Upgrade mypy to version 0.812 (#55712)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/54211

This was a little more annoying than expected, because the `exclude = ` key in `mypy.ini` is weird. I'll file an upstream issue about that.

I ignored one file, `torch/distributed/elastic/agent/server/api.py` that had ~8 errors that were hard to figure out. This can be done in a follow-up.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/55712

Reviewed By: walterddr

Differential Revision: D27694976

Pulled By: malfet

fbshipit-source-id: 228d8be6af040343ce46595dabaca212e69ccc68

.circleci/docker/common/install_conda.sh

@@ -113,7 +113,7 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
     boto3==1.16.34 \
     coverage \
     hypothesis==4.53.2 \
-    mypy==0.770 \
+    mypy==0.812 \
     tb-nightly
 
   # Update scikit-learn to a python-3.8 compatible version

.github/workflows/lint.yml

@@ -293,7 +293,7 @@ jobs:
         run: |
           set -eux
           pip install -r requirements.txt
-          pip install mypy==0.770
+          pip install mypy==0.812
       - name: Run autogen
         run: |
           set -eux

.github/workflows/test_tools.yml

@@ -25,6 +25,6 @@ jobs:
         run: |
           set -eux
           pip install -r requirements.txt
-          pip install boto3==1.16.34 mypy==0.770
+          pip install boto3==1.16.34 mypy==0.812
       - name: Run tests
         run: python -m unittest discover -vs tools/test -p 'test_*.py'

caffe2/contrib/fakelowp/test/test_chunking.py

@@ -29,7 +29,7 @@ class Fusions(serial.SerializedTestCase):
         tensor_min = min(0, np.min(tensor))
         scale = np.float32(np.float16((tensor_max - tensor_min) / 255.0))
         if scale < 1e-6:
-            scale = 1e-6
+            scale = np.float32(1e-6)
         zero_point = 0 - tensor_min / scale
         zero_point = int(round(np.clip(zero_point, 0, 255.0)))
         return (scale, zero_point)

caffe2/contrib/fakelowp/test/test_fc_nnpi_fp16.py

@@ -186,8 +186,7 @@ class FCTest(serial.SerializedTestCase):
                         0.0680542, 0.4255371, -0.42895508, -0.4128418,
                         -0.47973633, 0.33251953,
                         0.27807617, 0.3701172]], dtype=np.float32)
-        b0 = [0.47851562]
-        b0 = np.array(b0, dtype=np.float32)
+        b0 = np.array([0.47851562], dtype=np.float32)
 
         workspace.FeedBlob("W0", W0)
         workspace.FeedBlob("b0", b0)
@@ -252,7 +251,6 @@ class FCTest(serial.SerializedTestCase):
                     "m": m,
                     "k": k,
                     "n": n,
-                    "X": X0,
                     "W0": W0,
                     "b0": b0,
                     "Y_glow": Y_glow,

caffe2/contrib/fakelowp/test/test_int8_ops_nnpi.py

@@ -21,7 +21,7 @@ class Int8OpsTest(serial.SerializedTestCase):
         tensor_min = min(0, np.min(tensor))
         scale = np.float32(np.float16((tensor_max - tensor_min) / 255.0))
         if scale < 1e-6:
-            scale = 1e-6
+            scale = np.float32(1e-6)
         zero_point = 0 - tensor_min / scale
         zero_point = int(round(np.clip(zero_point, 0, 255.0)))
         return (scale, zero_point)

caffe2/contrib/fakelowp/test/test_layernorm_nnpi_fp16.py

@@ -122,7 +122,7 @@ class LayerNorm(serial.SerializedTestCase):
         tensor_min = min(0, np.min(tensor))
         scale = np.float32(np.float16((tensor_max - tensor_min) / 255.0))
         if scale < 1e-6:
-            scale = 1e-6
+            scale = np.float32(1e-6)
         zero_point = 0 - tensor_min / scale
         zero_point = int(round(np.clip(zero_point, 0, 255.0)))
         return (scale, zero_point)

caffe2/contrib/fakelowp/test/test_sls_4bit_nnpi_fp16.py

@@ -122,10 +122,10 @@ class SparseLengthsSum4BitFakeNNPIFp16Test(serial.SerializedTestCase):
         data = data * 1e-3
 
