Merge branch 'main' of https://osredm.com/p04798526/LLaMA-Factory-310P3

batch_run.sh

@@ -0,0 +1 @@
+bash run_once.sh lora_sft Qwen-7B 4 50

gpu_status.py

@@ -1,16 +1,17 @@
 import json
+import sys
 import pynvml 
 import time
 import psutil
 
+UNIT = 1024 * 1024 * 1024
 
-def main():
-    UNIT = 1024 * 1024 * 1024
-
+def gpu_status(output_path = "./results/gpu_status", print_status = False, sleep_time = 60):
     pynvml.nvmlInit()
     gpuDeviceCount = pynvml.nvmlDeviceGetCount()
     start_time = time.time()
-
+    first_loop = True
+    
     while time.time() - start_time < 3600 *24:
         # print(time.time() - start_time)
         all_gpu_status = []
@@ -43,14 +44,26 @@ def main():
             all_gpu_status = all_gpu_status,
             all_processes_status =  all_processes_status
         )
-        formatted_time = time.strftime('%Y%m%d%H%M%S', time.localtime())
-        with open(f"./results/gpu_status/gpu_status_{formatted_time}.json", "a", encoding="utf-8") as f:
-            f.write(json.dumps(logs) + "\n")
-        print(logs)
 
-        time.sleep(60)
+        with open(f"{output_path}/gpu_status.json", "a", encoding="utf-8") as f:
+            f.write(json.dumps(logs) + "\n")
+        
+        if first_loop:
+            print("Start run gpu_status.py")
+            first_loop = False
+            
+        if print_status:
+            print(logs)
+
+        time.sleep(sleep_time)
+        
     pynvml.nvmlShutdown()
 
+def main():
+    output_path = sys.argv[1]
+    print_status = sys.argv[2]
+    sleep_time = sys.argv[3]
+    gpu_status(output_path, print_status, sleep_time)
 
 if __name__ == "__main__":
     main()

prepare_yaml_file.py

@@ -0,0 +1,54 @@
+import sys
+import yaml
+
+def main():
+    run_type = sys.argv[1]
+    model = sys.argv[2]
+    max_steps = sys.argv[3]
+    run_name = sys.argv[4]
+    output_dir = sys.argv[5]
+
+    if run_type == "lora_sft":
+        yaml_file = './results/lora_sft_template.yaml'
+    elif run_type == "inference":
+        yaml_file = './results/predict_template.yaml'
+    
+    model_name_or_path = ""
+    template = ""
+    if model == "9g-8B":
+        model_name_or_path = "../../models/sft_8b_v2"
+        template = ""
+    elif model == "Baichuan2-7B":
+        model_name_or_path = "../../models/Baichuan2-7B-Base"
+        template = "baichuan2"
+    elif model == "ChatGLM2-6B":
+        model_name_or_path = "../../models/chatglm2-6b"
+        template = "chatglm2"
+    elif model == "Llama2-7B":
+        model_name_or_path = "../../models/llama-2-7b-ms"
+        template = "llama2"
+    elif model == "Qwen-7B":
+        model_name_or_path = "../../models/Qwen-7B"
+        template = "qwen"
+    else:
+        print("ERROR: model not supported.")
+        sys.exit()
+        
+    config = None
+    with open(yaml_file, 'r', encoding='utf-8') as f:
+        config = yaml.load(f.read(), Loader=yaml.FullLoader)
+    
+    
+    config['model_name_or_path'] = model_name_or_path
+    config['template'] = template
+    config['output_dir'] = output_dir
+    if run_type == "lora_sft":
+        config['max_steps'] = int(max_steps)
+
+    with open(f'{output_dir}/{run_name}.yaml', 'w', encoding='utf-8') as f:
+        yaml.dump(data=config, stream=f, allow_unicode=True)
+    
+    print(f"yaml file saved to {output_dir}/{run_name}.yaml")
+
+if __name__ == "__main__":
+    main()

results/lora_sft_Qwen-7B_4_gpu_50_step_20240905070656/lora_sft_Qwen-7B_4_gpu_50_step_20240905070656.yaml

