Skip to content

[Relax][PyTorch] Fix masked_select VM build #19937

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
V-aerus wants to merge 1 commit into
base: main
Choose a base branch
from
+30 −4
Open

[Relax][PyTorch] Fix masked_select VM build #19937

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
10 changes: 10 additions & 0 deletions python/tvm/relax/frontend/torch/base_fx_graph_translator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2619,6 +2619,16 @@ def _masked_select(self, node: fx.Node) -> relax.Var:
data_flat = self.block_builder.emit(relax.op.reshape(data, [-1]))
mask_flat = self.block_builder.emit(relax.op.reshape(mask, [-1]))
indices = self.block_builder.emit(relax.op.nonzero(mask_flat))
tensor_meta = node.meta.get("tensor_meta")
if tensor_meta is not None and len(tensor_meta.shape) == 1:
num_selected = tensor_meta.shape[0]
if not isinstance(num_selected, int):
num_selected = tirx.Var(str(num_selected), "int64")
else:
num_selected = tirx.Var(f"{node.name}_num_selected", "int64")
indices = self.block_builder.match_cast(
indices, relax.TensorType([1, num_selected], "int64")
)
Comment on lines +2629 to +2631

@gemini-code-assist gemini-code-assist Bot Jul 4, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

medium

In TVM Relax, match_cast expects a StructInfo (such as TensorStructInfo) rather than a Type (such as TensorType). Additionally, relax.TensorType's constructor takes ndim and dtype rather than a shape list, so passing [1, num_selected] as the first argument is incorrect. Please use relax.TensorStructInfo instead.

Suggested change
indices = self.block_builder.match_cast(
indices, relax.TensorType([1, num_selected], "int64")
)
indices = self.block_builder.match_cast(
indices, relax.TensorStructInfo([1, num_selected], "int64")
)

@V-aerus V-aerus Jul 4, 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.

Thanks for the review. I checked the current TVM Relax API and existing code patterns. BlockBuilder.match_cast is used with relax.TensorType(shape, dtype) in the current codebase, for example in tests/python/relax/test_blockbuilder_core.py, and the ONNX frontend has a similar nonzero -> match_cast pattern using relax.TensorType([1, num_nonzero], "int64").

Also, in this TVM version, relax.TensorType accepts a shape and dtype. I don't see relax.TensorStructInfo being used in the current tree. The current code passes the targeted masked_select tests, so I will keep this as TensorType unless maintainers prefer a different API.

indices_1d = self.block_builder.emit(relax.op.squeeze(indices, axis=[0]))

result = self.block_builder.emit(relax.op.take(data_flat, indices_1d, axis=0))
Expand Down
24 changes: 20 additions & 4 deletions tests/python/relax/test_frontend_from_exported_program.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6425,16 +6425,19 @@ def main(
data: R.Tensor((2, 3), dtype="float32"), mask: R.Tensor((2, 3), dtype="bool")
) -> R.Tuple(R.Tensor(dtype="float32", ndim=1)):
R.func_attr({"tir_var_lower_bound": {"u0": 0}, "tir_var_upper_bound": {"u0": 6}})
u0 = T.int64()
with R.dataflow():
lv: R.Tensor((6,), dtype="float32") = R.reshape(data, R.shape([6]))
lv1: R.Tensor((6,), dtype="bool") = R.reshape(mask, R.shape([6]))
lv2: R.Tensor(dtype="int64", ndim=2) = R.nonzero(lv1)
lv3: R.Tensor(dtype="int64", ndim=1) = R.squeeze(lv2, axis=[0])
lv4: R.Tensor(dtype="float32", ndim=1) = R.take(lv, lv3, axis=0, mode="fast")
lv5: R.Tensor((), dtype="int64") = R.const(0, "int64")
lv3: R.Tensor((1, u0), dtype="int64") = R.match_cast(
lv2, R.Tensor((1, u0), dtype="int64")
)
lv4: R.Tensor((u0,), dtype="int64") = R.squeeze(lv3, axis=[0])
lv5: R.Tensor((u0,), dtype="float32") = R.take(lv, lv4, axis=0, mode="fast")
lv6: R.Tensor((), dtype="bool") = R.const(True, "bool")
lv7: R.Tensor((), dtype="bool") = R.const(True, "bool")
gv: R.Tuple(R.Tensor(dtype="float32", ndim=1)) = (lv4,)
gv: R.Tuple(R.Tensor((u0,), dtype="float32")) = (lv5,)
R.output(gv)
return gv

Expand All @@ -6445,6 +6448,19 @@ def main(
verify_model(MaskedSelect(), example_args, {}, Expected)


@pytest.mark.skipif(not tvm.testing.device_enabled("llvm"), reason="llvm not enabled")
def test_masked_select_numerically():
class MaskedSelect(Module):
def forward(self, data: torch.Tensor, mask: torch.Tensor):
return torch.masked_select(data, mask)

example_args = (
torch.tensor([[1, 2, 3], [4, 5, 6]], dtype=torch.float32),
torch.tensor([[True, False, True], [False, True, False]]),
)
verify_model_numerically(MaskedSelect(), example_args)


def test_new_ones():
class NewOnes(Module):
def forward(self, x):
Expand Down