[torchlib] Fix implementation for clamp_max / clamp_min (#1765)

Update clamp_max and clamp_min. Remove support for size-0 inputs to
simplify the implementations. Fixed registration to make the operators
discoverable.

onnxscript/function_libs/torch_lib/ops/core.py

@@ -1660,40 +1660,32 @@ def aten_clamp(self: TReal, min: Optional[TReal] = None, max: Optional[TReal] =
     return clamped
 
 
-@torch_op("aten::clamp_max", traceable=True)
+@torch_op(("aten::clamp_max", "aten::clamp_max.Tensor"), traceable=True)
 def aten_clamp_max(self: TReal, max_: TReal) -> TReal:
     """clamp_max(Tensor self, Tensor max) -> Tensor"""
 
-    self_size = op.Size(self)
-    max_shape = op.Shape(max_)
-    max_rank = op.Size(max_shape)
-    if self_size == 0:
-        result = op.Expand(self, max_shape)
+    # This implementation does not intent to handle when self is an empty tensor
+    max_rank = Rank(max_)
+    if max_rank == 0:
+        max_ = op.CastLike(max_, self)
+        result = op.Clip(self, None, max_)
     else:
-        if max_rank == 0:
-            max_ = op.CastLike(max_, self)
-            result = op.Clip(self, None, max_)
-        else:
-            result = op.Min(self, max_)
+        result = op.Min(self, max_)
 
     return result
 
 
-@torch_op("aten::clamp_min", traceable=True)
+@torch_op(("aten::clamp_min", "aten::clamp_min.Tensor"), traceable=True)
 def aten_clamp_min(self: TReal, min_: TReal) -> TReal:
     """clamp_min(Tensor self, Tensor min) -> Tensor"""
 
-    self_size = op.Size(self)
-    min_shape = op.Shape(min_)
-    min_rank = op.Size(min_shape)
-    if self_size == 0:
-        result = op.Expand(self, min_shape)
+    # This implementation does not intent to handle when self is an empty tensor
+    min_rank = Rank(min_)
+    if min_rank == 0:
+        min_ = op.CastLike(min_, self)
+        result = op.Clip(self, min_, None)
     else:
-        if min_rank == 0:
-            min_ = op.CastLike(min_, self)
-            result = op.Clip(self, min_, None)
-        else:
-            result = op.Max(self, min_)
+        result = op.Max(self, min_)
 
     return result
 

tests/function_libs/torch_lib/ops_test_data.py

@@ -39,7 +39,6 @@
 import copy
 import dataclasses
 import functools
-import sys
 from typing import Any, Callable, Collection, Optional
 
 import numpy as np
@@ -713,19 +712,25 @@ def _where_input_wrangler(
         dtypes=(torch.bool,),
         reason="fixme: ORT does not implement SplitToSequence for bool inputs: https://github.com/microsoft/onnxruntime/issues/16905",
     ),
-    TorchLibOpInfo("clamp_max", core_ops.aten_clamp).skip(
-        enabled_if=sys.version_info[:2] >= (3, 9) or sys.platform != "win32",
-        reason="fails in this particular case",
-    ),
-    TorchLibOpInfo("clamp_max", core_ops.aten_clamp_max).skip(
+    TorchLibOpInfo("clamp_max", core_ops.aten_clamp_max)
+    .skip(
         matcher=lambda sample: len(sample.input.shape) == 0,
         enabled_if=version_utils.onnxruntime_older_than("1.16"),
         reason="fixme (core dump): ORT aborts on scalar inputs to Reduce*-18. https://github.com/microsoft/onnxruntime/issues/16492",
+    )
+    .skip(
+        reason="Size 0 inputs are not handled by design",
+        matcher=lambda sample: sample.input.numel() == 0,
     ),
-    TorchLibOpInfo("clamp_min", core_ops.aten_clamp_min).skip(
+    TorchLibOpInfo("clamp_min", core_ops.aten_clamp_min)
+    .skip(
         matcher=lambda sample: len(sample.input.shape) == 0,
         enabled_if=version_utils.onnxruntime_older_than("1.16"),
         reason="fixme (core dump): ORT aborts on scalar inputs to Reduce*-18. https://github.com/microsoft/onnxruntime/issues/16492",
+    )
+    .skip(
+        reason="Size 0 inputs are not handled by design",
+        matcher=lambda sample: sample.input.numel() == 0,
     ),
     TorchLibOpInfo("clone", core_ops.aten_clone),
     TorchLibOpInfo("complex", core_ops.aten_complex),