From 90eb894564ae1016c00775bc33d81024d42bd64e Mon Sep 17 00:00:00 2001
From: ZeyuanChen2000 <806126546@qq.com>
Date: Sat, 30 May 2026 17:10:58 +0800
Subject: [PATCH] [NPU] fix model llava-onevision-qwen2-7b-ov torch compiles
 error in npu case (#26573)

---
 python/sglang/srt/sampling/penaltylib/repetition_penalty.py | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

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,