Upload train_v0.4_0629_steps.sh

train_v0.4_0629_steps.sh

@@ -0,0 +1,55 @@
+export WANDB_MODE=online
+
+torchrun --nproc_per_node=8 --master_port=20001 /workspace/medvicuna/fastchat/train/train_mem.py \
+    --model_name_or_path eachadea/vicuna-7b-1.1 \
+    --data_path /workspace/medvicuna/medvicuna_v0.4_590k.json \
+    --push_to_hub False \
+    --bf16 True \
+    --output_dir output_medvicuna_v0.4_7b \
+    --num_train_epochs 8 \
+    --per_device_train_batch_size 32 \
+    --per_device_eval_batch_size 16 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "steps" \
+    --eval_steps 1150 \
+    --save_strategy "steps" \
+    --save_steps 1150 \
+    --save_total_limit 100 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.04 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --fsdp "full_shard auto_wrap" \
+    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True &>> output_medvicuna_v0.4_7b.log
+
+sleep 1200s
+
+torchrun --nproc_per_node=8 --master_port=20001 /workspace/medvicuna/fastchat/train/train_mem.py \
+    --model_name_or_path eachadea/vicuna-13b-1.1 \
+    --data_path /workspace/medvicuna/medvicuna_v0.4_590k.json \
+    --push_to_hub False \
+    --bf16 True \
+    --output_dir output_medvicuna_v0.4_13b \
+    --num_train_epochs 8 \
+    --per_device_train_batch_size 32 \
+    --per_device_eval_batch_size 16 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "steps" \
+    --eval_steps 1150 \
+    --save_strategy "steps" \
+    --save_steps 1150 \
+    --save_total_limit 100 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.04 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --fsdp "full_shard auto_wrap" \
+    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True &>> output_medvicuna_v0.4_13b.log