nevla

import py "torch"

// check propagates: the caller decides
fn logits(n int) (str, error?) {
    w := check torch.randn([784, 10],
        requires_grad: true)
    x := check torch.randn([n, 784])
    y := check (x @ w)
    return check str(y.shape), none
}

// main can act, so it handles
fn main() {
    shape, err := logits(32)
    if err != none {
        print("torch failed: " + err.msg)
        return
    }
    print("logits: " + shape)
}

That is real torch, one import away, with Python's exceptions arriving as typed error values instead of tracebacks. The whole program is checked before any of it runs; no crash originates in nevla.

Errors are values and dropping one is a compile error. check propagates to the caller; v, err := handles at the layer that can act. There is no nil. Dependencies are declared or the program does not compile. The twenty-minute crash dies at nevla check, in milliseconds, or not at all.