Llama-2-7B.w4a16

Creation

import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor import oneshot
from llmcompressor.modifiers.awq import AWQModifier
from llmcompressor.utils import dispatch_for_generation

MODEL_ID = "meta-llama/Llama-2-7b-hf"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "meta-llama/Llama-2-7b-chat-hf", trust_remote_code=True
)

DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"
NUM_CALIBRATION_SAMPLES = 256
MAX_SEQUENCE_LENGTH = 2048
ds = load_dataset(DATASET_ID, split=f"{DATASET_SPLIT}[:{NUM_CALIBRATION_SAMPLES}]")
ds = ds.shuffle(seed=42)


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

ds = ds.map(preprocess)

recipe = [
    AWQModifier(ignore=["lm_head"], scheme="W4A16_ASYM", targets=["Linear"]),
]

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

print("\n\n========== SAMPLE GENERATION ==============")
dispatch_for_generation(model)
input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(
    model.device
)
output = model.generate(input_ids, max_new_tokens=50)
print(tokenizer.decode(output[0]))
print("==========================================\n\n")

SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-awq-asym"

for name, param in model.named_parameters():
    if not param.is_contiguous():
        model._parameters[name] = param.contiguous()
try:
    model.save_pretrained(SAVE_DIR, save_compressed=True)
except ValueError:
    print("Non-contiguous tensor issue detected, saving with safe_serialization=False")
    model.save_pretrained(SAVE_DIR, safe_serialization=False)
tokenizer.save_pretrained(SAVE_DIR)