Skip to content

[Relax][PyTorch] Add support for decomposed operators and fix IR of ops tests(2) #18403

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
mshr-h merged 2 commits into from
Oct 30, 2025
Merged

[Relax][PyTorch] Add support for decomposed operators and fix IR of ops tests(2) #18403

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
6ab6067
finish1
tlopex Oct 29, 2025
07bed4e
finish2
tlopex Oct 29, 2025
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/exported_program_translator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -760,6 +760,14 @@ def _zeros(self, node: fx.Node) -> relax.Var:
)
return self.block_builder.emit(relax.op.zeros(size, dtype))

def _scalar_tensor(self, node: fx.Node) -> relax.Var:
args = self.retrieve_args(node)
scalar_value = args[0]
dtype = self._convert_data_type(
node.kwargs.get("dtype", torch.get_default_dtype()), self.env
)
return self.block_builder.emit(relax.const(scalar_value, dtype))

def _instance_norm(self, node: fx.Node):
import numpy as np

Expand Down Expand Up @@ -851,6 +859,7 @@ def create_convert_map(
"relu6_.default": self._unary_op(relax.op.nn.relu6),
"round.default": self._round,
"rsqrt.default": self._unary_op(relax.op.rsqrt),
"scalar_tensor.default": self._scalar_tensor,
"rsub.Tensor": self._rsub,
"rsub.Scalar": self._rsub,
"selu.default": self._unary_op(relax.op.nn.selu),
Expand All @@ -861,6 +870,7 @@ def create_convert_map(
"sin.default": self._unary_op(relax.op.sin),
"sinh.default": self._unary_op(relax.op.sinh),
"softmax.int": self._softmax,
"_softmax.default": self._softmax,
"softplus.default": self._softplus,
"softshrink.default": self._softshrink,
"softsign.default": self._softsign,
Expand Down
112 changes: 69 additions & 43 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 @@ -1103,8 +1103,8 @@ def main(
return gv

example_args = (torch.randn(1, 3, 10, 10, dtype=torch.float32),)
verify_model(Softmax(), example_args, {}, expected1)
verify_model(Softmax2(), example_args, {}, expected1)
verify_model(Softmax(), example_args, {}, expected1, run_ep_decomposition=True)
verify_model(Softmax2(), example_args, {}, expected1, run_ep_decomposition=True)


def test_softsign():
Expand Down Expand Up @@ -1135,8 +1135,8 @@ def main(
return gv

example_args = (torch.randn(1, 3, 10, 10, dtype=torch.float32),)
verify_model(Softsign(), example_args, {}, expected_softsign)
verify_model(Softsign2(), example_args, {}, expected_softsign)
verify_model(Softsign(), example_args, {}, expected_softsign, run_ep_decomposition=True)
verify_model(Softsign2(), example_args, {}, expected_softsign, run_ep_decomposition=True)


def test_softshrink():
Expand All @@ -1159,32 +1159,24 @@ def main(
input: R.Tensor((1, 3, 10, 10), dtype="float32"),
) -> R.Tuple(R.Tensor((1, 3, 10, 10), dtype="float32")):
with R.dataflow():
lv: R.Tensor((1, 3, 10, 10), dtype="float32") = R.subtract(
input, R.const(0.5, "float32")
)
lv1: R.Tensor((1, 3, 10, 10), dtype="bool") = R.greater(
input, R.const(0.5, "float32")
lv: R.Tensor((1, 3, 10, 10), dtype="float32") = R.abs(input)
lv1: R.Tensor((1, 3, 10, 10), dtype="bool") = R.greater(lv, R.const(0.5, "float32"))
lv2: R.Tensor((1, 3, 10, 10), dtype="float32") = R.sign(input)
lv3: R.Tensor((1, 3, 10, 10), dtype="float32") = R.multiply(
lv2, R.const(0.5, "float32")
)
lv2: R.Tensor((1, 3, 10, 10), dtype="float32") = R.astype(lv1, "float32")
lv3: R.Tensor((1, 3, 10, 10), dtype="float32") = R.multiply(lv, lv2)

lv4: R.Tensor((1, 3, 10, 10), dtype="float32") = R.add(
input, R.const(0.5, "float32")
lv4: R.Tensor((1, 3, 10, 10), dtype="float32") = R.subtract(input, lv3)
lv5: R.Tensor((1, 3, 10, 10), dtype="float32") = R.multiply(
input, R.const(0.0, "float32")
)
Comment on lines +1169 to 1171
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Oct 29, 2025

