ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Disable ring reduction for now

Browse Source
This commit is contained in:
Iwan Kawrakow
2026-01-06 13:14:30 +00:00
parent d9236392cf
commit 289aadb9d4
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-backend.cpp
View File

@@ -1111,7 +1111,7 @@ static bool ggml_is_view_op(enum ggml_op op) {
#endif #endif
#ifndef GGML_SCHED_MAX_COPIES #ifndef GGML_SCHED_MAX_COPIES
#define GGML_SCHED_MAX_COPIES 4 #define GGML_SCHED_MAX_COPIES 1
#endif #endif
struct ggml_backend_sched_split { struct ggml_backend_sched_split {

2
ggml/src/ggml-cuda/reduce.cu
View File

@@ -196,7 +196,7 @@ void ggml_cuda_op_reduce([[maybe_unused]] ggml_backend_cuda_context & ctx, ggml_
// i = 0, peer = 1, ichunk = 1 -> copy part 1 from device 1, device 0 now has parts 0, 1, 2, 3 // i = 0, peer = 1, ichunk = 1 -> copy part 1 from device 1, device 0 now has parts 0, 1, 2, 3
// etc. // etc.
// //
if (dst->ne[1] >= 32) { if (false && dst->ne[1] >= 32) {
auto nelem = ggml_nelements(dst); auto nelem = ggml_nelements(dst);
auto elem_size = ggml_element_size(dst); auto elem_size = ggml_element_size(dst);
auto nelem_per_device = (nelem + nhave - 1)/nhave; auto nelem_per_device = (nelem + nhave - 1)/nhave;