[Mamba1] - Kernel Level Chunk Alignment for Prefix Caching (#34798)
Signed-off-by: Josephasafg <ajgard7@gmail.com>
This commit is contained in:
@@ -2021,6 +2021,8 @@ def selective_scan_fwd(
|
||||
block_idx_first_scheduled_token: torch.Tensor | None = None,
|
||||
block_idx_last_scheduled_token: torch.Tensor | None = None,
|
||||
initial_state_idx: torch.Tensor | None = None,
|
||||
cu_chunk_seqlen: torch.Tensor | None = None,
|
||||
last_chunk_indices: torch.Tensor | None = None,
|
||||
):
|
||||
torch.ops._C.selective_scan_fwd(
|
||||
u,
|
||||
@@ -2041,6 +2043,8 @@ def selective_scan_fwd(
|
||||
block_idx_first_scheduled_token,
|
||||
block_idx_last_scheduled_token,
|
||||
initial_state_idx,
|
||||
cu_chunk_seqlen,
|
||||
last_chunk_indices,
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user