system_check.py

# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import logging
import os
import shutil
import subprocess
import time
from datetime import timedelta
from pathlib import Path

import torch
import torch.distributed
import torch.multiprocessing as mp
from lightning_utilities.core.imports import RequirementCache
from torch.multiprocessing.spawn import ProcessRaisedException

_psutil_available = RequirementCache("psutil")
_logger = logging.getLogger(__name__)
_system_check_dir = Path("./system_check")


def main(timeout: int = 60) -> None:
    _setup_logging()

    num_cuda_devices = torch.cuda.device_count()

    if num_cuda_devices == 0:
        _print0("Warning: Skipping system check because no GPUs were detected.")

    if num_cuda_devices == 1:
        _describe_nvidia_smi()
        # _check_cuda()

    if num_cuda_devices > 1:
        _describe_nvidia_smi()
        _describe_gpu_connectivity()

        success = _check_cuda_distributed(timeout)

        if not success:
            env = {
                "NCCL_P2P_DISABLE": "1",
                "NCCL_NET_PLUGIN": "none",
            }
            _print0(
                "The multi-GPU NCCL test did not succeed."
                " It looks like there is an issue with your multi-GPU setup."
                " Now trying to run again with NCCL features disabled."
            )
            os.environ.update(env)
            success = _check_cuda_distributed(timeout)
            if success:
                _print0("Disabling the following NCCL features seems to have fixed the issue:")
                _print_env_variables(env)
            else:
                _print0("Disabling NCCL features did not fix the issue.")

        if success:
            _print0("Multi-GPU test successful.")

    _print0(f"Find detailed logs at {_system_check_dir.absolute()}")


def _check_cuda_distributed(timeout: int) -> bool:
    if not _psutil_available:
        raise ModuleNotFoundError(str(_psutil_available))

    num_cuda_devices = torch.cuda.device_count()
    context = mp.spawn(
        _run_all_reduce_test,
        nprocs=num_cuda_devices,
        args=(num_cuda_devices,),
        join=False,
    )

    start = time.time()
    success = False
    while not success and (time.time() - start < timeout):
        try:
            success = context.join(timeout=5)
        except ProcessRaisedException as e:
            _logger.debug(str(e))
            success = False
            break

        time.sleep(1)

    if not success:
        for pid in context.pids():
            _kill_process(pid)
    return success


def _run_all_reduce_test(local_rank: int, world_size: int) -> None:
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "29500"
    os.environ["WORLD_SIZE"] = str(world_size)
    os.environ["RANK"] = str(local_rank)
    os.environ["LOCAL_RANK"] = str(local_rank)
    os.environ["NCCL_DEBUG"] = "INFO"
    os.environ["NCCL_DEBUG_FILE"] = str(_system_check_dir / f"nccl-rank-{local_rank}.txt")

    device = torch.device("cuda", local_rank)
    torch.cuda.set_device(local_rank)

    _print0("Setting up the process group ...")
    torch.distributed.init_process_group(
        backend="nccl",
        world_size=world_size,
        rank=local_rank,
        # NCCL gets initialized in the first collective call (e.g., barrier below),
        # which must be successful for this timeout to work.
        timeout=timedelta(seconds=10),
    )

    _print0("Synchronizing GPUs ... ")
    torch.distributed.barrier()

    payload = torch.rand(100, 100, device=device)
    _print0("Running all-reduce test ...")
    torch.distributed.all_reduce(payload)


def _setup_logging() -> None:
    if _system_check_dir.is_dir():
        shutil.rmtree(_system_check_dir)
    _system_check_dir.mkdir()

    _logger.setLevel(logging.DEBUG)
    file_handler = logging.FileHandler(str(_system_check_dir / "logs.txt"))
    file_handler.setLevel(logging.DEBUG)
    console_handler = logging.StreamHandler()
    console_handler.setLevel(logging.INFO)
    _logger.addHandler(file_handler)
    _logger.addHandler(console_handler)
    _logger.propagate = False


def _print0(string: str) -> None:
    if int(os.getenv("RANK", 0)) == 0:
        _logger.info(string)


def _print_env_variables(env: dict) -> None:
    for k, v in env.items():
        _print0(f"{k}={v}")


def _collect_nvidia_smi_topo() -> str:
    return subprocess.run(["nvidia-smi", "topo", "-m"], capture_output=True, text=True).stdout


def _collect_nvidia_smi() -> str:
    return subprocess.run(["nvidia-smi"], capture_output=True, text=True).stdout


def _describe_nvidia_smi() -> None:
    _logger.info(
        "Below is the output of `nvidia-smi`. It shows information about the GPUs that are installed on this machine,"
        " the driver version, and the maximum supported CUDA version it can run.\n"
    )
    _logger.info(_collect_nvidia_smi())


def _describe_gpu_connectivity() -> None:
    _logger.debug(
        "The matrix below shows how the GPUs in this machine are connected."
        " NVLink (NV) is the fastest connection, and is only available on high-end systems like V100, A100, etc.\n"
    )
    _logger.debug(_collect_nvidia_smi_topo())


def _kill_process(pid: int) -> None:
    import psutil

    try:
        process = psutil.Process(pid)
        if process.is_running():
            process.kill()
    except (psutil.NoSuchProcess, psutil.AccessDenied):
        pass


if __name__ == "__main__":
    main()