[Ascend]optimize Qwen3 on Ascend (#10574)

Co-authored-by: c30031083 <chenxu140@huawei.com>

python/sglang/srt/models/qwen3.py

@@ -30,12 +30,19 @@ from sglang.srt.model_loader.weight_utils import (
 )
 from sglang.srt.models.qwen2 import Qwen2MLP as Qwen3MLP
 from sglang.srt.models.qwen2 import Qwen2Model
-from sglang.srt.utils import add_prefix, is_cuda
+from sglang.srt.utils import (
+    add_prefix,
+    get_cmo_stream,
+    is_cuda,
+    is_npu,
+    wait_cmo_stream,
+)
 
 Qwen3Config = None
 
 logger = logging.getLogger(__name__)
 _is_cuda = is_cuda()
+_is_npu = is_npu()
 
 
 class Qwen3Attention(nn.Module):
@@ -235,9 +242,18 @@ class Qwen3DecoderLayer(nn.Module):
 
         # Fully Connected
         hidden_states, residual = self.layer_communicator.prepare_mlp(
-            hidden_states, residual, forward_batch
+            hidden_states,
+            residual,
+            forward_batch,
+            cache=(
+                [self.mlp.gate_up_proj.weight, self.mlp.down_proj.weight]
+                if _is_npu
+                else None
+            ),
         )
         hidden_states = self.mlp(hidden_states)
+        if _is_npu and get_cmo_stream():
+            wait_cmo_stream()
         hidden_states, residual = self.layer_communicator.postprocess_layer(
             hidden_states, residual, forward_batch
         )