Skip to content

introduce gradient clipping #3

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
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

introduce gradient clipping #3

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
16 changes: 10 additions & 6 deletions train_gpt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -367,11 +367,11 @@ def generate(self, class_labels, max_tokens=1024, temperature=1.0, top_k=None):
@click.option(
"--global_batch_size", default=32 * 8, help="Global batch size across all GPUs"
)
@click.option("--per_gpu_batch_size", default=32, help="Per GPU batch size")
@click.option("--per_gpu_batch_size", default=16, help="Per GPU batch size")
@click.option("--num_iterations", default=6004, help="Number of training iterations")
@click.option("--learning_rate", default=1e-3, help="Learning rate")
@click.option("--learning_rate", default=3e-3, help="Learning rate")
@click.option(
"--learning_rate_embed", default=1e-2, help="Learning rate for embeddings"
"--learning_rate_embed", default=3e-3, help="Learning rate for embeddings"
)
@click.option("--weight_decay", default=0.1, help="Weight decay")
@click.option("--warmup_iters", default=10, help="Warmup iterations")
Expand All @@ -381,7 +381,10 @@ def generate(self, class_labels, max_tokens=1024, temperature=1.0, top_k=None):
@click.option("--n_embed", default=768, help="Embedding dimension")
@click.option("--init_ckpt", default=None, help="Path to initial checkpoint")
@click.option("--vres", default=False, help="Use vres")
@click.option("--n_layer", default=12, help="Number of layers")
@click.option("--n_layer", default=8, help="Number of layers")
@click.option(
"--max_grad_norm", default=1.0, help="Maximum gradient norm for clipping"
)
def train(
run_name,
train_data,
Expand All @@ -400,6 +403,7 @@ def train(
init_ckpt,
vres,
n_layer,
max_grad_norm, # Added parameter
):
dist.init_process_group(backend="nccl")
ddp_rank = int(os.environ["RANK"])
Expand Down Expand Up @@ -502,7 +506,7 @@ def train(
weight_decay=weight_decay,
fused=True,
)

# enable_cudnn_sdp(True)
enable_flash_sdp(True)
# enable_mem_efficient_sdp(False)
Expand Down Expand Up @@ -574,6 +578,7 @@ def get_lr(it):
logits, loss = model(batch["input_ids"], batch["class_label"])

loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm) # add gradient clipping

optimizer.step()
optimizer.zero_grad()
Expand Down Expand Up @@ -608,7 +613,6 @@ def get_lr(it):
if master_process and step % save_every == 0:
checkpoint = {
"model": model.module.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
"step": step,
"config": model.module.config,
Expand Down