opt-6.7b-lora

Running opt-6.7b with added loras locally on windows!

bitsandbytes

I needed to get bitsandbytes working in my venv: I replaced the main.py in C:\Users\user\Desktop\test\peft\venv\Lib\site-packages\bitsandbytes\cuda_setup\main.py with the one here! I also added a .dll file here: C:\Users\user\Desktop\test\peft\venv\Lib\site-packages\bitsandbytes\libbitsandbytes_cuda116.dll

Training Script

(https://github.com/huggingface/peft/commit/df0e1fb59266c9903ddd6dbfe7339bcd2068d150) (It's from their notebook!)

#load


import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import torch
import torch.nn as nn
import bitsandbytes as bnb
from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "facebook/opt-6.7b", 
    load_in_8bit=True, 
    device_map='auto',
)

tokenizer = AutoTokenizer.from_pretrained("facebook/opt-6.7b")


#post-processing

for param in model.parameters():
  param.requires_grad = False  # freeze the model - train adapters later
  if param.ndim == 1:
    # cast the small parameters (e.g. layernorm) to fp32 for stability
    param.data = param.data.to(torch.float32)

model.gradient_checkpointing_enable()  # reduce number of stored activations
model.enable_input_require_grads()

class CastOutputToFloat(nn.Sequential):
  def forward(self, x): return super().forward(x).to(torch.float32)
model.lm_head = CastOutputToFloat(model.lm_head)

# apply lora

def print_trainable_parameters(model):
    """
    Prints the number of trainable parameters in the model.
    """
    trainable_params = 0
    all_param = 0
    for _, param in model.named_parameters():
        all_param += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    print(
        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
    )

# apply lora 2

from peft import LoraConfig, get_peft_model 

config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, config)
print_trainable_parameters(model)

# training

import transformers
from datasets import load_dataset
data = load_dataset("Abirate/english_quotes")
data = data.map(lambda samples: tokenizer(samples['quote']), batched=True)

trainer = transformers.Trainer(
    model=model, 
    train_dataset=data['train'],
    args=transformers.TrainingArguments(
        per_device_train_batch_size=4, 
        gradient_accumulation_steps=4,
        warmup_steps=100, 
        max_steps=200, 
        learning_rate=2e-4, 
        fp16=True,
        logging_steps=1, 
        output_dir='outputs'
    ),
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False)
)
model.config.use_cache = False  # silence the warnings. Please re-enable for inference!
trainer.train()

# push to huggingface txtloras
model.push_to_hub("Yoshiii/opt-6.7b-lora", use_auth_token=True)


# inference

batch = tokenizer("Two things are infinite: ", return_tensors='pt')

with torch.cuda.amp.autocast():
  output_tokens = model.generate(**batch, max_new_tokens=50)

print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=True))

Inference (loading this repo lora from hf)

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_model_id = "Yoshiii/opt-6.7b-lora"
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, return_dict=True, load_in_8bit=True, device_map='auto')
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load the Lora model
model = PeftModel.from_pretrained(model, peft_model_id)


batch = tokenizer("Two things are infinite: ", return_tensors='pt')

with torch.cuda.amp.autocast():
  output_tokens = model.generate(**batch, max_new_tokens=50)

print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=True))

Two things are infinite:  the universe and human stupidity; and I'm not sure about the universe.  -Albert Einstein     I'm not sure about the universe either.

This output is like the training data. If you run without applying the Lora, it will usually look worse. If you retrain the lora, know that your new lora is not going to output the same results, despite you using the same settings. Inference should usually be deterministic when using the same lora, or using without lora.

Also, If you want to download and use the loras from a visible folder, here's the inference script:

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_model_id = "./loramodel"
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, return_dict=True, load_in_8bit=True, device_map='auto')
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load the Lora model
model = PeftModel.from_pretrained(model, peft_model_id)


batch = tokenizer("Two things are infinite: ", return_tensors='pt')

with torch.cuda.amp.autocast():
  output_tokens = model.generate(**batch, max_new_tokens=50)

print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=True))

add your adapter_config.json and your adapter_model.bin to a folder in your current directory named loramodel, or whatever you choose.