vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-07 22:49:03 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

remove duplicate calls

Browse source
This commit is contained in:
Lizonghang 2024-12-29 22:00:17 +04:00
parent a7ec685eda
commit fa210d2034
1 changed files with 5 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
common/profiler.cpp
View file

@ -1494,17 +1494,17 @@ static float device_disk_access_delay(struct device_info & dev_info, struct llam
} }
} }
static float device_mem_copy_delay(struct llama_model * model, const struct llama_context_params cparams) { static float device_mem_copy_delay(struct device_info & dev_info, struct llama_model * model, const struct llama_context_params cparams) {
int n_layers = llama_model_n_layers(model); int n_layers = llama_model_n_layers(model);
int n_gpu_layers = std::min(static_cast<int>(cparams.n_gpu_layers), n_layers); int n_gpu_layers = std::min(static_cast<int>(cparams.n_gpu_layers), n_layers);
float layer_delay_cpu = device_cpu_mem_copy(model, cparams.n_threads); float layer_delay_cpu = dev_info.memory.mem_cpy_delay;
#ifdef GGML_USE_METAL #ifdef GGML_USE_METAL
float layer_delay_metal = device_metal_mem_copy(model); float layer_delay_metal = dev_info.gpu_props.metal_mem_cpy_delay;
return layer_delay_metal * n_gpu_layers + layer_delay_cpu * (n_layers - n_gpu_layers); return layer_delay_metal * n_gpu_layers + layer_delay_cpu * (n_layers - n_gpu_layers);
#elif GGML_USE_CUDA #elif GGML_USE_CUDA
float layer_delay_cuda = device_cuda_mem_copy(model); float layer_delay_cuda = dev_info.gpu_props.cuda_mem_cpy_delay;
return layer_delay_cuda * n_gpu_layers + layer_delay_cpu * (n_layers - n_gpu_layers); return layer_delay_cuda * n_gpu_layers + layer_delay_cpu * (n_layers - n_gpu_layers);
#else #else
(void)n_gpu_layers; (void)n_gpu_layers;
@ -1947,7 +1947,7 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
latency += device_compute_delay (dev_info_set[0], n_layers, cparams); latency += device_compute_delay (dev_info_set[0], n_layers, cparams);
latency += device_memory_access_delay(dev_info_set[0], model, cparams, n_layers); latency += device_memory_access_delay(dev_info_set[0], model, cparams, n_layers);
latency += device_disk_access_delay (dev_info_set[0], model, cparams); // if physical memory is not enough, some mapped data will be released and reloaded later latency += device_disk_access_delay (dev_info_set[0], model, cparams); // if physical memory is not enough, some mapped data will be released and reloaded later
latency += device_mem_copy_delay (model, cparams); // memory copy delay in kvcache latency += device_mem_copy_delay (dev_info_set[0], model, cparams); // memory copy delay in kvcache
LOG_INF("| Token latency (ms) "); LOG_INF("| Token latency (ms) ");
LOG_INF("| %-10.2f ", latency); LOG_INF("| %-10.2f ", latency);