Spaces:
Running
Running
feat: handle gradient checkpointing
Browse files- src/dalle_mini/model/modeling.py +2 -2
- tools/train/train.py +23 -1
src/dalle_mini/model/modeling.py
CHANGED
@@ -144,7 +144,7 @@ class FlaxBartEncoderLayerCollection(FlaxBartEncoderLayerCollection):
|
|
144 |
|
145 |
def setup(self):
|
146 |
layer_module = (
|
147 |
-
nn.remat(FlaxBartEncoderLayer)
|
148 |
if self.config.gradient_checkpointing
|
149 |
else FlaxBartEncoderLayer
|
150 |
)
|
@@ -211,7 +211,7 @@ class FlaxBartDecoderLayerCollection(FlaxBartDecoderLayerCollection):
|
|
211 |
|
212 |
def setup(self):
|
213 |
layer_module = (
|
214 |
-
nn.remat(FlaxBartDecoderLayer)
|
215 |
if self.config.gradient_checkpointing
|
216 |
else FlaxBartDecoderLayer
|
217 |
)
|
|
|
144 |
|
145 |
def setup(self):
|
146 |
layer_module = (
|
147 |
+
nn.remat(FlaxBartEncoderLayer, concrete=True)
|
148 |
if self.config.gradient_checkpointing
|
149 |
else FlaxBartEncoderLayer
|
150 |
)
|
|
|
211 |
|
212 |
def setup(self):
|
213 |
layer_module = (
|
214 |
+
nn.remat(FlaxBartDecoderLayer, concrete=True)
|
215 |
if self.config.gradient_checkpointing
|
216 |
else FlaxBartDecoderLayer
|
217 |
)
|
tools/train/train.py
CHANGED
@@ -18,6 +18,7 @@ Training DALL·E Mini.
|
|
18 |
Script adapted from run_summarization_flax.py
|
19 |
"""
|
20 |
|
|
|
21 |
import io
|
22 |
import logging
|
23 |
import os
|
@@ -531,6 +532,8 @@ def main():
|
|
531 |
# Set up our new model config
|
532 |
if model_args.config_name:
|
533 |
config = DalleBartConfig.from_pretrained(model_args.config_name)
|
|
|
|
|
534 |
else:
|
535 |
config = None
|
536 |
|
@@ -553,8 +556,27 @@ def main():
|
|
553 |
)
|
554 |
|
555 |
# update model config per training args
|
|
|
|
|
556 |
model.config.gradient_checkpointing = training_args.gradient_checkpointing
|
557 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
558 |
# get model metadata
|
559 |
model_metadata = model_args.get_metadata()
|
560 |
|
@@ -967,7 +989,7 @@ def main():
|
|
967 |
|
968 |
def compute_eval_loss(batch):
|
969 |
batch, labels = batch.pop("labels")
|
970 |
-
logits =
|
971 |
return loss_fn(logits, labels)
|
972 |
|
973 |
# calculate loss independently per dp_device
|
|
|
18 |
Script adapted from run_summarization_flax.py
|
19 |
"""
|
20 |
|
21 |
+
import copy
|
22 |
import io
|
23 |
import logging
|
24 |
import os
|
|
|
532 |
# Set up our new model config
|
533 |
if model_args.config_name:
|
534 |
config = DalleBartConfig.from_pretrained(model_args.config_name)
|
535 |
+
# initializing params with gradient checkpointing create issues
|
536 |
+
config.gradient_checkpointing = False
|
537 |
else:
|
538 |
config = None
|
539 |
|
|
|
556 |
)
|
557 |
|
558 |
# update model config per training args
|
559 |
+
# Done after initialization of weights to avoid issues with remat
|
560 |
+
# This is still considered correctly during training as function is pjitted
|
561 |
model.config.gradient_checkpointing = training_args.gradient_checkpointing
|
562 |
|
563 |
+
# eval model cannot use remat
|
564 |
+
eval_config = copy.deepcopy(model.config)
|
565 |
+
eval_config.gradient_checkpointing = False
|
566 |
+
|
567 |
+
if training_args.gradient_checkpointing:
|
568 |
+
eval_model = DalleBart(
|
569 |
+
eval_config,
|
570 |
+
seed=training_args.seed_model,
|
571 |
+
dtype=getattr(jnp, model_args.dtype),
|
572 |
+
abstract_init=True,
|
573 |
+
load_on_cpu=True,
|
574 |
+
)
|
575 |
+
del eval_model._params
|
576 |
+
eval_fn = eval_model.__call__
|
577 |
+
else:
|
578 |
+
eval_fn = model.__call__
|
579 |
+
|
580 |
# get model metadata
|
581 |
model_metadata = model_args.get_metadata()
|
582 |
|
|
|
989 |
|
990 |
def compute_eval_loss(batch):
|
991 |
batch, labels = batch.pop("labels")
|
992 |
+
logits = eval_fn(**batch, params=state.params, train=False)[0]
|
993 |
return loss_fn(logits, labels)
|
994 |
|
995 |
# calculate loss independently per dp_device
|