[NPU] Add NPU fallback for fused Triton gating kernels (#28293)

python/sglang/srt/models/qwen2_moe.py

@@ -540,6 +540,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
         use_fused_gate = (
             self.shared_expert_gate is not None
             and not use_intel_amx_backend(self.shared_expert_gate)
+            and not is_npu()
         )
 
         if hidden_states.shape[0] == 0:

python/sglang/srt/models/qwen3_5.py

@@ -983,7 +983,11 @@ class Qwen3_5AttentionDecoderLayer(nn.Module):
         attn_output = self.attn(q, k, v, forward_batch)
 
         if self.attn_output_gate:
-            attn_output = fused_sigmoid_mul(attn_output, gate, inplace=True)
+            if not _is_npu:
+                attn_output = fused_sigmoid_mul(attn_output, gate, inplace=True)
+            else:
+                gate_val = gate.reshape(gate.shape[0], -1) if gate.ndim == 3 else gate
+                attn_output.mul_(torch.sigmoid(gate_val))
 
         output, _ = self.o_proj(attn_output)
         return output