From c6a47dd35cc4f172476ac92e1f4f63153a4befcf Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Wed, 27 Jul 2022 13:14:21 +0100
Subject: [PATCH 01/18] Move half cast to occur on host instead of in fwd

---
 src/pytorch_lightning/strategies/ipu.py      | 28 ++++++++------------
 tests/tests_pytorch/accelerators/test_ipu.py |  2 +-
 2 files changed, 12 insertions(+), 18 deletions(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 001ad77fbb5cc..14d3ce7a3d9c5 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -43,24 +43,9 @@
 
 class LightningIPUModule(_LightningModuleWrapperBase):
     def __init__(
-        self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase], precision: Union[str, int]
+        self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase]
     ) -> None:
         super().__init__(pl_module)
-        self.precision = precision
-
-    def forward(self, *inputs: Any, **kwargs: Any) -> Any:
-        if self.precision in (PrecisionType.MIXED, PrecisionType.HALF):
-            inputs = self._move_float_tensors_to_half(inputs)
-
-        return super().forward(*inputs, **kwargs)
-
-    @staticmethod
-    def batch_to(data: Tensor) -> Tensor:
-        return data.half()
-
-    def _move_float_tensors_to_half(self, batch: Any) -> Any:
-        batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=self.batch_to)
-        return batch
 
 
 class IPUStrategy(ParallelStrategy):
@@ -142,7 +127,7 @@ def setup(self, trainer: "pl.Trainer") -> None:
         self._optimizer_zero_grad_original = self.lightning_module.optimizer_zero_grad
         self._disable_zero_grad()
 
-        model = LightningIPUModule(self.lightning_module, self.precision_plugin.precision)
+        model = LightningIPUModule(self.lightning_module)
         self.model = model
 
         # reset the backup
@@ -272,6 +257,15 @@ def to_tensor(x):
         args = apply_to_collection(args, dtype=(int, float), function=to_tensor)
         return args
 
+    @staticmethod
+    def batch_to(data: Tensor) -> Tensor:
+        return data.half()
+
+    def batch_to_device(self, batch: Any, device: Optional[torch.device] = None, dataloader_idx: int = 0) -> Any:
+        if self.precision_plugin.precision in (PrecisionType.MIXED, PrecisionType.HALF):
+            batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=self.batch_to)
+        return batch
+
     def _disable_zero_grad(self) -> None:
         lightning_module = self.lightning_module
         if is_overridden("optimizer_zero_grad", lightning_module):
diff --git a/tests/tests_pytorch/accelerators/test_ipu.py b/tests/tests_pytorch/accelerators/test_ipu.py
index 589ec7b29dd5b..a60a8aa330a37 100644
--- a/tests/tests_pytorch/accelerators/test_ipu.py
+++ b/tests/tests_pytorch/accelerators/test_ipu.py
@@ -205,7 +205,7 @@ def setup(self, trainer: Trainer, pl_module: LightningModule, stage: Optional[st
 def test_pure_half_precision(tmpdir):
     class TestCallback(Callback):
         def on_train_start(self, trainer: Trainer, pl_module: LightningModule) -> None:
-            assert trainer.strategy.model.precision == 16
+            assert trainer.strategy.precision_plugin.precision == 16
             for param in trainer.strategy.model.parameters():
                 assert param.dtype == torch.float16
             raise SystemExit

From 435d03fa1ce1b7fd8b122a31f901296c621d112f Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Wed, 27 Jul 2022 14:28:04 +0100
Subject: [PATCH 02/18] Update changelog

---
 src/pytorch_lightning/CHANGELOG.md | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/pytorch_lightning/CHANGELOG.md b/src/pytorch_lightning/CHANGELOG.md
index f8341248b20e8..f7a0acd14e182 100644
--- a/src/pytorch_lightning/CHANGELOG.md
+++ b/src/pytorch_lightning/CHANGELOG.md
@@ -113,6 +113,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
+- When training with `precision=16` on IPU, the cast has been moved off the IPU onto the host, making the copies from host to IPU cheaper ([#13880](https://github.com/Lightning-AI/lightning/pull/13880))
+
 - `accelerator="gpu"` now automatically selects an available GPU backend (CUDA and MPS currently) ([#13642](https://github.com/Lightning-AI/lightning/pull/13642))
 
