import ray
import subprocess
import threading
from ray.util.queue import Queue

print(ray.is_initialized())
ray.init(address='auto')

# ── 获取全部 Worker 节点 ──────────────────────────────────
nodes = [n for n in ray.nodes() if n['Alive']]
head_node_id = ray.get_runtime_context().get_node_id()
worker_nodes = sorted(
    [n for n in nodes if n['NodeID'] != head_node_id],
    key=lambda x: x['NodeManagerAddress']  # 排序保证 rank 稳定
)

num_workers  = len(worker_nodes)   # 4
gpus_per_node = 8
master_addr  = worker_nodes[0]['NodeManagerAddress']  # rank0 作为 master
master_port  = 29600

print(f"Worker 节点数: {num_workers}，共 {num_workers * gpus_per_node} 块 GPU")
for i, n in enumerate(worker_nodes):
    print(f"  [rank{i}] IP: {n['NodeManagerAddress']}")
print(f"Master: {master_addr}:{master_port}\n{'─'*60}")

# ── 远程函数：每个 Worker 各占 8 GPU ─────────────────────
@ray.remote(num_gpus=8)
def run_cmd_stream(command, queue, node_rank):
    import subprocess, threading, os

    env = os.environ.copy()
    # ✅ 注入分布式训练必要的环境变量
    env['MASTER_ADDR']  = master_addr
    env['MASTER_PORT']  = str(master_port)
    env['NODE_RANK']    = str(node_rank)
    env['NNODES']       = str(num_workers)
    env['NPROC_PER_NODE'] = str(gpus_per_node)

    print(f"[rank{node_rank}] CUDA_VISIBLE_DEVICES = {env.get('CUDA_VISIBLE_DEVICES', 'NOT SET')}")
    print(f"[rank{node_rank}] MASTER={env['MASTER_ADDR']}:{env['MASTER_PORT']}")

    process = subprocess.Popen(
        command,
        shell=True,
        stdout=subprocess.PIPE,
        stderr=subprocess.PIPE,
        text=True,
        bufsize=1,
        env=env
    )

    def push_stream(stream, tag):
        for line in iter(stream.readline, ''):
            queue.put((tag, f"[rank{node_rank}] {line}"))
        stream.close()

    t_out = threading.Thread(target=push_stream, args=(process.stdout, 'OUT'))
    t_err = threading.Thread(target=push_stream, args=(process.stderr, 'ERR'))
    t_out.start()
    t_err.start()
    t_out.join()
    t_err.join()

    returncode = process.wait()
    queue.put(('DONE', returncode))   # 每个 worker 发送一个 DONE
    return returncode

# ── 创建共享队列 ──────────────────────────────────────────
output_queue = Queue(maxsize=10000)

# ── 在全部 Worker 上同时启动 ──────────────────────────────
COMMAND = "redaccel-cli train /mnt/tidal-alsh01/usr/dawo/qinshengqian/LLaMA-Factory/qs_ray_ide.yaml"

futures = []
for rank, worker in enumerate(worker_nodes):

    future = run_cmd_stream.options(
        scheduling_strategy=ray.util.scheduling_strategies.NodeAffinitySchedulingStrategy(
            node_id=worker['NodeID'],
            soft=False
        )
    ).remote(COMMAND, output_queue, rank)
    futures.append(future)
    print(f"▶ 已提交 rank{rank} → {worker['NodeManagerAddress']}")

print(f"\n{'─'*60} 开始输出 {'─'*60}\n")

# ── 主进程消费队列，等待全部 Worker 完成 ──────────────────
done_count = 0
while done_count < num_workers:
    tag, content = output_queue.get()

    if tag == 'DONE':
        done_count += 1
        print(f"\n✅ Worker 完成 ({done_count}/{num_workers})，返回码: {content}")
    elif tag == 'OUT':
        print(content, end='', flush=True)
    elif tag == 'ERR':
        print(f"\033[33m{content}\033[0m", end='', flush=True)

# ── 汇总结果 ──────────────────────────────────────────────
print(f"\n{'─'*60}")
returncodes = ray.get(futures)
print(f"所有 Worker 返回码: {returncodes}")

if all(rc == 0 for rc in returncodes):
    print("🎉 全部成功！")
else:
    print("❌ 部分 Worker 失败，请检查上方日志")