stop_mode='rel' seems to be ignored (solver iterates until 1e-5 despite 5% tolerance)

[QuietNoviceCoder]

Hi there,

I am encountering an issue where the solver does not seem to respect the stop_mode='rel' configuration.

I have configured get_deq with a relative tolerance of 0.05 (5%), aiming for a looser stop condition to speed up training. However, looking at the logs, the solver continues to iterate way past this threshold, only stopping when the error reaches the 1e-5 level (which looks like a default absolute tolerance).

Here is my configuration:

get_deq(
ift = True,
f_solver=deq_solver, # e.g., 'anderson'
b_solver=deq_solver,
f_max_iter=deq_f_max_iter,
b_max_iter=deq_b_max_iter,
f_tol=0.05, # Expecting stop at 5% relative error
b_tol=0.05,
stop_mode='rel',
f_anderson_m=deq_anderson_m if deq_solver == 'anderson' else None,
b_anderson_m=deq_anderson_m if deq_solver == 'anderson' else None,
)
Observed Behavior: Even though f_tol is set to 0.05 with stop_mode='rel', the solver does not terminate early. It keeps running until the residual drops to approximately 1e-5.