what is the version of python

[learn01one]

hello,can i ask the version of the python,i met the problem like ValueError: mutable default <class 'trainer.accelerators.base_accelerator.DebugConfig'> for field debug is not allowed: use default_factory

[yuvalkirstain]

Seems like the issue is hydra version. I'm not which version it was, but try to use an older one.

[learn01one]

hello,The code running environment is torch=2.2.0，python=3.11,Others are installed provided versions，Many similar questions arise，like ,ValueError: mutable default <class 'trainer.configs.configs.DebugConfig'> for field debug is not allowed: use default_factory,and ValueError: mutable default <class 'trainer.accelerators.deepspeed_accelerator.DeepSpeedConfig'> for field deepspeed is not allowed: use default_factory

It should be an environmental problem. Can you provide a more detailed operating environment? Thanks.

[yuvalkirstain]

I was running with python 3.8

[yuvalkirstain]

I don't have access to the env I ran with at the moment. If you continue to have this issue I'll help you debug it.

[learn01one]

hello，All the problems before using python=3.8 have been solved. The training seems to be almost successful， just encountered a small problem.，like

fp16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=True)
│ 23 │ bf16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=False)
│ 24 │ optimizer: dict = field(default_factory=lambda: {
│ 25 │ │ "type": "AdamW",
TypeError: MixedPrecisionConfig() takes no arguments

Are there any additional parameters required during operation?thanks

[yuvalkirstain]

Can you please show here the entire error trace? Also, what cmd you are running.

[learn01one]

ok，running it by: accelerate launch --dynamo_backend no --gpu_ids all --num_processes 8 --num_machines 1 --use_deepspeed trainer/scripts/train.py +experiment=clip_h output_dir=output

the entire error trace like: ────────────────────────────────╮ │ /PickScore/trainer/scripts/train.py:13 in │ │ │ │ 10 from torch import nn │ │ 11 import sys │ │ 12 │ │ ❱ 13 from trainer.accelerators.base_accelerator import BaseAccelerator │ │ 14 from trainer.configs.configs import TrainerConfig, instantiate_with_cfg │ │ 15 │ │ 16 import pdb │ │ │ │ /PickScore/trainer/accelerators/init.py:4 in │ │ │ │ 1 from hydra.core.config_store import ConfigStore │ │ 2 │ │ 3 from trainer.accelerators.debug_accelerator import DebugAcceleratorConfig │ │ ❱ 4 from trainer.accelerators.deepspeed_accelerator import DeepSpeedAcceleratorConfig │ │ 5 │ │ 6 ACCELERATOR_GROUP_NAME = "accelerator" │ │ 7 │ │ │ │ /PickScore/trainer/accelerators/deepspeed_accelerator.py:20 in │ │ │ │ │ │ 17 │ │ 18 │ │ 19 @dataclass │ │ ❱ 20 class DeepSpeedConfig: │ │ 21 │ #pdb.set_trace() │ │ 22 │ fp16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=True) │ │ 23 │ bf16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=False) │ │ │ │ /PickScore/trainer/accelerators/deepspeed_accelerator.py:22 in │ │ DeepSpeedConfig │ │ │ │ 19 @dataclass │ │ 20 class DeepSpeedConfig: │ │ 21 │ #pdb.set_trace() │ │ ❱ 22 │ fp16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=True) │ │ 23 │ bf16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=False) │ │ 24 │ optimizer: dict = field(default_factory=lambda: { │ │ 25 │ │ "type": "AdamW", │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────╯ TypeError: MixedPrecisionConfig() takes no arguments

[yuvalkirstain]

Really strange.. I am able to run this code with no problem:

>>> from omegaconf import OmegaConf, MISSING, II
>>>
>>> from dataclasses import dataclass, field
>>>
>>> @dataclass
... class MixedPrecisionConfig:
...     enabled: bool = MISSING
...
>>> @dataclass
... class DeepSpeedConfig:
...     fp16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=False)
...     bf16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=False)
...
>>>

With Python 3.10.13. Perhaps try to use 3.10?

[learn01one]

hell0,Using python=3.10, the same problem still occurs. Can you show your deepspeed default_config.yaml?thanks

[yuvalkirstain]

what does the deepspeed config has to do with it? Can you locally run the code that fails?

from omegaconf import OmegaConf, MISSING, II
from dataclasses import dataclass, field

@dataclass
class MixedPrecisionConfig:
    enabled: bool = MISSING

@dataclass
class DeepSpeedConfig:
    fp16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=False)
    bf16: MixedPrecisionConfig = MixedPrecisionConfig(enabled=False)