diff --git a/tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py b/tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
index adb0e6ab..b6d8b669 100644
--- a/tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
+++ b/tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
@@ -1,8 +1,5 @@
 from __future__ import annotations
 
-import os
-from unittest.mock import patch
-
 import pytest
 from vllm import SamplingParams
 from vllm.config import CompilationConfig, CUDAGraphMode
@@ -111,19 +108,3 @@ def test_mtp2_correctness_full_graph(
     model_name: str,
 ):
     mtp_correctness(sampling_config, model_name, 2, CUDAGraphMode.FULL)
-
-
-@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_MLAPO": "1"})
-def test_mtp_correctness_piecewise_graph_with_mlapo_kernel(
-    sampling_config: SamplingParams,
-    model_name: str,
-):
-    mtp_correctness(sampling_config, model_name, 1)
-
-
-@patch.dict(os.environ, {"VLLM_ASCEND_ENABLE_MLAPO": "1"})
-def test_mtp_correctness_full_graph_with_mlapo_kernel(
-    sampling_config: SamplingParams,
-    model_name: str,
-):
-    mtp_correctness(sampling_config, model_name, 1, CUDAGraphMode.FULL)
diff --git a/vllm_ascend/attention/mla_v1.py b/vllm_ascend/attention/mla_v1.py
index bcdb6161..86418720 100644
--- a/vllm_ascend/attention/mla_v1.py
+++ b/vllm_ascend/attention/mla_v1.py
@@ -737,7 +737,7 @@ class AscendMLAImpl(MLAAttentionImpl):
         self.qb_qt_bias = qb_qt_bias.reshape(
             self.num_heads * (self.qk_nope_head_dim + self.qk_rope_head_dim))
 
-        device = self.q_a_proj.weight.device
+        device = self.q_proj.weight.device
         self.gamma1 = self.q_a_layernorm.weight.data
         self.beta1 = self.q_a_layernorm.bias.data
         self.gamma2 = self.kv_a_layernorm.weight.data