diff --git a/deepspeed/runtime/zero/stage2.py b/deepspeed/runtime/zero/stage2.py
index f502a6807d7f..92c3e230a48f 100755
--- a/deepspeed/runtime/zero/stage2.py
+++ b/deepspeed/runtime/zero/stage2.py
@@ -770,10 +770,7 @@ def report_ipg_memory_usage(self, tag, param_elems):
 
     # create a flat tensor aligned at the alignment boundary
     def flatten_dense_tensors_aligned(self, tensor_list, alignment):
-        num_elements = 0
-        for tensor in tensor_list:
-            num_elements = num_elements + tensor.numel()
-
+        num_elements = sum(t.numel() for t in tensor_list)
         remaining = num_elements % alignment
 
         if remaining:
@@ -782,8 +779,6 @@ def flatten_dense_tensors_aligned(self, tensor_list, alignment):
                                      device=tensor_list[0].device,
                                      dtype=tensor_list[0].dtype)
             padded_tensor_list = tensor_list + [pad_tensor]
-
-            num_elements = num_elements + elements_to_add
         else:
             padded_tensor_list = tensor_list
 
diff --git a/deepspeed/utils/zero_to_fp32.py b/deepspeed/utils/zero_to_fp32.py
index 357328b10ed9..337c21abd7e6 100755
--- a/deepspeed/utils/zero_to_fp32.py
+++ b/deepspeed/utils/zero_to_fp32.py
@@ -220,16 +220,27 @@ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
                           unpartitioned_numel).view(shape)
             offset += partitioned_numel
 
-    if zero_stage == 3:
-        offset *= world_size
+    if zero_stage == 2:
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
 
-    def align_to_4(x):
-        return 4 * math.ceil(x / 4)
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
 
-    if zero_stage == 2:
-        # Z2 started to align to 4 to improve nccl performance
-        offset = align_to_4(offset)
-        avail_numel = align_to_4(avail_numel)
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+    elif zero_stage == 3:
+        offset *= world_size
 
     # Sanity check
     if offset != avail_numel: