fix typo with colossalai/trainer utils zero (#3908)

colossalai/zero/gemini/chunk/chunk.py

@@ -416,7 +416,7 @@ class Chunk:
         Copy data slice to the memory space indexed by the input tensor in the chunk.
 
         Args:
-            tensor (torch.Tensor): the tensor used to retrive meta information
+            tensor (torch.Tensor): the tensor used to retrieve meta information
             data_slice (torch.Tensor): the tensor to be copied to the chunk
         """
         # sanity check

colossalai/zero/gemini/chunk/manager.py

@@ -157,7 +157,7 @@ class ChunkManager:
         Copy data to the chunk.
 
         Args:
-            tensor (torch.Tensor): the tensor used to retrive meta information
+            tensor (torch.Tensor): the tensor used to retrieve meta information
             data (torch.Tensor): the tensor to be copied to the chunk
         """
         chunk = self.tensor_chunk_map[tensor]

colossalai/zero/gemini/memory_tracer/chunk_memstats_collector.py

@@ -25,7 +25,7 @@ class ChunkMemStatsCollector(MemStatsCollector):
     # override
     def record_model_data_volume(self) -> None:
         """
-        record model data volumn on cuda and cpu.
+        record model data volume on cuda and cpu.
         """
         if self._start_flag and not self.use_outside_memstats:
             cuda_mem = self._chunk_manager.total_mem['cuda']

colossalai/zero/gemini/memory_tracer/memory_monitor.py

@@ -45,7 +45,7 @@ class MemoryMonitor:
 
 class AsyncMemoryMonitor(MemoryMonitor):
     """
-    An Async Memory Monitor runing during computing. Sampling memory usage of the current GPU
+    An Async Memory Monitor running during computing. Sampling memory usage of the current GPU
     at interval of `1/(10**power)` sec.
 
     The idea comes from Runtime Memory Tracer of PatrickStar
@@ -67,7 +67,7 @@ class AsyncMemoryMonitor(MemoryMonitor):
         async_mem_monitor.save('log.pkl')
 
     Args:
-        power (int, optional): the power of time interva. Defaults to 10.
+        power (int, optional): the power of time interval. Defaults to 10.
 
     .. _PatrickStar: Parallel Training of Pre-trained Models via Chunk-based Memory Management:
         https://arxiv.org/abs/2108.05818

colossalai/zero/gemini/utils.py

@@ -73,7 +73,7 @@ def get_static_torch_model(zero_ddp_model,
         zero_ddp_model (ZeroDDP): a zero ddp model
         device (torch.device): the device of the final torch model
         dtype (torch.dtype): the dtype of the final torch model
-        only_rank_0 (bool): if True, only rank0 has the coverted torch model
+        only_rank_0 (bool): if True, only rank0 has the converted torch model
 
     Returns:
         torch.nn.Module: a static torch model used for saving checkpoints or numeric checks

colossalai/zero/legacy/gemini/ophooks/utils.py

@@ -88,7 +88,7 @@ def register_ophooks_recursively(module: torch.nn.Module,
                                  ophook_list: List[BaseOpHook],
                                  name: str = "",
                                  filter_fn: Optional[Callable] = None):
-    r"""Recursilvely register pre/post hooks for all submodules in the module in FWD and BWD."""
+    r"""Recursively register pre/post hooks for all submodules in the module in FWD and BWD."""
     assert isinstance(module, torch.nn.Module)
     assert isinstance(ophook_list, (list, tuple))
     assert len(ophook_list) > 0, 'expected at least 1 hook in the argument ophook_list but found 0'
@@ -103,7 +103,7 @@ def register_ophooks_recursively(module: torch.nn.Module,
     if len(list(module.parameters(recurse=False))) == 0:
         return
 
-    # return from flitered module
+    # return from filtered module
     if filter_fn is not None and filter_fn(module):
         return
 

colossalai/zero/legacy/gemini/tensor_utils.py

@@ -77,7 +77,7 @@ def colo_model_data_tensor_move_inline(t: Union[StatefulTensor, torch.Tensor], t
     move a tensor to the target_device
     Args:
         t (Union[StatefulTensor, torch.Tensor]): the tensor be moved
-        target_device: a traget device, if type is int, it the index of cuda card.
+        target_device: a target device, if type is int, it the index of cuda card.
     """
     if not isinstance(target_device, torch.device):
         target_device = torch.device(f'cuda:{target_device}')