Remove dynamo+nvfuser (pytorch#105789)

This PR removes unmaintained Dynamo+nvFuser.

Pull Request resolved: pytorch#105789
Approved by: https://github.com/jansel, https://github.com/jjsjann123, https://github.com/albanD

docs/source/torch.compiler.rst

@@ -52,10 +52,6 @@ Some of the most commonly used backends include:
      - Description
    * - ``torch.compile(m, backend="inductor")``
      - Uses the TorchInductor backend. `Read more <https://dev-discuss.pytorch.org/t/torchinductor-a-pytorch-native-compiler-with-define-by-run-ir-and-symbolic-shapes/747>`__
-   * - ``torch.compile(m, backend="aot_ts_nvfuser")``
-     - nvFuser with AOT Autograd/TorchScript. `Read more <https://dev-discuss.pytorch.org/t/tracing-with-primitives-update-1-nvfuser-and-its-primitives/593>`__
-   * - ``torch.compile(m, backend="nvprims_nvfuser")``
-     - Tracing with nvFuser and its primitives. `Read more <https://dev-discuss.pytorch.org/t/tracing-with-primitives-update-1-nvfuser-and-its-primitives/593>`__
    * - ``torch.compile(m, backend="cudagraphs")``
      - CUDA graphs with AOT Autograd. `Read more <https://github.com/pytorch/torchdynamo/pull/757>`__
 

docs/source/torch.compiler_get_started.rst

@@ -93,7 +93,7 @@ hub.
 
 And that is not the only available backend, you can run in a REPL
 ``torch.compile.list_backends()`` to see all the available backends. Try out the
-``cudagraphs`` or ``nvfuser`` next as inspiration.
+``cudagraphs`` next as inspiration.
 
 Using a pretrained model
 ~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/torch.compiler_troubleshooting.rst

@@ -202,7 +202,7 @@ If the error does not occur with the ``"eager"`` backend, then the
 backend compiler is the source of the error (`example
 error <https://gist.github.com/mlazos/2f13681e3cc6c43b3911f336327032de%5D>`__).
 There are `different choices <./torch.compiler.rst>`__
-for backend compilers for TorchDynamo, with TorchInductor or nvfuser
+for backend compilers for TorchDynamo, with TorchInductor
 fitting the needs of most users. This section focuses on TorchInductor
 as the motivating example, but some tools can also be used with other
 backend compilers.

test/dynamo/test_backends.py

@@ -10,7 +10,6 @@
 from torch._dynamo.backends.onnxrt import has_onnxruntime
 from torch._dynamo.backends.tvm import has_tvm
 from torch._dynamo.testing import same
-from torch.testing._internal.common_utils import IS_FBCODE, skipIfRocm
 from torch.testing._internal.inductor_utils import HAS_CUDA
 
 requires_cuda = functools.partial(unittest.skipIf, not HAS_CUDA, "requires cuda")
@@ -123,21 +122,6 @@ def test_aot_ts(self):
     def test_aot_cudagraphs(self):
         self._check_backend_works("cudagraphs")
 
-    @skipIfRocm
-    @requires_cuda()
-    def test_aot_ts_nvfuser(self):
-        self._check_backend_works("aot_ts_nvfuser")
-
-    @requires_cuda()
-    @unittest.skipIf(IS_FBCODE, "BackendCompilerError")
-    def test_nvprims_nvfuser(self):
-        self._check_backend_works("nvprims_nvfuser")
-
-    @requires_cuda()
-    @unittest.skipIf(IS_FBCODE, "BackendCompilerError")
-    def test_nvprims_aten(self):
-        self._check_backend_works("nvprims_aten")
-
     @unittest.skipIf(not has_onnxruntime(), "requires onnxruntime")
     def test_onnxrt(self):
         self._check_backend_works("onnxrt")

test/test_ops.py

@@ -484,26 +484,9 @@ def test_python_ref_torch_fallback(self, device, dtype, op):
     @unittest.skipIf(TEST_WITH_ASAN, "Skipped under ASAN")
     @onlyCUDA
     @ops(python_ref_db)
-    @parametrize('executor', ['aten', 'nvfuser'])
+    @parametrize('executor', ['aten',])
     @skipIfTorchInductor("Takes too long for inductor")
     def test_python_ref_executor(self, device, dtype, op, executor):
-        # TODO: Not all dtypes are supported with nvfuser
-        from torch._prims_common import _torch_dtype_to_nvfuser_dtype_map
-        if executor == "nvfuser" and dtype not in _torch_dtype_to_nvfuser_dtype_map:
-            raise unittest.SkipTest(f"nvfuser doesn't support dtype {dtype}")
-
-        # nvFuser tests are rather slow so we only run int32 and float32 types
-        if executor == "nvfuser" and dtype not in [torch.int32, torch.float32]:
-            raise unittest.SkipTest("skipped for speed")
-
-        if executor == "nvfuser" and not op.supports_nvfuser:
-            raise unittest.SkipTest(f"{op.name} doesn't support nvfuser")
-
-        # nvFuser doesn't support reduction operations on 0-dim tensors yet
-        skip_zero_dim = False
-        if executor == "nvfuser" and isinstance(op, ReductionPythonRefInfo):
-            skip_zero_dim = True
-
         # skip zero-dim tensors for some composites of reduction operations and view
         skip_zero_dim_ops = [
             "_refs.logsumexp",
@@ -513,25 +496,16 @@ def test_python_ref_executor(self, device, dtype, op, executor):
             "_refs.sum_to_size",
             "ops.nvprims.view",
         ]
-        if executor == "nvfuser" and op.name in skip_zero_dim_ops:
-            skip_zero_dim = True
 
         from torch._prims.executor import make_traced
         from copy import copy
         op = copy(op)
-        executor = "strictly_nvfuser" if executor == "nvfuser" else executor
         op.op = partial(make_traced(op.op), executor=executor)
         self._ref_test_helper(
             contextlib.nullcontext,
             device,
             dtype,
             op,
-            skip_zero_numel=("nvfuser" in executor),  # nvfuser doesn't support zero-sized tensors
-            skip_zero_dim=skip_zero_dim,
-            skip_bfloat=("nvfuser" in executor),  # nvfuser doesn't support bfloat tensors for pre-11 cuda TK
-            # # nvfuser doesn't support view consistency
-            # https://github.com/pytorch/pytorch/issues/84863
-            skip_view_consistency=("nvfuser" in executor),
         )
 
     @skipMeta