import subprocess
import os
import time
import torch
import utils
model_name = "hf-models/Qwen2.5-7B-Instruct"

GITEE_ACCESS_TOKEN = os.environ.get("GITEE_ACCESS_TOKEN", "")

# print("开始下载 lora")
# git_clone_docs_lora_command = ["git", "clone",
#                                f"https://oauth2:{GITEE_ACCESS_TOKEN}@gitee.com/stringify/glm4-lora-gitee-docs-9b.git",  "--depth=1", "--single-branch", "./glm4-lora-gitee-docs-9b"]
# subprocess.run(
#     git_clone_docs_lora_command, text=True)
# print("lora 下载完成")

# subprocess.Popen(['pip','uninstall','-y','flash_attn'])


os.environ['HF_HOME'] = '/data'
api_server_command = [
    "python",
    "-m",
    "vllm.entrypoints.openai.api_server",
    "--model",
    model_name,
    "--dtype",
    "bfloat16",
    "--api-key",
    "",
    "--tensor-parallel-size",
    str(torch.cuda.device_count() or 2),
    "--trust-remote-code",
    "--gpu-memory-utilization",
    "0.71",
    "--max-num-batched-tokens",
    # "--enable-lora",
    # "--lora-modules",
    # "gitee-docs-lora=./glm4-lora-gitee-docs-9b", # vllm 0.3.3 不支持 glm4 qwen2 lora: ValueError: Model Qwen2ForCausalLM does not support LoRA, but LoRA is enabled. Support for this model may be added in the future. If this is important to you, please open an issue on github.
    # ValueError: max_num_batched_tokens (55000) is smaller than max_model_len (131072). This effectively limits the maximum sequence length to max_num_batched_tokens and makes vLLM reject longer sequences. Please increase max_num_batched_tokens or decrease max_model_len.
    "21000",
    "--max-model-len",
    "21000",
    "--disable-log-requests",
    "--disable-log-stats",
    "--port",
    "8000",
    "--block-size",
    "16",  # 0.33 只支持16 默认 16
    "--max-num-seqs",
    "1024"
    # "--enable-chunked-prefill",  # 0.3.3 不支持启用分块预填充
    # 多卡跑多 模型时, vllm GPU blocks: 0 https://github.com/vllm-project/vllm/issues/2248
    # "--enforce-eager"
]

chainlit_ui_process = subprocess.Popen(
    ['python', '-m', 'chainlit', 'run', 'chainlit_ui_crawler.py', '--host', '0.0.0.0', '--port', '7860', "--ci", "--headless"])


def wait_for_service(url):
    while True:
        try:
            server_ready = utils.is_port_open(url)
            if server_ready:
                return True
        except:
            time.sleep(5)


if (wait_for_service("http://127.0.0.1:7860")):
    api_process = subprocess.Popen(
        api_server_command, text=True)
    print("UI 服务已启动，开始启动 API 服务...")

print("开始启动 api 服务")

# device = torch.cuda.current_device()
# props = torch.cuda.get_device_properties(device)

# print(f"Device Name: {props.name}")
# print(f"Total Memory: {props.total_memory / (1024 ** 3)} GB")


try:
    api_process.wait()
    chainlit_ui_process.wait()
# except KeyboardInterrupt:
#     print("Shutting down servers.")
#     chainlit_ui_process.terminate()
#     api_process.terminate()
#     api_process.wait()
#     chainlit_ui_process.wait()
finally:
    api_process.kill()
    chainlit_ui_process.kill()
    print("Servers shut down.")