[Bugfix][Mamba] Clear deferred init metadata before speculative decode (#37165)
This commit is contained in:
@@ -3345,6 +3345,10 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
|
||||
# prefill-time tensor so it doesn't leak into ForwardBatch.
|
||||
self.input_embeds = None
|
||||
|
||||
self.mamba_cow_src_indices = None
|
||||
self.mamba_cow_dst_indices = None
|
||||
self.mamba_clear_indices = None
|
||||
|
||||
# Clear context parallel metadata - CP is only for prefill, not decode
|
||||
if hasattr(self, "attn_cp_metadata") and self.attn_cp_metadata is not None:
|
||||
self.attn_cp_metadata = None
|
||||
|
||||
Reference in New Issue
Block a user