vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-10 13:44:33 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

use highs to solve the allocation program

Browse source
This commit is contained in:
Lizonghang 2025-01-15 10:04:04 +04:00
parent b577c10d25
commit 5d9aadf3d5
6 changed files with 614 additions and 86 deletions
Download patch file Download diff file Expand all files Collapse all files

16
common/profiler.cpp
View file

@ -1312,8 +1312,8 @@ static float device_memory_access_delay(struct device_info & dev_info, struct ll
auto n_bytes = dev_info.model_bytes;
int n_gpu_layers = std::min(static_cast<int>(cparams.n_gpu_layers), n_layers);
uint64_t cpu_kv_size;
uint64_t gpu_kv_size;
int64_t cpu_kv_size;
int64_t gpu_kv_size;
#if defined(GGML_USE_METAL) || defined(GGML_USE_CUDA)
llama_kv_size(&cpu_kv_size, &gpu_kv_size, model, cparams, true);
@ -1428,17 +1428,17 @@ static float device_disk_access_delay(struct device_info & dev_info, struct llam
cpu_total_bytes += n_bytes.nb_output;
uint64_t cpu_kv_size;
uint64_t gpu_kv_size;
uint64_t cpu_compute_buf;
uint64_t gpu_compute_buf;
int64_t cpu_kv_size;
int64_t gpu_kv_size;
int64_t cpu_compute_buf;
int64_t gpu_compute_buf;
#if defined(GGML_USE_METAL) || defined(GGML_USE_CUDA)
llama_kv_size(&cpu_kv_size, &gpu_kv_size, model, cparams, true);
llama_model_compute_buf_size(&cpu_compute_buf, &gpu_compute_buf, model, cparams, true);
llama_model_compute_buf_size(&cpu_compute_buf, &gpu_compute_buf, model, cparams, true, true, n_layers, n_gpu_layers);
#else
llama_kv_size(&cpu_kv_size, &gpu_kv_size, model, cparams, false);
llama_model_compute_buf_size(&cpu_compute_buf, &gpu_compute_buf, model, cparams, false);
llama_model_compute_buf_size(&cpu_compute_buf, &gpu_compute_buf, model, cparams, false, true, n_layers, n_gpu_layers);
#endif
double cpu_kv_size_gib = static_cast<double>(cpu_kv_size) / 1024.0 / 1024.0 / 1024.0; // convert to GiB