"RuntimeError: Invalid device argument : did you call init? "When setting CUDA_VISIBLE_DEVICES

[zhanghanxing2022]

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

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

Expected Behavior

Generally, there is no need to actively torch.cuda.init()

Current behaviour

When CUDA_VISIBLE_DEVICES is set, a Runtime Error occurs: Invalid device argument : did you call init?

When I manually initialize before '`component_trace = _Fire(component, args, parsed_flag_args, context, name)', the problem still persists.

When CUDA_VISIBLE_DEVICES is not specified, this issue will not occur.

Steps to reproduce

CUDA_VISIBLE_DEVICES=2,3 accelerate launch -m axolotl.cli.train Mistral-7B-Instruct-v0.3.yml

Config yaml

base_model: ./mistralai
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer

strict: false

chat_template: llama3

datasets:
  - path: /tmp/5014fa1bc5b682bb_train_data.json
    type: 
      # 请根据需要调整这些字段名称以匹配实际的JSON字段
      field_instruction: question_statement  # 显式指定指令字段
      field_output: text            # 显式指定输出字段

      # 定义格式
      format: |-
        User: {question_statement} 
        Assistant: {text}
dataset_prepared_path:
val_set_size: 0.05
output_dir: id_24

sequence_len: 4056
sample_packing: false
pad_to_sequence_len: true
# trust_remote_code: true
# [2024-12-15 16:08:07,077] [WARNING] [axolotl.utils.config.models.input.hint_trust_remote_code:403] [PID:1119962] [RANK:1] `trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model.
adapter: lora
# # If you already have a lora model trained that you want to load, put that here.
# lora_model_dir: ./mistralai-lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out: false

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: false
fp16: true
tf32: false

gradient_checkpointing: false
# Stop training after this many evaluation losses have increased in a row
# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
early_stopping_patience: 
# Resume from a specific checkpoint dir
# resume_from_checkpoint: ./mistralai-resume_from_checkpoint
# Don't mess with this, it's here for accelerate and torchrun
local_rank:
logging_steps: 1
# xformers_attention:
flash_attention: true
# s2_attention:

wandb_project: Gradients-On-Demand
wandb_entity:
wandb_mode: offline
wandb_run: your_name
wandb_runid: default

# hub_model_id:
# hub_repo:
# hub_strategy: checkpoint
# hub_token: false


saves_per_epoch: 4
warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
max_steps: 10
debug:
deepspeed: /data/deepspeed_configs/zero3.json
weight_decay: 0.0
fsdp:
fsdp_config:

Possible solution

No response

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

Python 3.11.11

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.

[NanoCode012]

Which GPUs are you using? I just used the CUDA_VISIBLE_DEVICES yesterday, and it seemed to not have this issue.

[zhanghanxing2022]

H200