diff --git a/pytorch_lightning/accelerators/ddp2_backend.py b/pytorch_lightning/accelerators/ddp2_backend.py
index 7414535350f64..1b55a983c894e 100644
--- a/pytorch_lightning/accelerators/ddp2_backend.py
+++ b/pytorch_lightning/accelerators/ddp2_backend.py
@@ -16,13 +16,9 @@
 
 import torch
 
-from pytorch_lightning import _logger as log
-from pytorch_lightning.utilities import AMPType
-from pytorch_lightning.utilities.distributed import rank_zero_only
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.core.step_result import Result
 from pytorch_lightning.accelerators.ddp_base_backend import DDPBase
-from pytorch_lightning.plugins.apex import ApexPlugin
 
 try:
     from hydra.utils import to_absolute_path, get_original_cwd
@@ -38,7 +34,6 @@ class DDP2Backend(DDPBase):
     def __init__(self, trainer):
         super().__init__(trainer)
         self.task_idx = None
-        self.precision_backend = None
 
     def setup(self, model):
         self._resolve_task_idx()
diff --git a/pytorch_lightning/accelerators/ddp_backend.py b/pytorch_lightning/accelerators/ddp_backend.py
index 614d38bc4dc41..17739a605fa2b 100644
--- a/pytorch_lightning/accelerators/ddp_backend.py
+++ b/pytorch_lightning/accelerators/ddp_backend.py
@@ -22,9 +22,7 @@
 import numpy as np
 import torch
 
-from pytorch_lightning import _logger as log
-from pytorch_lightning.utilities import AMPType
-from pytorch_lightning.utilities.distributed import rank_zero_only, find_free_network_port
+from pytorch_lightning.utilities.distributed import find_free_network_port
 from pytorch_lightning.accelerators.ddp_base_backend import DDPBase
 
 try:
@@ -35,11 +33,6 @@
 else:
     HYDRA_AVAILABLE = True
 
-try:
-    from apex import amp
-except ImportError:
-    amp = None
-
 
 class DDPBackend(DDPBase):
 
diff --git a/pytorch_lightning/accelerators/ddp_base_backend.py b/pytorch_lightning/accelerators/ddp_base_backend.py
index 10e5cab3dcbb9..b0349ac6bbd4d 100644
--- a/pytorch_lightning/accelerators/ddp_base_backend.py
+++ b/pytorch_lightning/accelerators/ddp_base_backend.py
@@ -22,7 +22,6 @@
 from pytorch_lightning.utilities.cloud_io import atomic_save
 from pytorch_lightning.utilities.distributed import rank_zero_warn, rank_zero_only
 from pytorch_lightning import _logger as log
-from pytorch_lightning.plugins.apex import ApexPlugin
 
 try:
     from hydra.utils import to_absolute_path, get_original_cwd
@@ -37,7 +36,6 @@ class DDPBase(Accelerator):
 
     def __init__(self, trainer):
         super().__init__(trainer)
-        self.precision_backend = None
 
     def training_step(self, args):
         if self.trainer.amp_backend == AMPType.NATIVE:
@@ -151,9 +149,7 @@ def ddp_train_tmp(self, process_idx, mp_queue, model, is_master=False, proc_offs
 
         # AMP -
         # run through amp wrapper before going to distributed DP
-        if self.trainer.amp_backend == AMPType.APEX:
-            self.precision_backend = ApexPlugin(self.trainer)
-            model, optimizers = self.precision_backend._init(model)
+        model, optimizers = self.trainer.precision_connector.connect(model, optimizers)
 
         # device ids change depending on the DDP setup
         device_ids = self.get_device_ids()
diff --git a/pytorch_lightning/accelerators/ddp_spawn_backend.py b/pytorch_lightning/accelerators/ddp_spawn_backend.py
index 467712a121fbb..fc2fc88563e1d 100644
--- a/pytorch_lightning/accelerators/ddp_spawn_backend.py
+++ b/pytorch_lightning/accelerators/ddp_spawn_backend.py
@@ -19,11 +19,6 @@
 from pytorch_lightning.utilities.distributed import find_free_network_port
 from pytorch_lightning.accelerators.ddp_base_backend import DDPBase
 