 

From 8379f0143bf6d5673aae8c83a16ffa94d1b02304 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Wed, 27 Jul 2022 13:30:36 +0000
Subject: [PATCH 03/18] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 src/pytorch_lightning/strategies/ipu.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 14d3ce7a3d9c5..417bc28baab2d 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -42,9 +42,7 @@
 
 
 class LightningIPUModule(_LightningModuleWrapperBase):
-    def __init__(
-        self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase]
-    ) -> None:
+    def __init__(self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase]) -> None:
         super().__init__(pl_module)
 
 

From 423d63ac6f10c703ce6ba56b0a993b7a040c86c1 Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 09:47:39 +0100
Subject: [PATCH 04/18] batch_to -> to_half, now local to batch_to_device

---
 src/pytorch_lightning/strategies/ipu.py | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 417bc28baab2d..49a0daee79ca4 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -255,13 +255,11 @@ def to_tensor(x):
         args = apply_to_collection(args, dtype=(int, float), function=to_tensor)
         return args
 
-    @staticmethod
-    def batch_to(data: Tensor) -> Tensor:
-        return data.half()
-
     def batch_to_device(self, batch: Any, device: Optional[torch.device] = None, dataloader_idx: int = 0) -> Any:
         if self.precision_plugin.precision in (PrecisionType.MIXED, PrecisionType.HALF):
-            batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=self.batch_to)
+            def to_half(data: Tensor) -> Tensor:
+                return data.half()
+            batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=to_half)
         return batch
 
     def _disable_zero_grad(self) -> None:

From 28016af9a279cbf921b2467494b3d3b9d57fca30 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Thu, 28 Jul 2022 08:49:37 +0000
Subject: [PATCH 05/18] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 src/pytorch_lightning/strategies/ipu.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 49a0daee79ca4..5ba8c1aa48454 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -257,8 +257,10 @@ def to_tensor(x):
 
     def batch_to_device(self, batch: Any, device: Optional[torch.device] = None, dataloader_idx: int = 0) -> Any:
         if self.precision_plugin.precision in (PrecisionType.MIXED, PrecisionType.HALF):
+
             def to_half(data: Tensor) -> Tensor:
                 return data.half()
+
             batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=to_half)
         return batch
 

From 2cd738db3f4682ec21918823bf195754405f5e80 Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 10:00:24 +0100
Subject: [PATCH 06/18] Add comments to overloaded function

---
 src/pytorch_lightning/strategies/ipu.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 5ba8c1aa48454..6a32485997c10 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -256,12 +256,17 @@ def to_tensor(x):
         return args
 
     def batch_to_device(self, batch: Any, device: Optional[torch.device] = None, dataloader_idx: int = 0) -> Any:
+        # This override is necessary because the cast must occur before the data
+        # is moved to the device to prevent wasteful host->device copies.
         if self.precision_plugin.precision in (PrecisionType.MIXED, PrecisionType.HALF):
 
             def to_half(data: Tensor) -> Tensor:
                 return data.half()
 
             batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=to_half)
+        # We don't call `super().batch_to_device` because `data.to(device)` is not
+        # currently necessary for IPUs. The movement of data from host<->IPU is
+        # currently handled by PopTorch.
         return batch
 
     def _disable_zero_grad(self) -> None:

From b9ca3d9656457b45c1f2f77066e9bdcf4145a2c0 Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 10:13:11 +0100
Subject: [PATCH 07/18] Remove `LightningIPUModule`

---
 src/pytorch_lightning/strategies/ipu.py | 22 ++++++----------------
 1 file changed, 6 insertions(+), 16 deletions(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 6a32485997c10..54daa808369be 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -41,11 +41,6 @@
     poptorch = None
 
 
-class LightningIPUModule(_LightningModuleWrapperBase):
-    def __init__(self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase]) -> None:
-        super().__init__(pl_module)
-
-
 class IPUStrategy(ParallelStrategy):
     """Plugin for training on IPU devices."""
 
