Drop torch 1.6 testing (#10390)

* Drop torch 1.6 support

* Drop 1.6 support

* Update CHANGELOG

* Fixes

* Split change

* Undo change

* 1.7 -> 1.7.1

pytorch/pytorch#47354

* Force trigger nightly

* Update .github/workflows/events-nightly.yml

Co-authored-by: Aki Nitta <[email protected]>

* Revert 1.7.1 change - try wildcard

* Update adjust versions and test it

* Undo test changes

* Revert "Undo test changes"

This reverts commit 3a6acad.

* Update CHANGELOG.md

Co-authored-by: Aki Nitta <[email protected]>

Author: carmocca

pytorch_lightning/callbacks/quantization.py

@@ -28,7 +28,7 @@
 if _TORCH_GREATER_EQUAL_1_8:
     from torch.quantization import FakeQuantizeBase
 else:
-    # For torch 1.6 and 1.7.
+    # For torch 1.7.
     from torch.quantization import FakeQuantize as FakeQuantizeBase
 
 import pytorch_lightning as pl

pytorch_lightning/distributed/dist.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 from typing import Any
 
-from pytorch_lightning.overrides.torch_distributed import broadcast_object_list
+import torch.distributed
+
 from pytorch_lightning.utilities import rank_zero_deprecation
 from pytorch_lightning.utilities.distributed import group as _group
 
@@ -40,6 +41,6 @@ def broadcast(self, obj: Any, group=_group.WORLD):
         if self.rank != 0:
             obj = [None] * len(obj)
 
-        broadcast_object_list(obj, 0, group=group or _group.WORLD)
+        torch.distributed.broadcast_object_list(obj, 0, group=group or _group.WORLD)
 
         return obj[0]

pytorch_lightning/overrides/torch_distributed.py

@@ -34,7 +34,6 @@
 from pytorch_lightning.core.optimizer import LightningOptimizer
 from pytorch_lightning.overrides import LightningDistributedModule
 from pytorch_lightning.overrides.distributed import prepare_for_backward
-from pytorch_lightning.overrides.torch_distributed import broadcast_object_list
 from pytorch_lightning.plugins.environments.cluster_environment import ClusterEnvironment
 from pytorch_lightning.plugins.io.checkpoint_plugin import CheckpointIO
 from pytorch_lightning.plugins.training_type.parallel import ParallelPlugin
@@ -43,7 +42,6 @@
     _FAIRSCALE_AVAILABLE,
     _HYDRA_AVAILABLE,
     _IS_WINDOWS,
-    _TORCH_GREATER_EQUAL_1_7,
     _TORCH_GREATER_EQUAL_1_8,
     _TORCH_GREATER_EQUAL_1_9,
     _TORCH_GREATER_EQUAL_1_10,
@@ -255,15 +253,13 @@ def pre_configure_ddp(self):
         # when not all parameter backward hooks are fired by the autograd engine even if require_grad is set to True.
         # This flag does come with a performance hit, so it is suggested to disable in cases where it is possible.
         self._ddp_kwargs["find_unused_parameters"] = self._ddp_kwargs.get("find_unused_parameters", True)
-        # todo: PyTorch 1.7.0 DDP introduces `self.reducer._rebuild_buckets()` breaking manual_optimization
-        if (
-            _TORCH_GREATER_EQUAL_1_7
-            and not self.lightning_module.automatic_optimization
-            and not self._ddp_kwargs.get("find_unused_parameters", False)
+        if not self.lightning_module.automatic_optimization and not self._ddp_kwargs.get(
+            "find_unused_parameters", False
         ):
+            # TODO: PyTorch 1.7.0 DDP introduces `self.reducer._rebuild_buckets()` breaking manual_optimization
             rank_zero_warn(
-                "From PyTorch 1.7.0, Lightning ``manual_optimization`` needs to set ``find_unused_parameters=True`` "
-                "to properly work with DDP."
+                "From PyTorch 1.7.0, Lightning `manual_optimization` needs to set `find_unused_parameters=True` to"
+                " properly work with DDP. Using `find_unused_parameters=True`."
             )
             self._ddp_kwargs["find_unused_parameters"] = True
 
