ValueError: Expected a cuda device, but got: cpu when using Deepspeed zero3

[l3utterfly]

Please check that this issue hasn't been reported before.

• I searched previous Bug Reports didn't find any similar reports.

Expected Behavior

Works with DP3 out of the box

Current behaviour

Got this error:

[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/fire/core.py", line 477, in _Fire                                                                                       [45/1935]
[rank5]:     component, remaining_args = _CallAndUpdateTrace(
[rank5]:                                 ^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/fire/core.py", line 693, in _CallAndUpdateTrace
[rank5]:     component = fn(*varargs, **kwargs)
[rank5]:                 ^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/src/axolotl/src/axolotl/cli/train.py", line 38, in do_cli
[rank5]:     return do_train(parsed_cfg, parsed_cli_args)
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/src/axolotl/src/axolotl/cli/train.py", line 66, in do_train
[rank5]:     return train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/src/axolotl/src/axolotl/train.py", line 170, in train
[rank5]:     trainer.train(resume_from_checkpoint=resume_from_checkpoint)
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/transformers/trainer.py", line 1885, in train
[rank5]:     return inner_training_loop(
[rank5]:            ^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/transformers/trainer.py", line 2216, in _inner_training_loop
[rank5]:     tr_loss_step = self.training_step(model, inputs)
[rank5]:                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/transformers/trainer.py", line 3250, in training_step
[rank5]:     self.accelerator.backward(loss)
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/accelerate/accelerator.py", line 2117, in backward
[rank5]:     self.deepspeed_engine_wrapped.backward(loss, **kwargs)
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/accelerate/utils/deepspeed.py", line 175, in backward
[rank5]:     self.engine.step()
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/deepspeed/runtime/engine.py", line 2169, in step
[rank5]:     self._take_model_step(lr_kwargs)
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/deepspeed/runtime/engine.py", line 2075, in _take_model_step
[rank5]:     self.optimizer.step()
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/deepspeed/utils/nvtx.py", line 15, in wrapped_fn
[rank5]:     ret_val = func(*args, **kwargs)
[rank5]:               ^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/deepspeed/runtime/zero/stage3.py", line 2050, in step
[rank5]:     self._optimizer_step(sub_group_id)
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/deepspeed/runtime/zero/stage3.py", line 939, in _optimizer_step
[rank5]:     cpu_loss = self.optimizer.step()
[rank5]:                ^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/src/pytorch/torch/optim/lr_scheduler.py", line 129, in wrapper
[rank5]:     return func.__get__(opt, opt.__class__)(*args, **kwargs)
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/src/pytorch/torch/optim/optimizer.py", line 483, in wrapper
[rank5]:     out = func(*args, **kwargs)
[rank5]:           ^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/src/pytorch/torch/utils/_contextlib.py", line 115, in decorate_context
[rank5]:     return func(*args, **kwargs)
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/bitsandbytes/optim/optimizer.py", line 287, in step
[rank5]:     self.update_step(group, p, gindex, pindex)
[rank5]:   File "/home/layla/src/pytorch/torch/utils/_contextlib.py", line 115, in decorate_context
[rank5]:     return func(*args, **kwargs)
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/bitsandbytes/optim/optimizer.py", line 542, in update_step
[rank5]:     F.optimizer_update_8bit_blockwise(
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/bitsandbytes/functional.py", line 1770, in optimizer_update_8bit_blockwise
[rank5]:     prev_device = pre_call(g.device)
[rank5]:                   ^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/miniconda3/envs/axolotl/lib/python3.11/site-packages/bitsandbytes/functional.py", line 459, in pre_call
[rank5]:     torch.cuda.set_device(device)
[rank5]:   File "/home/layla/src/pytorch/torch/cuda/__init__.py", line 414, in set_device
[rank5]:     device = _get_device_index(device)
[rank5]:              ^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/home/layla/src/pytorch/torch/cuda/_utils.py", line 34, in _get_device_index
[rank5]:     raise ValueError(f"Expected a cuda device, but got: {device}")
[rank5]: ValueError: Expected a cuda device, but got: cpu

Steps to reproduce

1. Use deepspeed zero3 config in this repo
2. installed fixed deepspeed: pip install "deepspeed @ git+https://github.com/microsoft/DeepSpeed.git@bc48371c5e1fb8fd70fc79285e66201dbb65679b"

Config yaml

base_model: models/Meta-Llama-3-8B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: train.jsonl
    ds_type: json # see other options below
    type: sharegpt
    conversation: chatml
    roles:
      input: ['User', 'Information']
      output: ['Layla']

chat_template: chatml
default_system_message: The following is a conversation. Embody the character and personality completely.

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/out

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-7

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
auto_resume_from_checkpoints: false
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 5
evals_per_epoch: 10
eval_table_size:
saves_per_epoch: 10
debug:
deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_all.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>
tokens: # these are delimiters
  - "<|im_start|>"
  - "<|im_end|>"

Possible solution

No response

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.11

axolotl branch-commit

main

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this bug has not been reported yet.
• I am using the latest version of axolotl.
• I have provided enough information for the maintainers to reproduce and diagnose the issue.