@@ -125,8 +120,7 @@ def setup(self, trainer: "pl.Trainer") -> None:
         self._optimizer_zero_grad_original = self.lightning_module.optimizer_zero_grad
         self._disable_zero_grad()
 
-        model = LightningIPUModule(self.lightning_module)
-        self.model = model
+        self.model = _LightningModuleWrapperBase(self.lightning_module)
 
         # reset the backup
         self.poptorch_models = {}
@@ -139,22 +133,22 @@ def setup(self, trainer: "pl.Trainer") -> None:
             training_opts = self.training_opts
             inference_opts = self.inference_opts
             optimizer = self.lightning_module.trainer.optimizers[0]
-            model = poptorch.trainingModel(model=model, options=training_opts, optimizer=optimizer)
+            model = poptorch.trainingModel(model=self.model, options=training_opts, optimizer=optimizer)
             self.poptorch_models[RunningStage.TRAINING] = model
 
             if self.lightning_module.trainer.enable_validation:
-                model = poptorch.inferenceModel(model=model, options=inference_opts)
+                model = poptorch.inferenceModel(model=self.model, options=inference_opts)
                 self.poptorch_models[RunningStage.VALIDATING] = model
                 if self.lightning_module.trainer.num_sanity_val_steps > 0:
                     self.poptorch_models[RunningStage.SANITY_CHECKING] = model
         elif trainer_fn == TrainerFn.VALIDATING:
-            model = poptorch.inferenceModel(model=model, options=self.inference_opts)
+            model = poptorch.inferenceModel(model=self.model, options=self.inference_opts)
             self.poptorch_models[RunningStage.VALIDATING] = model
         elif trainer_fn == TrainerFn.TESTING:
-            model = poptorch.inferenceModel(model=model, options=self.inference_opts)
+            model = poptorch.inferenceModel(model=self.model, options=self.inference_opts)
             self.poptorch_models[RunningStage.TESTING] = model
         elif trainer_fn == TrainerFn.PREDICTING:
-            model = poptorch.inferenceModel(model=model, options=self.inference_opts)
+            model = poptorch.inferenceModel(model=self.model, options=self.inference_opts)
             self.poptorch_models[RunningStage.PREDICTING] = model
 
     def setup_optimizers(self, trainer: "pl.Trainer") -> None:
@@ -202,10 +196,6 @@ def inference_opts(self) -> "poptorch.Options":
             self._inference_opts = self._create_opts(training=False)
         return self._inference_opts
 
-    @property
-    def lightning_module(self) -> Optional["pl.LightningModule"]:
-        return self.model.module if isinstance(self.model, LightningIPUModule) else self.model
-
     def _convert_to_poptorch_loader(
         self, dataloader: DataLoader, sampler, mode: Optional[RunningStage] = None
     ) -> "poptorch.DataLoader":

From daf99497bb39f4074aca12376ea54d157f8172c3 Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 11:13:56 +0100
Subject: [PATCH 08/18] Validate precision in IPU tests

---
 tests/tests_pytorch/accelerators/test_ipu.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/tests/tests_pytorch/accelerators/test_ipu.py b/tests/tests_pytorch/accelerators/test_ipu.py
index a60a8aa330a37..0908c87085a46 100644
--- a/tests/tests_pytorch/accelerators/test_ipu.py
+++ b/tests/tests_pytorch/accelerators/test_ipu.py
@@ -40,16 +40,19 @@
 
 class IPUModel(BoringModel):
     def training_step(self, batch, batch_idx):
+        assert self.precision == torch.finfo(batch.dtype).bits
         output = self(batch)
         loss = self.loss(batch, output)
         return loss
 
     def validation_step(self, batch, batch_idx):
+        assert self.precision == torch.finfo(batch.dtype).bits
         output = self(batch)
         loss = self.loss(batch, output)
         return loss
 
     def test_step(self, batch, batch_idx):
+        assert self.precision == torch.finfo(batch.dtype).bits
         output = self(batch)
         loss = self.loss(batch, output)
         return loss

From 300056ee1750c1ac077eb5651814e937eccdc2d6 Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 11:37:08 +0100
Subject: [PATCH 09/18] Deprecate `LightningIPUModel`

