search

DeepLink-org / deeplink.framework

BSD 3-Clause "New" or "Revised" License
59 stars 28 forks source link

[Fix] fix ascend CI test-one-iter bug & update test-one-iter traditional model list #915

Closed DoorKickers closed 3 months ago

DoorKickers commented 3 months ago

summary

该pr主要用于：

  1. 解决CI ascend test-one-iter pin-memory相关bug
  2. 添加CI ascend test-one-iter 部分传统模型（9个）

问题描述

CI ascend test-one-iter会出现以下报错

Traceback (most recent call last):
2024-07-25T06:30:46.3968189Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmpretrain/tools/train.py", line 166, in <module>
2024-07-25T06:30:46.3969123Z     main()
2024-07-25T06:30:46.3970030Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmpretrain/tools/train.py", line 162, in main
2024-07-25T06:30:46.3970936Z     runner.train()
2024-07-25T06:30:46.3971950Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmengine/mmengine/runner/runner.py", line 1721, in train
2024-07-25T06:30:46.3973000Z     model = self.train_loop.run()  # type: ignore
2024-07-25T06:30:46.3974102Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmengine/mmengine/runner/loops.py", line 96, in run
2024-07-25T06:30:46.3975398Z     self.run_epoch()
2024-07-25T06:30:46.3976429Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmengine/mmengine/runner/loops.py", line 111, in run_epoch
2024-07-25T06:30:46.3977464Z     for idx, data_batch in enumerate(self.dataloader):
2024-07-25T06:30:46.3978179Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/dataloader.py", line 437, in __iter__
2024-07-25T06:30:46.3978878Z     self._iterator = self._get_iterator()
2024-07-25T06:30:46.3979590Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/dataloader.py", line 388, in _get_iterator
2024-07-25T06:30:46.3980325Z     return _MultiProcessingDataLoaderIter(self)
2024-07-25T06:30:46.3981029Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/dataloader.py", line 1055, in __init__
2024-07-25T06:30:46.3981842Z     current_device = torch.cuda.current_device()  # choose cuda for default
2024-07-25T06:30:46.3982634Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/cuda/__init__.py", line 674, in current_device
2024-07-25T06:30:46.3983247Z     _lazy_init()
2024-07-25T06:30:46.3983762Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/cuda/__init__.py", line 239, in _lazy_init
2024-07-25T06:30:46.3984460Z     raise AssertionError("Torch not compiled with CUDA enabled")
Traceback (most recent call last):
2024-07-25T06:30:46.9071670Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmpretrain/tools/train.py", line 166, in <module>
2024-07-25T06:30:46.9075870Z     main()
2024-07-25T06:30:46.9079095Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmpretrain/tools/train.py", line 162, in main
2024-07-25T06:30:46.9082117Z     runner.train()
2024-07-25T06:30:46.9085814Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmengine/mmengine/runner/runner.py", line 1721, in train
2024-07-25T06:30:46.9089244Z     model = self.train_loop.run()  # type: ignore
2024-07-25T06:30:46.9093014Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmengine/mmengine/runner/loops.py", line 96, in run
2024-07-25T06:30:46.9096128Z     self.run_epoch()
2024-07-25T06:30:46.9099569Z   File "/mnt/cache/share/deeplinkci/github/DeepLink-org/deeplink.framework/6045/Build-Ascend-910b/dipu/mmlab_pack/mmengine/mmengine/runner/loops.py", line 111, in run_epoch
2024-07-25T06:30:46.9101040Z     for idx, data_batch in enumerate(self.dataloader):
2024-07-25T06:30:46.9101557Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/dataloader.py", line 634, in __next__
2024-07-25T06:30:46.9102014Z     data = self._next_data()
2024-07-25T06:30:46.9102461Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/dataloader.py", line 680, in _next_data
2024-07-25T06:30:46.9103016Z     data = _utils.pin_memory.pin_memory(data, self._pin_memory_device)
2024-07-25T06:30:46.9103576Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/_utils/pin_memory.py", line 60, in pin_memory
2024-07-25T06:30:46.9104335Z     return type(data)({k: pin_memory(sample, device) for k, sample in data.items()})  # type: ignore[call-arg]
2024-07-25T06:30:46.9105004Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/_utils/pin_memory.py", line 60, in <dictcomp>
2024-07-25T06:30:46.9105835Z     return type(data)({k: pin_memory(sample, device) for k, sample in data.items()})  # type: ignore[call-arg]
2024-07-25T06:30:46.9106489Z   File "/mnt/cache/share/platform/cienv/pytorch/torch/utils/data/_utils/pin_memory.py", line 55, in pin_memory
2024-07-25T06:30:46.9106982Z     return data.pin_memory(device)
2024-07-25T06:30:46.9107667Z   File "/home/autolink/.local/lib/python3.9/site-packages/torch_dipu-0.1-py3.9-linux-x86_64.egg/torch_dipu/dipu/device.py", line 99, in _proxyFuncInst
2024-07-25T06:30:46.9108258Z     return rawfunc(self, *args, **kwargs)
2024-07-25T06:30:46.9111819Z NotImplementedError: Could not run 'aten::_pin_memory' with arguments from the 'CUDA' backend.

