update to newest mcore code

Signed-off-by: Chen Cui <chcui@nvidia.com>

nemo/collections/nlp/modules/common/megatron/adapters/mcore_mixins.py

@@ -400,7 +400,7 @@ def forward(self, permuted_local_hidden_states, tokens_per_expert):
 
         cumsum_num_tokens = torch.cumsum(tokens_per_expert, dim=0)
         # Insert zero at the begining for offset index's convenience
-        zero_tensor = torch.zeros(1, dtype=torch.long)
+        zero_tensor = torch.zeros(1, dtype=torch.long, device=cumsum_num_tokens.device)
         cumsum_num_tokens = torch.cat((zero_tensor, cumsum_num_tokens))
         for expert_num, expert in enumerate(self.local_experts):
             start = cumsum_num_tokens[expert_num]