Skip to content

[feat] add rsLoRA support to MultiLoRA #187

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
tastelikefeet merged 3 commits into from
May 8, 2026
Merged

[feat] add rsLoRA support to MultiLoRA #187

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
61a0053
[feat] add rsLoRA support to MultiLoRA
xichengpro May 8, 2026
b446a62
Refactor scaling calculation for lora configurations
xichengpro May 8, 2026
d3a5b4c
apply yapf formatting and fix line-too-long errors
xichengpro May 8, 2026
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: 7 additions & 3 deletions src/twinkle/model/multi_lora.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -234,7 +234,8 @@ def _linear_forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
lora_A = self.lora_A[active_adapter]
lora_B = self.lora_B[active_adapter]
dropout = self.lora_dropout[_lora.adapter_name]
scaling = _lora.tenant_config.lora_alpha / _lora.tenant_config.r
scaling = _lora.tenant_config.lora_alpha / (
_lora.tenant_config.r**0.5 if _lora.tenant_config.use_rslora else _lora.tenant_config.r)
x = self._cast_input_dtype(x, lora_A.weight.dtype)
dropout_x = dropout(x)
lora_A_out = torch.nn.functional.linear(
Expand Down Expand Up @@ -266,7 +267,9 @@ def _embedding_forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:

embedding_A = self.lora_embedding_A[active_adapter]
embedding_B = self.lora_embedding_B[active_adapter]
scaling = _lora.tenant_config.lora_alpha / _lora.tenant_config.r

scaling = _lora.tenant_config.lora_alpha / (
_lora.tenant_config.r**0.5 if _lora.tenant_config.use_rslora else _lora.tenant_config.r)

embedding_A_T = embedding_A.T[:, :_lora.tenant_config.r]
embedding_B_T = embedding_B.T[:_lora.tenant_config.r, :]
Expand Down Expand Up @@ -329,7 +332,8 @@ def _megatron_forward(self, x: torch.Tensor, *args: Any, **kwargs: Any):
lora_A = self.lora_A[active_adapter]
lora_B = self.lora_B[active_adapter]
dropout = self.lora_dropout[_lora.adapter_name]
scaling = _lora.tenant_config.lora_alpha / _lora.tenant_config.r
scaling = _lora.tenant_config.lora_alpha / (
_lora.tenant_config.r**0.5 if _lora.tenant_config.use_rslora else _lora.tenant_config.r)

def _lora_A(x, *args, **kwargs):
if isinstance(lora_A, TEGroupedLinear):
Expand Down
Loading