Skip to content

Prevent upcasting norm layers in prepare_model_for_kbit_training #4457

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
qgallouedec merged 5 commits into from
Nov 14, 2025
Merged

Prevent upcasting norm layers in prepare_model_for_kbit_training #4457

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
277b735
Prevent upcasting layers in prepare_model_for_kbit_training
sergiopaniego Nov 5, 2025
e44f33b
Merge branch 'main' into prevent-upcast
sergiopaniego Nov 12, 2025
ee01a7e
Merge branch 'main' into prevent-upcast
sergiopaniego Nov 13, 2025
2e3360c
Merge branch 'main' into prevent-upcast
qgallouedec Nov 13, 2025
d6edfda
Merge branch 'main' into prevent-upcast
qgallouedec Nov 13, 2025
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
10 changes: 1 addition & 9 deletions trl/models/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -485,18 +485,10 @@ def prepare_model_for_kbit_training(model, use_gradient_checkpointing=True, grad
if gradient_checkpointing_kwargs is None:
gradient_checkpointing_kwargs = {}

n_upcasted = 0
for name, param in model.named_parameters():
for _, param in model.named_parameters():
# freeze all parameters
param.requires_grad = False

# upcast LayerNorm / Norm to float32 for numerical stability
if (param.dtype in [torch.float16, torch.bfloat16]) and (
"norm" in name.lower() or "layernorm" in name.lower()
):
param.data = param.data.to(torch.float32)
n_upcasted += 1

# Enable gradient checkpointing if needed
if (loaded_in_kbit or is_quantized) and use_gradient_checkpointing:
if hasattr(model, "enable_input_require_grads"):
Expand Down
Loading