2023-05-01 nightly release (e779a30)

pytorch · ezyang · Sep 11, 2022 · Sep 12, 2022 · Sep 13, 2022 · Sep 14, 2022
commit f38a04a5c5ce632ba8b0534e3689e89f98e89d73
diff --git a/.flake8 b/.flake8
@@ -20,7 +20,7 @@ ignore =
     # these ignores are from flake8-logging-format; please fix!
     G100,G101,G200,G201,G202
     # these ignores are from flake8-simplify. please fix or ignore with commented reason
-    SIM105,SIM108,SIM109,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
+    SIM105,SIM108,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
     # flake8-simplify code styles
     SIM102,SIM103,SIM106,SIM112,
 per-file-ignores =

@@ -250,10 +250,7 @@ def remove_disabled_jobs(
             )
             return filtered_test_matrix
 
-        if (
-            disabled_job_cfg == TEST_JOB_NAME
-            or disabled_job_cfg == BUILD_AND_TEST_JOB_NAME
-        ):
+        if disabled_job_cfg in (TEST_JOB_NAME, BUILD_AND_TEST_JOB_NAME):
             print(
                 f"Issue {disabled_url} created by {author} has disabled all the test jobs for {workflow} / {job_name}"
             )
@@ -263,7 +260,7 @@ def remove_disabled_jobs(
         if m:
             disabled_job = m.group("job")
             # Make sure that the job name is a valid test job name first before checking the config
-            if disabled_job == TEST_JOB_NAME or disabled_job == BUILD_AND_TEST_JOB_NAME:
+            if disabled_job in (TEST_JOB_NAME, BUILD_AND_TEST_JOB_NAME):
                 disabled_cfg = m.group("cfg")
                 # Remove the disabled config from the test matrix
                 filtered_test_matrix["include"] = [

diff --git a/benchmarks/dynamo/ci_expected_accuracy/inductor_huggingface_dynamic_inference.csv b/benchmarks/dynamo/ci_expected_accuracy/inductor_huggingface_dynamic_inference.csv
@@ -1,7 +1,7 @@
 name,accuracy,graph_breaks
 AlbertForMaskedLM,pass,0
 AlbertForQuestionAnswering,pass,0
-AllenaiLongformerBase,pass,152
+AllenaiLongformerBase,pass,136
 BartForCausalLM,pass,0
 BertForMaskedLM,pass,0
 BertForQuestionAnswering,pass,0

diff --git a/pyproject.toml b/pyproject.toml
@@ -30,7 +30,6 @@ ignore = [
     "B007", "B008", "B017",
     "B018", # Useless expression
     "B019", "B020",
-    "B022", # Allow empty context manager
     "B023", "B024", "B026",
     "B028", # No explicit `stacklevel` keyword argument found
     "B027", "B904", "B905",
@@ -49,7 +48,6 @@ ignore = [
     "SIM102", "SIM103", "SIM112", # flake8-simplify code styles
     "SIM105", # these ignores are from flake8-simplify. please fix or ignore with commented reason
     "SIM108",
-    "SIM109",
     "SIM110",
     "SIM114", # Combine `if` branches using logical `or` operator
     "SIM115",

diff --git a/test/distributed/algorithms/test_join.py b/test/distributed/algorithms/test_join.py
@@ -250,7 +250,7 @@ def _test_join_base(
         with self.assertRaisesRegex(
             RuntimeError,
             expected_msg
-        ) if throw_on_early_termination else contextlib.suppress():
+        ) if throw_on_early_termination else contextlib.nullcontext():
             with Join(
                 allreducers,
                 enable=enable,

diff --git a/test/distributed/fsdp/test_fsdp_checkpoint.py b/test/distributed/fsdp/test_fsdp_checkpoint.py
@@ -257,7 +257,7 @@ def test_basic_checkpoint_end_to_end(
                         offload_ctx = (
                             get_patched_save_on_cpu()(pin_memory=True)
                             if offload_activations
-                            else contextlib.suppress()
+                            else contextlib.nullcontext()
                         )
                         with offload_ctx:
                             out = checkpoint(m, inp, use_reentrant=True)

diff --git a/test/distributed/fsdp/test_fsdp_comm.py b/test/distributed/fsdp/test_fsdp_comm.py
@@ -1,7 +1,7 @@
 # Owner(s): ["oncall: distributed"]
 
 import sys
-from contextlib import suppress
+from contextlib import nullcontext
 from enum import auto, Enum
 from typing import Optional
 from unittest.mock import patch
@@ -76,7 +76,7 @@ def _init_model(
 
     def _run_iter(self, fsdp_model, batch, use_no_sync: bool):
         """Runs an iteration inside or outside the ``no_sync()`` context."""
-        context = fsdp_model.no_sync() if use_no_sync else suppress()
+        context = fsdp_model.no_sync() if use_no_sync else nullcontext()
         with context:
             output = fsdp_model(*batch)
             loss = fsdp_model.module.get_loss(batch, output)

diff --git a/test/distributed/fsdp/test_fsdp_exec_order.py b/test/distributed/fsdp/test_fsdp_exec_order.py
@@ -2,7 +2,7 @@
 
 import sys
 import warnings
-from contextlib import suppress
+from contextlib import nullcontext
 
 import torch
 from torch import distributed as dist
@@ -159,7 +159,7 @@ def test_invalid_later_iter_order(
                 expected_regex=regex,
             )
             if self.rank != 0
-            else suppress()
+            else nullcontext()
         )
         if self.rank != 0:
             fsdp_model.flip_path()

