[refactor] replace scattered business kwargs with typed request objects and explicit stage boundaries (#7024)

### What this PR does / why we need it?
Refactor `vllm_ascend/ops/fused_moe` to replace scattered MoE business
`**kwargs` with typed request objects and explicit stage boundaries.

- Prepare, dispatch, MLP, and quant stages now have clearer ownership.
- Main MoE path no longer depends on business `kwargs.get(...)` lookups.
- Comm and dispatcher interfaces are request-only on the main path.
- UTs can assert stage-level fields directly instead of inferring
behavior indirectly.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
CI passed.

---------

Signed-off-by: linfeng-yuan <1102311262@qq.com>

vllm_ascend/quantization/methods/w4a16.py

@@ -25,8 +25,9 @@ from vllm.config import get_current_vllm_config
 from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.ascend_forward_context import _EXTRA_CTX
 from vllm_ascend.ops.fused_moe.experts_selector import select_experts
+from vllm_ascend.ops.fused_moe.moe_runtime_args import build_fused_experts_input
 
-from .base import AscendMoEScheme
+from .base import AscendMoEScheme, QuantType
 from .registry import register_scheme
 
 
@@ -103,6 +104,8 @@ def pack_to_int32(weight: torch.Tensor) -> torch.Tensor:
 class AscendW4A16FusedMoEMethod(AscendMoEScheme):
     """FusedMoE method for Ascend W4A16."""
 
+    quant_type: QuantType = QuantType.W4A16
+
     def __init__(self) -> None:
         self.transpose_weight = True
         self.num_bits = 4  # dtype = torch.int4
@@ -192,7 +195,10 @@ class AscendW4A16FusedMoEMethod(AscendMoEScheme):
         enable_force_load_balance: bool = True,
         log2phy: torch.Tensor | None = None,
         global_redundant_expert_num: int = 0,
-        **kwargs,
+        pertoken_scale: Any | None = None,
+        activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
+        mc2_mask: torch.Tensor | None = None,
     ) -> torch.Tensor:
         assert router_logits.shape[1] == global_num_experts - global_redundant_expert_num, (
             "Number of global experts mismatch (excluding redundancy)"
@@ -217,20 +223,26 @@ class AscendW4A16FusedMoEMethod(AscendMoEScheme):
 
         moe_comm_method = _EXTRA_CTX.moe_comm_method
         return moe_comm_method.fused_experts(
-            hidden_states=x,
-            w1=layer.w13_weight_packed,
-            w2=layer.w2_weight_packed,
-            w1_scale=layer.w13_weight_scale,
-            w2_scale=layer.w2_weight_scale,
-            w1_offset=layer.w13_weight_offset,
-            w2_offset=layer.w2_weight_offset,
-            topk_weights=topk_weights,
-            topk_ids=topk_ids,
-            use_int4_w4a16=True,
-            expert_map=expert_map,
-            log2phy=log2phy,
-            dynamic_eplb=self.dynamic_eplb,
-            mc2_mask=kwargs.get("mc2_mask"),
+            fused_experts_input=build_fused_experts_input(
+                hidden_states=x,
+                topk_weights=topk_weights,
+                topk_ids=topk_ids,
+                w1=layer.w13_weight_packed,
+                w2=layer.w2_weight_packed,
+                quant_type=self.quant_type,
+                dynamic_eplb=self.dynamic_eplb,
+                expert_map=expert_map,
+                global_redundant_expert_num=global_redundant_expert_num,
+                mc2_mask=mc2_mask,
+                apply_router_weight_on_input=apply_router_weight_on_input,
+                log2phy=log2phy,
+                pertoken_scale=pertoken_scale,
+                activation=activation,
+                w1_scale=layer.w13_weight_scale,
+                w2_scale=layer.w2_weight_scale,
+                w1_offset=layer.w13_weight_offset,
+                w2_offset=layer.w2_weight_offset,
+            )
         )
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None: