[Feature] Support engine with NPU backend. (#572)

* init npu * Update mmengine/optim/optimizer/amp_optimizer_wrapper.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * Update mmengine/dist/dist.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * change to is_hccl_backend * Update mmengine/optim/optimizer/amp_optimizer_wrapper.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * add comment with AmpOptimWrapper * Update mmengine/runner/amp.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * Update mmengine/runner/amp.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * add npu fn in base_model * Update mmengine/optim/optimizer/amp_optimizer_wrapper.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * clean lint * Update mmengine/optim/optimizer/amp_optimizer_wrapper.py Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> * Update mmengine/model/base_model/base_model.py Co-authored-by: Mashiro <57566630+HAOCHENYE@users.noreply.github.com> * add is_npu_available * try to fix * Add comments * Refine grammar Co-authored-by: Zaida Zhou <58739961+zhouzaida@users.noreply.github.com> Co-authored-by: Mashiro <57566630+HAOCHENYE@users.noreply.github.com> Co-authored-by: HAOCHENYE <21724054@zju.edu.cn>
open-mmlab · Oct 24, 2022 · 601db12 · 601db12
1 parent d270516
commit 601db12
Show file tree

Hide file tree

Showing 11 changed files with 83 additions and 16 deletions.
diff --git a/docs/en/api/device.rst b/docs/en/api/device.rst
@@ -13,5 +13,6 @@ mmengine.device
    get_device
    get_max_cuda_memory
    is_cuda_available
+   is_npu_available
    is_mlu_available
    is_mps_available
diff --git a/docs/zh_cn/api/device.rst b/docs/zh_cn/api/device.rst
@@ -13,5 +13,6 @@ mmengine.device
    get_device
    get_max_cuda_memory
    is_cuda_available
+   is_npu_available
    is_mlu_available
    is_mps_available
diff --git a/mmengine/device/__init__.py b/mmengine/device/__init__.py
@@ -1,8 +1,8 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from .utils import (get_device, get_max_cuda_memory, is_cuda_available,
-                    is_mlu_available, is_mps_available)
+                    is_mlu_available, is_mps_available, is_npu_available)
 
 __all__ = [
     'get_max_cuda_memory', 'get_device', 'is_cuda_available',
-    'is_mlu_available', 'is_mps_available'
+    'is_mlu_available', 'is_mps_available', 'is_npu_available'
 ]
diff --git a/mmengine/device/utils.py b/mmengine/device/utils.py
@@ -32,6 +32,15 @@ def is_cuda_available() -> bool:
     return torch.cuda.is_available()
 
 
+def is_npu_available() -> bool:
+    """Returns True if Ascend PyTorch and npu devices exist."""
+    try:
+        import torch_npu  # noqa: F401
+    except Exception:
+        return False
+    return hasattr(torch, 'npu') and torch.npu.is_available()
+
+
 def is_mlu_available() -> bool:
     """Returns True if Cambricon PyTorch and mlu devices exist."""
     return hasattr(torch, 'is_mlu_available') and torch.is_mlu_available()
@@ -49,9 +58,11 @@ def get_device() -> str:
     """Returns the currently existing device type.
 
     Returns:
-        str: cuda | mlu | mps | cpu.
+        str: cuda | npu | mlu | mps | cpu.
     """
-    if is_cuda_available():
+    if is_npu_available():
+        return 'npu'
+    elif is_cuda_available():
         return 'cuda'
     elif is_mlu_available():
         return 'mlu'

diff --git a/mmengine/dist/dist.py b/mmengine/dist/dist.py
@@ -20,6 +20,7 @@
                     get_comm_device, cast_data_device)
 from mmengine.utils import digit_version
 from mmengine.utils.dl_utils import TORCH_VERSION
+from mmengine.device import is_npu_available
 