---
 src/pytorch_lightning/strategies/ipu.py       | 24 +++++++++++++++++++
 .../deprecated_api/test_remove_2-0.py         |  6 +++++
 2 files changed, 30 insertions(+)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 54daa808369be..f10198be23939 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -33,6 +33,7 @@
 from pytorch_lightning.utilities.enums import PrecisionType
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.model_helpers import is_overridden
+from pytorch_lightning.utilities.rank_zero import rank_zero_deprecation
 from pytorch_lightning.utilities.types import STEP_OUTPUT
 
 if _POPTORCH_AVAILABLE:
@@ -41,6 +42,29 @@
     poptorch = None
 
 
+class LightningIPUModule(_LightningModuleWrapperBase):
+    def __init__(
+        self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase], precision: Union[str, int]
+    ) -> None:
+        rank_zero_deprecation('`LightningIPUModule` is deprecated in v1.8 and will be removed in v2.0.0')
+        super().__init__(pl_module)
+        self.precision = precision
+
+    def forward(self, *inputs: Any, **kwargs: Any) -> Any:
+        if self.precision in (PrecisionType.MIXED, PrecisionType.HALF):
+            inputs = self._move_float_tensors_to_half(inputs)
+
+        return super().forward(*inputs, **kwargs)
+
+    @staticmethod
+    def batch_to(data: Tensor) -> Tensor:
+        return data.half()
+
+    def _move_float_tensors_to_half(self, batch: Any) -> Any:
+        batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=self.batch_to)
+        return batch
+
+
 class IPUStrategy(ParallelStrategy):
     """Plugin for training on IPU devices."""
 
diff --git a/tests/tests_pytorch/deprecated_api/test_remove_2-0.py b/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
index b39c6dafc1696..48954a5446cb0 100644
--- a/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
+++ b/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
@@ -19,6 +19,7 @@
 import pytorch_lightning
 from pytorch_lightning import Trainer
 from pytorch_lightning.demos.boring_classes import BoringModel
+from pytorch_lightning.strategies.ipu import LightningIPUModule
 from tests_pytorch.callbacks.test_callbacks import OldStatefulCallback
 from tests_pytorch.helpers.runif import RunIf
 
@@ -49,6 +50,11 @@ def test_v2_0_0_deprecated_ipus(_, monkeypatch):
     with pytest.deprecated_call(match=r"is deprecated in v1.7 and will be removed in v2.0."):
         _ = Trainer(ipus=4)
 
+@mock.patch("pytorch_lightning.accelerators.ipu.IPUAccelerator.is_available", return_value=True)
+def test_v2_0_0_deprecated_lightning_ipu_module(_, monkeypatch):
+    monkeypatch.setattr(pytorch_lightning.strategies.ipu, "_IPU_AVAILABLE", True)
+    with pytest.deprecated_call(match=r"is deprecated in v1.8 and will be removed in v2.0."):
+        _ = LightningIPUModule(BoringModel(), 32)
 
 def test_v2_0_resume_from_checkpoint_trainer_constructor(tmpdir):
     # test resume_from_checkpoint still works until v2.0 deprecation

From ee6a4cfc21b7c10b2aacc32ea1edb5365d052b30 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Thu, 28 Jul 2022 10:38:48 +0000
Subject: [PATCH 10/18] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 src/pytorch_lightning/strategies/ipu.py               | 2 +-
 tests/tests_pytorch/deprecated_api/test_remove_2-0.py | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index f10198be23939..a332d339a8daf 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -46,7 +46,7 @@ class LightningIPUModule(_LightningModuleWrapperBase):
     def __init__(
         self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase], precision: Union[str, int]
     ) -> None:
-        rank_zero_deprecation('`LightningIPUModule` is deprecated in v1.8 and will be removed in v2.0.0')
+        rank_zero_deprecation("`LightningIPUModule` is deprecated in v1.8 and will be removed in v2.0.0")
         super().__init__(pl_module)
         self.precision = precision
 
diff --git a/tests/tests_pytorch/deprecated_api/test_remove_2-0.py b/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
index 48954a5446cb0..82369a6412c7a 100644
--- a/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
+++ b/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
@@ -50,12 +50,14 @@ def test_v2_0_0_deprecated_ipus(_, monkeypatch):
     with pytest.deprecated_call(match=r"is deprecated in v1.7 and will be removed in v2.0."):
         _ = Trainer(ipus=4)
 
