reverted tokenizer fix, will be solved manually upon testing, current evals 5213c6c Yeyito commited on Dec 20, 2023
More aggressive OOM prevention, num_z reduced to 50 for gsm8k and mmlu 98282cd Yeyito commited on Dec 19, 2023
Avoiding re-loading already loaded models. Stated unload functionality as not-implemented. b28ad14 Yeyito commited on Dec 19, 2023