Can I finetune llama3.3 using axolotl?

[hahmad2008]

could you please provide any example for the config file to finetune llama3.3 on instruction dataset?
model: meta-llama/Llama-3.3-70B-Instruct

[NanoCode012]

Hey, you could use the existing llama3 configs and point to that new model https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/lora-8b.yml

[hahmad2008]

Thanks @NanoCode012 . can we finetune a quantized model? AWQ version? what the config need to be set?

[hahmad2008]

model config.yaml:

base_model: unsloth/Meta-Llama-3.1-70B-bnb-4bit
base_model_config: unsloth/Meta-Llama-3.1-70B-bnb-4bit
gptq: true
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
trust_remote_code: null
is_falcon_derived_model: null
is_llama_derived_model: null
is_mistral_derived_model: null
load_in_8bit: false
load_in_4bit: true
strict: false
push_dataset_to_hub: null
chat_template: null
datasets:
- path: mhenrichsen/alpaca_2k_test
  type: alpaca
dataset_prepared_path: prepared-dataset
val_set_size: 0.01
adapter: qlora
lora_model_dir: null
sequence_len: 512
sample_packing: true
eval_sample_packing: false
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules: null
lora_target_linear: true
lora_fan_in_fan_out: null
wandb_project: null
wandb_entity: null
wandb_watch: null
wandb_run_id: null
wandb_log_model: null
output_dir: model-finetuned
gradient_accumulation_steps: 1
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_bnb_8bit
torchdistx_path: null
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: null
resume_from_checkpoint: null
local_rank: null
logging_steps: 1
xformers_attention: null
flash_attention: true
gptq_groupsize: null
gptq_model_v1: null
warmup_steps: 10
eval_steps: 100
save_steps: 100
debug: null
deepspeed: null
weight_decay: 0.0
fsdp: null
fsdp_config: null
special_tokens:
  bos_token: <s>
  eos_token: </s>
  unk_token: <unk>
tokens: null

[hahmad2008]

@NanoCode012 when I change it to lora I get this error:

ValueError: model_config.quantization_config is not set or quant_method is not set to gptq. Please make sure to point to a GPTQ model.

[NanoCode012]

Please set load_in_4bit: true, turn off gptq. Use adapter: qlora. I thought we removed that validation check.

[lynkz-matt-psaltis]

Setting load in 4bit in the config is ignored by:

axolotl/src/axolotl/cli/merge_lora.py

Line 72 in a98526e

load_in_4bit=False,

Does anyone have a solid example of merging awq base with qlora?

[NanoCode012]

@lynkz-matt-psaltis , sorry for late reply. AWQ is a different thing. I'm not aware it can be tuned on, only loaded for inference.