+
 @mock.patch("pytorch_lightning.accelerators.ipu.IPUAccelerator.is_available", return_value=True)
 def test_v2_0_0_deprecated_lightning_ipu_module(_, monkeypatch):
     monkeypatch.setattr(pytorch_lightning.strategies.ipu, "_IPU_AVAILABLE", True)
     with pytest.deprecated_call(match=r"is deprecated in v1.8 and will be removed in v2.0."):
         _ = LightningIPUModule(BoringModel(), 32)
 
+
 def test_v2_0_resume_from_checkpoint_trainer_constructor(tmpdir):
     # test resume_from_checkpoint still works until v2.0 deprecation
     model = BoringModel()

From de82e00e278164f63daa5c1bfcc9ee16cc98cf3b Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 18:22:32 +0100
Subject: [PATCH 11/18] Update deprecation timeline

---
 src/pytorch_lightning/strategies/ipu.py               | 2 +-
 tests/tests_pytorch/deprecated_api/test_remove_1-8.py | 8 ++++++++
 tests/tests_pytorch/deprecated_api/test_remove_2-0.py | 8 --------
 3 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index a332d339a8daf..5abd829e84e0e 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -46,7 +46,7 @@ class LightningIPUModule(_LightningModuleWrapperBase):
     def __init__(
         self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase], precision: Union[str, int]
     ) -> None:
-        rank_zero_deprecation("`LightningIPUModule` is deprecated in v1.8 and will be removed in v2.0.0")
+        rank_zero_deprecation("`LightningIPUModule` is deprecated in v1.7 and will be removed in v1.8.0")
         super().__init__(pl_module)
         self.precision = precision
 
diff --git a/tests/tests_pytorch/deprecated_api/test_remove_1-8.py b/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
index 12aca123eacc1..f412d37730044 100644
--- a/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
+++ b/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
@@ -43,6 +43,7 @@
 from pytorch_lightning.profiler import AbstractProfiler, BaseProfiler
 from pytorch_lightning.profilers import AdvancedProfiler, Profiler, SimpleProfiler
 from pytorch_lightning.strategies import DDP2Strategy, ParallelStrategy
+from pytorch_lightning.strategies.ipu import LightningIPUModule
 from pytorch_lightning.trainer.configuration_validator import _check_datamodule_checkpoint_hooks
 from pytorch_lightning.trainer.states import RunningStage
 from pytorch_lightning.utilities import device_parser
@@ -1008,6 +1009,13 @@ def test_trainer_config_ipus(monkeypatch, trainer_kwargs, expected_ipus):
         trainer.ipus == expected_ipus
 
 
+@mock.patch("pytorch_lightning.accelerators.ipu.IPUAccelerator.is_available", return_value=True)
+def test_v2_0_0_deprecated_lightning_ipu_module(_, monkeypatch):
+    monkeypatch.setattr(pytorch_lightning.strategies.ipu, "_IPU_AVAILABLE", True)
+    with pytest.deprecated_call(match=r"is deprecated in v1.7 and will be removed in v1.8."):
+        _ = LightningIPUModule(BoringModel(), 32)
+
+
 @pytest.mark.parametrize(
     ["trainer_kwargs", "expected_num_processes"],
     [
diff --git a/tests/tests_pytorch/deprecated_api/test_remove_2-0.py b/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
index 82369a6412c7a..b39c6dafc1696 100644
--- a/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
+++ b/tests/tests_pytorch/deprecated_api/test_remove_2-0.py
@@ -19,7 +19,6 @@
 import pytorch_lightning
 from pytorch_lightning import Trainer
 from pytorch_lightning.demos.boring_classes import BoringModel
-from pytorch_lightning.strategies.ipu import LightningIPUModule
 from tests_pytorch.callbacks.test_callbacks import OldStatefulCallback
 from tests_pytorch.helpers.runif import RunIf
 
