[hotfix] fix unsafe async comm in zero (#4404)

* improve stablility of zero

* fix wrong index

* add record stream

colossalai/zero/low_level/bookkeeping/bucket_store.py

@@ -13,15 +13,20 @@ class BucketStore(BaseStore):
     def __init__(self, torch_pg: ProcessGroup):
         super().__init__(torch_pg)
 
-        # init and reset
+        # init
         self.current_group_id = 0
+        self._num_elements_in_bucket = 0
         # mapping gardient slices and parameter
         self.grad_to_param_mapping = dict()
 
+        self._grad_in_bucket = dict()
         self._param_list = []
         self._padding_size = []
+        for rank in range(self._world_size):
+            self._grad_in_bucket[rank] = []
 
-        self.reset()
+        # offset_list records number of tensors in the bucket before each reduction
+        self.offset_list = [0]
 
     def num_elements_in_bucket(self) -> int:
         """Return the total number of elements in bucket
@@ -32,6 +37,12 @@ class BucketStore(BaseStore):
 
         return self._num_elements_in_bucket
 
+    def reset_num_elements_in_bucket(self):
+        """Set the number of elements in bucket to zero.
+        """
+
+        self._num_elements_in_bucket = 0
+
     def add_param_grad(self, group_id: int, param: Tensor, padding_size: int):
         """Add a param to bucket and record the padding size of a param for gradient padding
 
@@ -46,28 +57,32 @@ class BucketStore(BaseStore):
         self._num_elements_in_bucket += (param.numel() + padding_size)
         self.current_group_id = group_id
 
+        # number of tensors in current bucket
+        self.offset_list[-1] += 1
+
     def build_grad_in_bucket(self):
         """Orgnize parameters' gradient(padding and split), follows the paramters' splitting method
 
         Data structure of self._grad_in_bucket:
         {
         rank0: [grad0_rank0, grad1_rank0, ...]
-        rank1: [grad1_rank1, grad1_rank1, ...]
+        rank1: [grad0_rank1, grad1_rank1, ...]
         }
         """
-
         for param, padding_size in zip(self._param_list, self._padding_size):
-            with torch.no_grad():
-                grad = param.grad.detach().flatten()
+            grad = param.grad.clone().detach().flatten()
             if padding_size > 0:
-                    grad = torch.nn.functional.pad(grad, [0, padding_size])
+                with torch.no_grad():
+                    grad = torch.nn.functional.pad(grad.view(-1), [0, padding_size])
             grad_list = grad.split(grad.numel() // self._world_size)
             for rank in range(self._world_size):
-                    grad_current_rank = grad_list[rank].detach()
+                grad_current_rank = grad_list[rank].clone().detach()
                 self.grad_to_param_mapping[id(grad_current_rank)] = id(param)
                 self._grad_in_bucket[rank].append(grad_current_rank)
             param.grad = None
 
+        self.offset_list.append(0)
+
     def get_grad(self) -> Dict:
         """Return the dictionary of gradients slices, of which the keys are ranks
 
@@ -104,10 +119,12 @@ class BucketStore(BaseStore):
         return self.grad_to_param_mapping[id(grad)]
 
     def reset(self):
-        self.grad_to_param_mapping = dict()
-        self._num_elements_in_bucket = 0
-        self._param_list = []
-        self._padding_size = []
-        self._grad_in_bucket = dict()
+        """Reset the bucket storage after reduction, only release the tensors have been reduced
+        """
+        cur_offset = self.offset_list.pop(0)
+        self._param_list = self._param_list[cur_offset:]
+        self._padding_size = self._padding_size[cur_offset:]
+        for _ in range(cur_offset):
+            del self.grad_to_param_mapping[next(iter(self.grad_to_param_mapping))]
         for rank in range(self._world_size):
-            self._grad_in_bucket[rank] = []
+            self._grad_in_bucket[rank] = self._grad_in_bucket[rank][cur_offset:]

colossalai/zero/low_level/low_level_optim.py

@@ -242,10 +242,19 @@ class LowLevelZeroOptimizer(OptimizerWrapper):
     def _run_reduction(self):
         if self._bucket_store.num_elements_in_bucket() > 0:
             self._bucket_store.build_grad_in_bucket()
+
             flat_grads = self._bucket_store.get_flatten_grad()
             flat_grads /= self._world_size
+
+            # ready to add other tensors to bucket
+            self._bucket_store.reset_num_elements_in_bucket()
+
             if self._overlap_communication:
                 stream = self._comm_stream
+                # in case of the memory being reused in the default stream
+                flat_grads.record_stream(stream)
+                # waiting for ops in the default stream finishing
+                stream.wait_stream(torch.cuda.current_stream())
             else:
                 stream = torch.cuda.current_stream()
 

tests/test_zero/test_low_level/test_zero1_2.py

@@ -137,7 +137,7 @@ def exam_zero_1_torch_ddp(world_size, dtype: torch.dtype):
     zero_optimizer = LowLevelZeroOptimizer(zero_optimizer,
                                            overlap_communication=True,
                                            initial_scale=1,
-                                           reduce_bucket_size=262144)
+                                           reduce_bucket_size=1024 * 1024)
 
     torch_optimizer = torch.optim.SGD(torch_model.parameters(), lr=1)