 
 def _get_reduce_op(name: str) -> torch_dist.ReduceOp:
@@ -411,7 +412,11 @@ def _broadcast_object_list(object_list: List[Any],
     group_backend = get_backend(group)
     is_nccl_backend = group_backend == torch_dist.Backend.NCCL
     current_device = torch.device('cpu')
-    if is_nccl_backend:
+    is_hccl_backend = group_backend == 'hccl'
+    if is_hccl_backend:
+        current_device = torch.npu.current_device()
+        object_sizes_tensor = object_sizes_tensor.to(current_device)
+    elif is_nccl_backend:
         # See note about using torch.cuda.current_device() here in
         # docstring. We cannot simply use my_rank since rank == device is
         # not necessarily true.
@@ -430,7 +435,7 @@ def _broadcast_object_list(object_list: List[Any],
             dtype=torch.uint8,
         )
 
-    if is_nccl_backend:
+    if is_nccl_backend or is_hccl_backend:
         object_tensor = object_tensor.to(current_device)
     torch_dist.broadcast(object_tensor, src=src, group=group)
     # Deserialize objects using their stored sizes.
@@ -504,7 +509,8 @@ def broadcast_object_list(data: List[Any],
         if group is None:
             group = get_default_group()
 
-        if digit_version(TORCH_VERSION) >= digit_version('1.8.0'):
+        if digit_version(TORCH_VERSION) >= digit_version(
+                '1.8.0') and not is_npu_available():
             torch_dist.broadcast_object_list(data, src, group)
         else:
             _broadcast_object_list(data, src, group)

diff --git a/mmengine/dist/utils.py b/mmengine/dist/utils.py
@@ -10,7 +10,7 @@
 from torch import Tensor
 from torch import distributed as torch_dist
 from torch.distributed import ProcessGroup
-from mmengine.device import is_mlu_available
+from mmengine.device import is_mlu_available, is_npu_available
 
 from collections.abc import Iterable, Mapping
 
@@ -80,6 +80,14 @@ def _init_dist_pytorch(backend, **kwargs) -> None:
             rank=rank,
             world_size=int(os.environ['WORLD_SIZE']),
             **kwargs)
+    elif is_npu_available():
+        import torch_npu  # noqa: F401
+        torch.npu.set_device(rank)
+        torch_dist.init_process_group(
+            backend='hccl',
+            rank=rank,
+            world_size=int(os.environ['WORLD_SIZE']),
+            **kwargs)
     else:
         num_gpus = torch.cuda.device_count()
         torch.cuda.set_device(rank % num_gpus)
@@ -437,7 +445,10 @@ def get_comm_device(group: Optional[ProcessGroup] = None) -> torch.device:
         torch.device: The device of backend.
     """
     backend = get_backend(group)
-    if backend == torch_dist.Backend.NCCL:
+    if backend == 'hccl':
+        import torch_npu  # noqa: F401
+        return torch.device('npu', torch.npu.current_device())
+    elif backend == torch_dist.Backend.NCCL:
         return torch.device('cuda', torch.cuda.current_device())
     elif backend == 'cncl':
         import torch_mlu  # noqa: F401

diff --git a/mmengine/model/base_model/base_model.py b/mmengine/model/base_model/base_model.py
@@ -210,6 +210,25 @@ def cuda(
         self._set_device(torch.device(device))
         return super().cuda(device)
 
+    def npu(
+        self,
+        device: Union[int, str, torch.device, None] = None,
+    ) -> nn.Module:
+        """Overrides this method to call :meth:`BaseDataPreprocessor.npu`
+        additionally.
+
+        Returns:
+            nn.Module: The model itself.
+
+        Note:
+            This generation of NPU(Ascend910) does not support
+            the use of multiple cards in a single process,
+            so the index here needs to be consistent with the default device
+        """
+        device = torch.npu.current_device()
+        self._set_device(device)
+        return super().npu()
+
     def cpu(self, *args, **kwargs) -> nn.Module:
         """Overrides this method to call :meth:`BaseDataPreprocessor.cpu`
         additionally.