-try:
-    from apex import amp
-except ImportError:
-    amp = None
-
 
 class DDPSpawnBackend(DDPBase):
 
diff --git a/pytorch_lightning/accelerators/dp_backend.py b/pytorch_lightning/accelerators/dp_backend.py
index 09a373c3d4ef0..4bced57e9458c 100644
--- a/pytorch_lightning/accelerators/dp_backend.py
+++ b/pytorch_lightning/accelerators/dp_backend.py
@@ -20,12 +20,6 @@
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.core.step_result import Result
 from pytorch_lightning.accelerators.base_backend import Accelerator
-from pytorch_lightning.plugins.apex import ApexPlugin
-
-try:
-    from apex import amp
-except ImportError:
-    amp = None
 
 
 class DataParallelBackend(Accelerator):
@@ -33,7 +27,6 @@ class DataParallelBackend(Accelerator):
     def __init__(self, trainer):
         super().__init__(trainer)
         self.model_autocast_original_forward = None
-        self.precision_backend = None
 
     def setup(self, model):
         # call setup after the ddp process has connected
@@ -91,8 +84,7 @@ def __init_nvidia_apex(self, model):
                 f' See this note from NVIDIA for more info: https://github.com/NVIDIA/apex/issues/227.'
                 f' We recommend you switch to ddp if you want to use amp')
         else:
-            self.precision_backend = ApexPlugin(self.trainer)
-            model, optimizers = self.precision_backend._init(model)
+            model, optimizers = self.trainer.precision_connector.connect(model, self.trainer.optimizers)
 
         return model
 
diff --git a/pytorch_lightning/accelerators/gpu_backend.py b/pytorch_lightning/accelerators/gpu_backend.py
index ec3e84e840b02..9a00d34e05293 100644
--- a/pytorch_lightning/accelerators/gpu_backend.py
+++ b/pytorch_lightning/accelerators/gpu_backend.py
@@ -15,7 +15,6 @@
 import torch
 from pytorch_lightning.utilities import AMPType
 from pytorch_lightning.accelerators.base_backend import Accelerator
-from pytorch_lightning.plugins.apex import ApexPlugin
 
 
 class GPUBackend(Accelerator):
@@ -23,7 +22,6 @@ class GPUBackend(Accelerator):
 
     def __init__(self, trainer):
         super().__init__(trainer)
-        self.precision_backend = None
 
     def setup(self, model):
 
@@ -40,9 +38,8 @@ def setup(self, model):
         self.trainer.lr_schedulers = lr_schedulers
         self.trainer.optimizer_frequencies = optimizer_frequencies
 
-        if self.trainer.amp_backend == AMPType.APEX:
-            self.precision_backend = ApexPlugin(self.trainer)
-            model, optimizers = self.precision_backend._init(model)
+        # init precision
+        model, optimizers = self.trainer.precision_connector.connect(model, optimizers)
 
         self.trainer.model = model
 
diff --git a/pytorch_lightning/accelerators/horovod_backend.py b/pytorch_lightning/accelerators/horovod_backend.py
index 76887a6f87214..ba4e927a4ce13 100644
--- a/pytorch_lightning/accelerators/horovod_backend.py
+++ b/pytorch_lightning/accelerators/horovod_backend.py
@@ -13,12 +13,10 @@
 # limitations under the License.
 from contextlib import ExitStack
 import torch
-from pytorch_lightning.core import LightningModule
 from pytorch_lightning.utilities import AMPType
 from pytorch_lightning.accelerators.base_backend import Accelerator
 from pytorch_lightning.utilities.distributed import rank_zero_only
 from torch.optim.lr_scheduler import _LRScheduler
-from pytorch_lightning.plugins.apex import ApexPlugin
 
 try:
     import horovod.torch as hvd
@@ -33,7 +31,6 @@ class HorovodBackend(Accelerator):
 
     def __init__(self, trainer):
         super().__init__(trainer)
-        self.precision_backend = None
 
     def setup(self, model):
         # call setup after the ddp process has connected
@@ -83,9 +80,8 @@ def filter_named_parameters(model, optimizer):
             for optimizer in self.trainer.optimizers
         ]
 
