How to speedup the process of quantization which takes almost 20 hours to quantize llama3-70B with w8a8

[moonlightian]

llmcompressor==0.2.0
compressed-tensors==0.7.1
4 NVIDIA A100 80GB PCIe

from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_ID = "/llama3_1_70B"
model = SparseAutoModelForCausalLM.from_pretrained(
MODEL_ID, device_map="auto", torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048
ds = ds.shuffle(seed=42).select(range(NUM_CALIBRATION_SAMPLES))

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

def tokenize(sample):
    return tokenizer("\n\n".join(sample['text']), return_tensors='pt')
ds = ds.map(tokenize, remove_columns=ds.column_names)

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

# Configure the quantization algorithms
recipe = [
    SmoothQuantModifier(smoothing_strength=0.8),
    GPTQModifier(targets="Linear", dampening_frac=0.1, scheme="W8A8", ignore=["lm_head"]),
]

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

[dsikka]

Hi @moonlightian:

1. Can you confirm you ran the above with 4 x A100s?
2. Can you also confirm the stub of the model that was used and the dataset name? I had some trouble processing the dataset listed with the preprocess functions provided.