#71946 Remove Python 3.6 references (#72211)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/71946

This commit removes some bits of code that were hard coded for Python 3.6 support from the `.circleci` and `torch` folders. It should only be merged if https://github.com/pytorch/pytorch/issues/66462 is complete.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/72211

Reviewed By: dagitses, seemethere

Differential Revision: D33982604

Pulled By: musebc

fbshipit-source-id: 8f453bf9909df615addd59538adb369c65484044

.circleci/README.md

@@ -71,9 +71,9 @@ A **binary configuration** is a collection of
 * release or nightly
     * releases are stable, nightlies are beta and built every night
 * python version
-    * linux: 3.5m, 3.6m 3.7m (mu is wide unicode or something like that. It usually doesn't matter but you should know that it exists)
-    * macos: 3.6, 3.7, 3.8
-    * windows: 3.6, 3.7, 3.8
+    * linux: 3.7m (mu is wide unicode or something like that. It usually doesn't matter but you should know that it exists)
+    * macos: 3.7, 3.8
+    * windows: 3.7, 3.8
 * cpu version
     * cpu, cuda 9.0, cuda 10.0
     * The supported cuda versions occasionally change
@@ -428,7 +428,7 @@ docker run \
 # possibly need are in .circleci/scripts/binary_populate_env.sh
 # You should probably always export at least these 3 variables
 export PACKAGE_TYPE=conda
-export DESIRED_PYTHON=3.6
+export DESIRED_PYTHON=3.7
 export DESIRED_CUDA=cpu
 
 # Call the entrypoint
@@ -476,7 +476,7 @@ conda activate binary
 # possibly need are in .circleci/scripts/binary_populate_env.sh
 # You should probably always export at least these 3 variables
 export PACKAGE_TYPE=conda
-export DESIRED_PYTHON=3.6
+export DESIRED_PYTHON=3.7
 export DESIRED_CUDA=cpu
 
 # Call the entrypoint you want

.circleci/cimodel/data/pytorch_build_definitions.py

@@ -334,13 +334,12 @@ def instantiate_configs(only_slow_gradcheck):
             build_only=build_only,
         )
 
-        # run docs builds on "pytorch-linux-xenial-py3.6-gcc5.4". Docs builds
+        # run docs builds on "pytorch-linux-xenial-py3.7-gcc5.4". Docs builds
         # should run on a CPU-only build that runs on all PRs.
-        # XXX should this be updated to a more modern build? Projects are
-        #     beginning to drop python3.6
+        # XXX should this be updated to a more modern build?
         if (
             distro_name == "xenial"
-            and fc.find_prop("pyver") == "3.6"
+            and fc.find_prop("pyver") == "3.7"
             and cuda_version is None
             and parallel_backend is None
             and not is_vulkan

.circleci/cimodel/data/simple/docker_definitions.py

@@ -26,7 +26,7 @@ def get_workflow_jobs(images=IMAGE_NAMES, only_slow_gradcheck=False):
             "name": quote(f"docker-{image_name}"),
             "image_name": quote(image_name),
         })
-        if image_name == "pytorch-linux-xenial-py3.6-gcc5.4":
+        if image_name == "pytorch-linux-xenial-py3.7-gcc5.4":
             # pushing documentation on tags requires CircleCI to also
             # build all the dependencies on tags, including this docker image
             parameters['filters'] = gen_filter_dict(branches_list=r"/.*/",

.circleci/docker/common/install_conda.sh

@@ -13,12 +13,7 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
       CONDA_FILE="Miniconda2-latest-Linux-x86_64.sh"
     ;;
     3)
-      if [ "$ANACONDA_PYTHON_VERSION" = "3.6" ]; then
-        # Latest release of Conda that still supports python-3.6
-        CONDA_FILE="Miniconda3-py37_4.10.3-Linux-x86_64.sh"
-      else
-        CONDA_FILE="Miniconda3-latest-Linux-x86_64.sh"
-      fi
+      CONDA_FILE="Miniconda3-latest-Linux-x86_64.sh"
     ;;
     *)
       echo "Unsupported ANACONDA_PYTHON_VERSION: $ANACONDA_PYTHON_VERSION"
@@ -61,9 +56,7 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
   pushd /opt/conda
 
   # Track latest conda update
-  if [ "$ANACONDA_PYTHON_VERSION" != "3.6" ]; then
-    as_jenkins conda update -y -n base conda
-  fi
+  as_jenkins conda update -y -n base conda
 
   # Install correct Python version
   as_jenkins conda install -y python="$ANACONDA_PYTHON_VERSION"

torch/_jit_internal.py

@@ -1217,11 +1217,8 @@ def _isinstance(obj, target_type) -> bool:
     if origin_type:
         return container_checker(obj, target_type)
 
-    # Check to handle weird python type behaviors
-    # 1. python 3.6 returns None for origin of containers without
-    #    contained type (intead of returning outer container type)
-    # 2. non-typed optional origin returns as none instead
-    #    of as optional in 3.6-3.8
+    # Check to handle non-typed optional origin returns as none instead
+    #    of as optional in 3.7-3.8
     check_args_exist(target_type)
 
     # handle non-containers

torch/csrc/utils/python_compat.h

@@ -2,33 +2,9 @@
 
 #include <torch/csrc/python_headers.h>
 
-#if PY_VERSION_HEX < 0x03070000
-// METH_FASTCALL was introduced in Python 3.7, so we wrap _PyCFunctionFast
-// signatures for earlier versions.
-
-template <PyObject* (*f)(PyObject*, PyObject *const *, Py_ssize_t)>
-PyObject* maybe_wrap_fastcall(PyObject *module, PyObject *args) {
-  return f(
-    module,
-
-    // _PyTuple_ITEMS
-    //   Because this is only a compat shim for Python 3.6, we don't have
-    //   to worry about the representation changing.
-    ((PyTupleObject *)args)->ob_item,
-    PySequence_Fast_GET_SIZE(args)
-  );
-}
-
-#define MAYBE_METH_FASTCALL METH_VARARGS
-#define MAYBE_WRAP_FASTCALL(f) maybe_wrap_fastcall<f>
-
-#else
-
 #define MAYBE_METH_FASTCALL METH_FASTCALL
 #define MAYBE_WRAP_FASTCALL(f) (PyCFunction)(void(*)(void))f
 
-#endif
-
 // PyPy 3.6 does not yet have PySlice_Unpack
 #if PY_VERSION_HEX < 0x03060100 || defined(PYPY_VERSION)
 

torch/distributed/algorithms/join.py

@@ -297,9 +297,6 @@ class Join():
         """
         ones = torch.ones(1, device=self._device)
         dist.all_reduce(ones, group=self._process_group)
-        # NOTE: Raising `StopIteration` does not throw an error in Python 3.6
-        # and throws a `RuntimeError` in Python 3.7+ (PEP 479), so we just
-        # raise a `RuntimeError` here
         raise RuntimeError(f"Rank {self._rank} exhausted all inputs.")
 
     @staticmethod

torch/distributed/elastic/multiprocessing/api.py

@@ -14,7 +14,7 @@ import signal
 import subprocess
 import sys
 import time
-from contextlib import AbstractContextManager
+from contextlib import nullcontext
 from dataclasses import dataclass, field
 from enum import IntFlag
 from multiprocessing import synchronize
@@ -334,22 +334,9 @@ class PContext(abc.ABC):
             self._stderr_tail.stop()
 
 
-class _nullcontext(AbstractContextManager):
-    # TODO remove and replace in favor of contextlib.nullcontext
-    # when torch drops support for python3.6
-    def __init__(self, enter_result=None):
-        self.enter_result = enter_result
-
-    def __enter__(self):
-        return self.enter_result
-
-    def __exit__(self, *excinfo):
-        pass
-
-
 def get_std_cm(std_rd: str, redirect_fn):
     if IS_WINDOWS or IS_MACOS or not std_rd:
-        return _nullcontext()
+        return nullcontext()
     else:
         return redirect_fn(std_rd)
 

torch/futures/__init__.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 from typing import cast, Callable, Generic, List, Optional, Type, TypeVar, Union
 
 import torch
@@ -90,8 +92,7 @@ class Future(torch._C.Future, Generic[T], metaclass=_PyFutureMeta):
         """
         return super().value()
 
