README.md

import time
from datasets import load_dataset
from transformers import AutoTokenizer
from rich import print
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot

# Select model and load it.
MODEL_ID = "Unbabel/TowerInstruct-7B-v0.1"
model = SparseAutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
    torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# Select calibration dataset.
DATASET_ID = "neuralmagic/LLM_compression_calibration"
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 = 756
MAX_SEQUENCE_LENGTH = 2048

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))


def preprocess(example):
    return {
        "text": tokenizer.apply_chat_template(
            example["messages"],
            tokenize=False,
        )
    }


ds = ds.map(preprocess)


# 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
# Note: to reduce GPU memory use `sequential_update=False`
recipe = GPTQModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"])
print(recipe)

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

# Confirm generations of the quantized model look sane and measure generation time.
print("\n\n")
print("========== SAMPLE GENERATION ==============")
input_text = "Translate the following text from Portuguese into English.\nPortuguese: Um grupo de investigadores lançou um novo modelo para tarefas relacionadas com tradução.\nEnglish:"
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")
start_time = time.time()
output = model.generate(input_ids, max_new_tokens=100)
end_time = time.time()
generation_time = end_time - start_time
print(tokenizer.decode(output[0]))
print(f"Generation time: {generation_time:.2f} seconds")
print("==========================================\n\n")

# Save to disk compressed.
SAVE_DIR = MODEL_ID.split("/")[1] + "-quantized.w4a16"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)