Training loop for LoRA

[khatrimann]

Hey all, I wanted to ask if there is any way to apply LoRA on model and then train model in training loop without

SFTTrainer

I am Using the following way, is it fine?

model_name = "meta-llama/Meta-Llama-3.1-8B-Instruct"  # or any other model
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    # load_in_8bit=True,
    # quantization_config=quant_config,
    device_map="auto",
    max_memory={0: "4GiB", 1: "50GiB", "cpu": "0.0GiB"},
    offload_state_dict=True,
    low_cpu_mem_usage=True
)

...

model.config.use_cache = False
model.config.pretraining_tp = 1

model.gradient_checkpointing_enable()
model = prepare_model_for_kbit_training(model)

peft_params = LoraConfig(
    r=32,
    lora_alpha=64,
    target_modules=[
        "q_proj",
        "v_proj",
    ],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, peft_params)
model.print_trainable_parameters()
        
optimizer = Adam(model.parameters(), lr=1e-5)

for epoch_idx in [1]:
    for batch_idx, batch in tqdm(enumerate(train_data), total=len(train_data)):
        
        optimizer.zero_grad()
        outputs = model(**inputs)
        ...

And then finally using

merge_and_unload()

to merge adapter back to the model

[BenjaminB]

Yes, PEFT is not dependent on the usage of Trainer or SFTTrainer. Here is just one example of using LoRA with a custom training loop:

huggingface/peft/blob/main/examples/sequence_classification/LoRA.ipynb

Depending on what you would like to do, custom training loops can get a bit complicated (think: mixed precision, gradient accumulation, etc.) but it’s absolutely possible.

[khatrimann]

Thanks mate. I hope they mention this somewhere the left or Lora page