[Feat] 310p support MoE W8A8 quantizaition (#6641)

### What this PR does / why we need it?
This PR introduces support for W8A8 dynamic quantization for
Mixture-of-Experts (MoE) models on Ascend 310P devices. This is achieved
by:
- Implementing a new quantization scheme
`AscendW8A8DynamicFusedMoEMethod310`.
- Adding a unified MLP implementation (`unified_apply_mlp`) for 310P
that handles both quantized and unquantized paths.
- Refactoring the MoE and quantization configuration logic to correctly
route to the new 310P-specific implementations.
- Adding new e2e and unit tests to verify the functionality of MoE W8A8
quantization.

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

### How was this patch tested?
- Added a new e2e test `test_qwen3_moe_tp2_w8a8` to test MoE W8A8
quantization in a multi-card setup.
- Added several new unit tests for the 310P-specific MoE components,
including `experts_selector`, `fused_moe`, `moe_comm_method`, `moe_mlp`,
and the new `w8a8_dynamic` quantization method.

- vLLM version: v0.15.0
- vLLM main:
d7e17aaacd

---------

Signed-off-by: pu-zhe <zpuaa@outlook.com>

vllm_ascend/_310p/fused_moe/fused_moe.py

@@ -58,7 +58,6 @@ class AscendUnquantizedFusedMoEMethod310(UnquantizedFusedMoEMethod):
         num_expert_group: int | None = None,
         custom_routing_function: Callable | None = None,
         scoring_func: str = "softmax",
-        routed_scaling_factor: float = 1.0,
         e_score_correction_bias: torch.Tensor | None = None,
         global_num_experts: int = -1,
         expert_map: torch.Tensor | None = None,
@@ -67,7 +66,6 @@ class AscendUnquantizedFusedMoEMethod310(UnquantizedFusedMoEMethod):
     ) -> torch.Tensor:
         zero_expert_num = getattr(layer, "zero_expert_num", 0)
         zero_expert_type = getattr(layer, "zero_expert_type", None)
-        assert routed_scaling_factor == 1.0
 
         topk_weights, topk_ids = select_experts(
             hidden_states=x,
@@ -195,44 +193,36 @@ class AscendFusedMoE310(FusedMoE):
 
         method = quant_method.quant_method
         quant_type = getattr(method, "quant_type", QuantType.NONE)
-        if quant_type != QuantType.NONE:
-            # TODO: w8a8 quantization will be supported soon, and only reject w4a8 here.
-            raise RuntimeError("W8A8 is not supported currently.")
-        return QuantType.NONE
+        if quant_type not in [QuantType.NONE, QuantType.W8A8]:
+            raise RuntimeError("Only Unquant and W8A8 is supported.")
+        return quant_type
 
     def forward_impl(  # type: ignore[override]
         self, hidden_states: torch.Tensor, router_logits: torch.Tensor
     ) -> torch.Tensor:
         assert self.quant_method is not None
+        assert self.routed_scaling_factor == 1.0, "routed_scaling_factor != 1.0 is not supported."
         forward_context = get_forward_context()
 
         hidden_states, router_logits, _, context_metadata = forward_context.moe_comm_method.prepare(
             hidden_states=hidden_states, router_logits=router_logits, quant_type=self.quant_type
         )
 
-        if isinstance(hidden_states, tuple):
-            hidden_states, pertoken_scale = hidden_states
-        else:
-            pertoken_scale = None
-
         # Matrix multiply.
         fused_experts_results: FusedExpertsResult = self.quant_method.apply(
             layer=self,
             x=hidden_states,
-            router_logits=router_logits,
-            pertoken_scale=pertoken_scale,
-            top_k=self.top_k,
-            renormalize=self.renormalize,
             use_grouped_topk=self.use_grouped_topk,
-            global_num_experts=self.global_num_experts,
-            expert_map=self.local_expert_map,
+            top_k=self.top_k,
+            router_logits=router_logits,
+            renormalize=self.renormalize,
             topk_group=self.topk_group,
             num_expert_group=self.num_expert_group,
             custom_routing_function=self.custom_routing_function,
             scoring_func=self.scoring_func,
-            routed_scaling_factor=self.routed_scaling_factor,
             e_score_correction_bias=self.e_score_correction_bias,
-            activation=self.activation,
+            global_num_experts=self.global_num_experts,
+            expert_map=self.local_expert_map,
             apply_router_weight_on_input=self.apply_router_weight_on_input,
         )