diff --git a/vllm/model_executor/kernels/linear/scaled_mm/cpu.py b/vllm/model_executor/kernels/linear/scaled_mm/cpu.py
index 4ef1af0a70..92ea073a1a 100644
--- a/vllm/model_executor/kernels/linear/scaled_mm/cpu.py
+++ b/vllm/model_executor/kernels/linear/scaled_mm/cpu.py
@@ -261,6 +261,8 @@ class CPUFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):
         return True, None
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        if getattr(layer, "_cpu_skip_gemm_dispatch", False):
+            return
         # Skip the base class process (FP8 padding / fnuz normalization)
         # which is GPU-oriented.  Instead, VNNI-prepack weights for AMX.
         params = self._get_layer_params(layer)
diff --git a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
index 44a5bda21e..f7d82c691b 100644
--- a/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
@@ -468,6 +468,7 @@ class CPUExpertsFp8(mk.FusedMoEExpertsModular):
             RoutingMethodType.Renormalize,
             RoutingMethodType.RenormalizeNaive,
             RoutingMethodType.DeepseekV4,
+            RoutingMethodType.DeepSeekV3,
         ]
 
     @staticmethod