On the energy consumption of model inference - replication package

Replication package for the empirical study on the energy consumption of LLM inference during code generation tasks. The package contains four subfolders: raw_data folder contains the LLM-generated results from HumanEval+ and the matching raw energy measurements from the NVIDIA Management Library.result_summary folder contains comma-separated tables with relevant information extracted from the HumanEval+ json reports in the raw_data folder.test_execution_reports folder contains the results of running HumanEval+ test cases on the LLM-generated code. This data is used to evaluate the pass@1 rate of each LLM configuration.aggregated_tables folder contains tables that were created by aggregating the data from the other folders. These tables are used to generate the figures and analysis reported in the paper.