diff --git a/python/sglang/srt/layers/quantization/fp8_kernel.py b/python/sglang/srt/layers/quantization/fp8_kernel.py
index 4a3d1093d..814c2650a 100644
--- a/python/sglang/srt/layers/quantization/fp8_kernel.py
+++ b/python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -1854,6 +1854,10 @@ if _is_cuda:
         ):
             return
 
-    @torch.library.register_fake("sgl_kernel::sgl_per_token_quant_fp8")
-    def _(input, output_q, output_s):
-        return
+    # FIXME: for some models, this fake registration will cause NaN outputs.
+    # So we gate the fake registration with an environment variable for them.
+    if not get_bool_env_var("SGLANG_DISABLE_SGL_KERNEL_FAKE_REGISTER"):
+
+        @torch.library.register_fake("sgl_kernel::sgl_per_token_quant_fp8")
+        def _(input, output_q, output_s):
+            return