electron-rare/prima-cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

use the recommended working set size as metal mem bound, rather than the total physical mem

Browse Source
This commit is contained in:
Lizonghang
2025-01-23 23:04:40 +04:00
parent fb05f80f89
commit 5fcf020cfb
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/common.cpp
+3 -3
View File
@@ -943,7 +943,7 @@ static void assign_device(
if ((is_macos && !dev.gpu_support.metal) || is_linux) {
mem_budget[m] = dev.memory.available_physical;
} else if (is_macos && dev.gpu_support.metal) {
mem_budget[m] = dev.memory.total_physical;
mem_budget[m] = dev.gpu_props.memory_free;
} else if (is_android) {
mem_budget[m] = dev.memory.available_physical + dev.memory.used_can_swap;
} else {
@@ -1146,14 +1146,14 @@ static void assign_device(
if (in_set(m, M1)) {
vec_z[m] = (double)(dev.memory.available_physical * GIGABYTE - b_cio) / (double)(n_layer * b_prime);
} else if (in_set(m, M2)) {
vec_z[m] = (double)(dev.memory.total_physical * GIGABYTE - b_cio - c_gpu[m]) / (double)(n_layer * b_prime);
vec_z[m] = (double)(dev.gpu_props.memory_free * GIGABYTE - b_cio - c_gpu[m]) / (double)(n_layer * b_prime);
} else if (in_set(m, M3)) {
vec_z[m] = (double)(dev.memory.available_physical * GIGABYTE + dev.memory.used_can_swap * GIGABYTE * int(is_android) - b_cio) / (double)(n_layer * b_prime);
} else {
if (is_macos && !dev.gpu_support.metal) {
vec_z[m] = - (double)(dev.memory.available_physical * GIGABYTE - b_cio) / (double)(n_layer * b_prime);
} else if (is_macos && dev.gpu_support.metal) {
vec_z[m] = - (double)(dev.memory.total_physical * GIGABYTE - b_cio - c_gpu[m]) / (double)(n_layer * b_prime);
vec_z[m] = - (double)(dev.gpu_props.memory_free * GIGABYTE - b_cio - c_gpu[m]) / (double)(n_layer * b_prime);
} else {
vec_z[m] = - (double)(dev.memory.available_physical * GIGABYTE + dev.memory.used_can_swap * GIGABYTE * int(is_android) - b_cio) / (double)(n_layer * b_prime);
}