[zero] refactor shard and gather operation (#773)

Author: 1SAA

colossalai/zero/shard_utils/commons.py

@@ -14,7 +14,9 @@ def get_shard(tensor: torch.Tensor, rank: int, world_size: int) -> Tuple[torch.T
     num_to_pad = chunks[0].numel() - chunks[rank].numel()
     assert num_to_pad >= 0, num_to_pad
 
-    shard = chunks[rank].clone()
-    if num_to_pad > 0:
-        shard = F.pad(shard, [0, num_to_pad])
+    shard = torch.zeros_like(chunks[0])
+    length = chunks[rank].size(0)
+    shard_temp = shard[:length]
+    shard_temp.copy_(chunks[rank])
+
     return shard, num_to_pad

colossalai/zero/shard_utils/tensor_shard_strategy.py

@@ -43,18 +43,16 @@ def _gather_tensor(self, t: ShardedTensor, process_group: Optional[dist.ProcessG
         if not t.is_sharded:
             return
         target_device = t.device
-        buffer_list = []
         payload_numel = t.payload.numel()
         world_size = dist.get_world_size(process_group)
         rank = dist.get_rank(process_group)
-        for i in range(world_size):
-            if i == rank:
-                buffer_list.append(t.payload.cuda(get_current_device()))
-            else:
-                buffer_list.append(torch.zeros(payload_numel, dtype=t.dtype, device=get_current_device()))
+
+        buffer = torch.empty(payload_numel * world_size, dtype=t.payload.dtype, device=get_current_device())
+        buffer_list = list(torch.chunk(buffer, chunks=world_size, dim=0))
+        buffer_list[rank].copy_(t.payload)
 
         dist.all_gather(buffer_list, buffer_list[rank], group=process_group, async_op=False)
-        gathered_payload = torch.narrow(torch.cat(buffer_list), 0, 0, t.origin_numel).reshape(t.origin_shape)
+        gathered_payload = torch.narrow(buffer, 0, 0, t.origin_numel).reshape(t.origin_shape)
         t.reset_payload(gathered_payload)
         colo_model_data_tensor_move_inline(t, target_device)
         t.is_sharded = False