refactor linear (#2867)

### What this PR does / why we need it?
The current linear.py has the following issues:

- There is redundant conditional logic in the `comm_group` and `forward`
selection for classes such as `AscendMergedColumnParallelLinear`.

- Inconsistent comm_group selection logic exists among
`AscendMergedColumnParallelLinear`, `AscendColumnParallelLinear`, and
`AscendQKVParallelLinear`.

To address these two issues, this PR encapsulates `comm_group` and
`forward` into classes and extracts the classes selection logic into
common functions. For future additions of custom communication groups or
forward methods, it will only be necessary to extend
`CustomColumnParallelOp` or `CustomRowParallelOp` and add new selection
logic.

### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?


- vLLM version: v0.10.2
- vLLM main:
vllm-project/vllm@dd39baf

---------

Signed-off-by: realliujiaxu <realliujiaxu@163.com>
Co-authored-by: weijinqian0 <weijinqian@huawei.com>

Author: realliujiaxu

tests/ut/models/test_qwen2_5_vl.py

@@ -295,7 +295,7 @@ def init_vision_transformer(
         mock_group.rank_in_group = 0
         mock_group.world_size = 2
         mocker.patch(
-            "vllm_ascend.ops.linear.get_tp_group",
+            "vllm_ascend.ops.linear_op.get_tp_group",
             return_value=mock_group,
         )
 

tests/ut/ops/test_linear.py

@@ -7,8 +7,7 @@
 
 from vllm_ascend import ascend_config
 from vllm_ascend.distributed import parallel_state
-from vllm_ascend.ops.linear import (AscendColumnParallelLinear,
-                                    AscendMergedColumnParallelLinear,
+from vllm_ascend.ops.linear import (AscendMergedColumnParallelLinear,
                                     AscendRowParallelLinear)
 
 
@@ -32,7 +31,7 @@ def setUp(self):
                   return_value=self.mock_group),
             patch("vllm_ascend.distributed.parallel_state.get_mlp_tp_group",
                   return_value=self.mock_group),
-            patch("vllm_ascend.ops.linear.get_tp_group",
+            patch("vllm_ascend.ops.linear_op.get_tp_group",
                   return_value=self.mock_group),
             patch("vllm_ascend.utils.mlp_tp_enable", return_value=True),
             patch("vllm_ascend.utils.oproj_tp_enable", return_value=True)
@@ -56,8 +55,7 @@ def test_mlp_optimize(self):
             output_size=8,
             prefix="down_proj",
         )
-        self.assertEqual(linear.comm_group, parallel_state._MLP_TP)
-        self.assertEqual(linear.forward_type, "mlp_tp")
+        self.assertEqual(linear.custom_op.comm_group, parallel_state._MLP_TP)
 
         input_tensor = torch.randn(16, 8)
         linear(input_tensor)
@@ -71,34 +69,23 @@ def test_oproj_tp(self):
             output_size=8,
             prefix="o_proj",
         )
-        self.assertEqual(linear.comm_group, parallel_state._OTP)
-        self.assertEqual(linear.forward_type, "oproj_tp")
+        self.assertEqual(linear.custom_op.comm_group, parallel_state._OTP)
 
         input_tensor = torch.randn(16, 8)
         linear(input_tensor)
 
 
-class TestAscendColumnParallelLinear(BaseLinearTest):
-
-    def test_mlp_tp_init(self):
-        linear = AscendColumnParallelLinear(
-            input_size=16,
-            output_size=8,
-            prefix="down_proj",
-        )
-        self.assertEqual(linear.comm_group, parallel_state._MLP_TP)
-
-
 class TestAscendMergedColumnParallelLinear(BaseLinearTest):
 
     def test_merged_mlp_tp_init(self):
+        os.environ["VLLM_ASCEND_ENABLE_MLP_OPTIMIZE"] = "1"
+
         linear = AscendMergedColumnParallelLinear(
             input_size=16,
             output_sizes=[8, 8],
             prefix="gate_up_proj",
         )
-        self.assertEqual(linear.comm_group, parallel_state._MLP_TP)
-        self.assertEqual(linear.forward_type, "mlp_tp")
+        self.assertEqual(linear.custom_op.comm_group, parallel_state._MLP_TP)
 
 
 if __name__ == '__main__':