A RuntimeError occur when call _Jac.fullmatrix()

[xitorch]

A RuntimeError occur when call _Jac.fullmatrix()

[mfkasim1]

Hi, thanks for raising the issue. This type of bugs usually occurs when a dependency on a variable disappear in the first grad. For example, if the function is f = a*x+b, the derivative df/da=x which does not depend on a nor b. In that case, taking the second derivative w.r.t. a or b will raise that error. The fix from the xitorch side is just to set allow_unused=False, but you can also do a workaround by adding more dependency on the vars, e.g., f = a*x + b + a^2*0 + b^2*0

[floatingCatty]

Hello, thanks for your help, but I noticed that here dfdy somehow is related to V, since variables V in xitorch is used to calculated the dfdy, as realized in jachess:
Here for dfdy,
with torch.enable_grad():
yout = fcn(*params) # (*nout)
v = torch.ones_like(yout).to(yout.device).requires_grad_() # (*nout)
dfdy, = torch.autograd.grad(yout, (yparam,), grad_outputs=v, create_graph=True) # (*nin)

and for grad to v.
for i in range(nbatch):
dfdyf, = torch.autograd.grad(dfdy, (v,), grad_outputs=gy1[i].reshape(self.inshape),
retain_graph=True, create_graph=torch.is_grad_enabled()) # (*nout)
I cannot figure out why here the second derivative to v is unavailable. Since v is not explicit available to fcn function, adding dependencies seems difficult to conduct. Is there anything I can do for this?

Thanks again for helping!