Port SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN from deepseek_v4_dev (#25419)

python/sglang/srt/environ.py

@@ -596,6 +596,7 @@ class Envs:
     SGLANG_OPT_SWA_RADIX_CACHE_COMPACT = EnvBool(False)
     SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT = EnvBool(False)
     SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW = EnvBool(False)
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN = EnvBool(False)
 
     # DeepGemm Mega MoE
     SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE = EnvBool(False)

python/sglang/srt/managers/schedule_batch.py

@@ -2724,9 +2724,13 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         # preserving cache reuse in multi-turn scenarios. Without this, leaf nodes
         # may become tombstoned, causing SWA memory leak.
         # See also: _insert_helper case 3 in swa_radix_cache.py (defensive counterpart).
+        if envs.SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN.get():
+            evict_threshold = pre_len - sliding_window_size
+        else:
+            evict_threshold = pre_len - sliding_window_size - self.tree_cache.page_size
         new_swa_evicted_seqlen = max(
             req.swa_evicted_seqlen,
-            pre_len - sliding_window_size - self.tree_cache.page_size,
+            evict_threshold,
         )
 
         if self.tree_cache.page_size > 1: