LoRA not working with accelerate + bfloat16 (without load_in_8_bit)

[Palmik]

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

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

Expected Behavior

I should be able to use load_in_8bit: False + load_in_4bit: False + adapter: lora.
This issue was already observed here: #456 but closed without resolution.

Current behaviour

When you try this today, the process gets stuck, e.g.:

[2023-10-27 11:00:34,160] [DEBUG] [axolotl.load_tokenizer:75] [PID:11028] [RANK:0] EOS: 2 / </s>
[2023-10-27 11:00:34,161] [DEBUG] [axolotl.load_tokenizer:76] [PID:11028] [RANK:0] BOS: 1 / <s>
[2023-10-27 11:00:34,161] [DEBUG] [axolotl.load_tokenizer:77] [PID:11028] [RANK:0] PAD: 2 / </s>
[2023-10-27 11:00:34,161] [DEBUG] [axolotl.load_tokenizer:78] [PID:11028] [RANK:0] UNK: 0 / <unk>
[2023-10-27 11:00:34,434] [INFO] [axolotl.load_tokenized_prepared_datasets:129] [PID:11028] [RANK:0] Loading prepared dataset from disk at last_run_prepared/2d315456dea8cf361f2683ece0705571...
[2023-10-27 11:00:34,437] [INFO] [axolotl.load_tokenized_prepared_datasets:131] [PID:11028] [RANK:0] Prepared dataset loaded from disk...
[2023-10-27 11:00:35,053] [INFO] [axolotl.calculate_total_num_steps:245] [PID:11028] [RANK:0] total_num_steps: 149
[2023-10-27 11:00:35,061] [INFO] [axolotl.train.train:46] [PID:11028] [RANK:0] loading tokenizer... mistralai/Mistral-7B-v0.1
[2023-10-27 11:00:35,196] [DEBUG] [axolotl.load_tokenizer:75] [PID:11028] [RANK:0] EOS: 2 / </s>
[2023-10-27 11:00:35,196] [DEBUG] [axolotl.load_tokenizer:76] [PID:11028] [RANK:0] BOS: 1 / <s>
[2023-10-27 11:00:35,198] [DEBUG] [axolotl.load_tokenizer:77] [PID:11028] [RANK:0] PAD: 2 / </s>
[2023-10-27 11:00:35,198] [DEBUG] [axolotl.load_tokenizer:78] [PID:11028] [RANK:0] UNK: 0 / <unk>
[2023-10-27 11:00:35,473] [INFO] [axolotl.train.train:54] [PID:11028] [RANK:0] loading model and (optionally) peft_config...
Loading checkpoint shards: 100%|________________________________________________________________________________________________________________________________________________________________________________________________________________________________________| 2/2 [00:25<00:00, 12.58s/it]
[2023-10-27 11:03:11,241] [INFO] [axolotl.load_model:438] [PID:11028] [RANK:0] converting modules to torch.bfloat16 for flash attention
[2023-10-27 11:03:11,850] [INFO] [axolotl.load_lora:547] [PID:11028] [RANK:0] found linear modules: ['q_proj', 'o_proj', 'v_proj', 'k_proj', 'down_proj', 'gate_proj', 'up_proj']
trainable params: 167,772,160 || all params: 7,409,504,256 || trainable%: 2.264283199029719
[2023-10-27 11:03:17,041] [INFO] [axolotl.load_model:474] [PID:11028] [RANK:0] GPU memory usage after adapters: 0.000GB ()

(after this, nothing happens)

Steps to reproduce

.

Config yaml

base_model: mistralai/Mistral-7B-v0.1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:

• path: ...
  type: completion
  dataset_prepared_path: last_run_prepared
  val_set_size: 0
  output_dir: ./mistral-7b-lora-out

adapter: lora
lora_model_dir:

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:

• gate_proj
• down_proj
• up_proj
• q_proj
• v_proj
• k_proj
• o_proj

wandb_project:
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

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

warmup_steps: 10

eval_steps: 20

eval_table_size:

eval_table_max_new_tokens: 128

save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
bos_token: ""
eos_token: ""
unk_token: ""

Possible solution

No response

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.10.13

axolotl branch-commit

latest docker

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.

[winglian]

This isn't a supported configuration. We only officially support LoRA with 8bit and qLoRA with 4bit. if you wish to submit a fix to enable 16 bit lora fine-tuning, we would definitely welcome that PR.