Improve torch elemwise operator

Author: Ch0ronomato

pytensor/link/pytorch/dispatch/elemwise.py

@@ -11,9 +11,21 @@ def pytorch_funcify_Elemwise(op, node, **kwargs):
     scalar_op = op.scalar_op
     base_fn = pytorch_funcify(scalar_op, node=node, **kwargs)
 
-    def elemwise_fn(*inputs):
-        Elemwise._check_runtime_broadcast(node, inputs)
-        return base_fn(*inputs)
+    if hasattr(scalar_op, "nfunc_spec") and hasattr(torch, scalar_op.nfunc_spec[0]):
+        # torch can handle this scalar
+        # broadcast, we'll let it.
+        def elemwise_fn(*inputs):
+            Elemwise._check_runtime_broadcast(node, inputs)
+            return base_fn(*inputs)
+    else:
+
+        def elemwise_fn(*inputs):
+            Elemwise._check_runtime_broadcast(node, inputs)
+            broadcast_inputs = torch.broadcast_tensors(*inputs)
+            ufunc = base_fn
+            for _ in range(broadcast_inputs[0].dim()):
+                ufunc = torch.vmap(ufunc)
+            return ufunc(*broadcast_inputs)
 
     return elemwise_fn
 

tests/link/pytorch/test_elemwise.py

@@ -1,10 +1,13 @@
 import numpy as np
 import pytest
 
+import pytensor
 import pytensor.tensor as pt
 import pytensor.tensor.math as ptm
 from pytensor.configdefaults import config
 from pytensor.graph.fg import FunctionGraph
+from pytensor.scalar.basic import ScalarOp, get_scalar_type
+from pytensor.tensor.elemwise import Elemwise
 from pytensor.tensor.special import SoftmaxGrad, log_softmax, softmax
 from pytensor.tensor.type import matrix, tensor, tensor3, vector
 from tests.link.pytorch.test_basic import compare_pytorch_and_py
@@ -150,3 +153,33 @@ def test_cast():
         fgraph, [np.arange(6, dtype="float32").reshape(2, 3)]
     )
     assert res.dtype == torch.int32
+
+
+def test_vmap_elemwise():
+    from pytensor.link.pytorch.dispatch.basic import pytorch_funcify
+
+    class TestOp(ScalarOp):
+        def __init__(self):
+            super().__init__(
+                output_types_preference=lambda *_: [get_scalar_type("float32")]
+            )
+            self.call_shapes = []
+            self.nin = 1
+
+        def perform(self, *_):
+            raise RuntimeError("In perform")
+
+    @pytorch_funcify.register(TestOp)
+    def relu(op, node, **kwargs):
+        def relu(row):
+            op.call_shapes.append(row.size())
+            return torch.max(torch.zeros_like(row), row)
+
+        return relu
+
+    x = matrix("x", shape=(2, 3))
+    op = TestOp()
+    f = pytensor.function([x], Elemwise(op)(x), mode="PYTORCH")
+    vals = torch.zeros(2, 3).normal_()
+    np.testing.assert_allclose(f(vals), torch.relu(vals))
+    assert op.call_shapes == [torch.Size([])], op.call_shapes