qwen3_moe/qwen25 support torchair graph (#2403)

### What this PR does / why we need it?
Added support for the TorchAir graph mode in qwen3_moe and qwen2.5
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
```bash
llm = LLM(
    model=model,
    tensor_parallel_size=GPUs_per_dp_rank,
    enforce_eager=False,
    enable_expert_parallel=True,
    max_model_len=4096,
    max_num_seqs=16,
    trust_remote_code=trust_remote_code,
    gpu_memory_utilization=0.4,
    additional_config={
             "torchair_graph_config": {
                 "enabled": True,
                 "use_cached_graph": False,
                 "graph_batch_sizes_init": False,
                 "graph_batch_sizes": [16]
             },
             "ascend_scheduler_config": {
                 "enabled": True,
                 "chunked_prefill_enabled":True,
             },
             "refresh": True,
    },
)
```

- vLLM version: v0.10.0
- vLLM main:
b87cb97a53

Signed-off-by: taoyuxiang <oui.nicholas.tao@gmail.com>

tests/ut/models/test_qwen3_moe.py

@@ -12,11 +12,15 @@
 # limitations under the License.
 # This file is a part of the vllm-ascend project.
 #
+import math
+import unittest
 
 import pytest
+import torch
 from vllm.model_executor.models.qwen3_moe import Qwen3MoeForCausalLM
 
 from vllm_ascend.models.qwen3_moe import CustomQwen3MoeForCausalLM
+from vllm_ascend.torchair.models.qwen3_moe import CustomQwen3MoeAttention
 
 
 class TestCustomQwen3MoeForCausalLM:
@@ -44,3 +48,51 @@ class TestCustomQwen3MoeForCausalLM:
             ]
         }
         assert CustomQwen3MoeForCausalLM.packed_modules_mapping == expected_mapping
+
+
+class DummyRMSNorm:
+
+    def __init__(self, dim: int, eps: float = 1e-6):
+        self.dim = dim
+        self.eps = eps
+
+    def __call__(self, x):
+        mean_sq = x.pow(2).mean(dim=-1, keepdim=True)
+        denom = (mean_sq + self.eps).sqrt()
+        return x / denom
+
+
+class TestCustomQwen3MoeAttention(unittest.TestCase):
+
+    def setUp(self):
+        self.batch = 2
+        self.seq_len = 3
+        self.q_size = 8
+        self.kv_size = 8
+        self.head_dim = 4
+        self.rms_eps = 1e-6
+
+        total_dim = self.q_size + 2 * self.kv_size
+
+        self.qkv = torch.arange(self.batch * self.seq_len * total_dim,
+                                dtype=torch.float32).reshape(
+                                    self.batch, self.seq_len, total_dim)
+
+    def test_constant_input_normalization(self):
+        ones_qkv = torch.ones((1, 1, self.q_size + 2 * self.kv_size),
+                              dtype=torch.float32)
+
+        q_norm = DummyRMSNorm(self.head_dim, self.rms_eps)
+        k_norm = DummyRMSNorm(self.head_dim, self.rms_eps)
+        q, k, v = CustomQwen3MoeAttention.normalize_qkv(
+            ones_qkv, self.q_size, self.kv_size, self.head_dim, q_norm, k_norm)
+
+        norm_val = 1.0 / math.sqrt(1.0 + self.rms_eps)
+
+        expected_q = torch.full((1, 1, self.q_size), norm_val)
+        expected_k = torch.full((1, 1, self.kv_size), norm_val)
+        expected_v = torch.ones((1, 1, self.kv_size), dtype=torch.float32)
+
+        self.assertTrue(torch.allclose(q, expected_q, atol=1e-6))
+        self.assertTrue(torch.allclose(k, expected_k, atol=1e-6))
+        self.assertTrue(torch.equal(v, expected_v))