diff --git a/python/sglang/srt/sampling/penaltylib/repetition_penalty.py b/python/sglang/srt/sampling/penaltylib/repetition_penalty.py
index fd03fb2b5..b9ad94f39 100644
--- a/python/sglang/srt/sampling/penaltylib/repetition_penalty.py
+++ b/python/sglang/srt/sampling/penaltylib/repetition_penalty.py
@@ -1,10 +1,12 @@
 import torch
 
 from sglang.srt.sampling.penaltylib.orchestrator import _BatchedPenalizer
-from sglang.srt.utils import get_compiler_backend
+from sglang.srt.utils import get_compiler_backend, is_npu
+
+_is_npu = is_npu()
 
 
-@torch.compile(dynamic=True, backend=get_compiler_backend())
+@torch.compile(dynamic=True, backend=get_compiler_backend(), disable=_is_npu)
 def apply_scaling_penalties(logits, scaling_penalties):
     logits[:] = torch.where(
         logits < 0,