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

# Select model and load it.
MODEL_ID = "/mnt/llms/models/zai-org/GLM-4.6"
SAVE_DIR = MODEL_ID.split("/")[-1] + "-W4A16-awq"

# Configure the quantization algorithm to run.
# GLM-4.6 is a MoE model with:
# - Layers 0-2: Dense layers
# - Layers 3-92: MoE layers (160 routed experts + 1 shared expert per layer)
# We need to ignore the MoE gate layers to preserve routing quality
recipe = [
    AWQModifier(
        ignore=["lm_head", "re:.*mlp\\.gate$"],  # Ignore output head and MoE gates
        scheme="W4A16",
        targets=["Linear"],
    ),
]

# Select calibration dataset.
# Using a chat dataset since GLM-4.6 is an instruct model
DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"

# Select number of samples. 256 samples is a good place to start.
# Increasing the number of samples can improve accuracy.
NUM_CALIBRATION_SAMPLES = 256
MAX_SEQUENCE_LENGTH = 2048


def get_calib_dataset(tokenizer):
    """Prepare calibration dataset for GLM-4.6"""
    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)
    return ds


if __name__ == "__main__":
    # Load model and tokenizer
    print(f"Loading model: {MODEL_ID}")
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        torch_dtype="auto",
        trust_remote_code=True
    )
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)

    # Apply AWQ quantization
    print("Starting AWQ quantization...")
    oneshot(
        model=model,
        dataset=get_calib_dataset(tokenizer),
        recipe=recipe,
        max_seq_length=MAX_SEQUENCE_LENGTH,
        num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    )

    # Confirm generations of the quantized model look sane.
    print("\n\n")
    print("========== SAMPLE GENERATION ==============")
    dispatch_for_generation(model)
    input_ids = tokenizer("Write a Python function to calculate fibonacci numbers", return_tensors="pt").input_ids.to(
        model.device
    )
    output = model.generate(input_ids, max_new_tokens=150)
    print(tokenizer.decode(output[0]))
    print("==========================================\n\n")

    # Save to disk compressed.
    print(f"Saving quantized model to: {SAVE_DIR}")
    model.save_pretrained(SAVE_DIR, save_compressed=True)
    tokenizer.save_pretrained(SAVE_DIR)
    print("Done!")