From 793c98afaf5e81cf14a9810fcd25db15b28bd20a Mon Sep 17 00:00:00 2001
From: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>
Date: Wed, 10 Dec 2025 14:56:16 +0800
Subject: [PATCH] handling incomplete rope_scaling config ci after transformers
 upgrade (#14784)

---
 python/sglang/srt/models/deepseek_v2.py | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 2b1c3c04d..94cdae898 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -2704,6 +2704,12 @@ class DeepseekV2DecoderLayer(nn.Module):
         self.config = config
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None:
+            # In transformers 5.0.0rc0+, rope_theta and rope_type are also included in rope_scaling.
+            # Therefore, if rope_scaling contains only these two keys,
+            # it effectively means there are no special rope_scaling parameters.
+            if set(rope_scaling.keys()) <= {"rope_theta", "rope_type"}:
+                rope_scaling = None
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
             get_global_server_args().speculative_algorithm