[Bugifx] fix quant_apply_mlp w1_scale type error & fix getting num_local_expert

vllm_ascend/eplb/adaptor/vllm_adaptor.py

@@ -107,8 +107,8 @@ def init_buffer_tensor(self, num_buffer_tensor):
                 self.buffer_tensor_list[buffer_id].append(buffer_tensor)
 
     def init_expert_param_per_layer(self):
-        num_local_expert = self.param_dict["model.layers." + str(self.num_dense_layers) + \
-                                           ".mlp.experts." + self.expert_weight_names[0]].data.shape[0]
+        key = f"model.layers.{self.num_dense_layers}.mlp.experts.{self.expert_weight_names[0]}"
+        num_local_expert = len(self.param_dict[key])
         for moe_layer_id in range(self.num_moe_layers):
             layer_idx = self.num_dense_layers + moe_layer_id
             self.expert_param_per_layer[layer_idx] = list()

vllm_ascend/ops/fused_moe/moe_mlp.py

@@ -129,7 +129,7 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
             # act_fn: swiglu
             hidden_states, swiglu_out_scale = torch_npu.npu_dequant_swiglu_quant(
                 x=hidden_states,
-                weight_scale=w1_scale,
+                weight_scale=w1_scale[0],
                 activation_scale=pertoken_scale,
                 bias=None,
                 quant_scale=None,

vllm_ascend/quantization/w8a8_dynamic.py

@@ -289,7 +289,7 @@ def process_weights_after_loading(self, layer):
             ]
             layer.w13_weight_scale_fp32_list = [
                 weight.clone()
-                for weight in layer.w13_weight_scale.data.unbind(dim=0)
+                for weight in layer.w13_weight_scale_fp32.data.unbind(dim=0)
             ]
             layer.w2_weight_scale_list = [
                 weight.clone()