Can't load BnB models

[Blaizzy]

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

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

Expected Behavior

I want to load a BnB quantized model.

Current behaviour

It throws a ValueError.

Steps to reproduce

Launch the config yaml.

Config yaml

base_model: unsloth/tinyllama-bnb-4bit
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: HuggingFaceH4/ultrachat_200k
    split: train_sft
    type: sharegpt
    conversation: chatml

dataset_prepared_path:
val_set_size: 0.05
output_dir: ./qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 1096
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name: tinyllama
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 1
max_steps: 20
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002

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

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

warmup_steps: 10
evals_per_epoch:
saves_per_epoch:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"

tokens:
- "<|im_start|>"

Possible solution

Extend or remove the fixed check of gptq introduced here: https://github.com/OpenAccess-AI-Collective/axolotl/pull/913

Which Operating Systems are you using?

• [X] Linux
• [ ] macOS
• [ ] Windows

Python Version

3.10

axolotl branch-commit

main

Acknowledgements

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

[NanoCode012]

More details regarding error please. Were you also the one who posted a bnb issue on discord?

[tsunayoshi21]

Any new update regarding this error? I have a similar issue

[Blaizzy]

@NanoCode012

Could you let me know what else are you looking for?

[NanoCode012]

Could someone post logs of the issue? Is it due to the check of quant_config?

[Blaizzy]

Ayt, got it!

I will post the logs later today

[tsunayoshi21]

@NanoCode012 Yes, for me the error is that the check of quant_config always raises a error because the quant_method is not gptq, and if i set gptq:false in the yaml, it raises a error that says i can't load a quantized model without gptq.

So if my model is previously BnB quantized i have no clue of how i can finetune with axolotl

[FrederikHandberg]

@Blaizzy what was your fix?

[Blaizzy]

@Blaizzy what was your fix?

I used a full precision model and set load_in_4bit: to true

Example:

base_model: meta/llama-7b-hf
load_in_4bit: true

Whilst, I actually wanted to load a prequantized model.

base_model: meta/llama-7b-hf-4bit

[FrederikHandberg]

Thanks

+1 id like to do the same (would be a nice addition)

[NanoCode012]

Hey, sorry for taking so long to get back. To follow up on this. I re-used the author's config with some modification (remove wandb, change dataset loading, use bf16, save embed_tokens lm_head), and it runs for me.

base_model: unsloth/tinyllama-bnb-4bit
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

chat_template: chatml
datasets:
  - path: HuggingFaceH4/ultrachat_200k
    split: train_sft
    type: chat_template

    field_messages: messages
    message_field_role: role
    message_field_content: content


dataset_prepared_path:
val_set_size: 0.05
output_dir: ./qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 1096
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:
lora_modules_to_save: ['embed_tokens', 'lm_head']

# wandb_project: axolotl
# wandb_entity:
# wandb_watch:
# wandb_name: tinyllama
# wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 1
max_steps: 20
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002

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

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

warmup_steps: 10
evals_per_epoch:
saves_per_epoch:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"

tokens:
- "<|im_start|>"

Would anyone of you be able to re-test and confirm it now works?