         lengths = np.random.choice(np.arange(1, num_rows), batch_size).astype(np.int32)
-        indices = []
+        _indices = []
         for length in lengths:
-            indices.extend(np.random.choice(np.arange(1, num_rows), length))
-        indices = np.asarray(indices).astype(np.int64)
+            _indices.extend(np.random.choice(np.arange(1, num_rows), length))
+        indices = np.asarray(_indices).astype(np.int64)
 
         weights = np.random.uniform(
             low=0,

caffe2/contrib/fakelowp/test/test_sls_8bit_nnpi_fp16.py

@@ -1,4 +1,5 @@
 import unittest
+from typing import Dict, Any
 
 # Must happen before importing caffe2.python.*
 import caffe2.python.fakelowp.init_shared_libs  # noqa
@@ -41,7 +42,7 @@ class SparseLengthsSum8BitFakeNNPIFp16Test(serial.SerializedTestCase):
         fp16_c2_net = core.Net("test_fp16_c2")
         fp16_c2_net.SparseLengthsSumFakeFP16AccFP16(["D", "I", "L"], "fp16_out")
 
-        input_dict = {}
+        input_dict : Dict[Any, Any] = {}
 
         pred_net = caffe2_pb2.NetDef()
         pred_net.name = "pred"
@@ -214,10 +215,10 @@ class SparseLengthsSum8BitFakeNNPIFp16Test(serial.SerializedTestCase):
         data = np.random.rand(num_rows, embedding_dim).astype(np.float32)
         lengths = np.random.choice(np.arange(1, num_rows), batch_size).astype(np.int32)
 
-        indices = []
+        _indices = []
         for length in lengths:
-            indices.extend(np.random.choice(np.arange(1, num_rows), length))
-        indices = np.asarray(indices).astype(np.int64)
+            _indices.extend(np.random.choice(np.arange(1, num_rows), length))
+        indices = np.asarray(_indices).astype(np.int64)
 
         weights = np.random.uniform(
             low=0,

caffe2/contrib/fakelowp/test/test_sls_8bit_nnpi_fp32.py

@@ -49,10 +49,10 @@ class SparseLengthsSum8BitFakeNNPIFp32Test(serial.SerializedTestCase):
         data = np.random.rand(num_rows, embedding_dim).astype(np.float32)
         lengths = np.random.choice(np.arange(1, num_rows), batch_size).astype(np.int32)
 
-        indices = []
+        _indices = []
         for length in lengths:
-            indices.extend(np.random.choice(np.arange(1, num_rows), length))
-        indices = np.asarray(indices).astype(np.int64)
+            _indices.extend(np.random.choice(np.arange(1, num_rows), length))
+        indices = np.asarray(_indices).astype(np.int64)
 
         weights = np.random.uniform(
             low=0,
@@ -246,9 +246,6 @@ class SparseLengthsSum8BitFakeNNPIFp32Test(serial.SerializedTestCase):
                 "test_small_sls_acc32",
                 {
                     "seed": seed,
-                    "num_rows": num_rows,
-                    "embedding_dim": embedding_dim,
-                    "batch_size": batch_size,
                     "indices": indices,
                     "data": data,
                     "quantized_data": quantized_data,

mypy.ini

@@ -38,6 +38,10 @@ files =
     tools/generate_torch_version.py,
     tools/stats_utils/*.py
 
+#
+# `exclude` is a regex, not a list of paths like `files` (sigh)
+#
+exclude = torch/include/|torch/csrc/|torch/distributed/elastic/agent/server/api.py
 
