Fix OptimizerRegistry.create_optimizer crash on layer_decay without an explicit min scale - #2754

What

Minor, but a hard crash: calling the registry's create_optimizer with layer_decay set and the min scale left at its default raises a TypeError.

from timm.optim._optim_factory import default_registry
model = torch.nn.Sequential(torch.nn.Linear(10, 5), torch.nn.ReLU(), torch.nn.Linear(5, 2))

default_registry.create_optimizer(model, 'adamw', lr=1e-3, weight_decay=0.05, layer_decay=0.75)
# TypeError: '>' not supported between instances of 'float' and 'NoneType'

The common entry point create_optimizer_v2 is not affected — it passes 0.0.

Why

OptimizerRegistry.create_optimizer defaults layer_decay_min_scale to None and forwards it unguarded as min_scale into param_groups_layer_decay, whose scale line is:

layer_scales = list(max(min_scale, layer_decay ** (layer_max - i)) for i in range(num_layers))

max(None, <float>) raises in Python 3. Both create_optimizer_v2 (layer_decay_min_scale: float = 0.0) and param_groups_layer_decay itself (min_scale: float = 0.) already default this to 0.0; only the registry method defaulted it to None.

Fix

Default layer_decay_min_scale to 0.0 in the registry method, matching its siblings. 0.0 is a no-op clamp (the scales are already >= 0), so this only removes the crash — it doesn't change behavior for callers that passed a value.

layer_decay_no_opt_scale keeps its Optional[float] = None default on purpose: it's consumed behind a truthiness guard (if no_opt_scale and ...), so None is safe there. Verified that the same call with only no_opt_scale still crashes on min_scale, confirming min_scale is the sole culprit.

Tests

Added test_registry_create_optimizer_layer_decay_default_min_scale, which builds layer-decay param groups through the registry path with the min scale omitted. It fails on main (the TypeError) and passes with the fix; the related layer-decay/param-group tests stay green.