tenferro_internal_ad_core/
value.rs

1use tenferro_internal_frontend_core::DynTensor;
2
3/// Canonical reverse-mode value handle for tenferro's hard cut to tidu.
4pub type DynValue = tidu::Value<DynTensor>;
5
6pub fn new_dyn_value(primal: DynTensor) -> DynValue {
7    tidu::Value::new(primal)
8}
9
10pub fn new_reverse_leaf(primal: DynTensor) -> DynValue {
11    tidu::Value::new(primal).with_requires_grad(true)
12}
13
14#[cfg(test)]
15mod tests {
16    use super::*;
17    use tenferro_internal_frontend_core::{DynTensor, ScalarType, StructuredTensor};
18    use tenferro_tensor::{MemoryOrder, Tensor};
19
20    fn dyn_tensor_from_slice(data: &[f64], dims: &[usize]) -> DynTensor {
21        let tensor = Tensor::<f64>::from_slice(data, dims, MemoryOrder::ColumnMajor).unwrap();
22        StructuredTensor::from(tensor).into()
23    }
24
25    #[test]
26    fn new_dyn_value_stays_detached() {
27        let primal = dyn_tensor_from_slice(&[1.0, 2.0, 3.0, 4.0], &[2, 2]);
28        let value = new_dyn_value(primal);
29
30        assert!(!value.requires_grad());
31        assert_eq!(value.primal().scalar_type(), ScalarType::F64);
32        assert_eq!(value.primal().dims(), &[2, 2]);
33    }
34
35    #[test]
36    fn new_reverse_leaf_enables_grad_tracking() {
37        let primal = dyn_tensor_from_slice(&[5.0, 6.0], &[2]);
38        let value = new_reverse_leaf(primal);
39
40        assert!(value.requires_grad());
41        assert_eq!(value.primal().scalar_type(), ScalarType::F64);
42        assert_eq!(value.primal().dims(), &[2]);
43    }
44}
tenferro_internal_ad_core/value.rs

tenferro_internal_ad_core/
value.rs
