[Fix] Fixing a bug in register_methods (#331)

Previously, an error was encountered during a model compilation attempt: > torch._dynamo.exc.BackendCompilerFailed: backend='hidet' raised: > RuntimeError: Can not interpreting max given arguments: > max(tensor(...)) > Possible candidates are: > torch_max_v3(x: hidet.Tensor, dim: Union[int, hidet.ir.expr.Expr], keepdim: bool = False, *, out: Union[hidet.Tensor, Tuple[hidet.Tensor, ...], List[hidet.Tensor]] = None) -> Tuple[hidet.Tensor, hidet.Tensor] > File "/home/bolin/Desktop/hidet/python/hidet/graph/frontend/torch/register_functions.py", line 1067 Despite we indeed have a [function](https://github.com/CentML/hidet/blob/13a806608d40de2de1fcc682adeea8d204189f3c/python/hidet/graph/frontend/torch/register_functions.py#L1056-L1060) that can be used to interpret the `torch.Tensor.max` with described arguments.
hidet-org · Jul 23, 2024 · c87c515 · c87c515
1 parent ff9445e
commit c87c515
Show file tree

Hide file tree

Showing 5 changed files with 101 additions and 20 deletions.
diff --git a/python/hidet/graph/frontend/torch/interpreter.py b/python/hidet/graph/frontend/torch/interpreter.py
@@ -176,7 +176,7 @@ def _raise_exception(exception: Exception, target, caused_callable: Any, args, k
         if isinstance(caused_callable, OverloadedFunction):
             dispatched = caused_callable.resolve(*args, **kwargs)
             if dispatched is None:
-                msg = ['Can not interpreting {} given arguments: '.format(target_name)]
+                msg = ['Can not interpret {} given arguments: '.format(target_name)]
                 msg.append('  {}({})'.format(target_name, args_string))
                 msg.append('Possible candidates are: ')
                 for overload, sig in zip(caused_callable.functions, caused_callable.signatures):

diff --git a/python/hidet/graph/frontend/torch/register_functions.py b/python/hidet/graph/frontend/torch/register_functions.py
@@ -366,7 +366,7 @@ def cat(tensors: List[Tensor], dim: int = 0):
 
 
 @register_function(torch.cat)
-def cat(tensors: List[Tensor], axis: int):  # PyTorch supports axis as well as the argument name
+def cat_v2(tensors: List[Tensor], axis: int):  # PyTorch supports axis as well as the argument name
     dtype = functools.reduce(promote_type, [t.dtype for t in tensors])
     tensors = [ops.cast(t, dtype) for t in tensors]
     return ops.concat(tensors, axis)
@@ -1063,15 +1063,19 @@ def minimum(x: Tensor, other: Tensor, *, out: Optional[Tensor] = None) -> Tensor
 
 
 @register_function(torch.max)
+@register_method(torch.Tensor.max)
 def torch_max(x: Tensor, *, out: Optional[Tensor] = None) -> Tensor:
     if out is not None:
         raise NotImplementedError("hidet: does not support torch.max(..., out=...)")
-    return ops.max(x, dims=list(range(len(x.shape))), keep_dim=True)
+
+    # According to the PyTorch documentation,
+    # calling torch.max(tensor(...)) or some_tensor.max() results in a singleton tensor with shape torch.Size([]).
+    return ops.max(x, dims=list(range(len(x.shape))), keep_dim=False)
 
 
 @register_function(torch.max)
 @register_method(torch.Tensor.max)
-def torch_max(
+def torch_max_v2(
     x: Tensor, other: Union[Tensor, int], *, out: Optional[Tensor] = None
 ) -> Union[Tensor, Tuple[Tensor, Tensor]]:
     if out is not None:
@@ -1085,7 +1089,7 @@ def torch_max(
 @register_function(torch.max)
 @register_method(torch.Tensor.max)
 def torch_max_v3(
-    x: Tensor, dim: Int, keepdim: bool = False, *, out: Union[Tensor, Tuple[Tensor, ...], List[Tensor]] = None
+    x: Tensor, dim: Int, keepdim: bool = False, *, out: Optional[Union[Tensor, Tuple[Tensor, ...], List[Tensor]]] = None
 ) -> Tuple[Tensor, Tensor]:
     if out is not None:
         raise NotImplementedError("hidet: does not support torch.max(..., out=...)")
@@ -1095,14 +1099,19 @@ def torch_max_v3(
 
 
 @register_function(torch.min)
+@register_method(torch.Tensor.min)
 def torch_min(x: Tensor, *, out: Optional[Tensor] = None) -> Tensor:
     if out is not None:
         raise NotImplementedError("hidet: does not support torch.min(..., out=...)")
-    return ops.min(x, dims=list(range(len(x.shape))), keep_dim=True)
+
+    # Same as torch.max,
+    # torch.min(tensor(...)) or some_tensor.min() results in a singleton tensor with shape torch.Size([]).
+    return ops.min(x, dims=list(range(len(x.shape))), keep_dim=False)
 
