update the project to work with torch==2.10

[jsvir]

Hi!
Thank you for your great project!

I was struggling with running it in my updated environment, so I share it with others - it works now with torch==2.10.
I didn't fully tested it but only with the next command:

mkdir -p output/sst2

python run_minus_training.py --output_dir output/sst2 --task_name sst2 --model_name_or_path roberta-base --do_train --do_eval --save_strategy no --eval_strategy steps --logging_strategy steps --logging_steps 100 --log_level info --log_level_replica info --eval_steps 210500 --max_seq_length 512 --num_train_epochs 50 --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --lr_scheduler_type linear --warmup_ratio 0.06 --learning_rate 2e-4 --weight_decay 0 --seed 128 --apply_lora --lora_alpha 4 --lora_r 2 --report_to none --pruning_batches 64 --pruning_batch_size 4 --mac_constraint 0.4 --pruning_scheduler cubic_gradual --sparsity_warmup_epochs 1 --param_allocation_strategy running_fisher --teacher_param_tuning_config q:0-11,v:0-11,i:0-11 --student_param_tuning_config q:0-11,v:0-11,i:0-11 --head_scorer_type gradient_l1 --intermediate_scorer_type gradient_l1 --pruner_type running_fisher --pruning_start -1 --pruning_stop 20 --num_prunings 8 --pruning_scheduler_strategy saliency --collect_salience --salience_collecting_start 200 --salience_collecting_end -1 --pre_pruning_tuning_steps 200 --mask_lr 0.01 --grafting_top_k -1 --param_resizing_strategy tophalf_limited --tuning_expanding_ratio 4.0 --max_lora_r 4 | tee output/sst2/log.txt

Thanks!