fix/feat: Add and repair multiple converters

- Focus on SD-performance-accelerating converters
- Add test cases for converters to avoid regressions

Author: gs-olive

py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py

@@ -1156,6 +1156,7 @@ def aten_ops_sub(
 @dynamo_tensorrt_converter(torch.ops.aten.div.Tensor_mode)  # type: ignore[misc]
 @dynamo_tensorrt_converter(torch.ops.aten.div.Scalar)  # type: ignore[misc]
 @dynamo_tensorrt_converter(torch.ops.aten.div.Scalar_mode)  # type: ignore[misc]
+@dynamo_tensorrt_converter(torch.ops.prims.div.default)  # type: ignore[misc]
 def aten_ops_div(
     network: TRTNetwork,
     target: Target,
@@ -1413,3 +1414,34 @@ def aten_ops_linear(
         weight=args[1],
         bias=args_bounds_check(args, 2, None),
     )
+
+
+# TODO: expand the scope of this converter with aten.expand implementation
+def broadcast_checker(broadcast_node: torch.fx.Node) -> bool:
+    # The current implementation of broadcast_in_dim can only handle unsqueeze
+    return all(
+        broadcast_node.args[1][i] == 1
+        for i in range(len(broadcast_node.args[1]))
+        if i not in broadcast_node.args[2]
+    )
+
+
+@dynamo_tensorrt_converter(
+    torch.ops.prims.broadcast_in_dim.default, capability_validator=broadcast_checker
+)  # type: ignore[misc]
+def aten_ops_broadcast_in_dim(
+    network: TRTNetwork,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> Union[TRTTensor, Sequence[TRTTensor]]:
+    return impl.unsqueeze.broadcast_in_dim(
+        network,
+        target,
+        SourceIR.ATEN,
+        name,
+        args[0],
+        args[1],
+        args[2],
+    )

py/torch_tensorrt/dynamo/conversion/impl/condition/ops.py

@@ -1,5 +1,6 @@
 from typing import Optional
 
+import numpy as np
 import tensorrt as trt
 import torch
 from torch.fx.node import Target
@@ -22,16 +23,6 @@ def where(
     other: TRTTensor,
     condition: TRTTensor,
 ) -> TRTTensor:
-    input_dim = len(tuple(input.shape))
-    other_dim = len(tuple(other.shape))
-    condition_dim = len(tuple(condition.shape))
-
-    if type(input) != TRTTensor:
-        assert type(input) is torch.Tensor, f"value {input} is not torch.Tensor!"
-
-    if type(other) != TRTTensor:
-        assert type(other) is torch.Tensor, f"value {other} is not torch.Tensor!"
-
     if not (broadcastable(input, other)):
         assert "The two torch tensors should be broadcastable"
 
@@ -48,33 +39,37 @@ def where(
     x_shape = list(input.shape)
     y_shape = list(other.shape)
     condition_shape = list(condition.shape)
+
     output_shape = list(torch.broadcast_shapes(condition_shape, x_shape, y_shape))
 
     # expand shape
-    if type(condition) != TRTTensor:
-        assert condition.dtype == torch.bool, "condition dtype is not bool"
+    if not isinstance(condition, TRTTensor):
+        assert condition.dtype in (torch.bool, np.bool_), "condition dtype is not bool"
         if condition_shape != output_shape:
-            condition.expand(output_shape)
-        condition = condition.to(torch.int32)
-        condition_const = get_trt_tensor(network, condition, f"{name}_condition")
-        condition_layer = network.add_identity(condition_const)
-        condition_layer.set_output_type(0, trt.bool)
-        set_layer_name(condition_layer, target, f"{name}_condition")
-        condition_val = condition_layer.get_output(0)
+            condition = (
+                condition.expand(output_shape)
+                if isinstance(condition, torch.Tensor)
+                else np.broadcast_to(condition, output_shape)
+            )
+        condition_val = get_trt_tensor(network, condition, f"{name}_condition")
     else:
         assert condition.dtype == trt.bool, "mask dtype is not bool!"
-        if len(condition_shape) != condition_dim:
+        if condition_shape != output_shape:
             condition_val = expand(
                 network, target, source_ir, f"{name}_expand", condition, output_shape
             )
         else:
             condition_val = condition
 
-    if type(input) != TRTTensor:
+    if not isinstance(input, TRTTensor):
         if x_shape != output_shape:
             # special case where 1 element in input
             if len(input.shape) == 0:
-                input = input.unsqueeze(0)
+                input = (
+                    input.unsqueeze(0)
+                    if isinstance(input, torch.Tensor)
+                    else np.expand_dims(input, axis=0)
+                )
             input = input.expand(output_shape)
         x_val = get_trt_tensor(network, input, f"{name}_x")
     else:
@@ -84,11 +79,15 @@ def where(
                 network, target, source_ir, f"{name}_x_expand", input, output_shape
             )
 
-    if type(other) != TRTTensor:
+    if not isinstance(other, TRTTensor):
         if y_shape != output_shape:
             # special case where 1 element in other
             if len(other.shape) == 0:
-                other = other.unsqueeze(0)
+                other = (
+                    other.unsqueeze(0)
+                    if isinstance(other, torch.Tensor)
+                    else np.expand_dims(other, axis=0)
+                )
             other = other.expand(output_shape)
         y_val = get_trt_tensor(network, other, f"{name}_y")
     else:

py/torch_tensorrt/dynamo/conversion/impl/reduce.py

@@ -49,7 +49,7 @@ def sum(
     ):
         input_val = cast_trt_tensor(network, input_val, trt.float32, name)
 