Choose a reason for hiding this comment

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

medium

Using R.zeros_like(input) would be more idiomatic and clearer to express the intent of creating a zero tensor with the same shape and dtype as input.

                lv5: R.Tensor((1, 3, 10, 10), dtype="float32") = R.zeros_like(input)

lv5: R.Tensor((), dtype="float32") = R.negative(R.const(0.5, "float32"))
lv6: R.Tensor((1, 3, 10, 10), dtype="bool") = R.less(input, lv5)
lv7: R.Tensor((1, 3, 10, 10), dtype="float32") = R.astype(lv6, "float32")
lv8: R.Tensor((1, 3, 10, 10), dtype="float32") = R.multiply(lv4, lv7)

lv9: R.Tensor((1, 3, 10, 10), dtype="float32") = R.add(lv3, lv8)

gv: R.Tuple(R.Tensor((1, 3, 10, 10), dtype="float32")) = (lv9,)
lv6: R.Tensor((1, 3, 10, 10), dtype="float32") = R.where(lv1, lv4, lv5)
gv: R.Tuple(R.Tensor((1, 3, 10, 10), dtype="float32")) = (lv6,)
R.output(gv)
return gv

example_args = (torch.randn(1, 3, 10, 10, dtype=torch.float32),)
verify_model(Softshrink(), example_args, {}, expected_softshrink)
verify_model(Softshrink2(), example_args, {}, expected_softshrink)
verify_model(Softshrink(), example_args, {}, expected_softshrink, run_ep_decomposition=True)
verify_model(Softshrink2(), example_args, {}, expected_softshrink, run_ep_decomposition=True)


def test_tril_triu():
Expand All @@ -1198,16 +1190,27 @@ def forward(self, input):
class expected_tril:
@R.function
def main(
input_1: R.Tensor((10, 10), dtype="float32")
input: R.Tensor((10, 10), dtype="float32")
) -> R.Tuple(R.Tensor((10, 10), dtype="float32")):
# block 0
with R.dataflow():
lv: R.Tensor((10, 10), dtype="float32") = R.tril(input_1, 1)
gv: R.Tuple(R.Tensor((10, 10), dtype="float32")) = (lv,)
lv: R.Tensor((10,), dtype="int64") = R.arange(
R.prim_value(0), R.prim_value(10), R.prim_value(1), dtype="int64"
)
lv1: R.Tensor((1, 10), dtype="int64") = R.expand_dims(lv, axis=[-2])
lv2: R.Tensor((10,), dtype="int64") = R.arange(
R.prim_value(0), R.prim_value(10), R.prim_value(1), dtype="int64"
)
lv3: R.Tensor((10, 1), dtype="int64") = R.expand_dims(lv2, axis=[-1])
Comment on lines +1201 to +1204
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Oct 29, 2025

Choose a reason for hiding this comment

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

medium

The arange call to create lv2 is redundant, as lv created on line 1197 is identical. You can reuse lv to create lv3 and remove the definition of lv2 to make the IR more concise and avoid a redundant operation. This same issue is present in expected_triu as well.

