EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

cherry pick from pr 4270 (#4285)

Browse Source 
### What this PR does / why we need it?
avoid mrope fusion op when running qwen25vl on x86 machine

---------

Signed-off-by: 李少鹏 <lishaopeng21@huawei.com>
This commit is contained in:
shaopeng-666
2025-11-19 22:32:02 +08:00
committed by GitHub
parent 277670730c
commit b6d59bdea2
2 changed files with 14 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
tests/ut/ops/test_rotary_embedding.py
View File

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

6
vllm_ascend/ops/rotary_embedding.py
View File

@@ -24,6 +24,7 @@ from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.rotary_embedding import ( from vllm.model_executor.layers.rotary_embedding import (
DeepseekScalingRotaryEmbedding, MRotaryEmbedding, RotaryEmbedding, DeepseekScalingRotaryEmbedding, MRotaryEmbedding, RotaryEmbedding,
YaRNScalingRotaryEmbedding) YaRNScalingRotaryEmbedding)
from vllm.platforms import CpuArchEnum
from vllm_ascend.platform import NPUPlatform from vllm_ascend.platform import NPUPlatform
from vllm_ascend.utils import enable_custom_op, is_310p from vllm_ascend.utils import enable_custom_op, is_310p
@@ -405,7 +406,10 @@ class AscendMRotaryEmbedding(MRotaryEmbedding):
query: torch.Tensor, query: torch.Tensor,
key: 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) return super().forward_oot(positions, query, key)
import torch_npu import torch_npu