-    if dim is None:
+    if dim is None or (isinstance(dim, (tuple, list)) and len(dim) == 0):
         dim = tuple(range(len(input_val.shape)))
     layer = network.add_reduce(
         input_val,

py/torch_tensorrt/dynamo/conversion/impl/unsqueeze.py

@@ -1,4 +1,4 @@
-from typing import Optional, cast
+from typing import List, Optional, Sequence, cast
 
 from torch.fx.node import Target
 from torch_tensorrt.dynamo._SourceIR import SourceIR
@@ -48,3 +48,42 @@ def unsqueeze(
     )
     set_layer_name(layer, target, name, source_ir)
     return layer.get_output(0)
+
+
+def broadcast_in_dim(
+    network: TRTNetwork,
+    target: Target,
+    source_ir: Optional[SourceIR],
+    name: str,
+    input_t: TRTTensor,
+    shape: Sequence[int],
+    broadcast_dimensions: Sequence[int],
+) -> TRTTensor:
+    augmented_shape_list: List[Optional[int]] = list(shape)
+
+    # For each dimension being broadcasted, set the augmented shape to None
+    for broadcast_dim in broadcast_dimensions:
+        augmented_shape_list[broadcast_dim] = None
+
+    # TODO: Expand support to arbitrary broadcasts
+    assert all(
+        dim in (1, None) for dim in augmented_shape_list
+    ), "broadcast_in_dim currently only supports unsqueeze broadcasting"
+
+    # Unsqueeze the shape repeatedly to broadcast
+    output = input_t
+    for idx, x in enumerate(augmented_shape_list):
+        # If the value is not None, that dimension is to be broadcasted
+        if x is not None:
+            output = unsqueeze(
+                network,
+                target,
+                source_ir,
+                name + f"_unsqueeze_for_broadcast_{idx}",
+                output,
+                idx,
+            )
+
+    assert tuple(output.shape) == tuple(shape), "broadcast_in_dim shapes don't match"
+
+    return output

tests/py/dynamo/conversion/test_div_aten.py

@@ -82,6 +82,24 @@ def forward(self, lhs_val):
             expected_ops={torch.ops.aten.div.Tensor_mode},
         )
 
+    @parameterized.expand(
+        [
+            ("2d", (2, 1)),
+            ("3d", (2, 1, 2)),
+        ]
+    )
+    def test_prims_div_tensor(self, _, shape):
+        class div(nn.Module):
+            def forward(self, lhs_val, rhs_val):
+                return torch.ops.prims.div.default(lhs_val, rhs_val)
+
+        inputs = [torch.randn(shape), torch.randn(shape)]
+        self.run_test(
+            div(),
+            inputs,
+            expected_ops={torch.ops.prims.div.default},
+        )
+
 
 if __name__ == "__main__":
     run_tests()

tests/py/dynamo/conversion/test_unsqueeze_aten.py

@@ -4,6 +4,7 @@
 from parameterized import parameterized
 from torch.testing._internal.common_utils import run_tests
 from torch_tensorrt import Input
+from torch_tensorrt.dynamo.conversion import UnsupportedOperatorException
 
 from .harness import DispatchTestCase
 
@@ -59,5 +60,51 @@ def forward(self, x):
         )
 
 
+class TestBroadcastInDim(DispatchTestCase):
+    def test_broadcast_in_dim_supported(
+        self,
+    ):
+        class Unsqueeze(nn.Module):
+            def forward(self, x):
+                return torch.ops.prims.broadcast_in_dim.default(
+                    x, [4, 5, 6, 1, 1], [0, 1, 2]
+                )
+
+        inputs = [torch.randn(4, 5, 6)]
+        self.run_test(
+            Unsqueeze(), inputs, expected_ops={torch.ops.prims.broadcast_in_dim.default}
+        )
+
+    def test_broadcast_in_dim_supported_singleton(
+        self,
+    ):
+        class Unsqueeze(nn.Module):
+            def forward(self, x):
+                return torch.ops.prims.broadcast_in_dim.default(x, [1, 1, 1], [0, 1])
+
+        inputs = [torch.randn(1, 1)]
+        self.run_test(
+            Unsqueeze(), inputs, expected_ops={torch.ops.prims.broadcast_in_dim.default}
+        )
+
+    # TODO: Remove this test when support is updated
+    def test_broadcast_in_dim_unsupported(
+        self,
+    ):
+        class Unsqueeze(nn.Module):
+            def forward(self, x):
+                return torch.ops.prims.broadcast_in_dim.default(
+                    x, [4, 5, 6, 7, 1], [0, 1, 2]
+                )
+
+        inputs = [torch.randn(4, 5, 6)]
+        with self.assertRaises(UnsupportedOperatorException):
+            self.run_test(
+                Unsqueeze(),
+                inputs,
+                expected_ops={torch.ops.prims.broadcast_in_dim.default},
+            )
+
+
 if __name__ == "__main__":
     run_tests()

tests/py/dynamo/conversion/test_where_aten.py

@@ -44,6 +44,24 @@ def forward(self, condition, x, y):
             expected_ops={torch.ops.aten.where.self},
         )
 
+    def test_const_input(self):
+        class Where(nn.Module):
+            def __init__(self, *args, **kwargs) -> None:
+                super().__init__(*args, **kwargs)
+                self.inputY = torch.randn((5, 6, 7))
+                self.inputX = torch.randn((5, 6, 7))
+
+            def forward(self, condition):
+                return torch.where(condition, self.inputX, self.inputY)
+
+        input1 = torch.randn((5, 6, 7))
+        condition = input1 < 0
+        self.run_test(
+            Where(),
+            (condition,),
+            expected_ops={torch.ops.aten.where.self},
+        )
+
 
 if __name__ == "__main__":
     run_tests()