compress.py

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
import re

from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot

# Select model and load it.
MODEL_ID = "ArliAI/Llama-3.1-70B-ArliAI-RPMax-v1.3"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# Select calibration dataset.
DATASET_ID = "openerotica/erotiquant3"
DATASET_SPLIT = "train"

# Select number of samples. 512 samples is a good place to start.
# Increasing the number of samples can improve accuracy.
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 4096

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)

def preprocess(example):
    result = []
    matches = re.findall(r'(SYSTEM|USER|ASSISTANT):\s*((?:(?!SYSTEM|USER|ASSISTANT:).|\n)+)', example['text'], re.DOTALL)

    # Loop through the matches and create a dictionary for each role and its content
    for role, content in matches:
        result.append({"role": role.lower(), "content": content.strip()})

    text = tokenizer.apply_chat_template(result, tokenize=False, add_generation_prompt=False)
    tokens = tokenizer.apply_chat_template(result, tokenize=True, add_generation_prompt=False)

    return {
        "chat": result,
        "text": text,
        "tokens": tokens,
    }

ds = ds.map(preprocess)

def filter_short_rows(example):
    result = len(example['tokens']) > MAX_SEQUENCE_LENGTH
    if result == False:
        print(f"length: {len(example['tokens'])}")
    return result

ds = ds.filter(filter_short_rows)

ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))


# Tokenize inputs.
def tokenize(sample):
    return tokenizer(
        sample["text"],
        padding=False,
        max_length=MAX_SEQUENCE_LENGTH,
        truncation=True,
        add_special_tokens=False,
    )

ds = ds.map(tokenize, remove_columns=ds.column_names)

# Configure the quantization algorithm to run.
#   * quantize the weights to 4 bit with GPTQ with a group size 128
recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"])

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
    torch_dtype="auto",
)

# Apply algorithms.
oneshot(
    model=model,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES
)

print('SAVING')

# Save to disk compressed.
SAVE_DIR = MODEL_ID.split("/")[1] + "-W4A16-G128"
model.save_pretrained(SAVE_DIR, save_compressed=True, skip_compression_stats=True)
tokenizer.save_pretrained(SAVE_DIR)

print('Saved')