@@ -51,13 +50,6 @@ def test_v2_0_0_deprecated_ipus(_, monkeypatch):
         _ = Trainer(ipus=4)
 
 
-@mock.patch("pytorch_lightning.accelerators.ipu.IPUAccelerator.is_available", return_value=True)
-def test_v2_0_0_deprecated_lightning_ipu_module(_, monkeypatch):
-    monkeypatch.setattr(pytorch_lightning.strategies.ipu, "_IPU_AVAILABLE", True)
-    with pytest.deprecated_call(match=r"is deprecated in v1.8 and will be removed in v2.0."):
-        _ = LightningIPUModule(BoringModel(), 32)
-
-
 def test_v2_0_resume_from_checkpoint_trainer_constructor(tmpdir):
     # test resume_from_checkpoint still works until v2.0 deprecation
     model = BoringModel()

From bb45874a5d4aa15ac8b8d370a0aeb97ca170bb71 Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 18:24:24 +0100
Subject: [PATCH 12/18] Moved changelog entry to fixed

---
 src/pytorch_lightning/CHANGELOG.md | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/src/pytorch_lightning/CHANGELOG.md b/src/pytorch_lightning/CHANGELOG.md
index f7a0acd14e182..dae416f4ecf11 100644
--- a/src/pytorch_lightning/CHANGELOG.md
+++ b/src/pytorch_lightning/CHANGELOG.md
@@ -113,8 +113,6 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
-- When training with `precision=16` on IPU, the cast has been moved off the IPU onto the host, making the copies from host to IPU cheaper ([#13880](https://github.com/Lightning-AI/lightning/pull/13880))
-
 - `accelerator="gpu"` now automatically selects an available GPU backend (CUDA and MPS currently) ([#13642](https://github.com/Lightning-AI/lightning/pull/13642))
 
 
@@ -417,6 +415,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- When training with `precision=16` on IPU, the cast has been moved off the IPU onto the host, making the copies from host to IPU cheaper ([#13880](https://github.com/Lightning-AI/lightning/pull/13880))
 - Fixed an issue causing zero-division error for empty dataloaders ([#12885](https://github.com/PyTorchLightning/pytorch-lightning/pull/12885))
 - Fixed mismatching default values for the types of some arguments in the DeepSpeed and Fully-Sharded strategies which made the CLI unable to use them ([#12989](https://github.com/PyTorchLightning/pytorch-lightning/pull/12989))
 - Avoid redundant callback restore warning while tuning ([#13026](https://github.com/PyTorchLightning/pytorch-lightning/pull/13026))

From 5ffbec0a51872a1d935faee56c3968a1be8ae12f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Carlos=20Mochol=C3=AD?= <carlossmocholi@gmail.com>
Date: Thu, 28 Jul 2022 19:28:06 +0200
Subject: [PATCH 13/18] Minor fixes

---
 src/pytorch_lightning/CHANGELOG.md                    | 4 +++-
 tests/tests_pytorch/deprecated_api/test_remove_1-8.py | 2 +-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/src/pytorch_lightning/CHANGELOG.md b/src/pytorch_lightning/CHANGELOG.md
index dae416f4ecf11..47d56c454648d 100644
--- a/src/pytorch_lightning/CHANGELOG.md
+++ b/src/pytorch_lightning/CHANGELOG.md
@@ -385,6 +385,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Used `global_step` while restoring logging step for old checkpoints ([#13645](https://github.com/Lightning-AI/lightning/pull/13645))
 
 
+- When training with `precision=16` on IPU, the cast has been moved off the IPU onto the host, making the copies from host to IPU cheaper ([#13880](https://github.com/Lightning-AI/lightning/pull/13880))
+
+
 - Fixed error handling in learning rate finder when not enough data points are available to give a good suggestion ([#13845](https://github.com/Lightning-AI/lightning/pull/13845))
 
