file

[waterpeople]

Thank you for your response! Can you provide the max_min_energy_stats.pkl file? How can I generate files similar to max_min_energy_stats.pkl,arch_energy_gpu_dataset.pkl and device_embeddings.pkl?Is it by using HW-GPT-Bench? I am very interested in this work.Thank you!

[rheasukthanker]

1. Missing Files: The max_min_energy_stats.pkl is uploaded here
2. How can I generate files?
   • arch_energy_gpu_dataset.pkl is a processed version of the HW-GPT-Bench dataset collected here
   • device_embeddings.pkl is created in this line , which you can uncomment if needed. Device embeddings are used by the energy meta-predictor (check overview in Figure 1 https://arxiv.org/pdf/2402.18213v2) , to condition the hardware predictor on the gpu device type.
   • max_min_energy_stats.pkl These are the max-min energy prediction statistics from the meta-predictor which is a component of MODNAS pipeline (check overview in Figure 1 https://arxiv.org/pdf/2402.18213v2). The file to compute this is here.

Let me know if you have more questions!

[waterpeople]

Hello，when I run train_llm_configurable.py, and it indicates "end training!", I will obtain logfile.txt and config.yml in the directory MODNAS/experiments/owt_small_modnas_25k/owt_small_modnas_25k/default_juls. I use the OpenWebText and TinyStories datasets for validation. logfile_openwebtext.txt logfile_tinystories.txt config_openwebtext.json config_tinystories.json Since I am using an RTX 3090, I changed max_sample_len to 512 and max_steps to 50000. Why do the values of ppl and acc in the logfile remain almost unchanged? What do ppl and acc represent, respectively? Thank for your response!