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
+1 -1
View File
@@ -2,7 +2,7 @@
#include <cuda_bf16.h>
#include <float.h>
#include "attn_common.h"
#include "attn_kv_source.cuh"
#include "attn_layout_policies.cuh"
#include "attn_warp_utils.cuh"
constexpr int DC_CHUNK = 64;