                lv3: R.Tensor((10, 1), dtype="int64") = R.expand_dims(lv, axis=[-1])

lv4: R.Tensor((10, 10), dtype="int64") = R.subtract(lv1, lv3)
lv5: R.Tensor((10, 10), dtype="bool") = R.less_equal(lv4, R.const(1, "int64"))
lv6: R.Tensor((), dtype="float32") = R.const(0.0, "float32")
lv7: R.Tensor((10, 10), dtype="float32") = R.where(lv5, input, lv6)
gv: R.Tuple(R.Tensor((10, 10), dtype="float32")) = (lv7,)
R.output(gv)
return gv

verify_model(Tril(), example_args, {}, expected_tril)
verify_model(Tril(), example_args, {}, expected_tril, run_ep_decomposition=True)

class Triu(Module):
def forward(self, input):
Expand All @@ -1217,16 +1220,27 @@ def forward(self, input):
class expected_triu:
@R.function
def main(
input_1: R.Tensor((10, 10), dtype="float32")
input: R.Tensor((10, 10), dtype="float32")
) -> R.Tuple(R.Tensor((10, 10), dtype="float32")):
# block 0
with R.dataflow():
lv: R.Tensor((10, 10), dtype="float32") = R.triu(input_1, 1)
gv: R.Tuple(R.Tensor((10, 10), dtype="float32")) = (lv,)
lv: R.Tensor((10,), dtype="int64") = R.arange(
R.prim_value(0), R.prim_value(10), R.prim_value(1), dtype="int64"
)
lv1: R.Tensor((1, 10), dtype="int64") = R.expand_dims(lv, axis=[-2])
lv2: R.Tensor((10,), dtype="int64") = R.arange(
R.prim_value(0), R.prim_value(10), R.prim_value(1), dtype="int64"
)
lv3: R.Tensor((10, 1), dtype="int64") = R.expand_dims(lv2, axis=[-1])
Comment on lines +1231 to +1234
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Oct 29, 2025

Choose a reason for hiding this comment

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

medium

Similar to expected_tril, the arange call to create lv2 is redundant. You can reuse lv (defined on line 1227) to create lv3 and remove the definition of lv2.

                lv3: R.Tensor((10, 1), dtype="int64") = R.expand_dims(lv, axis=[-1])

lv4: R.Tensor((10, 10), dtype="int64") = R.subtract(lv1, lv3)
lv5: R.Tensor((10, 10), dtype="bool") = R.greater_equal(lv4, R.const(1, "int64"))
lv6: R.Tensor((), dtype="float32") = R.const(0.0, "float32")
lv7: R.Tensor((10, 10), dtype="float32") = R.where(lv5, input, lv6)
gv: R.Tuple(R.Tensor((10, 10), dtype="float32")) = (lv7,)
R.output(gv)
return gv

verify_model(Triu(), example_args, {}, expected_triu)
verify_model(Triu(), example_args, {}, expected_triu, run_ep_decomposition=True)


operator_binary_1 = [
Expand Down Expand Up @@ -1501,7 +1515,7 @@ def main(
torch.randn(64, 64, dtype=torch.float32),
torch.randn(64, dtype=torch.float32),
)
verify_model(DivModel(), example_args, {}, expected_div)
verify_model(DivModel(), example_args, {}, expected_div, run_ep_decomposition=True)

# Case 2: Division with trunc rounding
class DivTruncModel(torch.nn.Module):
Expand All @@ -1521,7 +1535,7 @@ def main(
R.output(gv)
return gv

verify_model(DivTruncModel(), example_args, {}, expected_div_trunc)
verify_model(DivTruncModel(), example_args, {}, expected_div_trunc, run_ep_decomposition=True)

# Case 3: Division with floor rounding
class DivFloorModel(torch.nn.Module):
Expand All @@ -1540,7 +1554,7 @@ def main(
R.output(gv)
return gv

verify_model(DivFloorModel(), example_args, {}, expected_div_floor)
verify_model(DivFloorModel(), example_args, {}, expected_div_floor, run_ep_decomposition=True)


