attempted gemma2

llama-factory/config/{gemma2_9b_lora_sft-p1.yaml → gemma2-9b_lora_sft_bf16-p1.yaml}

@@ -13,7 +13,7 @@ upcast_layernorm: true
 
 ### dataset
 dataset: alpaca_mgtv_p1
-template: gemma2
+template: gemma
 cutoff_len: 4096
 max_samples: 25000
 overwrite_cache: true
@@ -44,4 +44,4 @@ eval_strategy: steps
 eval_steps: 2109
 
 report_to: wandb
-run_name: qwen2_72b_4bit_p1_full # optional
+run_name: gemma2_9b_p1_full # optional

scripts/tune-mgtv-gemma2_9b.sh

@@ -14,7 +14,9 @@ lscpu
 grep MemTotal /proc/meminfo
 
 #pip install -r requirements.txt
-#cd ../LLaMA-Factory && pip install -e .[torch,bitsandbytes]
+cd ../LLaMA-Factory && pip install -e .[torch,bitsandbytes] && cd $BASEDIR
+#pip install -U transformers
+#pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
 
 export LOGICAL_REASONING_DATA_PATH=datasets/mgtv
 

scripts/tune-mgtv.sh

@@ -1 +1 @@
-tune-mgtv-qwen2_72b.sh
+tune-mgtv-gemma2_9b.sh