[Bugfix] Free cross attention block table for preempted-for-recompute sequence group. (#10013)

kathyyu-google · web-flow · commit 2f385183f354 · 2025-01-02T10:28:09.000-08:00
Signed-off-by: Kathy Yu &lt;feiyangyu@google.com&gt;
diff --git a/vllm/core/scheduler.py b/vllm/core/scheduler.py
@@ -1579,6 +1579,7 @@ def _preempt_by_recompute(
             seq.status = SequenceStatus.WAITING
             self.free_seq(seq)
             seq.reset_state_for_recompute()
+        self._free_seq_group_cross_attn_blocks(seq_group)
 
     def _preempt_by_swap(
         self,
