Drop PyTorch 1.7 support (#12432)

Author: carmocca

CHANGELOG.md

@@ -175,7 +175,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
-- Drop PyTorch 1.7 support ([#12191](https://github.com/PyTorchLightning/pytorch-lightning/pull/12191))
+- Drop PyTorch 1.7 support ([#12191](https://github.com/PyTorchLightning/pytorch-lightning/pull/12191)), ([#12432](https://github.com/PyTorchLightning/pytorch-lightning/pull/12432))
 
 
 - Make `benchmark` flag optional and set its value based on the deterministic flag ([#11944](https://github.com/PyTorchLightning/pytorch-lightning/pull/11944))

docs/source/advanced/fault_tolerant_training.rst

@@ -13,7 +13,7 @@ This would make benchmarking non-reproducible as optimization has been interrupt
 With Fault Tolerant Training, when ``Trainer.fit()`` fails in the middle of an epoch during training or validation,
 Lightning will restart exactly where it failed, and everything will be restored.
 
-Fault Tolerance requires PyTorch 1.7 or higher and can be enabled as follows:
+Fault tolerance can be enabled as follows:
 
 .. code-block:: bash
 

pytorch_lightning/accelerators/gpu.py

@@ -23,7 +23,6 @@
 from pytorch_lightning.accelerators.accelerator import Accelerator
 from pytorch_lightning.utilities import device_parser
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
-from pytorch_lightning.utilities.imports import _TORCH_GREATER_EQUAL_1_8
 from pytorch_lightning.utilities.types import _DEVICE
 
 _log = logging.getLogger(__name__)
@@ -70,9 +69,7 @@ def get_device_stats(self, device: _DEVICE) -> Dict[str, Any]:
             FileNotFoundError:
                 If nvidia-smi installation not found
         """
-        if _TORCH_GREATER_EQUAL_1_8:
-            return torch.cuda.memory_stats(device)
-        return get_nvidia_gpu_stats(device)
+        return torch.cuda.memory_stats(device)
 
     @staticmethod
     def parse_devices(devices: Union[int, str, List[int]]) -> Optional[List[int]]:
@@ -102,7 +99,7 @@ def register_accelerators(cls, accelerator_registry: Dict) -> None:
         )
 
 
-def get_nvidia_gpu_stats(device: _DEVICE) -> Dict[str, float]:
+def get_nvidia_gpu_stats(device: _DEVICE) -> Dict[str, float]:  # pragma: no-cover
     """Get GPU stats including memory, fan speed, and temperature from nvidia-smi.
 
     Args:

pytorch_lightning/callbacks/quantization.py

@@ -22,14 +22,7 @@
 
 import torch
 from torch import Tensor
-
-from pytorch_lightning.utilities.imports import _TORCH_GREATER_EQUAL_1_8
-
-if _TORCH_GREATER_EQUAL_1_8:
-    from torch.quantization import FakeQuantizeBase
-else:
-    # For torch 1.7.
-    from torch.quantization import FakeQuantize as FakeQuantizeBase
+from torch.quantization import FakeQuantizeBase
 
 import pytorch_lightning as pl
 from pytorch_lightning.callbacks.base import Callback

pytorch_lightning/core/lightning.py

@@ -66,7 +66,7 @@ class LightningModule(
     CheckpointHooks,
     Module,
 ):
-    # Below is for property support of JIT in PyTorch 1.7
+    # Below is for property support of JIT
     # since none of these are important when using JIT, we are going to ignore them.
     __jit_unused_properties__ = (
         [

pytorch_lightning/profiler/xla.py

@@ -15,10 +15,10 @@
 from typing import Dict
 
 from pytorch_lightning.profiler.profiler import Profiler
-from pytorch_lightning.utilities import _TORCH_GREATER_EQUAL_1_8, _TPU_AVAILABLE
+from pytorch_lightning.utilities import _TPU_AVAILABLE
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 
-if _TPU_AVAILABLE and _TORCH_GREATER_EQUAL_1_8:
+if _TPU_AVAILABLE:
     import torch_xla.debug.profiler as xp
 
 log = logging.getLogger(__name__)
@@ -45,8 +45,6 @@ def __init__(self, port: int = 9012) -> None:
         """
         if not _TPU_AVAILABLE:
             raise MisconfigurationException("`XLAProfiler` is only supported on TPUs")
-        if not _TORCH_GREATER_EQUAL_1_8:
-            raise MisconfigurationException("`XLAProfiler` is only supported with `torch-xla >= 1.8`")
         super().__init__(dirpath=None, filename=None)
         self.port = port
         self._recording_map: Dict = {}

pytorch_lightning/strategies/ddp.py

@@ -42,12 +42,16 @@
     get_default_process_group_backend_for_device,
 )
 from pytorch_lightning.utilities.distributed import group as _group
-from pytorch_lightning.utilities.distributed import init_dist_connection, ReduceOp, sync_ddp_if_available
+from pytorch_lightning.utilities.distributed import (
+    init_dist_connection,
+    ReduceOp,
+    register_ddp_comm_hook,
+    sync_ddp_if_available,
+)
 from pytorch_lightning.utilities.exceptions import DeadlockDetectedException
 from pytorch_lightning.utilities.imports import (
     _FAIRSCALE_AVAILABLE,
     _IS_WINDOWS,
-    _TORCH_GREATER_EQUAL_1_8,
     _TORCH_GREATER_EQUAL_1_9,
     _TORCH_GREATER_EQUAL_1_10,
     _TORCH_GREATER_EQUAL_1_11,
@@ -58,8 +62,6 @@
 
 if _FAIRSCALE_AVAILABLE:
     from fairscale.optim import OSS
-if _TORCH_GREATER_EQUAL_1_8:
-    from pytorch_lightning.utilities.distributed import register_ddp_comm_hook
 if _TORCH_GREATER_EQUAL_1_10:
     from torch.distributed.algorithms.model_averaging.averagers import ModelAverager
 
@@ -213,9 +215,7 @@ def _register_ddp_hooks(self) -> None:
         log.detail(f"{self.__class__.__name__}: registering ddp hooks")
         # In 1.8, DDP communication hooks only work with NCCL backend and SPSD (single process single device) mode
         # Since 1.9, DDP communication hooks can work on all backends.
-        if _TORCH_GREATER_EQUAL_1_9 or (
-            _TORCH_GREATER_EQUAL_1_8 and self.root_device.type == "cuda" and self._is_single_process_single_device
-        ):
+        if _TORCH_GREATER_EQUAL_1_9 or (self.root_device.type == "cuda" and self._is_single_process_single_device):
             register_ddp_comm_hook(
                 model=self.model,
                 ddp_comm_state=self._ddp_comm_state,
@@ -302,7 +302,7 @@ def determine_ddp_device_ids(self):
     def barrier(self, *args, **kwargs) -> None:
         if not distributed_available():
             return
-        if _TORCH_GREATER_EQUAL_1_8 and torch.distributed.get_backend() == "nccl":
+        if torch.distributed.get_backend() == "nccl":
             torch.distributed.barrier(device_ids=self.determine_ddp_device_ids())
         else:
             torch.distributed.barrier()

pytorch_lightning/strategies/ddp_spawn.py

@@ -35,15 +35,17 @@
     get_default_process_group_backend_for_device,
 )
 from pytorch_lightning.utilities.distributed import group as _group
-from pytorch_lightning.utilities.distributed import init_dist_connection, ReduceOp, sync_ddp_if_available
-from pytorch_lightning.utilities.imports import _TORCH_GREATER_EQUAL_1_8, _TORCH_GREATER_EQUAL_1_11
+from pytorch_lightning.utilities.distributed import (
+    init_dist_connection,
+    ReduceOp,
+    register_ddp_comm_hook,
+    sync_ddp_if_available,
+)
+from pytorch_lightning.utilities.imports import _TORCH_GREATER_EQUAL_1_11
 from pytorch_lightning.utilities.rank_zero import rank_zero_info, rank_zero_only
 from pytorch_lightning.utilities.seed import reset_seed
 from pytorch_lightning.utilities.types import STEP_OUTPUT
 
-if _TORCH_GREATER_EQUAL_1_8:
-    from pytorch_lightning.utilities.distributed import register_ddp_comm_hook
-
 log = logging.getLogger(__name__)
 
