forked from AnswerDotAI/fsdp_qlora
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_sql.sh
28 lines (27 loc) · 803 Bytes
/
train_sql.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
# run script also show stdout and save to log file
python train.py \
--context_length 256 \
--model_name codellama/CodeLlama-34b-hf \
--train_type qlora \
--batch_size 4 \
--gradient_accumulation_steps 4 \
--dataset sql \
--save_model True \
--output_dir sql_model_qlora \
--apply_gradient_clipping True \
--project_name fsdp_qlora_sql \
--precision bf16_buffers_autocast \
--log_to wandb 2>&1 | tee ~/qlora_sql.log
python train.py \
--context_length 256 \
--model_name codellama/CodeLlama-34b-hf \
--train_type custom_qlora \
--batch_size 4 \
--gradient_accumulation_steps 4 \
--dataset sql \
--save_model True \
--output_dir sql_model_custom_qlora \
--apply_gradient_clipping True \
--project_name fsdp_qlora_sql \
--precision bf16_buffers_autocast \
--log_to wandb 2>&1 | tee ~/custom_qlora_sql.log