Skip to content

fix lora initialization problem. Original method leads lora-A initialization all zeros --> No training #15

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
+9 −8
Open

fix lora initialization problem. Original method leads lora-A initialization all zeros --> No training #15

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
17 changes: 9 additions & 8 deletions cosmos_predict2/_src/predict2/models/text2world_model_rectified_flow.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -184,14 +184,6 @@ def build_net(self):
with misc.timer("Creating PyTorch model"):
with torch.device(init_device):
net = lazy_instantiate(config.net)
if config.use_lora:
self.add_lora(
net,
lora_rank=config.lora_rank,
lora_alpha=config.lora_alpha,
lora_target_modules=config.lora_target_modules,
init_lora_weights=config.init_lora_weights,
)

self._param_count = count_params(net, verbose=False)

Expand All @@ -204,6 +196,15 @@ def build_net(self):
# IMPORTANT: (qsh) model init should not depends on current tensor shape, or it can handle Dtensor shape.
net.init_weights()

if config.use_lora:
self.add_lora(
Copy link

@arslananvidia arslananvidia Oct 23, 2025

Choose a reason for hiding this comment

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

Thank you for pointing it out, I agree adding loRA after weights initialization is better to preserve (random A, zero B) metrices. We will add the change in the code.

net,
lora_rank=config.lora_rank,
lora_alpha=config.lora_alpha,
lora_target_modules=config.lora_target_modules,
init_lora_weights=config.init_lora_weights,
)

if self.fsdp_device_mesh:
broadcast_dtensor_model_states(net, self.fsdp_device_mesh)
for name, param in net.named_parameters():
Expand Down