[Bugfix] BailingMoeV3 KDA: skip absent metadata during CUDA graph profiling (#53593)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -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
|
||||
|
||||
Reference in New Issue
Block a user