[Mamba1] - Kernel Level Chunk Alignment for Prefix Caching (#34798)
Signed-off-by: Josephasafg <ajgard7@gmail.com>
This commit is contained in:
@@ -497,6 +497,8 @@ def selective_scan_fn(
|
||||
block_idx_first_scheduled_token=None,
|
||||
block_idx_last_scheduled_token=None,
|
||||
initial_state_idx=None,
|
||||
cu_chunk_seqlen=None,
|
||||
last_chunk_indices=None,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
u: (dim, total_length) for varlen or (batch, dim, seqlen)
|
||||
@@ -588,6 +590,8 @@ def selective_scan_fn(
|
||||
block_idx_first_scheduled_token,
|
||||
block_idx_last_scheduled_token,
|
||||
initial_state_idx,
|
||||
cu_chunk_seqlen,
|
||||
last_chunk_indices,
|
||||
)
|
||||
|
||||
if z is None:
|
||||
|
||||
Reference in New Issue
Block a user