@wetron/torchscript

TorchScript model parser for wetron. Reads .pt files (ZIP-based and FlatBuffers Mobile format) and returns a ModelGraph IR. Graph structure only - no weight tensors are deserialized.

Install

pnpm add @wetron/torchscript

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

API

import { parseTorchscript } from "@wetron/torchscript";

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

Throws ParseError from @wetron/common/ir on malformed input.

Format variants

ZIP-based (torch.jit.save): detected by PK\x03\x04 magic, reads bytecode.pkl.

FlatBuffers Mobile (.ptl): detected by PTMF at bytes 4-7, reads operator calls from instruction bytecode.

Both produce a linear graph: input -> op_0 -> op_1 -> … -> output.

Notes

• ModelGraph.initializers and tensorShapes are always empty - weight data and shape inference are not performed.
• Non-fatal per-method errors are attached as warnings on the returned ModelGraph.