diff --git a/python/sglang/srt/hardware_backend/npu/attention/mla_preprocess.py b/python/sglang/srt/hardware_backend/npu/attention/mla_preprocess.py
index a88179568..42d5ca2f6 100644
--- a/python/sglang/srt/hardware_backend/npu/attention/mla_preprocess.py
+++ b/python/sglang/srt/hardware_backend/npu/attention/mla_preprocess.py
@@ -459,8 +459,9 @@ class NPUFusedMLAPreprocess(torch.nn.Module):
         # assert self.quant_config and self.quant_config.get_name() == "modelslim"
         # route by `qkv_a_proj` quant type as MTP layers can be unquantized
         _is_w8a8 = (
-            hasattr(self.qkv_a_proj.quant_method, "quant_config")
-            and self.qkv_a_proj.quant_method.quant_config.get_name() == "modelslim"
+            hasattr(self.qkv_a_proj.quant_method, "quantization_config")
+            and self.qkv_a_proj.quant_method.quantization_config.get_name()
+            == "modelslim"
         )
         # with the mlaprolog enabled, the kv_b_proj layers are unquantized
         _is_mlaprolog = hasattr(self.quant_config, "ignore") and any(
diff --git a/python/sglang/srt/managers/schedule_policy.py b/python/sglang/srt/managers/schedule_policy.py
index cbb32868b..7ebd5cd7b 100644
--- a/python/sglang/srt/managers/schedule_policy.py
+++ b/python/sglang/srt/managers/schedule_policy.py
@@ -725,6 +725,15 @@ class PrefillAdder:
     def add_one_req(
         self, req: Req, has_chunked_req: bool, truncation_align_size: Optional[int]
     ):
+        if (self.prefill_delayer_single_pass is not None) and (
+            not self.prefill_delayer_single_pass.negotiate_should_allow_prefill(
+                local_prefillable=True,
+                running_batch=self.running_batch.batch_size(),
+                max_prefill_bs=self.max_prefill_bs,
+                max_running_requests=self.max_running_requests,
+            )
+        ):
+            return AddReqResult.OTHER
         # TODO support cp with multiple requests
         # Enabling context parallelism currently presents precision issues;
         # therefore, the prefill-batch setting is temporarily set to 1.
@@ -771,15 +780,6 @@ class PrefillAdder:
 
             if input_tokens >= self.rem_input_tokens and len(self.can_run_list) != 0:
                 return AddReqResult.OTHER
-            if (self.prefill_delayer_single_pass is not None) and (
-                not self.prefill_delayer_single_pass.negotiate_should_allow_prefill(
-                    local_prefillable=True,
-                    running_batch=self.running_batch.batch_size(),
-                    max_prefill_bs=self.max_prefill_bs,
-                    max_running_requests=self.max_running_requests,
-                )
-            ):
-                return AddReqResult.OTHER
 
             if self.dllm_config is not None:
                 if self.rem_dllm_tokens <= 0: