perf: precompute ragged Q tile scheduling

This commit is contained in:
2026-08-16 23:32:46 +08:00
parent 0d0dc64884
commit 1bcd8f53ab
17 changed files with 283 additions and 141 deletions
+2
View File
@@ -99,6 +99,8 @@ class KVCache:
max_len: int = 0
kv_indptr: Optional[Tensor] = None
qo_indptr: Optional[Tensor] = None
q_tile_to_batch: Optional[Tensor] = None
q_tile_to_index: Optional[Tensor] = None
decode_o_part: Optional[Tensor] = None
decode_ml_part: Optional[Tensor] = None
decode_out: Optional[Tensor] = None