 
@@ -171,7 +173,7 @@ def pre_configure_ddp(self):
     def _register_ddp_hooks(self) -> None:
         # currently, DDP communication hooks only work with NCCL backend and SPSD (single process single device) mode
         # https://github.com/pytorch/pytorch/blob/v1.8.0/torch/nn/parallel/distributed.py#L1080-L1084
-        if _TORCH_GREATER_EQUAL_1_8 and self.root_device.type == "cuda" and self._is_single_process_single_device:
+        if self.root_device.type == "cuda" and self._is_single_process_single_device:
             register_ddp_comm_hook(
                 model=self.model,
                 ddp_comm_state=self._ddp_comm_state,
@@ -192,7 +194,7 @@ def determine_ddp_device_ids(self):
     def barrier(self, *args, **kwargs) -> None:
         if not distributed_available():
             return
-        if _TORCH_GREATER_EQUAL_1_8 and torch.distributed.get_backend() == "nccl":
+        if torch.distributed.get_backend() == "nccl":
             torch.distributed.barrier(device_ids=self.determine_ddp_device_ids())
         else:
             torch.distributed.barrier()

pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -79,13 +79,7 @@
     rank_zero_warn,
 )
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
-from pytorch_lightning.utilities.imports import (
-    _HOROVOD_AVAILABLE,
-    _HPU_AVAILABLE,
-    _IPU_AVAILABLE,
-    _TORCH_GREATER_EQUAL_1_8,
-    _TPU_AVAILABLE,
-)
+from pytorch_lightning.utilities.imports import _HOROVOD_AVAILABLE, _HPU_AVAILABLE, _IPU_AVAILABLE, _TPU_AVAILABLE
 
 log = logging.getLogger(__name__)
 
@@ -217,10 +211,7 @@ def __init__(
 
     def _init_deterministic(self, deterministic: bool) -> None:
         self.deterministic = deterministic
-        if _TORCH_GREATER_EQUAL_1_8:
-            torch.use_deterministic_algorithms(deterministic)
-        else:
-            torch.set_deterministic(deterministic)
+        torch.use_deterministic_algorithms(deterministic)
         if deterministic:
             # fixing non-deterministic part of horovod
             # https://github.com/PyTorchLightning/pytorch-lightning/pull/1572/files#r420279383

pytorch_lightning/trainer/connectors/data_connector.py

@@ -220,35 +220,19 @@ def _worker_check(self, dataloader: DataLoader, name: str) -> None:
 
         # ddp_spawn + num_workers > 0 don't mix! tell the user
         if dataloader.num_workers > 0 and using_spawn:
-            # checks for the attr persistent_workers available in pytorch >= 1.7
-            if hasattr(dataloader, "persistent_workers"):
-                if not dataloader.persistent_workers:
-                    rank_zero_warn(
-                        "num_workers>0, persistent_workers=False, and strategy=ddp_spawn"
-                        " may result in data loading bottlenecks."
-                        " Consider setting persistent_workers=True"
-                        " (this is a limitation of Python .spawn() and PyTorch)"
-                    )
-            else:
+            if not dataloader.persistent_workers:
                 rank_zero_warn(
-                    "num_workers>0 and strategy=ddp_spawn do not mix well"
-                    " and may result in data loading bottlenecks."
-                    " Consider setting strategy=ddp to use num_workers>0"
+                    "num_workers>0, persistent_workers=False, and strategy=ddp_spawn"
+                    " may result in data loading bottlenecks."
+                    " Consider setting persistent_workers=True"
                     " (this is a limitation of Python .spawn() and PyTorch)"
                 )
 
         elif dataloader.num_workers == 0 and using_spawn:
-            # checks for the attr persistent_workers available in pytorch >= 1.7
-            if hasattr(dataloader, "persistent_workers"):
-                if not dataloader.persistent_workers:
-                    rank_zero_warn(
-                        "strategy=ddp_spawn and num_workers=0 may result in data loading bottlenecks."
-                        " Consider setting num_workers>0 and persistent_workers=True"
-                    )
-            else:
+            if not dataloader.persistent_workers:
                 rank_zero_warn(
                     "strategy=ddp_spawn and num_workers=0 may result in data loading bottlenecks."
-                    " Consider setting strategy=ddp and set num_workers>0"
+                    " Consider setting num_workers>0 and persistent_workers=True"
                 )
 
         elif dataloader.num_workers <= 2 < num_cpus and not using_spawn: