[BugFix]Fix moe load problems in torchair when using dynamic eplb (#3381)

### What this PR does / why we need it?

When using dynamic eplb, moe load is not imported. We fix this problem
by modifying the return value of hidden states in torchair.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
DeepseekV3 in A3.


- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: daishixun <dsxsteven@sina.com>

vllm_ascend/torchair/ops/torchair_fused_moe.py

@@ -1279,13 +1279,15 @@ class TorchairAscendFusedMoE(FusedMoE):
         )
 
         if shared_experts:
-            if isinstance(e_hidden_states, tuple):
+            if isinstance(e_hidden_states,
+                          tuple) and len(e_hidden_states) == 2:
                 e_hidden_states, shared_hidden_states = e_hidden_states
 
         if self.dynamic_eplb and isinstance(
                 e_hidden_states, tuple) and len(e_hidden_states) == 3:
-            self.moe_load += e_hidden_states[2] if e_hidden_states[1] == 0 else \
-                torch.cat(e_hidden_states[2][:1], e_hidden_states[2][1:] - e_hidden_states[2][:-1])
+            e_hidden_states, group_list_type, expert_tokens = e_hidden_states
+            self.moe_load += expert_tokens if group_list_type else \
+                torch.cat([expert_tokens[:1], expert_tokens[1:] - expert_tokens[:-1]])
 
         if (fused_moe_state not in [
                 FusedMoEState.AllGather, FusedMoEState.AllGatherEP,

vllm_ascend/torchair/quantization/torchair_w8a8_dynamic.py

@@ -220,6 +220,7 @@ def torchair_fused_experts_with_mc2(
     shared_dequant_scale: Optional[Any] = None,
     w1_scale_bias: torch.Tensor = None,
     w2_scale_bias: torch.Tensor = None,
+    dynamic_eplb: bool = False,
 ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
     assert mc2_mask is not None
     if log2phy is not None:
@@ -354,6 +355,9 @@ def torchair_fused_experts_with_mc2(
     ) if enable_dispatch_v2 else torch_npu.npu_moe_distribute_combine(
         **kwargs_mc2)
 
+    if dynamic_eplb:
+        return (hidden_states, 1, expert_token_nums)
+
     if shared_experts is None:
         return hidden_states
     else:
@@ -832,6 +836,7 @@ class TorchairAscendW8A8DynamicFusedMoEMethod:
         self.ep_group = get_ep_group()
 
         ascend_config = get_ascend_config()
+        self.dynamic_eplb = ascend_config.dynamic_eplb
         self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
         self.enable_shared_expert_dp = ascend_config.enable_shared_expert_dp
 
@@ -994,7 +999,8 @@ class TorchairAscendW8A8DynamicFusedMoEMethod:
                 is_torchair=self.torchair_graph_enabled,
                 mc2_mask=kwargs.get("mc2_mask", None),
                 shared_gate_up=shared_gate_up,
-                shared_dequant_scale=shared_dequant_scale)
+                shared_dequant_scale=shared_dequant_scale,
+                dynamic_eplb=self.dynamic_eplb)
         elif fused_moe_state in [
                 FusedMoEState.AllGather, FusedMoEState.NaiveMulticast
         ]: