nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix race in CUDA FA for head sizes 192/128 (#1104) 

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow 2026-01-05 08:21:07 +02:00 committed by GitHub
parent ab50c6cdcb
commit 385fc14110
1 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-cuda/fattn-new-mma.cu
View File

@ -285,17 +285,17 @@ struct fattn_mma_f16_config;
template <>
struct fattn_mma_f16_config<192, 128> {
static constexpr int nbatch_fa = 64;
static constexpr int nbatch_fa = 32;
static constexpr int nwarps_max = 4;
static constexpr bool Q_in_reg = true;
static constexpr int nstages_target = 1;
static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
return 64;
return 96;
}
static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
return 64;
return 96;
}
static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {
@ -317,17 +317,17 @@ struct fattn_mma_f16_config<192, 128> {
template <>
struct fattn_mma_f16_config<192, 192> {
static constexpr int nbatch_fa = 64;
static constexpr int nbatch_fa = 32;
static constexpr int nwarps_max = 4;
static constexpr bool Q_in_reg = true;
static constexpr int nstages_target = 1;
static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
return 64;
return 96;
}
static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
return 64;
return 96;
}
static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {