[code style] Clean dead triton kernel code in fused_moe and useless vllm_ops import (#8310)

python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -28,15 +28,6 @@ if TYPE_CHECKING:
         CompressedTensorsConfig,
     )
 
-_is_cuda = is_cuda()
-_is_npu = is_npu()
-_is_cpu_amx_available = cpu_has_amx_support()
-_is_cpu = is_cpu()
-_is_hip = is_hip()
-
-if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available) or _is_hip):
-    from vllm import _custom_ops as vllm_ops
-    from vllm._custom_ops import scaled_fp8_quant
 
 try:
     import vllm
@@ -568,6 +559,8 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
                 requires_grad=False,
             )
 
+        from vllm import _custom_ops as vllm_ops
+
         marlin_w13_qweight = vllm_ops.gptq_marlin_moe_repack(
             layer.w13_weight_packed,
             layer.w13_g_idx_sort_indices,