add logsigmoid op

thunder/tests/opinfos.py

@@ -1678,6 +1678,23 @@ def gen(op, device, dtype, requires_grad):
 elementwise_unary_ops.append(leaky_relu_opinfo)
 
 
+logsigmoid_opinfo = OpInfo(
+    ltorch.logsigmoid,
+    dtypes=(datatypes.floating,),
+    sample_input_generator=elementwise_unary_generator,
+    torch_reference=torch.nn.functional.logsigmoid,
+    test_directives=(
+        # test tols are too tight for these half precision tests
+        DecorateInfo(
+            pytest.mark.skip,
+            "test_core_vs_torch_consistency",
+            dtypes=(datatypes.float16, datatypes.bfloat16),
+        ),
+    ),
+)
+elementwise_unary_ops.append(logsigmoid_opinfo)
+
+
 relu_opinfo = OpInfo(
     ltorch.relu,
     sample_input_generator=elementwise_unary_generator,

thunder/torch/__init__.py

@@ -1812,6 +1812,14 @@ def leaky_relu(a: TensorProxy, /, negative_slope: float = 0.01, inplace: bool =
 _inplace_to_out_of_place[leaky_relu] = leaky_relu, 2
 
 
+@torchsymbol(torch.nn.functional.logsigmoid, is_method=False)
+def logsigmoid(a: TensorProxy, /):
+    return log(sigmoid(a))
+
+
+_inplace_to_out_of_place[logsigmoid] = logsigmoid, -1
+
+
 # TODO Should this use clamp? -- Would that propagate NaNs properly?
 @torchsymbol(torch.relu, torch.nn.functional.relu, id="torch.relu", is_method=True)
 def relu(a: TensorLike, /, inplace: bool = False) -> TensorLike:

thunder/torch/default_torch_ops.py

@@ -354,7 +354,6 @@
         torch.nn.functional.kl_div,
         torch.nn.functional.l1_loss,
         torch.nn.functional.local_response_norm,
-        torch.nn.functional.logsigmoid,
         torch.nn.functional.lp_pool1d,
         torch.nn.functional.lp_pool2d,
         torch.nn.functional.lp_pool3d,