[Fix] Fix trtllm fp4 moe kernel not found error (#21343)

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -76,21 +76,11 @@ from sglang.srt.utils.custom_op import register_custom_op
 if is_flashinfer_available():
     from flashinfer import fp4_quantize
 
-# Try to import FP4 TRTLLM function if flashinfer is available
-trtllm_fp4_block_scale_moe = None
-if get_moe_runner_backend().is_flashinfer_trtllm():
-    try:
-        from flashinfer.fused_moe import trtllm_fp4_block_scale_moe
-    except ImportError:
-        trtllm_fp4_block_scale_moe = None
-
 _is_hip = is_hip()
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
-logger = logging.getLogger(__name__)
-
 
 def create_moe_dispatcher(moe_runner_config: MoeRunnerConfig) -> BaseDispatcher:
     a2a_backend = get_moe_a2a_backend()
@@ -1329,6 +1319,8 @@ class FlashInferFP4MoE(FusedMoE):
             hidden_states: Input tensor
             topk_output: TopKOutput object with Bypassed format
         """
+        from flashinfer.fused_moe import trtllm_fp4_block_scale_moe
+
         assert isinstance(self.quant_method, ModelOptNvFp4FusedMoEMethod)
 
         assert (