-        if self.trainer.amp_backend == AMPType.APEX:
-            self.precision_backend = ApexPlugin(self.trainer)
-            model, optimizers = self.precision_backend._init(model)
+        # 16-bit
+        model, self.trainer.optimizers = self.trainer.precision_connector.connect(model, self.trainer.optimizers)
 
         # Update logger rank info from Horovod to avoid race conditions from  different ranks
         # creating directories / writing files in the same locations.
diff --git a/pytorch_lightning/plugins/apex.py b/pytorch_lightning/plugins/apex.py
index 27ccff5592520..9b51a8a0ab1c5 100644
--- a/pytorch_lightning/plugins/apex.py
+++ b/pytorch_lightning/plugins/apex.py
@@ -23,8 +23,8 @@ class ApexPlugin:
     def __init__(self, trainer):
         self.trainer = trainer
 
-    def _init(self, model):
-        model, optimizers = self.configure_apex(model, self.trainer.optimizers, self.trainer.amp_level)
+    def connect(self, model, optimizers):
+        model, optimizers = self.configure_apex(model, optimizers, self.trainer.amp_level)
         self.trainer.optimizers = optimizers
         self.trainer.reinit_scheduler_properties(self.trainer.optimizers, self.trainer.lr_schedulers)
         return model, optimizers
diff --git a/pytorch_lightning/plugins/native_amp.py b/pytorch_lightning/plugins/native_amp.py
new file mode 100644
index 0000000000000..3089a0558f0a8
--- /dev/null
+++ b/pytorch_lightning/plugins/native_amp.py
@@ -0,0 +1,29 @@
+# Copyright The PyTorch Lightning team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+
+class NativeAMP:
+
+    def __init__(self, trainer):
+        self.trainer = trainer
+
+    def connect(self, model, optimizers):
+        self.trainer.optimizers = optimizers
+        return model, optimizers
+
+    def training_step(self, fx, args):
+        with torch.cuda.amp.autocast():
+            output = fx(*args)
+        return output
diff --git a/pytorch_lightning/trainer/connectors/precision_connector.py b/pytorch_lightning/trainer/connectors/precision_connector.py
index 55fb945caf09e..296b9a178ab43 100644
--- a/pytorch_lightning/trainer/connectors/precision_connector.py
+++ b/pytorch_lightning/trainer/connectors/precision_connector.py
@@ -13,12 +13,15 @@
 # limitations under the License.
 from pytorch_lightning import _logger as log
 from pytorch_lightning.utilities import APEX_AVAILABLE, NATIVE_AMP_AVALAIBLE, rank_zero_warn, AMPType
+from pytorch_lightning.plugins.native_amp import NativeAMP
+from pytorch_lightning.plugins.apex import ApexPlugin
 
 
 class PrecisionConnector:
 
     def __init__(self, trainer):
         self.trainer = trainer
+        self.backend = None
 
     def on_trainer_init(self, precision, amp_level, amp_backend):
         # AMP init
@@ -52,6 +55,8 @@ def _setup_amp_backend(self, amp_type: str):
             else:
                 log.info('Using native 16bit precision.')
                 self.trainer.amp_backend = AMPType.NATIVE
+                self.backend = NativeAMP(self.trainer)
+
         if amp_type == 'apex':
             if not APEX_AVAILABLE:
                 rank_zero_warn('You have asked for Apex AMP but you have not installed it yet.'
@@ -59,8 +64,16 @@ def _setup_amp_backend(self, amp_type: str):
             else:
                 log.info('Using APEX 16bit precision.')
                 self.trainer.amp_backend = AMPType.APEX
+                self.backend = ApexPlugin(self.trainer)
+
         if not self.trainer.amp_backend:
             raise ModuleNotFoundError(
                 f'You have asked for AMP support {amp_type}, but there is no support on your side yet.'
                 f' Consider installing torch >= 1.6 or NVIDIA Apex.'
             )
+
+    def connect(self, model, optimizers):
+        if self.backend:
+            model, optimizers = self.backend.connect(model, optimizers)
+
+        return model, optimizers