TFLite

import { parseTflite } from "@wetron/tflite";

const graph = parseTflite(bytes: Uint8Array): ModelGraph

Synchronous - no await needed.

What is parsed#

• The first subgraph (subgraphs[0]); secondary subgraphs referenced by If / While are not yet inlined
• Op types from the built-in BuiltinOperator enum; custom ops use their custom_code string
• Tensor shapes and dtypes for all tensors in the parsed subgraph
• Graph inputs and outputs from the primary subgraph
• Initializer bytes exposed via ModelGraph.weights.get(name) - tensors with a non-empty buffer reference

Detection#

Detects both:

• TFL3 at offset 4 (standard TFLite)
• ODLF at offset 4 (LiteRT / ODLF variant)

Dtype mapping#

Notes#

• Initializer buffers are exposed lazily through ModelGraph.weights; consumers call weights.get(name) for raw bytes and decodeWeight / computeStats from @wetron/core to inspect values. See Weights.
• Synchronous because FlatBuffers decoding requires no async I/O.