Skip to content

Revert "fix: preserve batch dimension in tokenizer and predictor training" #381

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

Open
7l7a7l7o wants to merge 1 commit into
base: master
Choose a base branch
from
+14 −14
Open

Revert "fix: preserve batch dimension in tokenizer and predictor training" #381

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
14 changes: 7 additions & 7 deletions finetune/train_predictor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,9 +92,9 @@ def train_model(model, tokenizer, device, config, save_dir, logger, rank, world_
train_dataset.set_epoch_seed(epoch_idx * 10000 + rank)
valid_dataset.set_epoch_seed(0)

for i, (batch_x, batch_x_stamp) in enumerate(train_loader):
batch_x = batch_x.to(device, non_blocking=True)
batch_x_stamp = batch_x_stamp.to(device, non_blocking=True)
for i, (batch_x, batch_x_stamp) in enumerate(train_loader):
batch_x = batch_x.squeeze(0).to(device, non_blocking=True)
batch_x_stamp = batch_x_stamp.squeeze(0).to(device, non_blocking=True)

# Tokenize input data on-the-fly
with torch.no_grad():
Expand Down Expand Up @@ -135,10 +135,10 @@ def train_model(model, tokenizer, device, config, save_dir, logger, rank, world_
model.eval()
tot_val_loss_sum_rank = 0.0
val_batches_processed_rank = 0
with torch.no_grad():
for batch_x, batch_x_stamp in val_loader:
batch_x = batch_x.to(device, non_blocking=True)
batch_x_stamp = batch_x_stamp.to(device, non_blocking=True)
with torch.no_grad():
for batch_x, batch_x_stamp in val_loader:
batch_x = batch_x.squeeze(0).to(device, non_blocking=True)
batch_x_stamp = batch_x_stamp.squeeze(0).to(device, non_blocking=True)

token_seq_0, token_seq_1 = tokenizer.encode(batch_x, half=True)
token_in = [token_seq_0[:, :-1], token_seq_1[:, :-1]]
Expand Down
10 changes: 5 additions & 5 deletions finetune/train_tokenizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,8 +123,8 @@ def train_model(model, device, config, save_dir, logger, rank, world_size):
train_dataset.set_epoch_seed(epoch_idx * 10000 + rank)
valid_dataset.set_epoch_seed(0) # Keep validation sampling consistent

for i, (ori_batch_x, _) in enumerate(train_loader):
ori_batch_x = ori_batch_x.to(device, non_blocking=True)
for i, (ori_batch_x, _) in enumerate(train_loader):
ori_batch_x = ori_batch_x.squeeze(0).to(device, non_blocking=True)

# --- Gradient Accumulation Loop ---
current_batch_total_loss = 0.0
Expand Down Expand Up @@ -174,9 +174,9 @@ def train_model(model, device, config, save_dir, logger, rank, world_size):
model.eval()
tot_val_loss_sum_rank = 0.0
val_sample_count_rank = 0
with torch.no_grad():
for ori_batch_x, _ in val_loader:
ori_batch_x = ori_batch_x.to(device, non_blocking=True)
with torch.no_grad():
for ori_batch_x, _ in val_loader:
ori_batch_x = ori_batch_x.squeeze(0).to(device, non_blocking=True)
zs, _, _, _ = model(ori_batch_x)
_, z = zs
val_loss_item = F.mse_loss(z, ori_batch_x)
Expand Down
4 changes: 2 additions & 2 deletions finetune_csv/finetune_tokenizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,7 +190,7 @@ def train_tokenizer(model, device, config, save_dir, logger):
train_sampler.set_epoch(epoch)

for batch_idx, (ori_batch_x, _) in enumerate(train_loader):
ori_batch_x = ori_batch_x.to(device, non_blocking=True)
ori_batch_x = ori_batch_x.squeeze(0).to(device, non_blocking=True)

current_batch_total_loss = 0.0
for j in range(accumulation_steps):
Expand Down Expand Up @@ -239,7 +239,7 @@ def train_tokenizer(model, device, config, save_dir, logger):

with torch.no_grad():
for ori_batch_x, _ in val_loader:
ori_batch_x = ori_batch_x.to(device, non_blocking=True)
ori_batch_x = ori_batch_x.squeeze(0).to(device, non_blocking=True)
zs, _, _, _ = (model.module if use_ddp else model)(ori_batch_x)
_, z = zs
val_loss_item = F.mse_loss(z, ori_batch_x)
Expand Down