 
 @register_function(torch.min)
-def torch_min(
+@register_method(torch.Tensor.min)
+def torch_min_v2(
     x: Tensor, other: Union[Tensor, int], *, out: Optional[Tensor] = None
 ) -> Union[Tensor, Tuple[Tensor, Tensor]]:
     if out is not None:
@@ -1114,6 +1123,7 @@ def torch_min(
 
 
 @register_function(torch.min)
+@register_method(torch.Tensor.min)
 def torch_min_v3(
     x: Tensor, dim: Int, keepdim: bool = False, *, out: Union[Tensor, Tuple[Tensor, ...], List[Tensor]] = None
 ) -> Tuple[Tensor, Tensor]:
@@ -1212,13 +1222,15 @@ def tensor_pow(self: Union[Tensor, Number], exponent: Union[Tensor, Number]) ->
 def torch_mean(x: Tensor, *, dtype: Optional[DataType] = None) -> Tensor:
     if dtype:
         x = x.astype(dtype_from_torch(dtype))
-    output = ops.mean(x, dims=list(range(len(x.shape))), keep_dim=True)
+
+    # turns out here keep_dim should be False too, similar to torch.max/min/sum
+    output = ops.mean(x, dims=list(range(len(x.shape))), keep_dim=False)
     return output
 
 
 @register_function(torch.mean)
 @register_method(torch.Tensor.mean)
-def torch_mean(
+def torch_mean_v2(
     x: Tensor, dim, keepdim=False, *, dtype: Optional[DataType] = None, out: Optional[Tensor] = None
 ) -> Tensor:
     if out is not None:
@@ -1254,13 +1266,13 @@ def torch_var(
 def torch_sum(x: Tensor, *, dtype: Optional[DataType] = None) -> Tensor:
     if dtype:
         x = x.astype(dtype_from_torch(dtype))
-    output = ops.sum(x, dims=list(range(len(x.shape))), keep_dim=True)
+    output = ops.sum(x, dims=list(range(len(x.shape))), keep_dim=False)
     return output
 
 
 @register_function(torch.sum)
 @register_method(torch.Tensor.sum)
-def torch_sum(
+def torch_sum_v2(
     x: Tensor, dim, keepdim=False, *, dtype: Optional[DataType] = None, out: Optional[Tensor] = None
 ) -> Tensor:
     if out is not None:

diff --git a/python/hidet/graph/frontend/torch/registry.py b/python/hidet/graph/frontend/torch/registry.py
@@ -96,7 +96,7 @@ def decorator(hidet_func):
 
 def register_method(method: Callable):
     def decorator(hidet_method):
-        if method not in Registry.registered_functions:
+        if method not in Registry.registered_methods:
             Registry.registered_methods[method] = OverloadedFunction()
         Registry.registered_methods[method].overload(hidet_method)
         return hidet_method

diff --git a/python/hidet/testing/torch_utils.py b/python/hidet/testing/torch_utils.py
@@ -50,6 +50,10 @@ def check_module(model: torch.nn.Module, args: Sequence[torch.Tensor], atol=1e-4
         raise ValueError('torch_outputs and hidet_outputs have different length')
 
     for torch_output, hidet_output in zip(torch_outputs, hidet_outputs):
+        # Turns out np.testing.assert_allclose sometimes can pass even if the shapes are different
+        assert (
+            torch_output.shape == hidet_output.shape
+        ), f"Shape mismatch --- eager: {torch_output.shape} vs hidet: {hidet_output.shape}"
         torch_output = torch_output.detach().cpu().numpy()
         hidet_output = hidet_output.detach().cpu().numpy()
         numpy.testing.assert_allclose(torch_output, hidet_output, atol=atol, rtol=rtol)

diff --git a/tests/frontends/torch/test_torch_reduce.py b/tests/frontends/torch/test_torch_reduce.py
@@ -36,27 +36,92 @@ def test_minimum(shape):
 
 @pytest.mark.parametrize('shape', [[2], [2, 3], [2, 3, 4]])
 def test_max(shape):
-    check_module(FunctionalModule(op=lambda x: torch.max(x)), args=[torch.randn(shape)], atol=1e-5, rtol=1e-5)
+    check_module(FunctionalModule(op=lambda x: torch.max(x)), args=[torch.randn(shape)], atol=0, rtol=0)
     check_module(
-        FunctionalModule(op=lambda x, y: torch.max(x, y)),
-        args=[torch.randn(shape), torch.randn(shape)],
-        atol=1e-5,
-        rtol=1e-5,
+        FunctionalModule(op=lambda x, y: torch.max(x, y)), args=[torch.randn(shape), torch.randn(shape)], atol=0, rtol=0
     )
+    check_module(FunctionalModule(op=lambda x, dim: torch.max(x, dim)), args=[torch.randn(shape), 0], atol=0, rtol=0)
+
+    # Do the same checks as above all over again, this time for torch.Tensor.max methods
+    check_module(FunctionalModule(op=lambda x: x.max()), args=[torch.randn(shape)], atol=0, rtol=0)
+    check_module(FunctionalModule(op=lambda x, dim: x.max(dim)), args=[torch.randn(shape), 0], atol=0, rtol=0)
     check_module(
-        FunctionalModule(op=lambda x, dim: torch.max(x, dim)), args=[torch.randn(shape), 0], atol=1e-5, rtol=1e-5
+        FunctionalModule(op=lambda x, y: x.max(y)), args=[torch.randn(shape), torch.randn(shape)], atol=0, rtol=0
     )
 
