diff --git a/vllm/model_executor/layers/fused_moe/fused_moe.py b/vllm/model_executor/layers/fused_moe/fused_moe.py
index b4f81527141a8..3cb0419404625 100644
--- a/vllm/model_executor/layers/fused_moe/fused_moe.py
+++ b/vllm/model_executor/layers/fused_moe/fused_moe.py
@@ -203,14 +203,15 @@ def moe_align_block_size(
     - The padding ensures that the total number of tokens is now divisible
         by block_size for proper block matrix operations.
     """
-    sorted_ids = torch.empty(
-        (topk_ids.numel() + num_experts * (block_size - 1), ),
-        dtype=torch.int32,
-        device=topk_ids.device)
-    expert_ids = torch.empty((topk_ids.numel() + num_experts, ),
+    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
+    sorted_ids = torch.empty((max_num_tokens_padded, ),
                              dtype=torch.int32,
                              device=topk_ids.device)
     sorted_ids.fill_(topk_ids.numel())
+    max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)
+    expert_ids = torch.empty((max_num_m_blocks, ),
+                             dtype=torch.int32,
+                             device=topk_ids.device)
     num_tokens_post_pad = torch.empty((1),
                                       dtype=torch.int32,
                                       device=topk_ids.device)