fix: use DeepEPDispatcher on CUDA (sgl-project#5180)

Author: zhyncs

python/sglang/srt/models/deepseek_v2.py

@@ -55,7 +55,6 @@
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, EPMoE
-from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.moe.topk import select_experts
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -88,6 +87,8 @@
 
 if _is_cuda:
     from sgl_kernel import awq_dequantize, bmm_fp8
+
+    from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
 else:
     from vllm import _custom_ops as ops