Skip to content

[Fix][Relax][Torch] Materialize runtime scalar shape values #20138

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
akaashrp wants to merge 1 commit into
base: main
Choose a base branch
from
+84 −5
Open

[Fix][Relax][Torch] Materialize runtime scalar shape values #20138

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
41 changes: 36 additions & 5 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 @@ -2498,6 +2498,13 @@ def _fill(self, node: fx.Node) -> relax.Var:
value = args[1] if isinstance(args[1], relax.Expr) else relax.const(args[1], dtype)
return self.block_builder.emit(relax.op.full(x.ty.shape, value, dtype))

def _prim_value_to_scalar_tensor(self, value: relax.Expr, dtype: str) -> relax.Var:
"""Materialize an integer primitive value as a rank-zero tensor."""
value_tensor = self.block_builder.emit(relax.op.shape_to_tensor(relax.ShapeExpr([value])))
if dtype != "int64":
value_tensor = self.block_builder.emit(relax.op.astype(value_tensor, dtype))
return self.block_builder.emit(relax.op.squeeze(value_tensor, axis=[0]))

def _inplace_fill(self, node: fx.Node) -> relax.Var:
args = self.retrieve_args(node)
x = args[0]
Expand All @@ -2515,7 +2522,12 @@ def _full(self, node: fx.Node) -> relax.Var:
dtype = self._convert_data_type(
node.kwargs.get("dtype", torch.get_default_dtype()), self.env
)
value = args[1] if isinstance(args[1], relax.expr.Constant) else relax.const(args[1], dtype)
if isinstance(getattr(args[1], "ty", None), PrimType):
value = self._prim_value_to_scalar_tensor(args[1], dtype)
elif isinstance(args[1], relax.Expr):
value = args[1]
else:
value = relax.const(args[1], dtype)
return self.block_builder.emit(
relax.op.full(
size,
Expand All @@ -2525,11 +2537,17 @@ def _full(self, node: fx.Node) -> relax.Var:
)

def _full_like(self, node: fx.Node) -> relax.Var:
x = self.env[node.args[0]]
value = node.args[1]
fill_value = relax.const(value)
args = self.retrieve_args(node)
x = args[0]
value = args[1]
x_dtype = str(x.ty.dtype)
if isinstance(getattr(value, "ty", None), PrimType):
fill_value = self._prim_value_to_scalar_tensor(value, x_dtype)
elif isinstance(value, relax.Expr):
fill_value = value
else:
fill_value = relax.const(value)

x_dtype = x.ty.dtype.dtype
fill_dtype = None
if isinstance(value, int | float) and (math.isinf(value) or math.isnan(value)):
if not ("float" in x_dtype or "bfloat16" in x_dtype):
Expand Down Expand Up @@ -2800,6 +2818,19 @@ def _getitem(self, node: fx.Node) -> relax.Var:

def _item(self, node: fx.Node) -> relax.Var:
x = self.env[node.args[0]]
shape = self.shape_of(x)
if shape is not None and len(shape) == 0:
dtype = str(x.ty.dtype)
if dtype in ("int32", "int64"):
scalar = x
if dtype != "int64":
scalar = self.block_builder.emit(relax.op.astype(scalar, "int64"))
scalar = self.block_builder.emit(relax.op.reshape(scalar, [1]))
shape_value = self.block_builder.emit(relax.op.tensor_to_shape(scalar))
dim = tirx.Var(f"{node.name}_dim", "int64")
self.block_builder.match_cast(shape_value, relax.ShapeType([dim]))
return dim
return x
return self.block_builder.emit(relax.op.take(x, relax.const(0, "int64"), axis=0))

def _sym_size_int(self, node: fx.Node) -> relax.Expr:
Expand Down
3 changes: 3 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 @@ -1792,6 +1792,7 @@ def create_convert_map(
"triu.default": self._tril_triu(relax.op.triu),
"trunc.default": self._unary_op(relax.op.trunc),
# binary
"add": self._binary_op(relax.op.add, operator.add),
"add.Tensor": self._binary_op(relax.op.add, operator.add),
"add.Scalar": self._binary_op(relax.op.add, operator.add),
"add_.Tensor": self._binary_op(relax.op.add, operator.add),
Expand Down Expand Up @@ -1847,6 +1848,7 @@ def create_convert_map(
"pow.Scalar": self._binary_op(relax.op.power, operator.pow),
"pow.Tensor_Scalar": self._pow,
"pow.Tensor_Tensor": self._binary_op(relax.op.power, operator.pow),
"sub": self._binary_op(relax.op.subtract, operator.sub),
"sub.Tensor": self._binary_op(relax.op.subtract, operator.sub),
"sub.Scalar": self._binary_op(relax.op.subtract, operator.sub),
"__and__.Tensor": self._binary_op(relax.op.bitwise_and, operator.and_),
Expand Down Expand Up @@ -2038,6 +2040,7 @@ def create_convert_map(
"le": self._symbolic_comparison,
"gt": self._symbolic_comparison,
"lt": self._symbolic_comparison,
"eq": self._symbolic_comparison,
# higher-order ops
"cond": self._cond,
}
Expand Down
45 changes: 45 additions & 0 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 @@ -5507,6 +5507,51 @@ def main(x: R.Tensor((2, 8, 4), dtype="float32")) -> R.Tuple(
verify_model(SliceStaticModel(), example_args_static, {}, ExpectedStatic)


def test_dynamic_scalar_item_in_shape_operations():
class DynamicShapeOps(torch.nn.Module):
def forward(self, x):
lengths = torch.full(
(x.shape[0],),
x.shape[1],
device=x.device,
dtype=torch.int64,
)
max_len = lengths.max().item()
positions = torch.arange(max_len, device=x.device)
mask = positions.unsqueeze(0).expand(x.shape[0], -1) == lengths.unsqueeze(1)
filled = torch.full(
(x.shape[0], x.shape[1]),
x.shape[1],
device=x.device,
dtype=torch.int64,
)
shifted = torch.arange(x.shape[1] + 1, device=x.device)
shortened = torch.full_like(lengths, x.shape[1] - 1)
return mask, filled, shifted, shortened

example_args = (torch.randn(1, 4, 3, dtype=torch.float32),)
tokens = torch.export.Dim("tokens", min=1, max=8)
exported_program = export(
DynamicShapeOps(),
args=example_args,
dynamic_shapes={"x": {1: tokens}},
)
mod = from_exported_program(exported_program)

script = mod.script()
assert "R.tensor_to_shape" in script
assert "R.shape_to_tensor" in script

executable = relax.build(mod, tvm.target.Target("llvm"))
vm = relax.VirtualMachine(executable, tvm.cpu())
for token_count in (4, 6):
torch_input = torch.randn(1, token_count, 3)
expected = DynamicShapeOps()(torch_input)
actual = vm["main"](tvm.runtime.tensor(torch_input.numpy()))
for actual_value, expected_value in zip(actual, expected):
np.testing.assert_array_equal(actual_value.numpy(), expected_value.numpy())


def test_split():
class Chunk(Module):
def forward(self, input):
Expand Down
Loading