[Bugfix][Core] add seq_id_to_seq_group clearing to avoid memory leak when s… (vllm-project#16472)

Signed-off-by: 开哲 <[email protected]>
Co-authored-by: 开哲 <[email protected]>
Signed-off-by: Agata Dobrzyniewicz <[email protected]>

Author: 2 people

vllm/outputs.py

@@ -179,6 +179,13 @@ def from_seq_group(
                 group.finish_seq(seq_group)
             if assembled_seq_group is None:
                 return None
+
+            # clear finished seq in seq_id_to_seq_group
+            if len(group.to_be_finished) == 0:
+                for sub_request_id in list(group.seq_id_to_index.keys()):
+                    if sub_request_id in seq_id_to_seq_group:
+                        del seq_id_to_seq_group[sub_request_id]
+
             return cls.from_seq_group(assembled_seq_group, use_cache,
                                       seq_id_to_seq_group)