Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add more rigerous non-slow grad accum tests #35668

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Add more rigerous non-slow grad accum tests #35668

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
776758b
Add more rigerous non-slow grad accum tests
muellerzr Jan 13, 2025
7306624
Further nits
muellerzr Jan 13, 2025
b0be2ed
Re-add space
muellerzr Jan 13, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 21 additions & 22 deletions tests/trainer/test_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -762,35 +762,34 @@ def test_model_init(self):
trainer.train()
self.check_trained_model(trainer.model, alternate_seed=True)

@slow
def test_gradient_accumulation_loss_alignment_with_model_loss(self):
set_seed(42)
import datasets

model_name = "nickypro/tinyllama-110M"
model_name = "nickypro/tinyllama-15M"
dataset_name = "wikitext"
dataset_config = "wikitext-2-raw-v1"
dataset = datasets.load_dataset(dataset_name, dataset_config, split="train[:500]")
dataset = dataset.train_test_split(test_size=0.2)
dataset = datasets.load_dataset(dataset_name, dataset_config, split="train[:40]")
tokenizer = AutoTokenizer.from_pretrained(model_name)

tokenizer.pad_token = tokenizer.eos_token

def tokenize_function(examples):
return tokenizer(examples["text"], max_length=128, padding="max_length", truncation=True)
return tokenizer(examples["text"], max_length=16, padding="max_length", truncation=True)

tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset["train"].column_names)
tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset.column_names)

data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)

model = AutoModelForCausalLM.from_pretrained(model_name)
state_dict = model.state_dict()

base_loss_callback = StoreLossCallback()

args_kwargs = {
"report_to": "none",
"logging_steps": 1,
"max_steps": 20,
"max_steps": 5,
"learning_rate": 3e-4,
"disable_tqdm": True,
}
Expand All @@ -803,7 +802,7 @@ def tokenize_function(examples):
trainer = Trainer(
model,
args,
train_dataset=tokenized_dataset["train"],
train_dataset=tokenized_dataset,
callbacks=[base_loss_callback],
data_collator=data_collator,
)
Expand All @@ -823,19 +822,19 @@ def tokenize_function(examples):
trainer = Trainer(
model,
args,
train_dataset=tokenized_dataset["train"],
train_dataset=tokenized_dataset,
callbacks=[grad_accum_loss_callback],
data_collator=data_collator,
)
trainer.train()

set_seed(42)
model = AutoModelForCausalLM.from_pretrained(model_name)
model.load_state_dict(state_dict)
broken_loss_callback = StoreLossCallback()
trainer = Trainer(
model,
args,
train_dataset=tokenized_dataset["train"],
train_dataset=tokenized_dataset,
callbacks=[broken_loss_callback],
data_collator=data_collator,
)
Expand All @@ -855,24 +854,24 @@ def tokenize_function(examples):
self.assertLess(max(diff_truth), 0.01, f"Difference {max(diff_truth)} is not within 0.01")

# max diff broken should be very off
self.assertGreater(max(diff_broken), 3, f"Difference {max(diff_broken)} is not greater than 3")
self.assertGreater(max(diff_broken), 2, f"Difference {max(diff_broken)} is not greater than 2")

@slow
def test_gradient_accumulation_loss_alignment_with_loss_func(self):
set_seed(42)
import datasets

model_name = "roneneldan/TinyStories-33M"
model_name = "nickypro/tinyllama-15M"
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

When testing compute_loss_func, I think it's better to use a model that doesn't accept loss kwargs. Maybe TinyStories-33M is ok?"

dataset_name = "wikitext"
dataset_config = "wikitext-2-raw-v1"
dataset = datasets.load_dataset(dataset_name, dataset_config, split="train[:500]")
dataset = dataset.train_test_split(test_size=0.2)
dataset = datasets.load_dataset(dataset_name, dataset_config, split="train[:40]")
tokenizer = AutoTokenizer.from_pretrained(model_name)

tokenizer.pad_token = tokenizer.eos_token

def tokenize_function(examples):
return tokenizer(examples["text"])
return tokenizer(examples["text"], max_length=16, padding="max_length", truncation=True)

tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset["train"].column_names)
tokenized_dataset = dataset.map(tokenize_function, batched=True)

tokenizer.pad_token = tokenizer.eos_token
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
Expand All @@ -891,7 +890,7 @@ def compute_loss(logits, labels, vocab_size, num_items_in_batch, disable_num_ite
args_kwargs = {
"report_to": "none",
"logging_steps": 1,
"max_steps": 20,
"max_steps": 5,
"learning_rate": 3e-4,
"disable_tqdm": True,
}
Expand All @@ -904,7 +903,7 @@ def compute_loss(logits, labels, vocab_size, num_items_in_batch, disable_num_ite
trainer = Trainer(
model,
args,
train_dataset=tokenized_dataset["train"],
train_dataset=tokenized_dataset,
callbacks=[base_loss_callback],
compute_loss_func=loss_fn,
data_collator=data_collator,
Expand All @@ -924,7 +923,7 @@ def compute_loss(logits, labels, vocab_size, num_items_in_batch, disable_num_ite
trainer = Trainer(
model,
args,
train_dataset=tokenized_dataset["train"],
train_dataset=tokenized_dataset,
callbacks=[grad_accum_loss_callback],
compute_loss_func=loss_fn,
data_collator=data_collator,
Expand All @@ -938,7 +937,7 @@ def compute_loss(logits, labels, vocab_size, num_items_in_batch, disable_num_ite
trainer = Trainer(
model,
args,
train_dataset=tokenized_dataset["train"],
train_dataset=tokenized_dataset,
callbacks=[broken_loss_callback],
compute_loss_func=loss_fn,
data_collator=data_collator,
Expand Down
Loading