update readme for merge model

tests/deltatuner/finetune/merge_model/readme.md

@@ -27,7 +27,6 @@ python instruction_tuning_pipeline/finetune_clm.py \
         --output_dir "$DATA_PATH/llama2-7b-ssf-denas-bf16" \
         --delta ssf \
         --denas True \
-        --bf16 True \
         | tee llama2-7b-ssf-denas-bf16-1epoch.log
 ```
 
@@ -39,13 +38,10 @@ python instruction_tuning_pipeline/finetune_clm.py \
         --model_name_or_path "$DATA_PATH/Llama-2-7b-hf" \
         --train_file "$DATA_PATH/alpaca_data.json" \
         --dataset_concatenation \
-        --per_device_train_batch_size 8 \
         --per_device_eval_batch_size 8 \
         --gradient_accumulation_steps 1 \
         --validation_split_percentage 30 \
         --do_eval \
-        --learning_rate 1e-4 \
-        --num_train_epochs 1 \
         --logging_steps 100 \
         --save_total_limit 1 \
         --log_level info \
@@ -54,29 +50,27 @@ python instruction_tuning_pipeline/finetune_clm.py \
         --no_cuda \
         --output_dir "$DATA_PATH/llama2-7b-ssf-denas-bf16-merge" \
         --delta ssf \
-        --bf16 True \
         --resume_peft "$DATA_PATH/llama2-7b-ssf-denas-bf16" \
         --save_merged_model True \
-        --denas "$DATA_PATH/llama2-7b-ssf-denas-bf16/best_model_structure.txt" \
+        --merge_model_code_dir "instruction_tuning_pipeline/models/llama2-ssf" \
         --debugs
 ```
 
-### 3. Evaluate merged model
 As ssf will enable bias, while the default Llama2 disable all the bias, to enable the full parameters of the adapter, we need to change the model definition.
 
-First copy the updated model codes with the merged weights.
-```shell
-cp instruction_tuning_pipeline/models/llama2-ssf/* $DATA_PATH/llama2-7b-ssf-denas-bf16-merge/merged_model
-```
+First, specify the `merge_model_code_dir` args, it will copy the updated model codes with the merged weights.
 
-Then update the "best_model_structure" and "target_modules" setting in config.json. if not enable denas and not change "target_modules" default settings, can skip correpsonding setting.
+Then it will automatically update the "best_model_structure" and "target_modules" setting in `config.json`. if "denas" or "target_modules" are not enabled/changed, it will skip the corresponding setting.
+The changed `config.json` looks like this:
 ```shell
 ...
 "best_model_structure":  {"num_hidden_layers": [1, 1, 1, 0, 1, 0, 0, 0, 0, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 1, 1, 1, 1, 0, 1, 1]}, # change to your best structure, skip to keep default
 "target_modules": ["q_proj", "v_proj"], #change to your setting, skip to keep default
 ...
 ```
 
+### 3. Evaluate merged model
+
 Finally we can directly evalute the merged model.
 ```shell
 python instruction_tuning_pipeline/finetune_clm.py \
@@ -93,5 +87,4 @@ python instruction_tuning_pipeline/finetune_clm.py \
         --trust_remote_code True \
         --no_cuda \
         --output_dir "$DATA_PATH/llama2-7b-ssf-denas-bf16-merge/eval_merge" \
-        --bf16 True
 ```

tests/deltatuner/finetune/merge_model/ssf-merge-test.sh

@@ -19,6 +19,7 @@ python instruction_tuning_pipeline/finetune_clm.py \
         --delta ssf \
         --resume_peft "$DATA_PATH/mpt-7b-ssf-allmodules-denas-bf16" \
         --save_merged_model True \
+        --merge_model_code_dir instruction_tuning_pipeline/models/llama2-ssf \
         --debugs