Principle:LaurentMazare Tch rs JIT Forward Pass
| Knowledge Sources | |
|---|---|
| Domains | Model_Inference, Interoperability |
| Last Updated | 2026-02-08 14:00 GMT |
Overview
Execution of a TorchScript model's forward method with tensor or IValue inputs, returning model predictions.
Description
JIT forward pass executes the loaded TorchScript model's forward method by passing input data through the preserved computation graph. Two variants exist: forward_ts accepts a slice of Tensor inputs (suitable for standard models), while forward_is accepts IValue inputs (for models with complex I/O types like tuples or dictionaries). The CModule also implements the Module trait, enabling use with Tensor::apply for single-tensor input models.
Usage
Use forward_ts for models with simple tensor inputs. Use forward_is for models that expect or return non-tensor types (tuples, lists, dicts). For single-input models, Tensor::apply works via the Module trait implementation.
Theoretical Basis
CModule Forward Variants:
forward_ts(&[Tensor]) → Result<Tensor> — For tensor-only I/O
forward_is(&[IValue]) → Result<IValue> — For complex I/O types
Module::forward(&Tensor) → Tensor — For single tensor input
IValue types: Tensor, Tuple, List, Dict, Int, Double, String, Bool, None