[zero] find miss code (#378)

colossalai/utils/commons/__init__.py

@@ -0,0 +1,3 @@
+from .bucket_tensor_copy import BucketizedTensorCopy
+
+__all__ = ['BucketizedTensorCopy']

colossalai/utils/commons/bucket_tensor_copy.py

@@ -0,0 +1,61 @@
+import torch
+from colossalai.zero.sharded_param import ShardedParamV2
+from colossalai.utils import get_current_device
+from typing import List
+
+
+class BucketizedTensorCopy(object):
+
+    def __init__(
+        self,
+        chunk_size: int,
+    ):
+        r""" 
+        torch.nn.Parameter CPU (fp32) -> ShardedParam GPU (fp16)
+        TODO(jiaruifang) The class is a little bit hardcoded
+        I will make it more general later.
+        """
+
+        self.chunk_size = chunk_size
+        self._offset = 0
+        self._cpu_buffer = torch.empty(chunk_size, dtype=torch.float, device=torch.device("cpu:0"), pin_memory=True)
+        self._cuda_buffer = torch.empty(chunk_size,
+                                        dtype=torch.half,
+                                        device=torch.device(f"cuda:{get_current_device()}"))
+
+        self._buffered_param_list: List[ShardedParamV2] = []
+        self._numel_list = []
+
+    def copy(self, src_param: torch.nn.Parameter, target_param: ShardedParamV2):
+        assert isinstance(target_param, ShardedParamV2)
+        assert isinstance(src_param, torch.nn.Parameter)
+
+        numel = src_param.numel()
+
+        if self._offset + numel > self.chunk_size:
+            self.flush()
+
+        assert src_param.data.device.type == 'cpu'
+        self._cpu_buffer.narrow(0, self._offset, numel).copy_(src_param.data.view(-1))
+
+        self._buffered_param_list.append(target_param)
+        self._numel_list.append(numel)
+
+        self._offset += numel
+
+    def flush(self):
+        """
+        flush to cuda memory
+        """
+        self._cuda_buffer.copy_(self._cpu_buffer)
+        flush_offset = 0
+        for sparam, numel in zip(self._buffered_param_list, self._numel_list):
+            sparam.data.copy_payload(self._cpu_buffer.narrow(0, flush_offset, numel))
+            flush_offset += numel
+
+        self.reset()
+
+    def reset(self):
+        self._buffered_param_list = []
+        self._numel_list = []
+        self._offset = 0

colossalai/zero/sharded_optim/sharded_optim_v2.py

@@ -88,14 +88,19 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
             self.zero_grad()
             return
 
-        # Write master param to p.data
+        # assign master param pointers to p.data.
+        # We will not trigger data copy here.
         for group in self.optim.param_groups:
             for p in group['params']:
                 p.data = self.master_params[p]
                 # Now p.data is sharded
                 # So optimizer states are sharded naturally
+
         ret = self.optim.step(*args, **kwargs)
-        # Write master param to payload
+
+        # Copy master param data (fp32) to payload of col_attr (fp16)
+        # TODO() improve efficiency by gathering tensors into a chunk and transfering
+        # a chunk.
         for group in self.optim.param_groups:
             for p in group['params']:
                 is_param_sharded = p.col_attr.data.is_sharded
@@ -108,7 +113,10 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
                     self.shard_strategy.shard([p.col_attr.data])
                 # We have to use `copy_payload` instead of `reset_payload`
                 # Since p.data is fp32 and p.col_attr.data is fp16
+
+                # TODO() optimize this line
                 p.col_attr.data.copy_payload(p.data)
+
                 if not is_param_sharded:
                     # We gather full fp16 param here
                     self.shard_strategy.gather([p.col_attr.data])

colossalai/zero/sharded_param/sharded_tensor.py

@@ -14,7 +14,6 @@ class ShardedTensor(object):
         self.world_size = dist.get_world_size(self.process_group)
         self.local_rank = dist.get_rank(self.process_group)
         self._is_sharded = False
-        self._payload = tensor
 
         self._origin_shape = tensor.shape
         self._origin_numel = tensor.numel()
@@ -41,7 +40,7 @@ class ShardedTensor(object):
         return self._payload
 
     def copy_payload(self, tensor):
-        self._payload.copy_(tensor)
+        self._payload.view(-1).copy_(tensor.view(-1))
 
     def reset_payload(self, tensor):
         del self._payload