Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-10-29 22:20:15 +01:00
Code Issues Packages Projects Releases Wiki Activity

metal : utilize max shared memory for mul_mat_id

Browse Source
This commit is contained in:
Georgi Gerganov 2024-06-14 13:02:25 +03:00
parent 172c825684
commit eaf34ba0cd
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
ggml-metal.m
View File

@ -1862,9 +1862,10 @@ static enum ggml_status ggml_metal_graph_compute(
// ne21 = n_rows
const int dst_rows = ne20*ne21;
const int dst_rows_min = n_as;
const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4;
// max size of the rowids array in the kernel shared buffer
GGML_ASSERT(dst_rows <= 2048);
GGML_ASSERT(dst_rows <= dst_rows_max);
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel