vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-10 22:04:33 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add disk access delay while swapping

Browse source
This commit is contained in:
Lizonghang 2024-12-08 20:07:30 +04:00
parent f5243571d1
commit aad604ec8b
1 changed files with 6 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

11
common/profiler.cpp
View file

@ -291,7 +291,7 @@ float device_inp_embd_delay(struct llama_model * model, enum ggml_type src0t, in
matrix_B = malloc(embd_size * sizeof(float)); matrix_B = malloc(embd_size * sizeof(float));
float * matrix_B_f32 = static_cast<float *>(matrix_B); float * matrix_B_f32 = static_cast<float *>(matrix_B);
for (size_t i = 0; i < embd_size; ++i) { for (size_t i = 0; i < embd_size; ++i) {
matrix_B_f32[i] = static_cast<float>(rand()) / RAND_MAX; matrix_B_f32[i] = static_cast<float>(rand() / RAND_MAX);
} }
break; break;
} }
@ -299,7 +299,7 @@ float device_inp_embd_delay(struct llama_model * model, enum ggml_type src0t, in
matrix_B = malloc(embd_size * sizeof(ggml_fp16_t)); matrix_B = malloc(embd_size * sizeof(ggml_fp16_t));
std::vector<float> temp_f32(embd_size); std::vector<float> temp_f32(embd_size);
for (size_t i = 0; i < embd_size; ++i) { for (size_t i = 0; i < embd_size; ++i) {
temp_f32[i] = static_cast<float>(rand()) / RAND_MAX; temp_f32[i] = static_cast<float>(rand() / RAND_MAX);
} }
ggml_fp32_to_fp16_row(temp_f32.data(), static_cast<ggml_fp16_t *>(matrix_B), embd_size); ggml_fp32_to_fp16_row(temp_f32.data(), static_cast<ggml_fp16_t *>(matrix_B), embd_size);
break; break;
@ -1229,13 +1229,14 @@ static float device_disk_access_delay(struct device_info & dev_info, struct llam
double total_bytes_gib = static_cast<double>(cpu_total_bytes + gpu_total_bytes) / 1024.0 / 1024.0 / 1024.0; double total_bytes_gib = static_cast<double>(cpu_total_bytes + gpu_total_bytes) / 1024.0 / 1024.0 / 1024.0;
double total_kv_size_gib = cpu_kv_size_gib + gpu_kv_size_gib; double total_kv_size_gib = cpu_kv_size_gib + gpu_kv_size_gib;
double total_compute_buf_gib = cpu_compute_buf_gib + gpu_compute_buf_gib; double total_compute_buf_gib = cpu_compute_buf_gib + gpu_compute_buf_gib;
float disk_read_bw = dev_info.disk.read_rnd_bw;
if (total_bytes_gib + total_kv_size_gib + total_compute_buf_gib < dev_info.memory.total_physical) { if (total_bytes_gib + total_kv_size_gib + total_compute_buf_gib < dev_info.memory.total_physical) {
float disk_read_bw = dev_info.disk.read_rnd_bw;
// each time one new row of lookup table will be loaded // each time one new row of lookup table will be loaded
return static_cast<double>(input_bytes) / 1e9 / disk_read_bw * 1000; // convert to ms return static_cast<double>(input_bytes) / 1e9 / disk_read_bw * 1000; // convert to ms
} else { } else {
LOG_INF("swap occurs, not allowed\n"); // warn: OOM error may occur if -ngl is set large
return 1e9; float weight_reload_delay = total_bytes_gib * 1024.0 * 1024.0 * 1024.0 / 1e6 / disk_read_bw; // ms
return weight_reload_delay;
} }
} }
#endif #endif