Skip to content

[Relax][PyTorch] Simplify tensor args conversion in Dynamo #18726

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
mshr-h wants to merge 3 commits into
base: main
Choose a base branch
from
+4 −9
Open

[Relax][PyTorch] Simplify tensor args conversion in Dynamo #18726

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
194cc58
use dlpack to convert torch tensor to tvm tensor
mshr-h Feb 8, 2026
89ca79b
simplify torch tensor to tvm tensor conversion
mshr-h Feb 8, 2026
7fadb76
remove
mshr-h Feb 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
13 changes: 4 additions & 9 deletions python/tvm/relax/frontend/torch/dynamo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,14 +64,6 @@ def to_torch_tensor(nd_tensor):
else:
raise ValueError(f"Unsupported type {type(nd_tensor)}")

def to_tvm_tensor(torch_tensor):
"""A helper function to transfer a torch.tensor to Tensor."""
if not isinstance(torch_tensor, torch._subclasses.fake_tensor.FakeTensor):
return tvm.runtime.tensor(torch_tensor.numpy())
# Fake Tensor
real_tensor = torch.randn(torch_tensor.shape, dtype=torch_tensor.dtype)
return tvm.runtime.tensor(real_tensor.numpy())

graph_module.graph.eliminate_dead_code()

device = device_from_inputs(example_inputs)
Expand Down Expand Up @@ -139,7 +131,10 @@ def exec_tvm(*i_args):
for arg in args:
if arg.requires_grad:
arg = arg.detach()
vm_args.append(to_tvm_tensor(arg))
if isinstance(arg, torch._subclasses.fake_tensor.FakeTensor):
# Materialize a real (eager) Tensor
arg = torch.randn(arg.shape, dtype=arg.dtype, device=device)
vm_args.append(arg)
outputs = vm["main"](*vm_args)
return to_torch_tensor(outputs)

Expand Down
Loading