diff --git a/python/sglang/srt/mem_cache/kv_cache_configurator.py b/python/sglang/srt/mem_cache/kv_cache_configurator.py
index 33b58fe59..fffed010b 100644
--- a/python/sglang/srt/mem_cache/kv_cache_configurator.py
+++ b/python/sglang/srt/mem_cache/kv_cache_configurator.py
@@ -1246,7 +1246,7 @@ class KVCacheConfigurator:
         elif self.use_mla_backend and is_dsa_model and not self.mambaish_config:
             token_to_kv_pool = self._build_dsa_kv_pool(
                 max_total_num_tokens=sizes.max_total_num_tokens,
-                max_running_requests=sizes.max_running_requests,
+                max_running_requests=req_to_token_pool.req_to_token.shape[0],
             )
         elif self.use_mla_backend and not self.mambaish_config:
             assert not is_dsa_model