 
 @pytest.mark.parametrize('shape', [[2], [2, 3], [2, 3, 4]])
 def test_min(shape):
-    check_module(FunctionalModule(op=lambda x: torch.min(x)), args=[torch.randn(shape)], atol=1e-5, rtol=1e-5)
+    check_module(FunctionalModule(op=lambda x: torch.min(x)), args=[torch.randn(shape)], atol=0, rtol=0)
     check_module(
         FunctionalModule(op=lambda x, y: torch.min(x, y)),
         args=[torch.randn(shape), torch.randn(shape)],
         atol=1e-5,
         rtol=1e-5,
     )
+    check_module(FunctionalModule(op=lambda x, dim: torch.min(x, dim)), args=[torch.randn(shape), 0], atol=0, rtol=0)
+
+    # Doing the same checks as above again, this time for `torch.Tensor.min` method.
+    check_module(FunctionalModule(op=lambda x: x.min()), args=[torch.randn(shape)], atol=0, rtol=0)
+
+    check_module(FunctionalModule(op=lambda x, dim: x.min(dim)), args=[torch.randn(shape), 0], atol=0, rtol=0)
+
+    check_module(
+        FunctionalModule(op=lambda x, y: x.min(y)), args=[torch.randn(shape), torch.randn(shape)], atol=0, rtol=0
+    )
+
+
+@pytest.mark.parametrize('shape', [[2], [2, 3], [2, 3, 4]])
+def test_sum(shape):
+    # Similar idea as test_max and test_min
+    check_module(FunctionalModule(op=lambda x: torch.sum(x)), args=[torch.randn(shape)], atol=1e-5, rtol=1e-5)
+
+    check_module(
+        FunctionalModule(op=lambda x, dim: torch.sum(x, dim)), args=[torch.randn(shape), 0], atol=1e-5, rtol=1e-5
+    )
+
+    check_module(FunctionalModule(op=lambda x: x.sum()), args=[torch.randn(shape)], atol=1e-5, rtol=1e-5)
+
+    check_module(FunctionalModule(op=lambda x, dim: x.sum(dim)), args=[torch.randn(shape), 0], atol=1e-5, rtol=1e-5)
+
     check_module(
-        FunctionalModule(op=lambda x, dim: torch.min(x, dim)), args=[torch.randn(shape), 0], atol=1e-5, rtol=1e-5
+        FunctionalModule(op=lambda x, dim: x.sum(dim)),
+        args=[torch.randn(shape), list(range(len(shape)))],
+        atol=1e-5,
+        rtol=1e-5,
+    )
+
+    check_module(
+        FunctionalModule(op=lambda x, dim: x.sum(dim, keepdim=True)),
+        args=[torch.randn(shape), None],
+        atol=1e-5,
+        rtol=1e-5,
+    )
+
+
+@pytest.mark.parametrize('shape', [[2], [2, 3], [2, 3, 4]])
+def test_mean(shape):
+    # Similar idea as test_sum
+    check_module(FunctionalModule(op=lambda x: torch.mean(x)), args=[torch.randn(shape)], atol=1e-5, rtol=1e-5)
+
+    check_module(
+        FunctionalModule(op=lambda x, dim: torch.mean(x, dim)), args=[torch.randn(shape), 0], atol=1e-5, rtol=1e-5
+    )
+
+    check_module(
+        FunctionalModule(op=lambda x, dim: torch.mean(x, dim)),
+        args=[torch.randn(shape), list(range(len(shape)))],
+        atol=1e-5,
+        rtol=1e-5,
+    )
+
+    check_module(FunctionalModule(op=lambda x: x.mean()), args=[torch.randn(shape)], atol=1e-5, rtol=1e-5)
+
+    check_module(FunctionalModule(op=lambda x, dim: x.mean(dim)), args=[torch.randn(shape), 0], atol=1e-5, rtol=1e-5)
+
+    check_module(
+        FunctionalModule(op=lambda x, dim: x.mean(dim)),
+        args=[torch.randn(shape), list(range(len(shape)))],
+        atol=1e-5,
+        rtol=1e-5,
     )