#!/usr/bin/env python3

# bash commands to run first:
    # pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com nvidia-ammo
    # pip install llmcompressor
    # MAX_JOBS=18 pip install flash-attn --no-build-isolation
    # apt update
    # apt install nano tmux -y
    # export HF_HOME=/workspace/huggingface
    # huggingface-cli download Qwen/Qwen2.5-72B-Instruct

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

from llmcompressor import oneshot

# Select model and load it.
MODEL_ID = "Qwen/Qwen2.5-72B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="cuda",
    attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# Select calibration dataset.
DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"

# Select number of samples. 512 samples is a good place to start.
# Increasing the number of samples can improve accuracy.
NUM_CALIBRATION_SAMPLES = 2048
MAX_SEQUENCE_LENGTH = 20000

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=DATASET_SPLIT)
ds = ds.shuffle(seed=42069).select(range(NUM_CALIBRATION_SAMPLES))


def process_and_tokenize(example):
    text = tokenizer.apply_chat_template(example["messages"], tokenize=False)
    return tokenizer(
        text,
        padding=False,
        max_length=MAX_SEQUENCE_LENGTH,
        truncation=True,
        add_special_tokens=False,
    )


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

# Configure the quantization algorithm and scheme.
# In this case, we:
#   * quantize the weights to fp8 with per-channel scales
#   * quantize the activations to fp8 with dynamic per-token scales
#   * quantize the kv cache to fp8 with per-tensor scales
recipe = """
quant_stage:
    quant_modifiers:
        QuantizationModifier:
            ignore: ["lm_head"]
            config_groups:
                group_0:
                    weights:
                        num_bits: 8
                        type: float
                        strategy: channel
                        dynamic: false
                        symmetric: true
                    input_activations:
                        num_bits: 8
                        type: float
                        strategy: token
                        dynamic: true
                        symmetric: true
                    targets: ["Linear"]
            kv_cache_scheme:
                num_bits: 8
                type: float
                strategy: tensor
                dynamic: false
                symmetric: true
"""

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

# Save to disk compressed.
SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

print(
    "Note: Inference with the quantized kv_cache is not supported. ",
    "Please use vLLM for inference with the quantized kv_cache.",
)
# Confirm generations of the quantized model look sane.

print("\n\n")
print("========== SAMPLE GENERATION ==============")
input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
output = model.generate(input_ids, max_new_tokens=100)
print(tokenizer.decode(output[0]))
print("==========================================\n\n")

# Save to disk compressed.
SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-KV-2"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)