[main] Optimize rope in Qwen Models (#2571)

### What this PR does / why we need it?
Optimize rope by caching sin and cos at the first layer in Qwen Models.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
CI passed with new added/existing test.


- vLLM version: v0.10.1.1
- vLLM main:
562663a044

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
Signed-off-by: ZYang6263 <zy626375@gmail.com>
Signed-off-by: rjg-lyh <1318825571@qq.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
Co-authored-by: ZYang6263 <51255902183@stu.ecnu.edu.cn>
Co-authored-by: ZYang6263 <zy626375@gmail.com>

vllm_ascend/models/qwen3_moe.py

@@ -20,6 +20,7 @@
 from typing import Optional, Union
 
 import torch
+import torch_npu
 from torch import nn
 from transformers import PretrainedConfig
 from vllm.compilation.decorators import support_torch_compile
@@ -280,6 +281,11 @@ class CustomQwen3MoeModel(Qwen3MoeModel):
         self.make_empty_intermediate_tensors = (
             make_empty_intermediate_tensors_factory(
                 ["hidden_states", "residual"], config.hidden_size))
+        # Call ATB matmul to warm up; otherwise, the first operation (ReshapeAndCache) may cause performance degradation at runtime.
+        x = torch.rand((2, 4), dtype=torch.float16).npu()
+        weight = torch.rand((2, 4), dtype=torch.float16).npu()
+        c = torch.rand((4, 4), dtype=torch.float32).npu()
+        torch_npu._npu_matmul_add_fp32(x, weight, c)
 
     def forward(
         self,