added t5 multipack

session/translation/end-to-end/nanot5-base-multipack.sh

@@ -0,0 +1,25 @@
+WANDB_PROJECT="nanot5-base-malaysian-cased-translation-v4-multipack" \
+/root/app/venv/bin/torchrun \
+--nproc_per_node 1 \
+-m run_t5_multipack \
+--model_name_or_path mesolitica/nanot5-base-malaysian-translation-v2 \
+--num_train_epochs 2 \
+--eval_steps 1000000000 \
+--logging_steps 2 \
+--save_steps 200 \
+--save_total_limit 3 \
+--do_train \
+--train_file malaysian-translation-v2-multipack-2048 \
+--output_dir nanot5-base-malaysian-cased-translation-v4-packing \
+--dataloader_num_workers=5 \
+--per_device_train_batch_size=1 \
+--per_device_eval_batch_size=3 \
+--gradient_accumulation_steps=16 \
+--max_source_length 2048 \
+--max_target_length 2048 \
+--learning_rate 2e-5 \
+--gradient_checkpointing false \
+--weight_decay 0.01 \
+--bf16 \
+--ddp_find_unused_parameters true \
+--dataloader_pin_memory false

session/translation/end-to-end/nanot5-small-multipack.sh

@@ -0,0 +1,24 @@
+WANDB_PROJECT="nanot5-small-malaysian-cased-translation-v4-multipack" \
+torchrun \
+--nproc_per_node 1 \
+-m run_t5_multipack \
+--model_name_or_path mesolitica/nanot5-small-malaysian-translation-v2 \
+--num_train_epochs 2 \
+--eval_steps 1000000000 \
+--logging_steps 2 \
+--save_steps 200 \
+--save_total_limit 3 \
+--do_train \
+--train_file /home/husein/mesolitica/t5-sdpa-multipack/packing \
+--output_dir nanot5-small-malaysian-cased-translation-v4-packing \
+--dataloader_num_workers=5 \
+--per_device_train_batch_size=1 \
+--per_device_eval_batch_size=3 \
+--gradient_accumulation_steps=16 \
+--max_source_length 2048 \
+--max_target_length 2048 \
+--learning_rate 2e-5 \
+--gradient_checkpointing false \
+--weight_decay 0.01 \
+--bf16 \
+--ddp_find_unused_parameters true