[Bugfix] BailingMoeV3 KDA: skip absent metadata during CUDA graph profiling (#53593)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Aug 24, 2026 at 18:02 UTC
09fddeb4ce7ee8cd0c46f1da9c45aafb9a17e89a
1 file changed
+4
-1
vllm/model_executor/models/bailing_moe_v3.py
+4
-1
@@ -813,7 +813,10 @@ class BailingMoeV3KimiDeltaAttention(PluggableLayer, MambaBase):
813
return
814
815
assert isinstance(attn_metadata_map, dict)
816
- attn_metadata = attn_metadata_map[self.prefix]
816
+ attn_metadata = attn_metadata_map.get(self.prefix)
817
+ if attn_metadata is None:
818
+ # Profile/warmup dummy runs skip mamba-family metadata.
819
+ return
820
assert isinstance(attn_metadata, GDNAttentionMetadata)
821
has_initial_state = attn_metadata.has_initial_state
822
spec_query_start_loc = attn_metadata.spec_query_start_loc