Skip to content

add configurable unique layer init, clean up lr and loss display #64

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

Merged
merged 5 commits into from
Feb 22, 2024
Merged

add configurable unique layer init, clean up lr and loss display #64

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
267c6b7
unique layer init, clean up lr display
lessw2020 Feb 16, 2024
36ceea5
lint fixes
lessw2020 Feb 16, 2024
ae49794
fix line too long lint, clean up current loss display
lessw2020 Feb 16, 2024
c7b7bc9
Merge branch 'pytorch-labs:main' into configurable_init
lessw2020 Feb 22, 2024
91038a7
change name to depth_init
lessw2020 Feb 22, 2024
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
9 changes: 8 additions & 1 deletion torchtrain/models/llama/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,9 @@ class ModelArgs:

max_batch_size: int = 32
max_seq_len: int = 32768
depth_init: bool = (
True # initialization uses each unique layer_id or total model layer count
)


class RMSNorm(torch.nn.Module):
Expand Down Expand Up @@ -392,7 +395,11 @@ def __init__(self, layer_id: int, model_args: ModelArgs):
self.num_layers = model_args.n_layers
self.attention_norm = RMSNorm(model_args.dim, eps=model_args.norm_eps)
self.ffn_norm = RMSNorm(model_args.dim, eps=model_args.norm_eps)
self.weight_init_std = 0.02 / (2 * self.num_layers) ** 0.5

if model_args.depth_init:
self.weight_init_std = 0.02 / (2 * (self.layer_id + 1)) ** 0.5
else:
self.weight_init_std = 0.02 / (2 * self.num_layers) ** 0.5

def forward(
self,
Expand Down
15 changes: 9 additions & 6 deletions train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -207,12 +207,14 @@ def main(args):

# log metrics
if (train_state.step - 1) % args.log_freq == 0:
avg_loss, max_loss = np.mean(losses_since_last_log), np.max(
losses_since_last_log
avg_loss, max_loss = (
np.mean(losses_since_last_log),
np.max(losses_since_last_log),
)
global_avg_loss, global_max_loss = (
dist_mean(avg_loss, world_mesh),
dist_max(max_loss, world_mesh),
)
global_avg_loss, global_max_loss = dist_mean(
avg_loss, world_mesh
), dist_max(max_loss, world_mesh)

time_delta = timer() - time_last_log
wps = nwords_since_last_log / (
Expand All @@ -239,7 +241,8 @@ def main(args):
time_last_log = timer()

rank0_log(
f"step: {train_state.step}, current loss: {train_state.current_loss}, lr: {scheduler.get_last_lr()}"
f"step: {train_state.step}, current loss: {round(train_state.current_loss,4)},"
f" lr: {round(float(scheduler.get_last_lr()[0]), 8)}"
)
scheduler.step()

Expand Down