guanghulab/scripts/auto_pipeline.py

84 lines
3.2 KiB
Python
Raw Normal View History

#!/usr/bin/env python3
"""自动训练流水线 - 铸渊自管理系统
母模型完成 上传COS 启动代码模型训练 上传COS
使用方法
export ZY_OSS_KEY=<SecretId>
export ZY_OSS_SECRET=<SecretKey>
nohup python3 auto_pipeline.py > pipeline.log 2>&1 &
"""
import os, json, time, sys, subprocess
sys.stdout.reconfigure(line_buffering=True)
MOTHER_LOG = "/root/autodl-tmp/train_mother.log"
MOTHER_OUT = "/root/autodl-tmp/output/qwen25-7b-sft/final"
COS_BUCKET = "sy-finetune-corpus-1317346199"
COS_REGION = "ap-guangzhou"
# 密钥通过环境变量读取(不硬编码)
COS_SECRET_ID = os.environ.get("ZY_OSS_KEY") or os.environ.get("COS_SECRET_ID")
COS_SECRET_KEY = os.environ.get("ZY_OSS_SECRET") or os.environ.get("COS_SECRET_KEY")
if not COS_SECRET_ID or not COS_SECRET_KEY:
print("❌ 需要环境变量export ZY_OSS_KEY=... ZY_OSS_SECRET=...")
sys.exit(1)
CODE_MODEL = "Qwen/Qwen2.5-Coder-7B"
CODE_CACHE = "/root/autodl-tmp/cache"
CODE_OUT = "/root/autodl-tmp/output/qwen25-coder-7b-sft"
CODE_SCRIPT = "/root/autodl-tmp/train_coder.py"
DATA_FILE = "/root/autodl-tmp/data/sft.jsonl"
POLL_INTERVAL = 300
def run_cmd(cmd):
r = subprocess.run(cmd, shell=True, capture_output=True, text=True)
return r.stdout.strip(), r.returncode
def log(msg):
t = time.strftime("%Y-%m-%d %H:%M:%S")
print(f"[{t}] {msg}"); sys.stdout.flush()
def upload_to_cos(local_path, cos_key):
log(f"Uploading {local_path} → cos://{COS_BUCKET}/{cos_key}")
s = "from qcloud_cos import CosConfig, CosS3Client; "
s += f"c=CosS3Client(CosConfig(Region='{COS_REGION}',SecretId='{COS_SECRET_ID}',SecretKey='{COS_SECRET_KEY}')); "
s += f"c.upload_file(Bucket='{COS_BUCKET}',Key='{cos_key}',LocalFilePath='{local_path}')"
out, rc = run_cmd(f"python3 -c \"{s}\"")
log(out); return rc == 0
def check_mother_done():
if not os.path.exists(MOTHER_LOG): return False
out, _ = run_cmd(f"grep -a 'DONE' {MOTHER_LOG} | tail -1")
return "DONE" in out
def check_mother_output():
return os.path.isdir(MOTHER_OUT) and any(f.endswith('.safetensors') for f in os.listdir(MOTHER_OUT))
def generate_coder_script():
s = open("/root/autodl-tmp/train_mother.py").read()
s = s.replace("Qwen2___5-7B", "Qwen2___5-Coder-7B")
s = s.replace("qwen25-7b-sft", "qwen25-coder-7b-sft")
return s
def main():
log("=== 自动训练流水线启动 ===")
while True:
if check_mother_done() or check_mother_output():
break
time.sleep(POLL_INTERVAL)
log("✅ 母模型训练完成")
if upload_to_cos(MOTHER_OUT, "models/qwen25-7b-sft/final"):
log("✅ 母模型上传COS完成")
with open(CODE_SCRIPT, 'w') as f: f.write(generate_coder_script())
run_cmd(f"nohup python3 -u {CODE_SCRIPT} > train_coder.log 2>&1 &")
log("✅ 代码模型训练已启动")
while True:
out, _ = run_cmd("grep -a 'DONE' /root/autodl-tmp/train_coder.log | tail -1")
if "DONE" in out:
log("✅ 代码模型训练完成"); break
time.sleep(POLL_INTERVAL)
if upload_to_cos(f"{CODE_OUT}/final", "models/qwen25-coder-7b-sft/final"):
log("✅ 代码模型上传COS完成")
log("🎉 全部训练完成!")
if __name__ == "__main__":
main()