diff --git a/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py b/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py
index 5f95bf20f..3ccdfd66f 100644
--- a/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py
+++ b/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py
@@ -361,7 +361,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
             symm_output = torch.empty(
                 hidden_states.shape[0],
                 hidden_states.shape[1],
-                dtype=torch.bfloat16,
+                dtype=hidden_states.dtype,
                 device=hidden_states.device,
             )
 
@@ -441,9 +441,11 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
                 tune_max_num_tokens=next_power_of_2(a_q.shape[0]),
                 fp8_quantization_type=int(fp8_quantization_type),
             )
+        # TODO: Once https://github.com/flashinfer-ai/flashinfer/issues/2703 is fixed, pass output to moe kernel and remove this copy.
         symm_output.copy_(output)
         output = symm_output
     else:
+        assert TopKOutputChecker.format_is_bypassed(topk_output)
         assert quant_info.w13_input_scale is not None
         assert quant_info.output1_scales_scalar is not None
         assert quant_info.output1_scales_gate_scalar is not None
diff --git a/python/sglang/srt/layers/quantization/fp8.py b/python/sglang/srt/layers/quantization/fp8.py
index 217b3fce7..9d0d15716 100644
--- a/python/sglang/srt/layers/quantization/fp8.py
+++ b/python/sglang/srt/layers/quantization/fp8.py
@@ -1369,7 +1369,10 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 self.process_weights_hip_scale_padding(layer)
 
             # Align FP8 weights to FlashInfer per-tensor kernel layout if enabled
-            if get_moe_runner_backend().is_flashinfer_trtllm():
+            if (
+                get_moe_runner_backend().is_flashinfer_trtllm()
+                or get_moe_runner_backend().is_flashinfer_trtllm_routed()
+            ):
                 from sglang.srt.layers.moe.moe_runner.flashinfer_trtllm import (
                     align_fp8_moe_weights_for_flashinfer_trtllm,
                 )
@@ -1619,7 +1622,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 local_num_experts=num_local_experts,
                 intermediate_size=layer.w2_weight.shape[2],
                 routing_method_type=int(
-                    getattr(layer, "routing_method_type", RoutingMethodType.DeepSeekV3)
+                    getattr(layer, "routing_method_type", None)
+                    or RoutingMethodType.DeepSeekV3
                 ),
                 block_quant=self.block_quant,
                 use_mxfp8=getattr(self.quant_config, "use_mxfp8", False),
diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py
index 80b0e959b..297e4bbd7 100644
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -2046,6 +2046,8 @@ class ModelRunner(ModelRunnerKVCacheMixin):
 
         if backend_str not in [
             "flashinfer_trtllm",
+            # TODO: Enable for flashinfer_trtllm_routed once https://github.com/flashinfer-ai/flashinfer/issues/2749 is fixed.
+            # "flashinfer_trtllm_routed",
             "flashinfer_mxfp4",
             # TODO: flashinfer_cutlass will cause some flashinfer compilation errors. To be fixed.
             # "flashinfer_cutlass",