switching NNC as default for TorchScript support (#105185)

Disable nvfuser by default in TorchScript
Add deprecation warning for nvfuser usage via TorchScript and PrimTorch

Pull Request resolved: #105185
Approved by: https://github.com/malfet, https://github.com/davidberard98

third_party/nvfuser/python_tests/test_dynamo.py

@@ -136,10 +136,7 @@ def func(a):
             # nvfuser would promote output to fp32 in math, FusionDefinition should cast output dtype back
             return torch.where(tmp > 0, tmp, 0.0)
 
-        # No warnings and no errors
-        with warnings.catch_warnings(record=True) as w:
-            nvfuser_result = func(input1)
-            self.assertEqual(len(w), 0)
+        nvfuser_result = func(input1)
         eager_result = func.__wrapped__(input1)
         self.assertEqual(eager_result, nvfuser_result)
 

torch/_prims/nvfuser_executor.py

@@ -297,6 +297,7 @@ def nvfuser_execute(gm: GraphModule, *args, executor_parameters=None):
                     ),
                 )
             )
+        warn("nvfuser integration in primTorch is deprecated")
         result = tree_unflatten(
             fusion.execute(concrete_fusion_inputs),  # type: ignore[has-type]
             unflatten_spec,  # type: ignore[has-type]

torch/csrc/jit/codegen/cuda/interface.cpp

@@ -151,11 +151,8 @@ class NVFuserEnabler {
       return *getCachedFuserEnabledEnvVar();
     }
     // 3. default value
-#if defined(USE_ROCM) || defined(FBCODE_CAFFE2)
+    // default off since TorchScript integration isn't maintained any longer
     return false;
-#else
-    return nvfuserCanBeEnabled();
-#endif
   }
 
  public:
@@ -235,6 +232,7 @@ void fuseGraph(std::shared_ptr<Graph>& graph) {
     return;
   }
 
+  TORCH_WARN_ONCE("nvfuser integration in TorchScript is deprecated.");
   TORCH_CHECK(
       getFuserInterface()->fn_fuse_graph != nullptr,
       "Running the CUDA fuser requires a CUDA build.");