Init working

Author: DelVel

.gitignore

@@ -0,0 +1,7 @@
+**/__pycache__
+/data/nli_for_simcse.csv
+/e5v-8b*
+/merged-e5v-8b*
+/models
+/wandb
+/test.ipynb

ft_llm.py

@@ -337,6 +337,8 @@ def train(
         bf16: bool = False,
         not_eol: bool = False,
         org_attn: bool = False,
+        *arg,
+        **kwarg
 ):
     # set NCCL_DEBUG
 

requirements.txt

@@ -6,6 +6,6 @@ bitsandbytes==0.43.1
 scipy
 datasets==2.20.0
 safetensors==0.4.3
-pandas==1.1.5
+pandas
 scikit-learn
 prettytable

retrieval.py

@@ -190,7 +190,7 @@ def init_model_and_transform(lora_path, bf16, fp32, use_e5v=False):
     model_name = "llava-hf/llava-v1.6-mistral-7b-hf"
 
     if MODEL_TYPE == 'llava_llama3':
-        model_name = "./llava-llama-3-8b"
+        model_name = "models/llava-llama-3-8b"
 
     if lora_path is not None:
         merge_path = 'merged-' + lora_path.replace('/', '-').replace('.', '')

run.sh

@@ -26,7 +26,7 @@ NUM_NODES=4
 wandb online
 
 
-NCCL_DEBUG=ERROR deepspeed --num_gpus=$GPUS --num_nodes=$NUM_NODES ft_llm.py \
+NCCL_DEBUG=ERROR deepspeed ft_llm.py \
         --base_model   $BASE_MODEL \
         --data_path 'data/nli_for_simcse.csv' \
         --batch_size $BATCH_SIZE \
@@ -42,5 +42,6 @@ NCCL_DEBUG=ERROR deepspeed --num_gpus=$GPUS --num_nodes=$NUM_NODES ft_llm.py \
         --deepspeed ds.config \
         --lora_target_modules q_proj,k_proj,v_proj,o_proj,gate_proj,down_proj,up_proj  --logging_steps 1 --grad_checkpoint \
          --load_kbit $BIT \
+         --bf16 true \
          ${args[@]}