Trouble resuming from checkpoint

[asaluja]

Hello,

I'm running bitsandbytes==0.41.1 in a Python 3.10 miniconda environment, 8xA100 GPU (using accelerate for multi-GPU), Cuda 12.2.

I'm having problems resuming training (DPO) from a checkpoint:

Error invalid argument at line 393 in file /mmfs1/gscratch/zlab/timdettmers/git/bitsandbytes/csrc/pythonInterface.c
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 3777589) of binary: /mnt/miniconda3/envs/synlm/bin/python3.10
Traceback (most recent call last):
  File "/mnt/miniconda3/envs/synlm/bin/accelerate", line 8, in <module>
    sys.exit(main())
  File "/mnt/miniconda3/envs/synlm/lib/python3.10/site-packages/accelerate/commands/accelerate_cli.py", line 45, in main
    args.func(args)
  File "/mnt/miniconda3/envs/synlm/lib/python3.10/site-packages/accelerate/commands/launch.py", line 970, in launch_command
    multi_gpu_launcher(args)
  File "/mnt/miniconda3/envs/synlm/lib/python3.10/site-packages/accelerate/commands/launch.py", line 646, in multi_gpu_launcher
    distrib_run.run(args)
  File "/mnt/miniconda3/envs/synlm/lib/python3.10/site-packages/torch/distributed/run.py", line 785, in run
    elastic_launch(
  File "/mnt/miniconda3/envs/synlm/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 134, in __call__
    return launch_agent(self._config, self._entrypoint, list(args))
  File "/mnt/miniconda3/envs/synlm/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
    raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

I'm not quite sure how to debug this error, any ideas?

[hiyouga]

same error for bitsandbytes==0.41.1 in 8 * A100 GPUs

update: this problem was caused by the paged_adamw_32bit optimizer. I solved this problem by using the adamw_torch optimizer and processing the optimizer.pt file using the following script.

import os
import torch

if __name__ == "__main__":
    os.rename("optimizer.pt", "optimizer.paged.pt")
    state_dict = torch.load("optimizer.paged.pt")
    for i in state_dict["state"].keys():
        exp_avg = state_dict["state"][i].pop("state1")
        exp_avg_sq = state_dict["state"][i].pop("state2")
        state_dict["state"][i]["exp_avg"] = exp_avg
        state_dict["state"][i]["exp_avg_sq"] = exp_avg_sq

    for i in range(len(state_dict["param_groups"])):
        state_dict["param_groups"][i]["amsgrad"] = False
        state_dict["param_groups"][i]["foreach"] = None
        state_dict["param_groups"][i]["maximize"] = False
        state_dict["param_groups"][i]["capturable"] = False
        state_dict["param_groups"][i]["differentiable"] = False
        state_dict["param_groups"][i]["fused"] = None

    torch.save(state_dict, "optimizer.pt")

[fmmoret]

I'm getting this issue too on T4 and L4. I'm going to dig into it more, but do you have more insight into it? Is the state dict serialization wrong after a change? Or the state dict loading?

[markrmiller]

I'm seeing this same issue on resume with paged_adamw_8bit.

[Andcircle]

@markrmiller I'm facing the same issue with paged_adamw_8bit, did you find any solution for it? thanks 🙏

[markrmiller]

Only to start training over with standard Adam. Later I used adam_paged_32bit and hit the same thing - so now I'm just avoiding all these optimizers.

[Andcircle]

@markrmiller thanks for advice

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

[k21993]

I am facing the same issue on a 8xA100 machine with bitsandbytes==0.42.0. I am using the paged_adamw_32bit optimizer. Did any of you find a solution? Please help, thank you 🙏

[Syzygianinfern0]

I am facing the same issue on a 8xA100 machine with bitsandbytes==0.42.0. I am using the paged_adamw_32bit optimizer. Did any of you find a solution? Please help, thank you 🙏

I have the exact same issue

[Kyeongpil]

I got a similar error when trying to save optimizer states (PagedAdam8Bit)

Traceback (most recent call last):
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/home/ubuntu/projects/scripts/train_ranker.py", line 404, in <module>
    train(accelerator, args)
  File "/home/ubuntu/projects/scripts/train_ranker.py", line 357, in train
    save_training_artifacts(
  File "/home/ubuntu/projects/scripts/train_ranker.py", line 74, in save_training_artifacts
    accelerator.save_state(save_dir)
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/accelerate/accelerator.py", line 2958, in save_state
    save_fsdp_optimizer(self.state.fsdp_plugin, self, opt, self._models[i], output_dir, i)
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/accelerate/utils/fsdp_utils.py", line 168, in save_fsdp_optimizer
    optim_state = FSDP.optim_state_dict(model, optimizer)
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py", line 1840, in optim_state_dict
    return FullyShardedDataParallel._optim_state_dict_impl(
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py", line 1263, in _optim_state_dict_impl
    return _optim_state_dict(
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/distributed/fsdp/_optim_utils.py", line 1971, in _optim_state_dict
    fsdp_osd_state = convert_fn(
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/distributed/fsdp/_optim_utils.py", line 1794, in _convert_state_with_orig_params
    _gather_all_orig_param_state(
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/distributed/fsdp/_optim_utils.py", line 1688, in _gather_all_orig_param_state
    output_states = _allgather_orig_param_states(
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/distributed/fsdp/_optim_utils.py", line 1518, in _allgather_orig_param_states
    dtype, state_buffers = _convert_all_state_info(
  File "/home/ubuntu/miniconda3/envs/python-3.10/lib/python3.10/site-packages/torch/distributed/fsdp/_optim_utils.py", line 1377, in _convert_all_state_info
    assert dtype == info.dtype
AssertionError

[geeky-programer]

I am facing the same issue. Is there any solution other than changing optimizers ?

[nighting0le01]

@Kyeongpil were you able to solve this issue? i get the same issue when using fsdp

[Kyeongpil]

@nighting0le01 I apologize for being unable to access my code due to changing my job.

[ExcitingFrog]

same error for bitsandbytes==0.41.1 in 8 * A100 GPUs

update: this problem was caused by the paged_adamw_32bit optimizer. I solved this problem by using the adamw_torch optimizer and processing the optimizer.pt file using the following script.

import os
import torch

if __name__ == "__main__":
    os.rename("optimizer.pt", "optimizer.paged.pt")
    state_dict = torch.load("optimizer.paged.pt")
    for i in state_dict["state"].keys():
        exp_avg = state_dict["state"][i].pop("state1")
        exp_avg_sq = state_dict["state"][i].pop("state2")
        state_dict["state"][i]["exp_avg"] = exp_avg
        state_dict["state"][i]["exp_avg_sq"] = exp_avg_sq

    for i in range(len(state_dict["param_groups"])):
        state_dict["param_groups"][i]["amsgrad"] = False
        state_dict["param_groups"][i]["foreach"] = None
        state_dict["param_groups"][i]["maximize"] = False
        state_dict["param_groups"][i]["capturable"] = False
        state_dict["param_groups"][i]["differentiable"] = False
        state_dict["param_groups"][i]["fused"] = None

    torch.save(state_dict, "optimizer.pt")

fyi i have same err when resuming from checkpoint, use this flow can't fix my problem. 8*a6000, other training task running on slurm and use half gpus , it may cause this err. if anyone have this err, you can try to reboot or stop slurm task

[cryotheta]

Is there a fix? Facing the same issue with 4*A100. Using accelerate to run multigpu training