@@ -371,7 +367,7 @@ def broadcast(self, obj: object, src: int = 0) -> object:
         obj = [obj]
         if self.global_rank != src:
             obj = [None]
-        broadcast_object_list(obj, src, group=_group.WORLD)
+        torch.distributed.broadcast_object_list(obj, src, group=_group.WORLD)
         return obj[0]
 
     def pre_backward(self, closure_loss: torch.Tensor) -> None:

pytorch_lightning/plugins/training_type/ddp.py

@@ -27,12 +27,11 @@
 import pytorch_lightning as pl
 from pytorch_lightning.overrides import LightningDistributedModule
 from pytorch_lightning.overrides.distributed import prepare_for_backward
-from pytorch_lightning.overrides.torch_distributed import broadcast_object_list
 from pytorch_lightning.plugins.environments.cluster_environment import ClusterEnvironment
 from pytorch_lightning.plugins.io.checkpoint_plugin import CheckpointIO
 from pytorch_lightning.plugins.training_type.parallel import ParallelPlugin
 from pytorch_lightning.trainer.states import TrainerFn
-from pytorch_lightning.utilities import _TORCH_GREATER_EQUAL_1_7, _TORCH_GREATER_EQUAL_1_8, rank_zero_warn
+from pytorch_lightning.utilities import _TORCH_GREATER_EQUAL_1_8, rank_zero_warn
 from pytorch_lightning.utilities.apply_func import apply_to_collection, move_data_to_device
 from pytorch_lightning.utilities.cloud_io import atomic_save
 from pytorch_lightning.utilities.cloud_io import load as pl_load
@@ -238,15 +237,13 @@ def pre_configure_ddp(self):
         # when not all parameter backward hooks are fired by the autograd engine even if require_grad is set to True.
         # This flag does come with a performance hit, so it is suggested to disable in cases where it is possible.
         self._ddp_kwargs["find_unused_parameters"] = self._ddp_kwargs.get("find_unused_parameters", True)
-        # todo: PyTorch 1.7.0 DDP introduces `self.reducer._rebuild_buckets()` breaking manual_optimization
-        if (
-            _TORCH_GREATER_EQUAL_1_7
-            and not self.lightning_module.automatic_optimization
-            and not self._ddp_kwargs.get("find_unused_parameters", False)
+        if not self.lightning_module.automatic_optimization and not self._ddp_kwargs.get(
+            "find_unused_parameters", False
         ):
+            # TODO: PyTorch 1.7.0 DDP introduces `self.reducer._rebuild_buckets()` breaking manual_optimization
             rank_zero_warn(
-                "From PyTorch 1.7.0, Lightning ``manual_optimization`` needs to set ``find_unused_parameters=True`` "
-                "to properly work with DDP."
+                "From PyTorch 1.7.0, Lightning `manual_optimization` needs to set `find_unused_parameters=True` to"
+                " properly work with DDP. Using `find_unused_parameters=True`."
             )
             self._ddp_kwargs["find_unused_parameters"] = True
 
@@ -323,7 +320,7 @@ def broadcast(self, obj: object, src: int = 0) -> object:
         obj = [obj]
         if self.global_rank != src:
             obj = [None]
-        broadcast_object_list(obj, src, group=_group.WORLD)
+        torch.distributed.broadcast_object_list(obj, src, group=_group.WORLD)
         return obj[0]
 
     def model_to_device(self):

pytorch_lightning/plugins/training_type/ddp_spawn.py

@@ -74,7 +74,6 @@
 from pytorch_lightning.utilities.imports import (
     _HOROVOD_AVAILABLE,
     _IPU_AVAILABLE,
-    _TORCH_GREATER_EQUAL_1_7,
     _TORCH_GREATER_EQUAL_1_8,
     _TPU_AVAILABLE,
 )
@@ -190,10 +189,8 @@ def _init_deterministic(self, deterministic: bool) -> None:
         self.deterministic = deterministic
         if _TORCH_GREATER_EQUAL_1_8:
             torch.use_deterministic_algorithms(deterministic)
-        elif _TORCH_GREATER_EQUAL_1_7:
+        else:
             torch.set_deterministic(deterministic)
-        else:  # the minimum version Lightning supports is PyTorch 1.6
-            torch._set_deterministic(deterministic)
         if deterministic:
             # fixing non-deterministic part of horovod
             # https://github.com/PyTorchLightning/pytorch-lightning/pull/1572/files#r420279383

pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -44,7 +44,6 @@
     _OMEGACONF_AVAILABLE,
     _POPTORCH_AVAILABLE,
     _RICH_AVAILABLE,
-    _TORCH_GREATER_EQUAL_1_7,
     _TORCH_GREATER_EQUAL_1_8,
     _TORCH_GREATER_EQUAL_1_9,
     _TORCH_GREATER_EQUAL_1_10,

pytorch_lightning/utilities/__init__.py

@@ -305,9 +305,6 @@ def _wrap_generator_samplers(self) -> None:
         # access wrapped dataset attributes
         dataset_dict = self.dataset.__dict__
 
