Skip to content

[https://nvbugs/5606166][fix] AutoDeploy: use tuples for cudagraph shape lookup #8772

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
MartinMarciniszyn merged 1 commit into from
Oct 31, 2025
Merged

[https://nvbugs/5606166][fix] AutoDeploy: use tuples for cudagraph shape lookup #8772

Show file tree
Hide file tree
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
4 changes: 2 additions & 2 deletions tensorrt_llm/_torch/auto_deploy/compile/backends/torch_cudagraph.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,7 +124,7 @@ def capture_graph(self, *args, **kwargs):
args, kwargs = self._in_spec.unflatten(inputs_truncated + args_static)

# capture graph for truncated inputs
combined_shape = sum((input.shape for input in inputs_truncated), start=())
combined_shape = sum((tuple(input.shape) for input in inputs_truncated), start=())
self.graphs[combined_shape] = self._capture_one_graph(*args, **kwargs)

def forward(self, *args, **kwargs) -> Any:
Expand All @@ -142,7 +142,7 @@ def forward(self, *args, **kwargs) -> Any:

# Calculate rounded-up shapes for each input
rounded_shapes = [
(self.round_to_cuda_batch_size(input.shape[0]),) + input.shape[1:]
(self.round_to_cuda_batch_size(input.shape[0]),) + tuple(input.shape[1:])
for input in args_batched
]
combined_shape = sum(rounded_shapes, start=())
Expand Down
12 changes: 11 additions & 1 deletion tests/unittest/_torch/auto_deploy/unit/singlegpu/compile/test_cuda_graph_batch_sizes.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,10 @@
class TestCudaGraphBatchSizes:
"""Test class for CUDA graph batch size handling."""

@staticmethod
def _raise_error_for_forward(*args, **kwargs):
raise RuntimeError("forward method should not be called")

@pytest.fixture
def simple_model_and_inputs(self):
"""Create a simple model and inputs for testing."""
Expand Down Expand Up @@ -197,7 +201,13 @@ def test_forward_uses_cuda_graph_for_valid_batch_sizes(self, simple_model_and_in
test_input = data["input_tensor"][:batch_size]

with torch.inference_mode():
output = captured_graph.forward(test_input)
# temporarily remove model forward to ensure that the captured graph is used
original_forward = captured_graph.model.forward
captured_graph.model.forward = self._raise_error_for_forward
try:
output = captured_graph.forward(test_input)
finally:
captured_graph.model.forward = original_forward

# Should get valid output
assert output is not None
Expand Down