diff --git a/mmengine/optim/optimizer/amp_optimizer_wrapper.py b/mmengine/optim/optimizer/amp_optimizer_wrapper.py
@@ -3,13 +3,18 @@
 
 import torch
 import torch.nn as nn
-from torch.cuda.amp import GradScaler
 
+from mmengine.device import is_cuda_available, is_npu_available
 from mmengine.registry import OPTIM_WRAPPERS
 from mmengine.utils import digit_version
 from mmengine.utils.dl_utils import TORCH_VERSION
 from .optimizer_wrapper import OptimWrapper
 
+if is_npu_available():
+    from torch.npu.amp import GradScaler
+else:
+    from torch.cuda.amp import GradScaler
+
 
 @OPTIM_WRAPPERS.register_module()
 class AmpOptimWrapper(OptimWrapper):
@@ -44,8 +49,8 @@ class AmpOptimWrapper(OptimWrapper):
     def __init__(self, loss_scale='dynamic', **kwargs):
         assert digit_version(TORCH_VERSION) >= digit_version('1.6.0'), (
             '`torch.cuda.amp` is only available when pytorch version >= 1.6')
-        assert torch.cuda.is_available(), (
-            '``AmpOptimizerWrapper`` is only available training on gpu')
+        assert is_cuda_available() or is_npu_available(), (
+            '``AmpOptimizerWrapper`` is only available training on gpu or npu')
         super().__init__(**kwargs)
         self._scale_update_param = None
         if loss_scale == 'dynamic':

diff --git a/mmengine/optim/optimizer/builder.py b/mmengine/optim/optimizer/builder.py
@@ -7,6 +7,7 @@
 import torch.nn as nn
 
 from mmengine.config import Config, ConfigDict
+from mmengine.device import is_npu_available
 from mmengine.registry import OPTIM_WRAPPER_CONSTRUCTORS, OPTIMIZERS
 from .optimizer_wrapper import OptimWrapper
 
@@ -53,6 +54,13 @@ def build_optim_wrapper(model: nn.Module,
     constructor_type = optim_wrapper_cfg.pop('constructor',
                                              'DefaultOptimWrapperConstructor')
     paramwise_cfg = optim_wrapper_cfg.pop('paramwise_cfg', None)
+
+    # Since the current generation of NPU(Ascend 910) only supports
+    # mixed precision training, here we turn on mixed precision by default
+    # on the NPU to make the training normal
+    if is_npu_available():
+        optim_wrapper_cfg['type'] = 'AmpOptimWrapper'
+
     optim_wrapper_constructor = OPTIM_WRAPPER_CONSTRUCTORS.build(
         dict(
             type=constructor_type,

diff --git a/mmengine/runner/amp.py b/mmengine/runner/amp.py
@@ -5,7 +5,7 @@
 
 import torch
 
-from mmengine.device import get_device
+from mmengine.device import get_device, is_cuda_available, is_npu_available
 from mmengine.logging import print_log
 from mmengine.utils import digit_version
 from mmengine.utils.dl_utils import TORCH_VERSION
@@ -86,7 +86,10 @@ def autocast(device_type: Optional[str] = None,
                 logger='current',
                 level=logging.WARNING)
 
-        if torch.cuda.is_available():
+        if is_npu_available():
+            with torch.npu.amp.autocast(enabled=enabled):
+                yield
+        elif is_cuda_available():
             with torch.cuda.amp.autocast(enabled=enabled):
                 yield
         else:

diff --git a/tests/test_device/test_device.py b/tests/test_device/test_device.py
@@ -1,11 +1,13 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from mmengine.device import (get_device, is_cuda_available, is_mlu_available,
-                             is_mps_available)
+                             is_mps_available, is_npu_available)
 
 
 def test_get_device():
     device = get_device()
-    if is_cuda_available():
+    if is_npu_available():
+        assert device == 'npu'
+    elif is_cuda_available():
         assert device == 'cuda'
     elif is_mlu_available():
         assert device == 'mlu'