Remove qwen3 moe MC2 cumsum & cast (#3126)

What this PR does / why we need it?
The Qwen3 moe MC2 graph currently has two redundant computational
operator implementations. After npu_moe_distribute_dispatch_v2, the
cumsum and cast operations have been added. By using
expert_token_nums_type=0 and not converting weight_scale to float32,
these two operators can be eliminated, thereby improving inference
performance.

Does this PR introduce any user-facing change?
No

How was this patch tested?
No need

vLLM version: v0.10.2
vLLM main:
vllm-project/vllm@f225ea7

- vLLM version: v0.10.2
- vLLM main:
vllm-project/vllm@f225ea7

---------

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

Author: florenceCH

tests/ut/ops/test_token_dispatcher.py

@@ -98,7 +98,7 @@ def test_token_permutation_dispatch(self):
                                                     self.row_idx, expert_map)
             mock_dispatch.assert_called_once()
             self.assertEqual(output["group_list_type"],
-                             1)  # group_list_type == 1
+                             0)  # group_list_type == 0
 
     def test_token_dispatch_with_shared_experts_and_quant(self):
         self.shared_experts = MagicMock()

vllm_ascend/ops/moe/moe_mlp.py

@@ -79,8 +79,6 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
 
     is_mc2 = get_forward_context().moe_comm_type == MoECommType.MC2
     if w1_scale_bias is None and is_mc2:
-        if w1_scale.dtype != torch.float32:
-            w1_scale = w1_scale.to(torch.float32)
         if fusion:
             # gmm1: gate_up_proj & act_fn: swiglu
             hidden_states, swiglu_out_scale, _ = torch_npu.npu_grouped_matmul_swiglu_quant(
@@ -90,6 +88,8 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
                 weight_scale=w1_scale,
                 x_scale=pertoken_scale)
         else:
+            if w1_scale.dtype != torch.float32:
+                w1_scale = w1_scale.to(torch.float32)
             # gmm1: gate_up_proj
             hidden_states = torch_npu.npu_grouped_matmul(
                 x=[hidden_states],

vllm_ascend/ops/moe/token_dispatcher.py

@@ -133,6 +133,7 @@ def get_dispatch_mc2_kwargs(
             "shared_expert_rank_num": 0,
             "moe_expert_num": moe_expert_num,
             "global_bs": 0,
+            "expert_token_nums_type": 0,
         }
 
         stage1_kwargs = {
@@ -204,7 +205,7 @@ def token_dispatch(self,
             if shared_experts is not None:
                 shared_gate_up, _ = shared_experts.gate_up_proj(hidden_states)
                 self.shared_act = shared_experts.act_fn(shared_gate_up)
-        group_list_type = 1
+        group_list_type = 0
         return {
             "group_list_type": group_list_type,
             "hidden_states": expand_x,