@@ -0,0 +1,31 @@
+bf16: true
+cutoff_len: 1024
+dataset: belle_1m
+ddp_timeout: 180000000
+do_train: true
+eval_steps: 500
+eval_strategy: steps
+finetuning_type: lora
+gradient_accumulation_steps: 8
+include_num_input_tokens_seen: true
+include_tokens_per_second: true
+learning_rate: 0.0001
+logging_steps: 3
+lora_target: all
+lr_scheduler_type: cosine
+max_samples: 10000
+max_steps: '50'
+model_name_or_path: ../../models/Qwen-7B
+num_train_epochs: 10.0
+output_dir: ./results/lora_sft_Qwen-7B_4_gpu_50_step_20240905070656
+overwrite_cache: true
+overwrite_output_dir: true
+per_device_eval_batch_size: 2
+per_device_train_batch_size: 2
+plot_loss: true
+preprocessing_num_workers: 16
+save_steps: 500
+stage: sft
+template: qwen
+val_size: 0.1
+warmup_ratio: 0.1

results/lora_sft_template.yaml

@@ -0,0 +1,42 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+stage: sft
+do_train: true
+finetuning_type: lora
+lora_target: all
+
+### dataset
+dataset: belle_1m
+template: baichuan
+cutoff_len: 1024
+max_samples: 10000
+overwrite_cache: true
+preprocessing_num_workers: 16
+
+### output
+output_dir: ./results/lora_sft_2/Baichuan2-7B/Baichuan2_lora_sft_1_single_step500
+logging_steps: 3
+save_steps: 500
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 8
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+fp16: true
+ddp_timeout: 180000000
+max_steps: 500
+include_num_input_tokens_seen: true
+include_tokens_per_second: true
+
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500

results/predict_template.yaml

@@ -0,0 +1,23 @@
+### model
+model_name_or_path: ../../llm/baichuan
+
+### method
+do_predict: true
+
+### dataset
+eval_dataset: alpaca_gpt4_zh
+template: baichuan
+cutoff_len: 1024
+max_samples: 50
+overwrite_cache: true
+preprocessing_num_workers: 16
+include_tokens_per_second: true
+
+### output
+output_dir: ./results/inference/Baichuan2-7B/Baichuan2_predict_1
+overwrite_output_dir: true
+
+### eval
+per_device_eval_batch_size: 2
+predict_with_generate: true
+ddp_timeout: 180000000

run_once.sh

@@ -0,0 +1,52 @@
+#!/bin/bash
+
+run_type="$1"
+model="$2"
+gpu_cnt="$3"
+max_steps="$4"
+
+current_datetime=$(date +%Y%m%d%H%M%S)
+
+if [ "${run_type}"="lora_sft" ]; then
+    run_name="${run_type}_${model}_${gpu_cnt}_gpu_${max_steps}_step_${current_datetime}"
+    
+else
+    run_name="${run_type}_${model}_${gpu_cnt}_gpu_${current_datetime}"
+fi
+
+output_dir="./results/${run_name}"
+
+if [ ! -d "$output_dir" ]; then
+    mkdir -p "$output_dir"
+    echo "output_dir created: $output_dir"
+else
+    echo "output_dir exists: $output_dir"
+fi
+
+# echo "${run_type} ${model} ${gpu_cnt} ${max_steps} ${run_name} ${output_dir}"
+python prepare_yaml_file.py ${run_type} ${model} ${max_steps} ${run_name} ${output_dir}
+
+export USE_MODELSCOPE_HUB=1
+
+# echo "Start recording gpu status "
+# # 0 means not printing gpu status
+# python gpu_status.py ${output_dir} 1 10 &
+# gpu_status_pid=$!
+# echo "${gpu_status_pid}"
+
+if [ "${gpu_cnt}"="1" ]; then
+   ASCEND_RT_VISIBLE_DEVICES=0 llamafactory-cli train ${output_dir}/${run_name}.yaml | tee "${output_dir}/log.txt" &
+   train_pid=$!
+   echo "Start train"
+else
+   FORCE_TORCHRUN=1 llamafactory-cli train ${output_dir}/${run_name}.yaml | tee "${output_dir}/log.txt" &
+   train_pid=$!
+   echo "Start train"
+fi
+
+wait $train_pid
+echo "Train ended"
+
+# sleep 90
+# kill $gpu_status_pid
+# echo "Gpu status ended"