Deepspeed zero3 + LoRA: RuntimeError: Only Tensors of floating point and complex dtype can require gradients

[bursteratom]

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

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

Expected Behavior

I expect deepspeed zero3 and 8-bit LoRA to be compatible and runs without error

Current behaviour

When loading model with deepspeed zero3 and 8-bit LoRA enabled, I ran into the error RuntimeError: Only Tensors of floating point and complex dtype can require gradients :

However, if you use zero3 in tandem with 4-bit qLoRA, or just do full fine-tuning with zero3 enabled, it works fine.

Steps to reproduce

1. Set up LoRA
2. enable deepspeed zero3
3. RuntimeError: Only Tensors of floating point and complex dtype can require gradients

Config yaml

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

load_in_8bit: true
load_in_4bit: false
strict: false

datasets:
  - path: mhenrichsen/alpaca_2k_test
    type: alpaca
dataset_prepared_path:
val_set_size: 0
output_dir: ./outputs/lora-out

sequence_len: 4096
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_modules_to_save:
  - embed_tokens
  - lm_head

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

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

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
   pad_token: <|end_of_text|>

Possible solution

(Putting on a tinfoil hat) I think it's a bug within axolotl code base as opposed to some deeper issue with deepspeed zero3, seeing as it works with qlora.

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.11.10

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.

[bursteratom]

this is being worked on via PR#1852 and huggingface/transformers#32943