vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-06 14:19:02 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix bugs on cpu

Browse source
This commit is contained in:
Zonghang Li 2024-11-29 22:06:02 +04:00
parent 81fd77033e
commit 7e4bb65eab
1 changed files with 6 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
common/profiler.cpp
View file

@ -556,6 +556,7 @@ time_based=1
runtime=2 runtime=2
size=500M size=500M
group_reporting=1 group_reporting=1
cpus_allowed_policy=shared
[read-job] [read-job]
rw=%s rw=%s
@ -901,6 +902,7 @@ static float device_compute_delay(struct device_info & dev_info, int n_layers, c
// estimate the memory access delay, except for the input embedding because it has been considered in n_flops.inp_embd_ms // estimate the memory access delay, except for the input embedding because it has been considered in n_flops.inp_embd_ms
static float device_memory_access_delay(struct device_info & dev_info, const struct llama_context_params cparams, int n_layers) { static float device_memory_access_delay(struct device_info & dev_info, const struct llama_context_params cparams, int n_layers) {
struct model_params n_params = dev_info.model_params; struct model_params n_params = dev_info.model_params;
int n_gpu_layers = cparams.n_gpu_layers;
int64_t layer_bytes = int64_t layer_bytes =
n_params.layer_f32 * 4 + n_params.layer_f32 * 4 +
@ -917,8 +919,8 @@ static float device_memory_access_delay(struct device_info & dev_info, const str
n_params.output_q80; n_params.output_q80;
#if defined(GGML_USE_CUDA) || defined(GGML_USE_METAL) #if defined(GGML_USE_CUDA) || defined(GGML_USE_METAL)
int64_t vram_bytes = layer_bytes * cparams.n_gpu_layers; int64_t vram_bytes = layer_bytes * n_gpu_layers;
int64_t ram_bytes = layer_bytes * (n_layers - cparams.n_gpu_layers) + output_bytes; int64_t ram_bytes = layer_bytes * (n_layers - n_gpu_layers) + output_bytes;
#ifdef GGML_USE_CUDA #ifdef GGML_USE_CUDA
double vram_access_delay = (double)(vram_bytes) / 1e6 / dev_info.gpu_props.cuda_read_vram_bw; double vram_access_delay = (double)(vram_bytes) / 1e6 / dev_info.gpu_props.cuda_read_vram_bw;
@ -930,7 +932,8 @@ static float device_memory_access_delay(struct device_info & dev_info, const str
return static_cast<float>(vram_access_delay + ram_access_delay); // ms return static_cast<float>(vram_access_delay + ram_access_delay); // ms
#else #else
int64_t ram_bytes = layer_bytes * n_layers; (void)n_gpu_layers;
int64_t ram_bytes = layer_bytes * n_layers + output_bytes;
double ram_access_delay = (double)(ram_bytes) / 1e6 / dev_info.memory.cpu_read_ram_bw; double ram_access_delay = (double)(ram_bytes) / 1e6 / dev_info.memory.cpu_read_ram_bw;
return static_cast<float>(ram_access_delay); // ms return static_cast<float>(ram_access_delay); // ms
#endif #endif