| 463 | } |
| 464 | |
| 465 | json get_formated_timings() { |
| 466 | return json |
| 467 | { |
| 468 | {"prompt_n", num_prompt_tokens_processed}, |
| 469 | {"prompt_ms", t_prompt_processing}, |
| 470 | {"prompt_per_token_ms", t_prompt_processing / num_prompt_tokens_processed}, |
| 471 | {"prompt_per_second", 1e3 / t_prompt_processing * num_prompt_tokens_processed}, |
| 472 | |
| 473 | {"predicted_n", n_decoded}, |
| 474 | {"predicted_ms", t_token_generation}, |
| 475 | {"predicted_per_token_ms", t_token_generation / n_decoded}, |
| 476 | {"predicted_per_second", 1e3 / t_token_generation * n_decoded}, |
| 477 | }; |
| 478 | } |
| 479 | |
| 480 | void print_timings() { |
| 481 | LOG_TEE("\n"); |