Skip to content

turn off dynamic shape for torch.compile #297

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 2 commits into from
May 3, 2024
Merged

turn off dynamic shape for torch.compile #297

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b8deb9e
turn off dynamic shape for torch.compile
wanchaol May 2, 2024
88b3e15
switch 8b toml to use rmsnorm python by default
wanchaol May 2, 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
4 changes: 3 additions & 1 deletion train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,7 +230,9 @@ def loss_fn(pred, labels):
True
)
logger.info("Compiling model with torch.compile")
model = torch.compile(model)
# Dynamic shape have issues with distributed, turn dynamic off as Transformer
# training is static_shape TODO: resolve dynamic shape issue and restore defaults
model = torch.compile(model, dynamic=False)

train_state = TrainState()

Expand Down
2 changes: 1 addition & 1 deletion train_configs/llama3_8b.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,7 @@ save_tb_folder = "tb"
[model]
name = "llama3"
flavor = "8B"
norm_type = "fused_rmsnorm" # [layernorm / np_layernorm / rmsnorm / fused_rmsnorm]
norm_type = "rmsnorm" # [layernorm / np_layernorm / rmsnorm / fused_rmsnorm]
Copy link
Collaborator

Choose a reason for hiding this comment

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

Is this change intentional?

Copy link
Collaborator Author

Choose a reason for hiding this comment

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

yeah I feel it's a bit painful to switch between rmsnorm and fused_rmsnorm locally when switching between 1-D and 2-D sharding, so I changed rmsnorm to default, we can always default to fused_rmsnorm once @tianyu-l changes the fused_rmsnorm registration #296.

Let me know if you think we should default to fused_rmsnorm for 8B

Copy link
Collaborator

Choose a reason for hiding this comment

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

Keeping it as rmsnorm sounds good to me!

tokenizer_path = "./torchtitan/datasets/tokenizer/original/tokenizer.model"

[optimizer]
Expand Down