TypeError: unhashable type: 'Literal'

[DanChai22]

Issue Report: TypeError with Constant Multiplication in network When Using kfac_jax

Summary

When using kfac_jax with a network, introducing a scaling factor (e.g., geo_scale) for lattice parameters in the computational graph causes a TypeError related to the use of Literal. This occurs whether geo_scale is passed as a parameter or defined as a constant multiplier in the computation.

---

Steps to Reproduce

Here’s an example illustrating the issue(get_jacobian is part of the network):

1. Working Example: No Scaling
   The following works without errors:

   def get_jacobian(params):
       p_cell = params['cell'].ravel()  # No scaling applied
       return jnp.diag(p_cell)
   

2. Failing Example 1: Direct Multiplication Adding a constant multiplier to params['cell'] causes a TypeError:

   def get_jacobian(params):
       p_cell = params['cell'].ravel() * 1e-3  # Multiplying with a constant
       return jnp.diag(p_cell)
   

   Error Raised:

   TypeError: unhashable type: 'Literal'
   

3. Failing Example 2: Adding geo_scale Parameter
   Introducing a geo_scale parameter also causes the same TypeError:

   def get_jacobian(params, geo_scale=1e-3):
       p_cell = params['cell'].ravel() * geo_scale
       return jnp.diag(p_cell)
   

   Error Raised:

   TypeError: unhashable type: 'Literal'
   

---

Questions

1. Is there a recommended approach for handling constants or scaling factors like geo_scale in kfac_jax workflows to avoid such issues?

---