Enable interpolate VJP tests, add PrimID mapping for ceil and floor

Author: aobolensk

thunder/core/transforms.py

@@ -2574,6 +2574,8 @@ def uniform_backward(primal, minval, maxval, g):
     prims.PrimIDs.BITWISE_XOR,
     prims.PrimIDs.SIGNBIT,
     prims.PrimIDs.FULL,
+    prims.PrimIDs.FLOOR,
+    prims.PrimIDs.CEIL,
 }
 
 

thunder/tests/opinfos.py

@@ -9839,20 +9839,6 @@ def interpolate_error_generator(op, device, dtype=torch.float32, **kwargs):
     error_input_generator=interpolate_error_generator,
     torch_reference=torch.nn.functional.interpolate,
     dtypes=(datatypes.floating,),
-    test_directives=(
-        # PyTorch does not support CPU Half upsample used in interpolate
-        DecorateInfo(
-            pytest.mark.xfail,
-            "test_core_vs_torch_consistency",
-            dtypes=(datatypes.float16,),
-            devicetypes=(devices.DeviceType.CPU,),
-        ),
-        # This should be fixed now; TODO re-enable and test
-        DecorateInfo(
-            pytest.mark.xfail,
-            "test_vjp_correctness",
-        ),
-    ),
 )
 nn_ops.append(interpolate_opinfo)