root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-10 18:51:45 +00:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

Moving Block_release to the deallocation code

Browse Source
This commit is contained in:
Paul Tsochantaris 2024-10-07 12:09:16 +01:00
parent 7403c05c06
commit 5e6358398c
1 changed files with 2 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
ggml/src/ggml-metal.m
View File

@ -239,8 +239,6 @@ struct ggml_backend_metal_context {
struct ggml_cgraph * gf; struct ggml_cgraph * gf;
// the callback given to the thread pool // the callback given to the thread pool
// TODO: ideally, this should be created once, utilizing the command buffer state above
// for some reason, doing it like this leads to a crash
void (^encode_async)(size_t ith); void (^encode_async)(size_t ith);
// n_cb command buffers + 1 used by the main thread // n_cb command buffers + 1 used by the main thread
@ -438,7 +436,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(void) {
ctx->capture_scope = nil; ctx->capture_scope = nil;
ctx->gf = nil; ctx->gf = nil;
Block_release(ctx->encode_async);
ctx->encode_async = nil; ctx->encode_async = nil;
for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) { for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
ctx->command_buffers[i] = nil; ctx->command_buffers[i] = nil;
@ -684,6 +681,8 @@ static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
[ctx->kernels[i].pipeline release]; [ctx->kernels[i].pipeline release];
} }
Block_release(ctx->encode_async);
[ctx->queue release]; [ctx->queue release];
[ctx->device release]; [ctx->device release];