vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-04 22:09:04 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix support for Q5_0

Browse source
This commit is contained in:
Lizonghang 2025-02-27 22:25:03 +04:00
parent c8e615d69c
commit 9cbdf01645
3 changed files with 23 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
common/common.cpp
View file

@ -1116,7 +1116,7 @@ static bool assign_layers_to_device(
dev.model_flops.layer_f32_f32 / (dev.cpu_props.flops_f32_f32 * 1e9 + EPS) +
dev.model_flops.layer_f16_f32 / (dev.cpu_props.flops_f16_f32 * 1e9 + EPS) +
dev.model_flops.layer_q4k_f32 / (dev.cpu_props.flops_q4k_f32 * 1e9 + EPS) +
dev.model_flops.layer_q5k_f32 / (dev.cpu_props.flops_q50_f32 * 1e9 + EPS) +
dev.model_flops.layer_q50_f32 / (dev.cpu_props.flops_q50_f32 * 1e9 + EPS) +
dev.model_flops.layer_q5k_f32 / (dev.cpu_props.flops_q5k_f32 * 1e9 + EPS) +
dev.model_flops.layer_q6k_f32 / (dev.cpu_props.flops_q6k_f32 * 1e9 + EPS) +
dev.model_flops.layer_q80_f32 / (dev.cpu_props.flops_q80_f32 * 1e9 + EPS)) * 1000; // in ms

20
common/profiler.cpp
View file

@ -1978,6 +1978,10 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_flops.output_q4k_f32);
LOG_INF("\n");
LOG_INF("| Model flops (output Q50xF32) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_flops.output_q50_f32);
LOG_INF("\n");
LOG_INF("| Model flops (output Q5KxF32) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_flops.output_q5k_f32);
LOG_INF("\n");
@ -2002,6 +2006,10 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_flops.layer_q4k_f32);
LOG_INF("\n");
LOG_INF("| Model flops (layer Q50xF32) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_flops.layer_q50_f32);
LOG_INF("\n");
LOG_INF("| Model flops (layer Q5KxF32) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_flops.layer_q5k_f32);
LOG_INF("\n");
@ -2026,6 +2034,10 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.input_q4k);
LOG_INF("\n");
LOG_INF("| Model params (input Q50) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.input_q50);
LOG_INF("\n");
LOG_INF("| Model params (input Q5K) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.input_q5k);
LOG_INF("\n");
@ -2050,6 +2062,10 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.layer_q4k);
LOG_INF("\n");
LOG_INF("| Model params (layer Q50) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.layer_q50);
LOG_INF("\n");
LOG_INF("| Model params (layer Q5K) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.layer_q5k);
LOG_INF("\n");
@ -2074,6 +2090,10 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.output_q4k);
LOG_INF("\n");
LOG_INF("| Model params (output Q50) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.output_q50);
LOG_INF("\n");
LOG_INF("| Model params (output Q5K) ");
LOG_INF("| %-10" PRId64 " ", dev_info_set[0].model_params.output_q5k);
LOG_INF("\n");

2
src/llama.cpp
View file

@ -21476,6 +21476,7 @@ void llama_model_n_flops(
n_flops->layer_f32_f32 = static_cast<int64_t>((double)n_flops->layer_f32_f32 / (double)n_layer);
n_flops->layer_f16_f32 = static_cast<int64_t>((double)n_flops->layer_f16_f32 / (double)n_layer);
n_flops->layer_q4k_f32 = static_cast<int64_t>((double)n_flops->layer_q4k_f32 / (double)n_layer);
n_flops->layer_q50_f32 = static_cast<int64_t>((double)n_flops->layer_q50_f32 / (double)n_layer);
n_flops->layer_q5k_f32 = static_cast<int64_t>((double)n_flops->layer_q5k_f32 / (double)n_layer);
n_flops->layer_q6k_f32 = static_cast<int64_t>((double)n_flops->layer_q6k_f32 / (double)n_layer);
n_flops->layer_q80_f32 = static_cast<int64_t>((double)n_flops->layer_q80_f32 / (double)n_layer);
@ -21483,6 +21484,7 @@ void llama_model_n_flops(
n_params->layer_f32 = static_cast<int64_t>((double)n_params->layer_f32 / (double)n_layer);
n_params->layer_f16 = static_cast<int64_t>((double)n_params->layer_f16 / (double)n_layer);
n_params->layer_q4k = static_cast<int64_t>((double)n_params->layer_q4k / (double)n_layer);
n_params->layer_q50 = static_cast<int64_t>((double)n_params->layer_q50 / (double)n_layer);
n_params->layer_q5k = static_cast<int64_t>((double)n_params->layer_q5k / (double)n_layer);
n_params->layer_q6k = static_cast<int64_t>((double)n_params->layer_q6k / (double)n_layer);
n_params->layer_q80 = static_cast<int64_t>((double)n_params->layer_q80 / (double)n_layer);