Skip to content

Fix LoraEmbedding casting outputs to the token ID dtype #67

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

Closed
Robertboy18 wants to merge 2 commits into from
+15 −2
Closed

Fix LoraEmbedding casting outputs to the token ID dtype #67

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
de8b4e2
Fix LoraEmbedding output dtype
Robertboy18 Jul 14, 2026
94240e9
Clarify LoraEmbedding output dtype variable
Robertboy18 Jul 16, 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
4 changes: 2 additions & 2 deletions src/neuronx_distributed/modules/lora/layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -314,8 +314,8 @@ def _embed(self, input: torch.Tensor, weight: torch.Tensor) -> torch.Tensor:
)

def forward(self, x: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor:
previous_dtype = x.dtype
result = self.base_layer(x, *args, **kwargs)
output_dtype = result.dtype
if not self.merged:
Comment on lines 316 to 319

@Robertboy18 Robertboy18 Jul 16, 2026

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

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

Good point. Renamed it to output_dtype in 94240e9 to make the intent explicit.

assert self.lora_embedding_A is not None and self.lora_embedding_B is not None
embedding_A = self.lora_embedding_A.T
Expand All @@ -324,7 +324,7 @@ def forward(self, x: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor:
after_A = self._embed(x, embedding_A)
result += (after_A @ embedding_B) * scaling

result = result.to(previous_dtype)
result = result.to(output_dtype)
return result


Expand Down
13 changes: 13 additions & 0 deletions test/unit_test/modules/lora/test_lora_layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,6 +71,19 @@ def test_torch_embedding_layer(self):
layer_str = str(lora_layer)
assert "lora" in layer_str

def test_torch_embedding_forward_preserves_output_dtype(self):
layer = torch.nn.Embedding(8, 3, dtype=torch.float32)
with torch.no_grad():
layer.weight[0] = torch.tensor([0.25, 0.5, 0.75])
lora_layer = LoraEmbedding(layer, get_lora_config())
input_ids = torch.tensor([0], dtype=torch.long)

expected = layer(input_ids)
actual = lora_layer(input_ids)

self.assertEqual(actual.dtype, expected.dtype)
torch.testing.assert_close(actual, expected)

@patch("neuronx_distributed.parallel_layers.layers.get_tensor_model_parallel_size", MagicMock(return_value=8))
@patch("neuronx_distributed.parallel_layers.layers.get_tensor_model_parallel_rank", MagicMock(return_value=1))
@patch("neuronx_distributed.parallel_layers.parallel_state.initialize_model_parallel", MagicMock(return_value=True))
Expand Down