[Fix] Merge adjacent KV-row frees so a mid-page split under DCP cannot double-free (#38941)

python/sglang/srt/mem_cache/base_prefix_cache.py

@@ -415,12 +415,14 @@ class BasePrefixCache(ABC, PrefixCacheTrait):
         """Give back ascending, disjoint, half-open row-position ranges
         of the ``kv`` record's row; one call keeps a shared page freed once.
         """
-        from sglang.srt.mem_cache.common import free_kv_row_segments
+        from sglang.srt.mem_cache.common import coalesce_ranges, free_kv_row_segments
 
         row = self.req_to_token_pool.req_to_token[kv.req_pool_idx]
+        # Adjacent pieces whose seam falls inside one (DCP-widened) page would
+        # free that page twice; the allocator rejects that, so merge them first.
         free_kv_row_segments(
             self.token_to_kv_pool_allocator,
-            [(row[start:end], start) for start, end in ranges],
+            [(row[start:end], start) for start, end in coalesce_ranges(ranges)],
             swa_evicted_seqlen=kv.swa_evicted_seqlen,
         )
 

python/sglang/srt/mem_cache/common.py

@@ -108,6 +108,17 @@ def free_swa_out_of_window_slots(
         req.kv.swa_evicted_seqlen = new_swa_evicted_seqlen
 
 
+def coalesce_ranges(ranges: list[tuple[int, int]]) -> list[tuple[int, int]]:
+    """Merge adjacent half-open ranges so a split that falls mid-page frees that page once."""
+    merged: list[tuple[int, int]] = []
+    for start, end in ranges:
+        if merged and start == merged[-1][1]:
+            merged[-1] = (merged[-1][0], end)
+        else:
+            merged.append((start, end))
+    return merged
+
+
 def free_kv_row_segments(
     allocator: BaseTokenToKVPoolAllocator,
     segments: list[tuple[torch.Tensor, int]],