vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-11 06:14:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

consider active pages compression in macos available memory estimation

Browse source
This commit is contained in:
Lizonghang 2025-01-29 20:33:13 +04:00
parent 27c996835d
commit cd758247e6
1 changed files with 13 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

13
common/profiler.cpp
View file

@ -517,6 +517,19 @@ static uint64_t device_host_physical_memory(bool available) {
if (host_statistics64(host, HOST_VM_INFO64, (host_info64_t)&vm_stats, &count) == KERN_SUCCESS) { if (host_statistics64(host, HOST_VM_INFO64, (host_info64_t)&vm_stats, &count) == KERN_SUCCESS) {
size_t page_size = get_page_size(); size_t page_size = get_page_size();
memory = (vm_stats.free_count + vm_stats.inactive_count + vm_stats.purgeable_count) * page_size; memory = (vm_stats.free_count + vm_stats.inactive_count + vm_stats.purgeable_count) * page_size;
// 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)
// because GPU is more likely to use the inactive memory
memory += vm_stats.active_count * 0.2 * page_size;
} else {
// assume 50% of active pages can be compressed on macOS x86_64 (an empirical value)
memory += vm_stats.active_count * 0.5 * page_size;
}
if (!is_uma_arch()) memory += vm_stats.speculative_count * page_size; if (!is_uma_arch()) memory += vm_stats.speculative_count * page_size;
} else { } else {
LOG_INF("host_statistics64 failed\n"); LOG_INF("host_statistics64 failed\n");