[FEATURE] 接入 LightOP 的 silu_and_mul 自定义算子并统一 OPT 路径

在 vllm/model_executor/layers/activation.py 中调整 SiluAndMul.forward_cuda：当 VLLM_USE_OPT_OP=1 时统一走 ops.silu_and_mul_opt_lightop(x)
在 vllm/_custom_ops.py 中新增并注册 silu_and_mul_opt_lightop（含 fake_impl），用于编译/非编译路径统一调用

vllm/_custom_ops.py

@@ -383,6 +383,24 @@ def fused_add_rms_norm_opt_fake(
 ) -> None:
     return None
 
+
+def silu_and_mul_opt_lightop(input: torch.Tensor) -> torch.Tensor:
+    return torch.ops.vllm.silu_and_mul_opt_lightop(input)
+
+
+def silu_and_mul_opt_lightop_impl(input: torch.Tensor) -> torch.Tensor:
+    d = input.shape[-1] // 2
+    output_shape = input.shape[:-1] + (d,)
+    out = torch.empty(output_shape, dtype=input.dtype, device=input.device)
+    op.silu_and_mul_opt(out, input)
+    return out
+
+
+def silu_and_mul_opt_lightop_fake(input: torch.Tensor) -> torch.Tensor:
+    d = input.shape[-1] // 2
+    output_shape = input.shape[:-1] + (d,)
+    return input.new_empty(output_shape)
+
 def fused_qk_norm_rope(
     qkv: torch.Tensor,
     num_heads_q: int,
@@ -3631,6 +3649,13 @@ direct_register_custom_op(
     fake_impl=fused_add_rms_norm_opt_fake,
 )
 
+direct_register_custom_op(
+    op_name="silu_and_mul_opt_lightop",
+    op_func=silu_and_mul_opt_lightop_impl,
+    mutates_args=[],
+    fake_impl=silu_and_mul_opt_lightop_fake,
+)
+
 """
 qwen3-vl-8b中LLM的修改 rms+mrope dim==1  2026/03/18 
 """
@@ -3749,4 +3774,4 @@ direct_register_custom_op(
     mutates_args=["query","key"],
     fake_impl=rms_mrope_fuse_fake,
 
-)
\ No newline at end of file
+)

vllm/model_executor/layers/activation.py

@@ -150,13 +150,15 @@ class SiluAndMul(CustomOp):
         return F.silu(x[..., :d]) * x[..., d:]
 
     def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
-        d = x.shape[-1] // 2
-        output_shape = x.shape[:-1] + (d,)
-        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
         if envs.VLLM_USE_OPT_OP:
-            self.op_opt(out, x)
+            from vllm import _custom_ops as ops
+
+            return ops.silu_and_mul_opt_lightop(x)
         else:
-            self.op(out, x) 
+            d = x.shape[-1] // 2
+            output_shape = x.shape[:-1] + (d,)
+            out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+            self.op(out, x)
         return out
 
     def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: