From 09fddeb4ce7ee8cd0c46f1da9c45aafb9a17e89a Mon Sep 17 00:00:00 2001 From: Nick Hill Date: Mon, 24 Aug 2026 18:02:50 -0700 Subject: [PATCH] [Bugfix] BailingMoeV3 KDA: skip absent metadata during CUDA graph profiling (#53593) Signed-off-by: Nick Hill --- vllm/model_executor/models/bailing_moe_v3.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/bailing_moe_v3.py b/vllm/model_executor/models/bailing_moe_v3.py index 6ddb277778..0c15f073a1 100644 --- a/vllm/model_executor/models/bailing_moe_v3.py +++ b/vllm/model_executor/models/bailing_moe_v3.py @@ -813,7 +813,10 @@ class BailingMoeV3KimiDeltaAttention(PluggableLayer, MambaBase): return assert isinstance(attn_metadata_map, dict) - attn_metadata = attn_metadata_map[self.prefix] + attn_metadata = attn_metadata_map.get(self.prefix) + if attn_metadata is None: + # Profile/warmup dummy runs skip mamba-family metadata. + return assert isinstance(attn_metadata, GDNAttentionMetadata) has_initial_state = attn_metadata.has_initial_state spec_query_start_loc = attn_metadata.spec_query_start_loc