support argmax converter (#2291)

Signed-off-by: Bo Wang <[email protected]>
Co-authored-by: gs-olive <[email protected]>

Author: bowang007

py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py

@@ -175,7 +175,7 @@ def aten_ops_group_norm(
     )
 
 
-@dynamo_tensorrt_converter(torch.ops.aten.cat.default)
+@dynamo_tensorrt_converter(torch.ops.aten.cat.default)  # type: ignore[misc]
 def aten_ops_cat(
     ctx: ConversionContext,
     target: Target,
@@ -1797,3 +1797,23 @@ def aten_ops_reshape(
         input=args[0],
         shape=args[1],
     )
+
+
+@enforce_tensor_types({0: (TRTTensor,)})  # type: ignore[misc]
+@dynamo_tensorrt_converter(torch.ops.aten.argmax.default)  # type: ignore[misc]
+def aten_ops_argmax(
+    ctx: ConversionContext,
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> Union[TRTTensor, Sequence[TRTTensor]]:
+    return impl.argmax.argmax(
+        ctx,
+        target,
+        SourceIR.ATEN,
+        name,
+        input=args[0],
+        dim=args_bounds_check(args, 1),
+        keep_dim=args_bounds_check(args, 2, False),
+    )

py/torch_tensorrt/dynamo/conversion/impl/__init__.py

@@ -3,6 +3,7 @@
 from . import (
     activation,
     attention,
+    argmax,
     cast,
     cat,
     condition,

py/torch_tensorrt/dynamo/conversion/impl/argmax.py

@@ -0,0 +1,74 @@
+from typing import Optional
+
+import tensorrt as trt
+from torch.fx.node import Target
+from torch_tensorrt.dynamo._SourceIR import SourceIR
+from torch_tensorrt.dynamo.conversion import impl
+from torch_tensorrt.dynamo.conversion._ConversionContext import ConversionContext
+from torch_tensorrt.dynamo.conversion.converter_utils import (
+    cast_trt_tensor,
+    flatten_dims,
+    get_axes_for_reduce_op,
+    get_positive_dim,
+)
+from torch_tensorrt.fx.converters.converter_utils import set_layer_name
+from torch_tensorrt.fx.types import TRTTensor
+
+
+def argmax(
+    ctx: ConversionContext,
+    target: Target,
+    source_ir: Optional[SourceIR],
+    name: str,
+    input: TRTTensor,
+    dim: Optional[int],
+    keep_dim: bool = False,
+) -> TRTTensor:
+    if input.dtype == trt.int32:
+        input = cast_trt_tensor(ctx, input, trt.float32, name, target, source_ir)
+
+    # Three different cases here:
+    # 1. dim == None, flatten input tensor first, keep_dim will be ignore and the output rank == input rank
+    # 2. input rank == 1: TopK layer does not support 1 dimensional topk operation. Broadcast input to rank == 2
+    # 3. normal cases, no additional handlings
+    out = input
+
+    if dim is None:
+        new_shape = (*flatten_dims(input, 0, -1), 1)
+        out = impl.shuffle.reshape(
+            ctx, target, source_ir, f"{name}_flatten", input, new_shape
+        )
+    elif len(input.shape) == 1:
+        new_shape = (*input.shape, 1)
+        out = impl.shuffle.reshape(
+            ctx, target, source_ir, f"{name}_broadcast", input, new_shape
+        )
+
+    # Reduce over the flattened input if the dimension is None, otherwise the specified dimension
+    reduce_mask = get_axes_for_reduce_op(
+        get_positive_dim(dim if dim is not None else 0, len(out.shape))
+    )
+
+    topk_layer = ctx.net.add_topk(out, trt.TopKOperation.MAX, 1, reduce_mask)
+    set_layer_name(topk_layer, target, name, source_ir)
+
+    out = topk_layer.get_output(1)
+
+    if dim is None:
+        new_shape = ((1,) * len(input.shape)) if keep_dim else ()
+        out = impl.shuffle.reshape(
+            ctx, target, source_ir, f"{name}_unflatten", out, new_shape
+        )
+    elif len(input.shape) == 1:
+        out = impl.squeeze.squeeze(
+            ctx,
+            target,
+            source_ir,
+            f"{name}_squeeze",
+            out,
+            1 if keep_dim else (0, 1),
+        )
+    elif not keep_dim:
+        out = impl.squeeze.squeeze(ctx, target, source_ir, f"{name}_squeeze", out, dim)
+
+    return out

tests/py/dynamo/conversion/test_argmax_aten.py

@@ -0,0 +1,41 @@
+import torch
+import torch.nn as nn
+from parameterized import parameterized
+from torch.testing._internal.common_utils import run_tests
+
+from .harness import DispatchTestCase
+
+
+class TestArgmaxConverter(DispatchTestCase):
+    @parameterized.expand(
+        [
+            # input dimension == 1
+            ("dim_1_keep_dim_true", (3,), 0, True),
+            ("dim_1_keep_dim_true", (3,), 0, False),
+            # dim == None
+            ("dim_none", (3,), None, True),
+            ("dim_none", (3, 3), None, True),
+            ("dim_none", (3, 3, 3), None, False),
+            # # common cases
+            ("dim_1_keep_dim_true", (3, 3), 1, True),
+            ("dim_1_keep_dim_false", (3, 3), 1, False),
+            ("dim_0_keep_dim_true", (4, 4, 4), 0, True),
+            ("dim_0_keep_dim_false", (4, 4, 4), 0, False),
+            ("dim_negative_keep_dim_true", (1, 2, 3), -1, True),
+        ]
+    )
+    def test_argmax(self, _, input_shape, dim, keep_dim):
+        class ArgMax(nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, input):
+                return torch.ops.aten.argmax.default(input, dim, keep_dim)
+
+        input = [torch.randn(*input_shape)]
+
+        self.run_test(ArgMax(), input)
+
+
+if __name__ == "__main__":
+    run_tests()