经过debug，发现这两个报错都与训练过程中 dataloader 中的pin-memory参数有关。 若在训练命令后加上--no-pin-memory参数，就不会出现上述报错。

通过查阅相关代码，我发现以上两种报错的根本原因是 one-iter CPU模式训练过程中尝试去使用并不存在的cuda device。 在one-iter CPU模式中，脚本将mock_cuda改成false，导致训练过程中的cuda device并没有落到厂商device上。

#!/bin/bash

if [ -z "$ONE_ITER_TOOL_STORAGE_PATH" ]; then
    export ONE_ITER_TOOL_STORAGE_PATH=$(pwd)/one_iter_data
fi

# 获取传递的参数
train_path=$1
config_path=$2
work_dir=$3
opt_arg=$4
workdirs="${work_dir##*/}"

echo "cpu input"
export DIPU_MOCK_CUDA=False // cpu模式将mock_cuda设为false

device_mode_list=()

if [ "$ONE_ITER_TOOL_DEVICE_COMPARE" == "cpu" ]; then
    ONE_ITER_TOOL_DEVICE=cpu ONE_ITER_TOOL_MODE=input python -u $train_path $config_path  $work_dir  $opt_arg
    device_mode_list=("cpu others" "dipu others")
elif [ "$ONE_ITER_TOOL_DEVICE_COMPARE" == "gpu" ]; then
    ONE_ITER_TOOL_DEVICE=gpu ONE_ITER_TOOL_MODE=input python -u $train_path $config_path  $work_dir  $opt_arg
    device_mode_list=("gpu others" "dipu others")
else
    echo "ONE_ITER_TOOL_DEVICE_COMPARE is not 'cpu' or 'gpu'"
fi
# 检查返回值
if [ $? -ne 0 ]; then
    echo "Error executing task: $config_path  cpu  input"
    exit 1
fi

同时，在import dipu时，会有一些逻辑绕过torch dataloader原生的device check：

torch/utils/data/dataloader.py: _BaseDataLoaderIter类中部分代码如下：

        # for other backends, pin_memory_device need to set. if not set
        # default behaviour is CUDA device. if pin_memory_device is selected
        # and pin_memory is not set, the default behaviour false.
        if len(loader.pin_memory_device) == 0:
            self._pin_memory = loader.pin_memory and torch.cuda.is_available()
            self._pin_memory_device = None
        else:
            if not loader.pin_memory:
                warn_msg = (
                    "pin memory device is set and pin_memory flag is not used then device pinned memory won't be used"
                    "please set pin_memory to true, if you need to use the device pin memory"
                )
                warnings.warn(warn_msg)

            self._pin_memory = loader.pin_memory
            self._pin_memory_device = loader.pin_memory_device

如果用户没有传入pin_memory_device，那么torch会默认pin_memory_device为cuda，同时会检查torch.cuda.is_available()，若当前cuda device不可用，torch会将用户传递的pin_memory_device设为false。

而torch_dipu实现的dataloader中，将pin_memory_device直接置为“cuda”，这会使得len(loader.pin_memory_device) != 0，即绕过了torch的cuda可用性检查。

class DIPUDataLoader(DataLoader):
    UNSUPPORTED_PINMEMORY_VENDORS = ["DROPLET"]

    def __init__(
        self,
        dataset: Dataset[T_co],
        batch_size: Optional[int] = 1,
        shuffle: Optional[bool] = None,
        sampler: Union[Sampler, Iterable, None] = None,
        batch_sampler: Union[Sampler[Sequence], Iterable[Sequence], None] = None,
        num_workers: int = 0,
        collate_fn: Optional[_collate_fn_t] = None,
        pin_memory: bool = False,
        drop_last: bool = False,
        timeout: float = 0,
        worker_init_fn: Optional[_worker_init_fn_t] = None,
        multiprocessing_context=None,
        generator=None,
        *,
        prefetch_factor: Optional[int] = None,
        persistent_workers: bool = False,
        pin_memory_device: str = "",
    ):
        if dipu.vendor_type in self.UNSUPPORTED_PINMEMORY_VENDORS:
            print(
                f"[DIPU] warning: {dipu.vendor_type} does not support pin_memory, continuing with pin_memory=False\n",
                flush=True,
                end=None,
            )
            pin_memory = False
        elif pin_memory:
            pin_memory_device = "cuda" # 此行造成torch检查失效

        super().__init__(
            dataset,
            batch_size,
            shuffle,
            sampler,
            batch_sampler,
            num_workers,
            collate_fn,
            pin_memory,
            drop_last,
            timeout,
            worker_init_fn,
            multiprocessing_context,
            generator,
            prefetch_factor=prefetch_factor,
            persistent_workers=persistent_workers,
            pin_memory_device=pin_memory_device,
        )

因此训练过程中会尝试使用并不存在的cuda device，从而出现错误。

同时，经过测试，若将寒武纪CI脚本run one iter中的srun命令改为bash，也会出现相同的报错this commit check。

    elif device == "camb":
        # For the inference of large language models, simply compare the inference results on the current device directly with the results generated on the GPU
        if "infer" in p2 and "infer" in p3:
            cmd_run_one_iter = f"srun --job-name={job_name} --partition={partition}  --gres={gpu_requests} --time=40 python {train_path}"
            cmd_cp_one_iter = ""
        else:
            cmd_run_one_iter = f"bash SMART/tools/one_iter_tool/run_one_iter.sh {train_path} {config_path} {work_dir} {opt_arg}"
            cmd_cp_one_iter = f"bash SMART/tools/one_iter_tool/compare_one_iter.sh {package_name} {atol} {rtol} {metric}"
            # cmd_run_one_iter = f"srun --job-name={job_name} --partition={partition}  --gres={gpu_requests} --time=40 sh SMART/tools/one_iter_tool/run_one_iter.sh {train_path} {config_path} {work_dir} {opt_arg}"
            # cmd_cp_one_iter = f"srun --job-name={job_name} --partition={partition}  --gres={gpu_requests} --time=30 sh SMART/tools/one_iter_tool/compare_one_iter.sh {package_name} {atol} {rtol} {metric}"

解决方法

在环境变量dipu_mock_cuda = False的情况下，绕过dipu dataloader中的pin_memory_device初始化赋值，从而在pin-memory = True时会走到torch dataloader原生的torch.cuda.is_available()检查