vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-14 22:59:40 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix latency estimation in set m1

Browse source
This commit is contained in:
Lizonghang 2025-02-03 07:56:02 +04:00
parent 83b3d01844
commit 64089236eb
2 changed files with 29 additions and 31 deletions
Download patch file Download diff file Expand all files Collapse all files

11
common/profiler.cpp
View file

@ -521,17 +521,20 @@ static uint64_t device_host_physical_memory(bool available) {
// active pages compression has higher priority than releasing the clean mmap-ed pages
// some of the active pages can be compressed to save memory for our mmap-ed model weights
if (is_uma_arch()) {
// assume 20% of active pages can be compressed on macOS UMA (an empirical value)
// assume 30% of active pages can be compressed on macOS UMA (an empirical value)
// because GPU is more likely to use the inactive memory
memory += vm_stats.active_count * 0.2 * page_size;
memory += vm_stats.active_count * 0.3 * page_size;
} else {
// assume 50% of active pages can be compressed on macOS NUMA (an empirical value)
memory += vm_stats.active_count * 0.5 * page_size;
}
if (!is_uma_arch()) {
memory += vm_stats.speculative_count * page_size;
memory += vm_stats.compressor_page_count * page_size;
memory += (vm_stats.speculative_count + vm_stats.compressor_page_count) * page_size;
} else {
// #ifndef GGML_USE_METAL
// memory += vm_stats.speculative_count * page_size;
// #endif
}
} else {
LOG_INF("host_statistics64 failed\n");