 
@@ -415,7 +418,6 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
-- When training with `precision=16` on IPU, the cast has been moved off the IPU onto the host, making the copies from host to IPU cheaper ([#13880](https://github.com/Lightning-AI/lightning/pull/13880))
 - Fixed an issue causing zero-division error for empty dataloaders ([#12885](https://github.com/PyTorchLightning/pytorch-lightning/pull/12885))
 - Fixed mismatching default values for the types of some arguments in the DeepSpeed and Fully-Sharded strategies which made the CLI unable to use them ([#12989](https://github.com/PyTorchLightning/pytorch-lightning/pull/12989))
 - Avoid redundant callback restore warning while tuning ([#13026](https://github.com/PyTorchLightning/pytorch-lightning/pull/13026))
diff --git a/tests/tests_pytorch/deprecated_api/test_remove_1-8.py b/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
index f412d37730044..3f8505cc8de82 100644
--- a/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
+++ b/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
@@ -1010,7 +1010,7 @@ def test_trainer_config_ipus(monkeypatch, trainer_kwargs, expected_ipus):
 
 
 @mock.patch("pytorch_lightning.accelerators.ipu.IPUAccelerator.is_available", return_value=True)
-def test_v2_0_0_deprecated_lightning_ipu_module(_, monkeypatch):
+def test_v1_8_0_deprecated_lightning_ipu_module(_, monkeypatch):
     monkeypatch.setattr(pytorch_lightning.strategies.ipu, "_IPU_AVAILABLE", True)
     with pytest.deprecated_call(match=r"is deprecated in v1.7 and will be removed in v1.8."):
         _ = LightningIPUModule(BoringModel(), 32)

From a5b1883e9f9639af20ddf6ac0bd3d0c169e855f3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Carlos=20Mochol=C3=AD?= <carlossmocholi@gmail.com>
Date: Thu, 28 Jul 2022 19:30:16 +0200
Subject: [PATCH 14/18] Refactor

---
 src/pytorch_lightning/strategies/ipu.py | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 5abd829e84e0e..ce5e97e6bbbc3 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -273,11 +273,7 @@ def batch_to_device(self, batch: Any, device: Optional[torch.device] = None, dat
         # This override is necessary because the cast must occur before the data
         # is moved to the device to prevent wasteful host->device copies.
         if self.precision_plugin.precision in (PrecisionType.MIXED, PrecisionType.HALF):
-
-            def to_half(data: Tensor) -> Tensor:
-                return data.half()
-
-            batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=to_half)
+            batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=Tensor.half)
         # We don't call `super().batch_to_device` because `data.to(device)` is not
         # currently necessary for IPUs. The movement of data from host<->IPU is
         # currently handled by PopTorch.

From 9f303a4b448ab7fb16aebb31a305268cc7a2b862 Mon Sep 17 00:00:00 2001
From: Harry Mellor <harrym@graphcore.ai>
Date: Thu, 28 Jul 2022 19:42:09 +0100
Subject: [PATCH 15/18] Update test

---
 tests/tests_pytorch/accelerators/test_ipu.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/tests/tests_pytorch/accelerators/test_ipu.py b/tests/tests_pytorch/accelerators/test_ipu.py
index 0c54e3a6707f3..9d510d9d60e9e 100644
--- a/tests/tests_pytorch/accelerators/test_ipu.py
+++ b/tests/tests_pytorch/accelerators/test_ipu.py
@@ -40,19 +40,16 @@
 
 class IPUModel(BoringModel):
     def training_step(self, batch, batch_idx):
-        assert self.precision == torch.finfo(batch.dtype).bits
         output = self(batch)
         loss = self.loss(batch, output)
         return loss
 
     def validation_step(self, batch, batch_idx):
-        assert self.precision == torch.finfo(batch.dtype).bits
         output = self(batch)
         loss = self.loss(batch, output)
         return loss
 
     def test_step(self, batch, batch_idx):
-        assert self.precision == torch.finfo(batch.dtype).bits
         output = self(batch)
         loss = self.loss(batch, output)
         return loss
@@ -222,6 +219,7 @@ def on_train_start(self, trainer: Trainer, pl_module: LightningModule) -> None:
     assert isinstance(trainer.strategy, IPUStrategy)
     assert isinstance(trainer.strategy.precision_plugin, IPUPrecisionPlugin)
     assert trainer.strategy.precision_plugin.precision == 16
+    assert trainer.strategy.batch_to_device(torch.zeros((1), dtype=torch.float)).dtype == torch.half
 
     with pytest.raises(SystemExit):
         trainer.fit(model)

From a9f12e94705c3ef18ac94535a76873bf6fe64326 Mon Sep 17 00:00:00 2001
From: HMellor <HMellor@users.noreply.github.com>
Date: Thu, 28 Jul 2022 19:43:33 +0100
Subject: [PATCH 16/18] Update src/pytorch_lightning/strategies/ipu.py

Co-authored-by: Rohit Gupta <rohitgr1998@gmail.com>
---
 src/pytorch_lightning/strategies/ipu.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index ce5e97e6bbbc3..8d9cc50b90350 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -46,7 +46,7 @@ class LightningIPUModule(_LightningModuleWrapperBase):
     def __init__(
         self, pl_module: Union["pl.LightningModule", _LightningPrecisionModuleWrapperBase], precision: Union[str, int]
     ) -> None:
-        rank_zero_deprecation("`LightningIPUModule` is deprecated in v1.7 and will be removed in v1.8.0")
+        rank_zero_deprecation("`LightningIPUModule` has been deprecated in v1.7.0 and will be removed in v1.8.0")
         super().__init__(pl_module)
         self.precision = precision
 

From af485fa73045dee3e7d28007aacffd55ce4b2b98 Mon Sep 17 00:00:00 2001
From: HMellor <HMellor@users.noreply.github.com>
Date: Thu, 28 Jul 2022 19:43:42 +0100
Subject: [PATCH 17/18] Update
 tests/tests_pytorch/deprecated_api/test_remove_1-8.py

Co-authored-by: Rohit Gupta <rohitgr1998@gmail.com>
---
 tests/tests_pytorch/deprecated_api/test_remove_1-8.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/tests_pytorch/deprecated_api/test_remove_1-8.py b/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
index 59ec7331fefe5..0d9fc1e7a2baf 100644
--- a/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
+++ b/tests/tests_pytorch/deprecated_api/test_remove_1-8.py
@@ -1010,7 +1010,7 @@ def test_trainer_config_ipus(monkeypatch, trainer_kwargs, expected_ipus):
 @mock.patch("pytorch_lightning.accelerators.ipu.IPUAccelerator.is_available", return_value=True)
 def test_v1_8_0_deprecated_lightning_ipu_module(_, monkeypatch):
     monkeypatch.setattr(pytorch_lightning.strategies.ipu, "_IPU_AVAILABLE", True)
-    with pytest.deprecated_call(match=r"is deprecated in v1.7 and will be removed in v1.8."):
+    with pytest.deprecated_call(match=r"has been deprecated in v1.7.0 and will be removed in v1.8."):
         _ = LightningIPUModule(BoringModel(), 32)
 
 

From 87d14266f5010a1a5bf18153d7a9e103410f932a Mon Sep 17 00:00:00 2001
From: HMellor <HMellor@users.noreply.github.com>
Date: Thu, 28 Jul 2022 19:45:22 +0100
Subject: [PATCH 18/18] Update src/pytorch_lightning/strategies/ipu.py

Co-authored-by: Rohit Gupta <rohitgr1998@gmail.com>
---
 src/pytorch_lightning/strategies/ipu.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/pytorch_lightning/strategies/ipu.py b/src/pytorch_lightning/strategies/ipu.py
index 8d9cc50b90350..82ba4ad227f7c 100644
--- a/src/pytorch_lightning/strategies/ipu.py
+++ b/src/pytorch_lightning/strategies/ipu.py
@@ -273,7 +273,7 @@ def batch_to_device(self, batch: Any, device: Optional[torch.device] = None, dat
         # This override is necessary because the cast must occur before the data
         # is moved to the device to prevent wasteful host->device copies.
         if self.precision_plugin.precision in (PrecisionType.MIXED, PrecisionType.HALF):
-            batch = apply_to_collection(batch, (FloatTensor, torch.cuda.FloatTensor), function=Tensor.half)
+            batch = apply_to_collection(batch, Tensor, function=Tensor.half)
         # We don't call `super().batch_to_device` because `data.to(device)` is not
         # currently necessary for IPUs. The movement of data from host<->IPU is
         # currently handled by PopTorch.