-    # Have to use string annotations because  PEP-0563 is not available in 3.6
-    def then(self, callback):  # type: (Callable[[Future[T]], S]) -> Future[S]
+    def then(self, callback: Callable[[Future[T]], S]) -> Future[S]:
         r"""
         Append the given callback function to this ``Future``, which will be run
         when the ``Future`` is completed.  Multiple callbacks can be added to
@@ -154,8 +155,7 @@ class Future(torch._C.Future, Generic[T], metaclass=_PyFutureMeta):
         """
         return cast(Future[S], super().then(callback))
 
-    # Have to use string annotations because  PEP-0563 is not available in 3.6
-    def add_done_callback(self, callback):  # type: (Callable[[Future[T]], None]) -> None
+    def add_done_callback(self, callback: Callable[[Future[T]], None]) -> None:
         r"""
         Append the given callback function to this ``Future``, which will be run
         when the ``Future`` is completed.  Multiple callbacks can be added to

torch/fx/node.py

@@ -49,11 +49,7 @@ def _type_repr(obj):
     typically enough to uniquely identify a type.  For everything
     else, we fall back on repr(obj).
     """
-    # HACK: In Python 3.6, type aliases from ``typing`` are instances of ``type``, but in
-    # later Python versions, type aliases are not instances of ``type``!! We want
-    # all type aliases to fall through to ``repr``, so if we have a type that is
-    # in the module typing, don't go down this path.
-    if isinstance(obj, type) and obj.__module__ != 'typing':
+    if isinstance(obj, type):
         if obj.__module__ == 'builtins':
             return obj.__qualname__
         return f'{obj.__module__}.{obj.__qualname__}'

torch/nn/modules/module.py

@@ -8,6 +8,7 @@ from ..parameter import Parameter
 import torch.utils.hooks as hooks
 
 from torch import Tensor, device, dtype
+import typing
 from typing import Union, Tuple, Any, Callable, Iterator, Set, Optional, overload, TypeVar, Mapping, Dict, List
 from ...utils.hooks import RemovableHandle
 
@@ -1288,7 +1289,7 @@ class Module:
     # TODO: Remove string escape once Python-3.6 no longer supported
     # See https://github.com/python/mypy/issues/6904#issuecomment-496207426
     @overload
-    def state_dict(self, prefix: str = ..., keep_vars: bool = ...) -> 'OrderedDict[str, Tensor]':
+    def state_dict(self, prefix: str = ..., keep_vars: bool = ...) -> typing.OrderedDict[str, Tensor]:
         ...
 
     def state_dict(self, destination=None, prefix='', keep_vars=False):

torch/overrides.py

@@ -1436,9 +1436,7 @@ has_torch_function_variadic = _add_docstr(
     _has_torch_function_variadic,
     r"""Special case of `has_torch_function` that skips tuple creation.
 
-    This uses the METH_FASTCALL protocol introduced in Python 3.7; for 3.6
-    and before it has roughly equivilent performance compared to
-    `has_torch_function`.
+    This uses the METH_FASTCALL protocol introduced in Python 3.7
 
     Instead of:
       `has_torch_function((a, b))`

torch/utils/data/_typing.py

@@ -258,7 +258,6 @@ class _DataPipeMeta(GenericMeta):
 
         return super().__new__(cls, name, bases, namespace, **kwargs)  # type: ignore[call-overload]
 
-        # For Python > 3.6
         cls.__origin__ = None
         if 'type' in namespace:
             return super().__new__(cls, name, bases, namespace, **kwargs)  # type: ignore[call-overload]