 # Minimum version supported - variable annotations were introduced
 # in Python 3.6

tools/print_test_stats.py

@@ -15,7 +15,7 @@ from glob import glob
 from pathlib import Path
 from typing import (Any, DefaultDict, Dict, Iterable, Iterator, List, Optional,
                     Set, Tuple, cast)
-from xml.dom import minidom  # type: ignore[import]
+from xml.dom import minidom
 
 import requests
 from typing_extensions import TypedDict

torch/distributed/benchmarks/benchmark_ddp_rpc.py

@@ -101,7 +101,7 @@ def _print_cont(msg):
 
 
 def _run_printable(cmd):
-    proc = subprocess.run(shlex.split(cmd), capture_output=True)
+    proc = subprocess.run(shlex.split(cmd), capture_output=True)  # type: ignore
     assert proc.returncode == 0
 
     buffer = io.BytesIO()
@@ -212,9 +212,12 @@ def run_worker(rank, world_size):
     if rank == (NUM_TRAINERS + NUM_PS):
 
         rpc.init_rpc(
-            "master", rank=rank, backend=BackendType.TENSORPIPE, world_size=world_size
+            "master", rank=rank,
+            backend=BackendType.TENSORPIPE,  # type: ignore[attr-defined]
+            world_size=world_size
         )
 
+
         # Build the Embedding tables on the Parameter Servers.
         emb_rref_list = []
         index = 0
@@ -278,7 +281,7 @@ def run_worker(rank, world_size):
             ps_name,
             rank=rank,
             world_size=world_size,
-            backend=BackendType.TENSORPIPE,
+            backend=BackendType.TENSORPIPE,  # type: ignore[attr-defined]
             rpc_backend_options=rpc_backend_options,
         )
         # parameter server do nothing

torch/distributed/elastic/agent/server/api.py

@@ -95,6 +95,7 @@ class WorkerSpec:
         if isinstance(self.entrypoint, str):
             return os.path.basename(self.entrypoint)
         else:
+            assert self.entrypoint is not None
             return self.entrypoint.__qualname__
 
 

torch/distributed/elastic/agent/server/local_elastic_agent.py

@@ -20,7 +20,7 @@ from torch.distributed.elastic.agent.server.api import (
     WorkerState,
 )
 from torch.distributed.elastic.metrics.api import prof
-from torch.distributed.elastic.multiprocessing import start_processes
+from torch.distributed.elastic.multiprocessing import start_processes, PContext
 
 
 log = logging.getLogger(__name__)
@@ -104,7 +104,7 @@ class LocalElasticAgent(SimpleElasticAgent):
     ):
         super().__init__(spec, exit_barrier_timeout)
         self._start_method = start_method
-        self._pcontext = None
+        self._pcontext : Optional[PContext] = None
         rdzv_run_id = spec.rdzv_handler.get_run_id()
         self._log_dir = self._make_log_dir(log_dir, rdzv_run_id)
 
@@ -161,6 +161,7 @@ class LocalElasticAgent(SimpleElasticAgent):
         shutil.rmtree(attempt_log_dir, ignore_errors=True)
         os.makedirs(attempt_log_dir)
 
+        assert spec.entrypoint is not None
         self._pcontext = start_processes(
             name=spec.role,
             entrypoint=spec.entrypoint,
@@ -184,6 +185,7 @@ class LocalElasticAgent(SimpleElasticAgent):
     def _monitor_workers(self, worker_group: WorkerGroup) -> RunResult:
         role = worker_group.spec.role
         worker_pids = {w.id for w in worker_group.workers}
+        assert self._pcontext is not None
         pc_pids = set(self._pcontext.pids().values())
         if worker_pids != pc_pids:
             log.error(

torch/jit/_recursive.py

@@ -15,7 +15,7 @@ from torch.nn import Module
 
 
 ScriptMethodStub = collections.namedtuple('ScriptMethodStub', ('resolution_callback', 'def_', 'original_method'))
-PropertyStub = collections.namedtuple('Property', ('resolution_callback', 'def_'))
+PropertyStub = collections.namedtuple('PropertyStub', ('resolution_callback', 'def_'))
 
 
 # TODO: there should be a more principled way of doing this.

torch/nn/intrinsic/qat/modules/conv_fused.py

@@ -191,8 +191,10 @@ class _ConvBnNd(nn.modules.conv._ConvNd, nni._FusedModule):
             Args: `mod` a float module, either produced by torch.quantization utilities
             or directly from user
         """
+        # The ignore is because _FLOAT_MODULE is a TypeVar here where the bound
+        # has no __name__ (code is fine though)
         assert type(mod) == cls._FLOAT_MODULE, 'qat.' + cls.__name__ + '.from_float only works for ' + \
-            cls._FLOAT_MODULE.__name__
+            cls._FLOAT_MODULE.__name__  # type: ignore
         assert hasattr(mod, 'qconfig'), 'Input float module must have qconfig defined'
         assert mod.qconfig, 'Input float module must have a valid qconfig'
         qconfig = mod.qconfig

torch/nn/modules/adaptive.py

@@ -12,7 +12,7 @@ from .module import Module
 from ..functional import log_softmax
 
