vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-06 09:49:02 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

ignore the first 5 evals due to preheat

Browse source
This commit is contained in:
Lizonghang 2025-01-31 08:53:51 +04:00
parent fdecd4b54c
commit dd632ee6df
2 changed files with 9 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
common/profiler.cpp
View file

@ -525,12 +525,14 @@ static uint64_t device_host_physical_memory(bool available) {
// because GPU is more likely to use the inactive memory
memory += vm_stats.active_count * 0.2 * page_size;
} else {
// assume 50% of active pages can be compressed on macOS x86_64 (an empirical value)
memory += vm_stats.active_count * 0.6 * page_size;
// assume 50% of active pages can be compressed on macOS NUMA (an empirical value)
memory += vm_stats.active_count * 0.5 * page_size;
}
if (!is_uma_arch()) memory += vm_stats.speculative_count * page_size;
if (!is_uma_arch()) {
memory += vm_stats.speculative_count * page_size;
memory += vm_stats.compressor_page_count * page_size;
}
} else {
LOG_INF("host_statistics64 failed\n");
}

6
src/llama.cpp
View file

@ -3344,7 +3344,7 @@ struct llama_context {
mutable int64_t n_queued_tokens = 0;
mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
mutable int32_t n_eval = -1; // number of eval calls, set to -1 to ignore the first eval
mutable int32_t n_eval = -5; // number of eval calls, set to -5 to ignore the first 5 evals
// host buffer for the model output (logits and embeddings)
ggml_backend_buffer_t buf_output = nullptr;
@ -22716,7 +22716,7 @@ void llama_synchronize(struct llama_context * ctx) {
// add the evaluation to the stats
if (ctx->n_queued_tokens == 1) {
if (!ctx->cparams.no_perf && ctx->n_eval >= 0) { // ignore the first two evals due to preheat
if (!ctx->cparams.no_perf && ctx->n_eval >= 0) { // ignore the first 5 evals due to preheat
ctx->t_eval_us += ggml_time_us() - ctx->t_compute_start_us;
}
ctx->n_eval++;
@ -23364,7 +23364,7 @@ void llama_perf_context_print(const struct llama_context * ctx) {
void llama_perf_context_reset(struct llama_context * ctx) {
ctx->t_start_us = ggml_time_us();
ctx->t_eval_us = 0;
ctx->n_eval = -1; // set to -1 to ignore the first eval due to preheat
ctx->n_eval = -5; // set to -5 to ignore the first 5 evals due to preheat
ctx->t_p_eval_us = ctx->n_p_eval = 0;
}