diff --git a/test/distributed/fsdp/test_fsdp_grad_acc.py b/test/distributed/fsdp/test_fsdp_grad_acc.py
@@ -183,7 +183,7 @@ def permute_tensor(x: torch.Tensor):
         batch_idx = 0
         for config in configs:
             sync_context = (
-                fsdp_model.no_sync() if config.use_no_sync else contextlib.suppress()
+                fsdp_model.no_sync() if config.use_no_sync else contextlib.nullcontext()
             )
             with sync_context:
                 for _ in range(config.num_iters):

diff --git a/test/distributed/fsdp/test_fsdp_misc.py b/test/distributed/fsdp/test_fsdp_misc.py
@@ -5,7 +5,7 @@
 import sys
 import warnings
 from collections import namedtuple
-from contextlib import suppress
+from contextlib import nullcontext
 from copy import deepcopy
 from typing import Any, Tuple
 
@@ -371,7 +371,7 @@ def test_module_device_mismatches_device_id(self):
         context = (
             self.assertRaisesRegex(ValueError, f"cuda:{self.rank} vs cuda:0")
             if self.rank != 0
-            else suppress()
+            else nullcontext()
         )
         with context:
             NestedWrappedModule.init(
@@ -427,7 +427,7 @@ def test_no_params(self):
                 )
             )
             if self.rank != 0
-            else suppress()
+            else nullcontext()
         )
         with context:
             module = FSDP(no_params, device_id=0)

diff --git a/test/distributed/fsdp/test_fsdp_state_dict.py b/test/distributed/fsdp/test_fsdp_state_dict.py
@@ -3,7 +3,7 @@
 import io
 import itertools
 import sys
-from contextlib import suppress
+from contextlib import nullcontext
 from copy import deepcopy
 from functools import partial
 from typing import Any, Dict
@@ -899,7 +899,7 @@ def test_state_dict_skip_module(self, state_dict_type, double_nest):
 
         def _create_module(wrap_fsdp=True):
             LINEAR_SKIP = "linear_skip"
-            ctx = enable_wrap(wrapper_cls=FSDP) if wrap_fsdp else suppress()
+            ctx = enable_wrap(wrapper_cls=FSDP) if wrap_fsdp else nullcontext()
             with ctx:
                 module = SkipModel(double_nest=double_nest)
                 # Full name of linear_skip param tensors in SkipModel, as would be

diff --git a/test/distributed/fsdp/test_fsdp_unshard_params.py b/test/distributed/fsdp/test_fsdp_unshard_params.py
@@ -522,7 +522,7 @@ def _check_grads(
 
         def _get_error_context(is_supported: bool):
             return (
-                contextlib.suppress()
+                contextlib.nullcontext()
                 if is_supported
                 else self.assertRaises(NotImplementedError)
             )  # some configs are not implemented yet

diff --git a/test/distributed/optim/test_zero_redundancy_optimizer.py b/test/distributed/optim/test_zero_redundancy_optimizer.py
@@ -9,7 +9,7 @@
 import os
 import sys
 import unittest
-from contextlib import suppress
+from contextlib import nullcontext
 from typing import Any, cast, List
 
 import numpy as np
@@ -301,7 +301,7 @@ def test_constructor(self):
             (list(m.parameters()), None),  # `params` as a list
         ]
         for ctor_input, error in ctor_inputs:
-            context = self.assertRaises(error) if error else suppress()
+            context = self.assertRaises(error) if error else nullcontext()
             with context:
                 ZeroRedundancyOptimizer(
                     ctor_input,
@@ -371,7 +371,7 @@ def world_size(self):
     @property
     def context(self):
         return (
-            suppress()
+            nullcontext()
             if not torch.cuda.is_available()
             else torch.cuda.device(self.rank)
         )

diff --git a/test/distributed/test_c10d_common.py b/test/distributed/test_c10d_common.py
@@ -7,7 +7,7 @@
 import tempfile
 import threading
 import time
-from contextlib import suppress
+from contextlib import nullcontext
 from dataclasses import dataclass
 from datetime import timedelta
 from itertools import product
@@ -545,7 +545,7 @@ def test_ddp_checkpointing_unused_params(self, use_reentrant):
         process_group = self._get_process_group()
         for use_bucket_view in (True, False):
             err_ctx = (
-                suppress() if not use_reentrant else
+                nullcontext() if not use_reentrant else
                 self.assertRaisesRegex(
                     RuntimeError,
                     "Expected to mark a variable ready only once."
@@ -577,7 +577,7 @@ def test_ddp_checkpointing_twice(self, use_reentrant):
         process_group = self._get_process_group()
         for use_bucket_view in (True, False):
             err_ctx = (
-                suppress() if not use_reentrant else
+                nullcontext() if not use_reentrant else
                 self.assertRaisesRegex(
                     RuntimeError,
                     "Expected to mark a variable ready only once."
@@ -678,7 +678,7 @@ def test_ddp_checkpointing_weight_sharing(self, use_reentrant):
                 self.assertRaisesRegex(
                     RuntimeError,
                     "Your training graph has changed in this iteration"
-                ) if static_graph and not use_reentrant else suppress()
+                ) if static_graph and not use_reentrant else nullcontext()
             )
             with err_ctx:
                 self._test_ddp_checkpointing(