 
-_ASMoutput = namedtuple('ASMoutput', ['output', 'loss'])
+_ASMoutput = namedtuple('_ASMoutput', ['output', 'loss'])
 
 
 class AdaptiveLogSoftmaxWithLoss(Module):

torch/nn/quantized/modules/conv.py

@@ -549,7 +549,7 @@ class _ConvTransposeNd(_ConvNd):
         """
         # derived classes override cls._FLOAT_MODULE attribute
         msg = ' nnq.' + cls.__name__ + '.from_float only works for ' + \
-              cls._FLOAT_MODULE.__name__
+              cls._FLOAT_MODULE.__name__  # type: ignore
         assert type(mod) == cls._FLOAT_MODULE, msg
         assert hasattr(mod, 'qconfig'), \
             'Input float module must have qconfig defined.'

torch/nn/utils/rnn.pyi

@@ -3,7 +3,7 @@ from typing import Any, Optional, overload, Union, TypeVar, Tuple, Sequence
 from torch import Tensor
 from torch.types import _dtype, _device
 
-PackedSequence_ = namedtuple('PackedSequence', ['data', 'batch_sizes', 'sorted_indices', 'unsorted_indices'])
+PackedSequence_ = namedtuple('PackedSequence_', ['data', 'batch_sizes', 'sorted_indices', 'unsorted_indices'])
 
 
 def bind(optional: Any, fn: Any): ...

torch/testing/_asserts.py

@@ -154,7 +154,7 @@ def _equalize_attributes(a: torch.Tensor, b: torch.Tensor) -> Tuple[torch.Tensor
     return a, b
 
 
-_Trace = namedtuple("Trace", ("total", "abs", "rel", "idx", "diff", "a", "b"))
+_Trace = namedtuple("_Trace", ("total", "abs", "rel", "idx", "diff", "a", "b"))
 
 
 def _trace_mismatches(a: torch.Tensor, b: torch.Tensor, mismatches: torch.Tensor) -> _Trace:

torch/testing/_internal/common_device_type.py

@@ -461,7 +461,8 @@ def filter_desired_device_types(device_type_test_bases, except_for=None, only_fo
 _TORCH_TEST_DEVICES = os.environ.get('TORCH_TEST_DEVICES', None)
 if _TORCH_TEST_DEVICES:
     for path in _TORCH_TEST_DEVICES.split(':'):
-        mod = runpy.run_path(path, init_globals=globals())
+        # runpy (a stdlib module) lacks annotations
+        mod = runpy.run_path(path, init_globals=globals())  # type: ignore[func-returns-value]
         device_type_test_bases.append(mod['TEST_CLASS'])
 
 

torch/utils/benchmark/examples/sparse/fuzzer.py

@@ -39,7 +39,7 @@ def main():
             [
                 benchmark_utils.FuzzedSparseTensor(
                     name=name,
-                    size=[f"k{i}" for i in range(3)],
+                    size=tuple([f"k{i}" for i in range(3)]),
                     min_elements=64 * 1024,
                     max_elements=128 * 1024,
                     sparse_dim="sparse_dim",

torch/utils/data/_typing.py

@@ -86,14 +86,15 @@ def _decompose_type(t, to_list=True):
             ts = [t.__bound__]
         else:
             # For T_co, __constraints__ is ()
-            ts = t.__constraints__
+            ts = list(t.__constraints__)
     elif hasattr(t, '__origin__') and t.__origin__ == Union:
         ts = t.__args__
     else:
         if not to_list:
             return None
         ts = [t]
-    ts = list(TYPE2ABC.get(_t, _t) for _t in ts)
+    # Ignored: Generator has incompatible item type "object"; expected "Type[Any]"
+    ts = list(TYPE2ABC.get(_t, _t) for _t in ts)  # type: ignore[misc]
     return ts