diff --git a/pyproject.toml b/pyproject.toml
index 1cbf03dac..95e95fed8 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -252,3 +252,4 @@ fullset = "fullset"
 Vai = "Vai"
 tockens = "tockens"
 CANN = "CANN"
+ND = "ND"
diff --git a/vllm_omni/platforms/npu/platform.py b/vllm_omni/platforms/npu/platform.py
index 8a30b4c7a..e371c73a6 100644
--- a/vllm_omni/platforms/npu/platform.py
+++ b/vllm_omni/platforms/npu/platform.py
@@ -31,6 +31,14 @@ class NPUOmniPlatform(OmniPlatform, NPUPlatform):
     _omni_enum = OmniPlatformEnum.NPU
     dist_backend: str = "hccl"
 
+    @classmethod
+    def set_device(cls, device: torch.device) -> None:
+        super().set_device(device)
+        # Ascend quantized weights are converted from ND to FRACTAL_NZ
+        # after loading. Enable internal format so the NZ storage layout
+        # is preserved for fused NPU kernels.
+        torch.npu.config.allow_internal_format = True
+
     @classmethod
     def get_omni_ar_worker_cls(cls) -> str:
         return "vllm_omni.platforms.npu.worker.npu_ar_worker.NPUARWorker"