vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-08 10:09:03 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

assume 10% of active pages can be compressed on macOS UMA

Browse source
This commit is contained in:
Lizonghang 2025-02-11 11:06:33 +04:00
parent 261c88f058
commit 24974a488c
1 changed files with 2 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
common/profiler.cpp
View file

@ -521,9 +521,9 @@ static uint64_t device_host_physical_memory(bool available) {
// active pages compression has higher priority than releasing the clean mmap-ed pages // 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 // some of the active pages can be compressed to save memory for our mmap-ed model weights
if (is_uma_arch()) { if (is_uma_arch()) {
// assume 30% of active pages can be compressed on macOS UMA (an empirical value) // assume 10% of active pages can be compressed on macOS UMA (an empirical value)
// because GPU is more likely to use the inactive memory // because GPU is more likely to use the inactive memory
memory += vm_stats.active_count * 0.3 * page_size; memory += vm_stats.active_count * 0.1 * page_size;
} else { } else {
// assume 50% of active pages can be compressed on macOS NUMA (an empirical value) // assume 50% of active pages can be compressed on macOS NUMA (an empirical value)
memory += vm_stats.active_count * 0.5 * page_size; memory += vm_stats.active_count * 0.5 * page_size;