avoid mrope fusion op when running qwen2.5-vl on a+x machine (#4270)

### What this PR does / why we need it?
avoid mrope fusion op when running qwen2.5-vl on a+x machine
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
Test text VQA accuracy on G8600 with aisbench

- vLLM version: v0.11.0
- vLLM main:
vllm-project/vllm@2918c1b

---------

Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>

Author: shaopeng-666

tests/ut/ops/test_rotary_embedding.py

@@ -7,6 +7,7 @@
 from vllm.config import ModelConfig, VllmConfig
 from vllm.model_executor.layers.rotary_embedding import (
     DeepseekScalingRotaryEmbedding, MRotaryEmbedding, RotaryEmbedding)
+from vllm.platforms import CpuArchEnum
 
 from tests.ut.base import TestBase
 from vllm_ascend.ascend_forward_context import set_ascend_forward_context
@@ -424,11 +425,14 @@ def _create_vllm_config(self):
         return vllm_config
 
     @patch('torch_npu.npu_mrope')
+    @patch('vllm_ascend.platform.NPUPlatform.get_cpu_architecture')
     @patch('vllm.config.ModelConfig.__post_init__', MagicMock())
     @patch('vllm.config.VllmConfig.__post_init__', MagicMock())
     @patch('vllm.distributed.parallel_state._DP', MagicMock(world_size=1))
     @patch('vllm.distributed.parallel_state._TP', MagicMock(world_size=1))
-    def test_forward_oot_1d_positions(self, mock_npu_mrope):
+    def test_forward_oot_1d_positions(self, mock_cpu_arc, mock_npu_mrope):
+        mock_cpu_arc.return_value = CpuArchEnum.ARM
+
         mock_npu_mrope.return_value = (torch.zeros_like(self.query),
                                        torch.zeros_like(self.key))
 
@@ -443,11 +447,14 @@ def test_forward_oot_1d_positions(self, mock_npu_mrope):
         self.assertEqual(result_q.shape, self.query.shape)
 
     @patch('torch_npu.npu_mrope')
+    @patch('vllm_ascend.platform.NPUPlatform.get_cpu_architecture')
     @patch('vllm.config.ModelConfig.__post_init__', MagicMock())
     @patch('vllm.config.VllmConfig.__post_init__', MagicMock())
     @patch('vllm.distributed.parallel_state._DP', MagicMock(world_size=1))
     @patch('vllm.distributed.parallel_state._TP', MagicMock(world_size=1))
-    def test_forward_oot_2d_positions(self, mock_npu_mrope):
+    def test_forward_oot_2d_positions(self, mock_cpu_arc, mock_npu_mrope):
+        mock_cpu_arc.return_value = CpuArchEnum.ARM
+
         mock_npu_mrope.return_value = (torch.zeros_like(self.query),
                                        torch.zeros_like(self.key))
 

vllm_ascend/ops/rotary_embedding.py

@@ -24,6 +24,7 @@
 from vllm.model_executor.layers.rotary_embedding import (
     DeepseekScalingRotaryEmbedding, MRotaryEmbedding, RotaryEmbedding,
     YaRNScalingRotaryEmbedding)
+from vllm.platforms import CpuArchEnum
 
 from vllm_ascend.platform import NPUPlatform
 from vllm_ascend.utils import enable_custom_op, is_310p
@@ -405,7 +406,10 @@ def forward_oot(
         query: torch.Tensor,
         key: torch.Tensor,
     ):
-        if self.mrope_section != [16, 24, 24]:
+        # TODO: This judgment will be removed once the mrope precision issue is fixed
+        if self.mrope_section != [
+                16, 24, 24
+        ] or NPUPlatform.get_cpu_architecture() == CpuArchEnum.X86:
             return super().forward_oot(positions, query, key)
 
         import torch_npu
@@ -428,4 +432,4 @@ def forward_oot(
                                          mrope_section=mrope_section,
                                          rotary_mode='half')
 
-        return query, key
+        return query, key