@wetron/executorch

ExecuTorch model parser for wetron. Reads .pte FlatBuffers files and returns a ModelGraph IR. Graph structure only - no weight tensors are deserialized.

Install

pnpm add @wetron/executorch

Included automatically when you install @wetron/core or @wetron/react.

API

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

const bytes = new Uint8Array(await file.arrayBuffer());
const graph = parseExecutorch(bytes); // synchronous

Throws ParseError from @wetron/common/ir on malformed input or missing ET12 identifier.

What gets parsed

• First ExecutionPlan from the root Program table
• Operator list -> op type strings
• EValue tensor entries -> shapes and dtypes
• KernelCall instructions -> operator calls

dtype mapping

| ExecuTorch enum | dtype | | --------------- | --------- | | 0 | uint8 | | 1 | int8 | | 2 | int16 | | 3 | int32 | | 4 | int64 | | 5 | float16 | | 6 | float32 | | 7 | float64 | | 11 | bool |

Notes

• Only the first execution plan is parsed.
• ModelGraph.initializers is always empty.
• Non-fatal per-instruction errors are attached as warnings.