def test_batchnorm2d():
Expand Down Expand Up @@ -1578,6 +1592,8 @@ def main(
epsilon=1e-05,
center=True,
scale=True,
momentum=1e-05,
training=False,
)
lv1: R.Tensor((1, 3, 10, 10), dtype="float32") = lv[0]
gv: R.Tuple(R.Tensor((1, 3, 10, 10), dtype="float32")) = (lv1,)
Expand All @@ -1593,7 +1609,7 @@ def main(
"w3": model.bn.running_mean.detach().numpy(),
"w4": model.bn.running_var.detach().numpy(),
}
verify_model(model, example_args, binding, expected1)
verify_model(model, example_args, binding, expected1, run_ep_decomposition=True)


def test_adaptive_avgpool1d():
Expand Down Expand Up @@ -1748,8 +1764,8 @@ def main(
torch.randn(10, 10, dtype=torch.float32),
)

verify_model(Addmm1(), example_args, {}, expected1)
verify_model(Addmm2(), example_args, {}, expected2)
verify_model(Addmm1(), example_args, {}, expected1, run_ep_decomposition=True)
verify_model(Addmm2(), example_args, {}, expected2, run_ep_decomposition=True)


def test_avg_pool1d():
Expand Down Expand Up @@ -2054,8 +2070,10 @@ def main(
inp_2: R.Tensor((4, 256, 512), dtype="float32"),
) -> R.Tuple(R.Tensor((4, 128, 512), dtype="float32")):
with R.dataflow():
lv: R.Tensor((4, 128, 512), dtype="float32") = R.matmul(inp_1, inp_2)
lv1: R.Tensor((4, 128, 512), dtype="float32") = R.add(lv, inp_0)
lv: R.Tensor((4, 128, 512), dtype="float32") = R.matmul(
inp_1, inp_2, out_dtype="float32"
)
lv1: R.Tensor((4, 128, 512), dtype="float32") = R.add(inp_0, lv)
gv: R.Tuple(R.Tensor((4, 128, 512), dtype="float32")) = (lv1,)
R.output(gv)
return gv
Expand All @@ -2076,7 +2094,9 @@ def main(
inp_2: R.Tensor((4, 256, 512), dtype="float32"),
) -> R.Tuple(R.Tensor((4, 128, 512), dtype="float32")):
with R.dataflow():
lv: R.Tensor((4, 128, 512), dtype="float32") = R.matmul(inp_1, inp_2)
lv: R.Tensor((4, 128, 512), dtype="float32") = R.matmul(
inp_1, inp_2, out_dtype="float32"
)
lv1: R.Tensor((4, 128, 512), dtype="float32") = R.multiply(
lv, R.const(2, "float32")
)
Expand All @@ -2100,14 +2120,16 @@ def main(
inp_2: R.Tensor((4, 256, 512), dtype="float32"),
) -> R.Tuple(R.Tensor((4, 128, 512), dtype="float32")):
with R.dataflow():
lv: R.Tensor((4, 128, 512), dtype="float32") = R.matmul(inp_1, inp_2)
lv: R.Tensor((4, 128, 512), dtype="float32") = R.matmul(
inp_1, inp_2, out_dtype="float32"
)
lv1: R.Tensor((4, 128, 512), dtype="float32") = R.multiply(
lv, R.const(2, "float32")
)
lv2: R.Tensor((4, 128, 512), dtype="float32") = R.multiply(
inp_0, R.const(3, "float32")
)
lv3: R.Tensor((4, 128, 512), dtype="float32") = R.add(lv1, lv2)
lv3: R.Tensor((4, 128, 512), dtype="float32") = R.add(lv2, lv1)
gv: R.Tuple(R.Tensor((4, 128, 512), dtype="float32")) = (lv3,)
R.output(gv)
return gv
Expand All @@ -2122,20 +2144,23 @@ def main(
example_args,
{},
Expected1,
run_ep_decomposition=True,
)

verify_model(
BAddBMM2(),
example_args,
{},
Expected2,
run_ep_decomposition=True,
)

verify_model(
BAddBMM3(),
example_args,
{},
Expected3,
run_ep_decomposition=True,
)


Expand Down Expand Up @@ -2172,6 +2197,7 @@ def main(
example_args,
{},
Expected,
run_ep_decomposition=True,
)


Expand Down
Loading