@@ -270,9 +270,9 @@ def create_optimizer(
270
270
271
271
if param_group_fn :
272
272
# run custom fn to generate param groups from nn.Module
273
- parameters = param_group_fn (model_or_params )
273
+ params = param_group_fn (model_or_params )
274
274
elif layer_decay is not None :
275
- parameters = param_groups_layer_decay (
275
+ params = param_groups_layer_decay (
276
276
model_or_params ,
277
277
weight_decay = weight_decay ,
278
278
layer_decay = layer_decay ,
@@ -281,17 +281,17 @@ def create_optimizer(
281
281
)
282
282
weight_decay = 0.
283
283
elif weight_decay and weight_decay_exclude_1d :
284
- parameters = param_groups_weight_decay (
284
+ params = param_groups_weight_decay (
285
285
model_or_params ,
286
286
weight_decay = weight_decay ,
287
287
no_weight_decay_list = no_weight_decay ,
288
288
)
289
289
weight_decay = 0.
290
290
else :
291
- parameters = model_or_params .parameters ()
291
+ params = model_or_params .parameters ()
292
292
else :
293
293
# pass parameters / parameter groups through to optimizer
294
- parameters = model_or_params
294
+ params = model_or_params
295
295
296
296
# Parse optimizer name
297
297
opt_split = opt .lower ().split ('_' )
@@ -330,7 +330,7 @@ def create_optimizer(
330
330
331
331
# Create optimizer
332
332
opt_class = self .get_optimizer_class (opt_info , bind_defaults = False )
333
- optimizer = opt_class (parameters , ** opt_args )
333
+ optimizer = opt_class (params , ** opt_args )
334
334
335
335
# Apply Lookahead if requested
336
336
if use_lookahead :
@@ -685,12 +685,14 @@ def _register_bnb_optimizers(registry: OptimizerRegistry) -> None:
685
685
'bnblion' ,
686
686
'bitsandbytes.optim.Lion' ,
687
687
description = 'bitsandbytes Lion' ,
688
+ has_eps = False ,
688
689
has_betas = True
689
690
),
690
691
OptimInfo (
691
692
'bnblion8bit' ,
692
693
'bitsandbytes.optim.Lion8bit' ,
693
694
description = 'bitsandbytes 8-bit Lion with dynamic quantization' ,
695
+ has_eps = False ,
694
696
has_betas = True
695
697
),
696
698
OptimInfo (
0 commit comments