Skip to content

Latest commit

 

History

History
43 lines (22 loc) · 1.94 KB

File metadata and controls

43 lines (22 loc) · 1.94 KB

code structure

benchmark_throughput.py:                end-to-end single model 

output_length_sampler.py:               sampling output length

fake_scheduling.py:                     doing fake scheduling to estimate cost


construct_cost_model.py:                run this file to collect data to build cost model

my_per_iter_latency_estimator.py:       about the cost model: linear functions, etc.

model_coeff_database.py:                stores the coefficients for each model to compute the flops
                                        obtained by running comp_model_size.py

comp_model_size.py:                     run this file (1) to get the model parameter sizes and store them in
                                        model_size_database.py
                                        and (2) to get the model coefficients to compute flops and store them in model_coeff_database.py

model_size_database.py:                 stores the model parameter sizes
                                        obtained by running comp_model_size.py

model_initcost_database.py              stores the model init costs
                                        obtained by running comp_model_size.py
                                        NOTE: every time introducing a new model (i.e., the models except those used in LLM-Blender), run comp_model_size.py 
                                            for that model.

search_exec_plans.py:                   generate multi-model schedule plan
                                        


schedule_multi_model.py:                schedule the end-to-end inference

my_llm_infer_worker_multiprocessing.py  inference code for each data parallel worker.
                                        will be called by bench_throughput.py run_vllm.


multimodel_scheduler.py                 the communication betwen multiple model processes.