TorchScript

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

const graph = parseTorchscript(bytes: Uint8Array): ModelGraph

Synchronous.

Format variants#

ZIP-based (torch.jit.save)#

Detected by PK\x03\x04 ZIP magic at offset 0. Reads bytecode.pkl from the archive and decodes the Python binary serialization stream (protocol 2/4) - extracts operator names and overloads as metadata. No code is executed.

Operators are extracted as (name, overload, n) tuples from the operators section of the bytecode.

FlatBuffers Mobile (.ptl)#

Detected by PTMF at bytes 4-7. Reads the Module root table, walks methods -> IValue references -> Function entries, and extracts operator calls from instruction bytecode. Opcode 0 (OP) indicates a call to a registered operator.

Falls back to scanning all IValues for Function type if the methods array yields nothing.

Graph structure#

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

Fidelity caveat. TorchScript bytecode does not record explicit dataflow at the level Wetron parses (it records only the ordered sequence of operator calls). The visualiser therefore chains operators in execution order and produces a straight line, regardless of branching, parallel paths, or shared tensors in the original model. Use this view to inspect the operator sequence, not the dataflow graph.

input
  ↓
op_0  (e.g. aten::conv2d)
  ↓
op_1  (e.g. aten::batch_norm)
  ↓
...
  ↓
output

The op type strings (aten::conv2d, aten::leaky_relu, …) are still accurate and remain useful for category classification, icon assignment, and operator inventory.

Notes#

• ModelGraph.initializers is empty and ModelGraph.weights is absent - TorchScript bytecode does not include tensor payloads at the level this parser walks.
• ModelGraph.tensorShapes is always empty - shape inference is not performed.
• Non-fatal per-method errors are attached as warnings on the returned graph.
• Uses flatbuffers for Mobile format and fflate for ZIP extraction.