an index -> output tensor of this node
• Subsequent occurrence -> input tensor

This is a heuristic. It matches the common SSA-style ExecuTorch programs where each value index is written exactly once, and tracks plan inputs as “pre-produced.” It is approximate for programs that use:

• Out-variant kernels (aten::add.out(a, b, *, out)) - out is passed in the args list as a pre-allocated buffer and may appear classified as an output here even though the kernel mutates it in place.
• Constants / parameters that are not plan inputs and never produced by a prior instruction - these may be classified as outputs of their first use rather than as inputs.

For most viewer purposes the result is correct enough to walk the operator chain, but the produced IR should not be relied on for graph rewriting.

Notes#

• Only the first execution plan is parsed (multi-plan programs are rare in practice).
• Non-fatal per-instruction errors are attached as warnings.
• ModelGraph.initializers is empty and ModelGraph.weights is absent - constant buffer payloads are not surfaced.
• Uses the flatbuffers npm package with raw vtable offset decoding.