Closed aksuhton closed 4 months ago
That PR will fix it.
1 side note
model = @compact(; potential=Dense(N => N, gelu)) do x
jac_pot = batched_jacobian(potential, AutoForwardDiff(), x)
return jac_pot, potential.st
end
should be
model = @compact(; potential=Dense(N => N, gelu)) do x
jac_pot = batched_jacobian(potential, AutoForwardDiff(), x)
return jac_pot
end
Thank you for the fix and the tip!
I'm getting an InvalidIRError when using batched_jacobian with CUDA if the feature array takes on certain sizes. Below is both an MWE and stacktrace. Thank you again for working with me so much on Jacobians!
MWE:
Stacktrace: