error with ReverseSplitWithPrimal and Flux.RNN
- Dominant language
- Julia
- Stars
- 586
- Forks
- 108
- Avg merge
- 1d 5h
- Merged PRs (30d)
- 44
Description
on julia 1.10:
```julia
using Flux, Enzyme, Statistics
loss(m, x...) = mean(m(x...))
m = RNN(3 => 2, tanh)
dm = Duplicated(m, Enzyme.make_zero(m))
x = randn(Float32, 3, 2, 2)
autodiff(Reverse, Const(loss), Active, dm, Const(x)) # OK
autodiff(ReverseWithPrimal, Const(loss), Active, dm, Const(x)) # OK
# ERROR
forward, reverse = autodiff_thunk(ReverseSplitWithPrimal, Const{typeof(loss)}, Active, typeof(dm), typeof(Const(x)))
```
produces a very long error with first lines
```
ERROR: AssertionError: ; Function Attrs: alwaysinline mustprogress
define internal "enzymejl_parmtype"="4844532784" "enzymejl_parmtype_ref"="0" void @diffejulia_loss_84364_inner.1([1 x { {}
```
Contributor guide
Assessment
This issue has not been assessed yet.