Initial approach results with GPU

These are the results for the first iteration of our proposal, where we expected to see different performance values for different demand in isolated individual requests. The requests were sent one-at-a-time, with a wait period of at most 1 minute between beginning and ending of request, as that is the frequency at which GPU is monitored; we needed the GPU to be monitored while the request was being processed. The prompts that we used can be found in the "experiment_queries_mini.csv" file. We decided to use a language other than English to test how the model performed with different languages, which was randomly chosen to be Spanish. Our conclusions were: The memory usage—both RAM and VRAM—caused by individual requests is negligible when compared to the baseline cost of loading and running the model, regardless of the specific prompt characteristics. As for GPU, the workload of individual requests was measurable, but there was no logical difference between prompts based on their demand; if two same prompts were tested over and over again, the results did not explain their demand differences. The GPU usage would always range between 84\% and 86\%.