ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

gradient checkpointing

Browse source
This commit is contained in:
Daniel Han 2025-06-29 03:19:20 -07:00
commit 603a9863ed
2 changed files with 26 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

21
unsloth/models/llama.py
View file

@ -909,12 +909,7 @@ def LlamaModel_fast_forward(
mask = self. GA_mask if use_static_mask else dynamic_GA_mask
pass
try:
is_gradient_checkpointing_layer = isinstance(decoder_layer, GradientCheckpointingLayer)
except:
is_gradient_checkpointing_layer = False
if gradient_checkpointing and not is_gradient_checkpointing_layer:
if gradient_checkpointing and not isinstance(decoder_layer, GradientCheckpointingLayer):
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs, past_key_value, output_attentions, padding_mask = padding_mask, position_embeddings = position_embeddings)
@ -2019,7 +2014,7 @@ class FastLlamaModel:
f" {chr(92)}{chr(92)} /| Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\n"\\
f"O^O/ {chr(92)}_/ {chr(92)} Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\n"\\
f"{chr(92)} / Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size:,}\\n"\\
f' "-____-" Trainable parameters = {get_model_param_count(model, trainable_only=True):,}/{get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_param_count(model)*100:.2f}% trained)'
f' "-____-" Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_param_count(model)*100:.2f}% trained)'
logger.warning(debug_info)
import gc
for _ in range(3):
@ -2842,6 +2837,12 @@ class FastLlamaModel:
m = m.model
_for_inference(m)
# Since transformers 4.53, must turn off explicitly
for module in model.modules():
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = False
pass
# Also disable training for embeddings for NEFTune
if hasattr(model, "get_input_embeddings"):
embeddings = model.get_input_embeddings()
@ -2880,6 +2881,12 @@ class FastLlamaModel:
m = m.model
_for_training(m)
# Since transformers 4.53, must turn on explicitly
for module in model.modules():
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = use_gradient_checkpointing
pass
# Also re-enable training for embeddings for NEFTune
if hasattr(model, "get_input_embeddings"):
embeddings = model.get_input_embeddings()

12
unsloth/models/vision.py
View file

@ -713,6 +713,12 @@ class FastBaseModel:
m = m.model
_for_inference(m)
# Since transformers 4.53, must turn off explicitly
for module in model.modules():
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = False
pass
# Also disable training for embeddings for NEFTune
if hasattr(model, "get_input_embeddings"):
embeddings = model.get_input_embeddings()
@ -755,6 +761,12 @@ class FastBaseModel:
m = m.model
_for_training(m)
# Since transformers 4.53, must turn on explicitly
for module in model.modules():
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = True
pass
# Also re-enable training for embeddings for NEFTune
if hasattr(model, "get_input_embeddings"):
embeddings = model.get_input_embeddings()