perf: precompute ragged Q tile scheduling
This commit is contained in:
@@ -610,6 +610,8 @@ class CudaBackend(AttentionBackend):
|
||||
kv_cache.req_pool_indices,
|
||||
kv_cache.kv_indptr,
|
||||
kv_cache.qo_indptr,
|
||||
kv_cache.q_tile_to_batch,
|
||||
kv_cache.q_tile_to_index,
|
||||
attn_mask,
|
||||
is_causal=is_causal,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user