-        # create a tuple of sampler names
-        samplers_names = tuple(v.__class__.__name__ for k, v in dataset_dict.items() if isinstance(v, Sampler))
-
         # create a dictionary of generator present within the dataset attributes
         dataset_sampler_generators = {k: v for k, v in dataset_dict.items() if isinstance(v, (Generator, Iterator))}
 
@@ -318,31 +315,17 @@ def _wrap_generator_samplers(self) -> None:
             if isinstance(generator, Sampler):
                 continue
 
-            # used to handle a weird behaviour from PyTorch 1.6
-            # where the sampler is converted to a list_iterator
-            is_legacy = False
-
-            if isinstance(generator, Generator):
-                # Generator name have the  the form `SamplerName.__iter__`
-                generator_name = generator.__qualname__.split(".")[0]
-            else:
-                # assume the retrieved iterator is coming from sampler.
-                is_legacy = True
-
-            # validate the base generator name matches a sampler name.
-            if is_legacy or any(sampler_name == generator_name for sampler_name in samplers_names):
-
-                # wrap the generator into a `FastForwardSampler`
-                sampler = FastForwardSampler(generator, attr_name=generator_attr_name)
+            # wrap the generator into a `FastForwardSampler`
+            sampler = FastForwardSampler(generator, attr_name=generator_attr_name)
 
-                # if `CaptureIterableDataset` was available, the sampler should reload its own state.
-                if self._state_dict is not None:
-                    sampler.load_state_dict(self._state_dict[generator_attr_name])
-                # store the samplers
-                self.samplers[generator_attr_name] = sampler
+            # if `CaptureIterableDataset` was available, the sampler should reload its own state.
+            if self._state_dict is not None:
+                sampler.load_state_dict(self._state_dict[generator_attr_name])
+            # store the samplers
+            self.samplers[generator_attr_name] = sampler
 
-                # replace generator with the generator from the `FastForwardSampler`.
-                dataset_dict[generator_attr_name] = iter(sampler)
+            # replace generator with the generator from the `FastForwardSampler`.
+            dataset_dict[generator_attr_name] = iter(sampler)
 
         self.reset_on_epoch()
 

pytorch_lightning/utilities/auto_restart.py

@@ -19,7 +19,6 @@
 import fsspec
 import torch
 from fsspec.implementations.local import AbstractFileSystem, LocalFileSystem
-from packaging.version import Version
 
 
 def load(
@@ -59,12 +58,6 @@ def atomic_save(checkpoint: Dict[str, Any], filepath: Union[str, Path]) -> None:
     """
 
     bytesbuffer = io.BytesIO()
-    # Can't use the new zipfile serialization for 1.6.0 because there's a bug in
-    # torch.hub.load_state_dict_from_url() that prevents it from loading the new files.
-    # More details can be found here: https://github.com/pytorch/pytorch/issues/42239
-    if Version(torch.__version__).release[:3] == (1, 6, 0):
-        torch.save(checkpoint, bytesbuffer, _use_new_zipfile_serialization=False)
-    else:
-        torch.save(checkpoint, bytesbuffer)
+    torch.save(checkpoint, bytesbuffer)
     with fsspec.open(filepath, "wb") as f:
         f.write(bytesbuffer.getvalue())

pytorch_lightning/utilities/cloud_io.py

@@ -70,7 +70,6 @@ def _compare_version(package: str, op: Callable, version: str, use_base_version:
 
 _IS_WINDOWS = platform.system() == "Windows"
 _IS_INTERACTIVE = hasattr(sys, "ps1")  # https://stackoverflow.com/a/64523765
-_TORCH_GREATER_EQUAL_1_7 = _compare_version("torch", operator.ge, "1.7.0")
 _TORCH_GREATER_EQUAL_1_8 = _compare_version("torch", operator.ge, "1.8.0")
 _TORCH_GREATER_EQUAL_1_8_1 = _compare_version("torch", operator.ge, "1.8.1")
 _TORCH_GREATER_EQUAL_1_9 = _compare_version("torch", operator.ge, "1.9.0")
@@ -112,4 +111,4 @@ def _compare_version(package: str, op: Callable, version: str, use_base_version:
 
 # experimental feature within PyTorch Lightning.
 def _fault_tolerant_training() -> bool:
-    return _TORCH_GREATER_EQUAL_1_7 and int(os.getenv("PL_FAULT_TOLERANT_TRAINING", 0))
+    return bool(int(os.getenv("PL_FAULT_TOLERANT_TRAINING", 0)))