[bugfix] Fix moe bug: allgather error. (#3279)

It will crash when deepseek model executed in A2.


- vLLM version: v0.11.0rc3
- vLLM main:
vllm-project/vllm@releases/v0.11.0

---------

Signed-off-by: weijinqian_v1 <[email protected]>
Co-authored-by: weijinqian_v1 <[email protected]>

Author: weijinqian0

tests/ut/ops/test_token_dispatcher.py

@@ -269,6 +269,7 @@ def test_token_dispatch_with_quant(self):
     def test_token_combine_with_expert_map(self):
         self.dispatcher.expert_map = torch.tensor([0, 1, 2, 3])
         self.dispatcher.sorted_token_indices = torch.tensor([0, 1, 1, 1, 1, 1])
+        self.dispatcher.expanded_row_idx = torch.tensor([0, 1, 1, 1, 1, 1])
         self.dispatcher.sorted_weights = torch.tensor(
             [0.5, 0.5, 0.5, 0.5, 0.5, 0.5])
         self.dispatcher.original_shape = (3, 128)

vllm_ascend/ops/moe/token_dispatcher.py

@@ -383,7 +383,7 @@ def token_combine(self,
         assert self.original_shape is not None
         final_hidden_states = torch_npu.npu_moe_token_unpermute(
             permuted_tokens=hidden_states,
-            sorted_indices=self.expanded_row_idx,
+            sorted_indices=torch.abs(self.expanded_row_idx),
             probs=self.topk_weights)
         if len(self.original_shape) == 3:
             final_hidden_states = final_hidden_states.view(self.original_shape)