[main] mlp weight prefetch in Qwen Dense Models (#2816)

### What this PR does / why we need it?
This PR prefetchs the weight of mlp layers in Qwen Dense Models to
optimize the performance in Decode phase mainly.

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

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

- vLLM version: main
- vLLM main:
a1213fae5f

Signed-off-by: rjg-lyh <1318825571@qq.com>
Co-authored-by: Shuming19 <313093131@qq.com>

vllm_ascend/torchair/ops/torchair_activation.py

@@ -0,0 +1,37 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+
+import torch
+
+
+def torchair_silu_and_mul_forward_oot(self, x: torch.Tensor) -> torch.Tensor:
+    """AscendSiluAndMul forward in torchair mode.
+
+    The key difference from the original implementation is the removal of operators
+    from the torch.ops.vllm class, as these operators only function in non-torchair
+    modes. Adding them back would cause the graph compilation to fail.
+    """
+
+    import torch_npu
+
+    from vllm_ascend.utils import is_310p
+
+    if is_310p():
+        out = torch_npu.npu_swiglu(x.to(torch.float32)).to(torch.float16)
+    else:
+        out = torch_npu.npu_swiglu(x)
+    return out

vllm_ascend/torchair/ops/torchair_layernorm.py

@@ -0,0 +1,51 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+
+from typing import Optional, Tuple, Union
+
+import torch
+
+
+def torchair_rmsnorm_forward_oot(
+    self,
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    """AscendRMSNorm forward in torchair mode.
+
+    The key difference from the original implementation is the removal of operators
+    from the torch.ops.vllm class, as these operators only function in non-torchair
+    modes. Adding them back would cause the graph compilation to fail.
+    """
+
+    import torch_npu
+
+    from vllm_ascend.utils import is_310p
+    if residual is not None:
+        if is_310p():
+            orig_dtype = residual.dtype
+            x = x + residual.to(x.dtype)
+            residual = x.to(orig_dtype)
+            x, _ = torch_npu.npu_rms_norm(x, self.weight,
+                                          self.variance_epsilon)
+        else:
+            x, _, residual = torch_npu.npu_add_rms_norm(
+                x, residual, self.weight, self.variance_epsilon)
+        return x, residual
+
+    x, residual = torch_npu.npu_rms_norm(x, self.weight, self.variance_epsilon)
+    return x

vllm_ascend/torchair/utils.py

@@ -199,8 +199,12 @@ def torchair_quant_method_register():
 
 
 def torchair_ops_patch():
+    from vllm_ascend.ops.activation import AscendSiluAndMul
+    from vllm_ascend.ops.layernorm import AscendRMSNorm
     from vllm_ascend.ops.rotary_embedding import (
         AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding)
+    from vllm_ascend.torchair.ops import (torchair_activation,
+                                          torchair_layernorm)
     from vllm_ascend.torchair.ops.torchair_rotary_embedding import (
         deepseek_rope_init_func, native_rope_deepseek_forward,
         qwen_rope_init_func, rope_forward)
@@ -210,3 +214,6 @@ def torchair_ops_patch():
 
     AscendDeepseekScalingRotaryEmbedding.__init__ = deepseek_rope_init_func  # type: ignore[method-assign]
     AscendDeepseekScalingRotaryEmbedding.forward = native_rope_deepseek_forward  # type: ignore[method-assign]
+
+    AscendRMSNorm.forward_oot = torchair_layernorm.torchair_rmsnorm_forward_oot  # type: ignore[method-assign]
+    AscendSiluAndMul.forward_oot = torchair_activation.torchair_silu_and_mul_forward_oot  # type: ignore[method-assign]