[Mamba] Fix long-prefill accuracy drop in radix prefix-cache state restore (#29368)
This commit is contained in:
@@ -241,11 +241,6 @@ class Mamba2Metadata(ForwardMetadata):
|
||||
context_lens_tensor = forward_batch.extend_prefix_lens
|
||||
assert context_lens_tensor is not None
|
||||
has_initial_states = context_lens_tensor > 0
|
||||
mamba_track_mask = getattr(forward_batch, "mamba_track_mask", None)
|
||||
if mamba_track_mask is not None:
|
||||
has_initial_states = (
|
||||
has_initial_states & mamba_track_mask[: has_initial_states.shape[0]]
|
||||
)
|
||||
prep_initial_states = torch.any(has_initial_states[:num_prefills]).item()
|
||||
|
||||
query_start_loc = forward_metadata.query_start_loc[: num_prefills + 1]
|
||||
|
||||
Reference in New Issue
Block a user