import json
import re
from huggingface_hub.hf_api import ModelInfo, get_safetensors_metadata, model_info as get_model_info, get_hf_file_metadata, hf_hub_url
from huggingface_hub import hf_hub_download

# Map model IDs to the number of bytes used for one parameter. So, 4 bytes for fp32, 2 bytes for fp16, etc.
# By default, we assume that the model is stored in fp32.
KNOWN_BYTES_PER_PARAM = {}


def get_model_size(model_info: ModelInfo):
    '''Get the size of the model in million of parameters.'''
    try:
        safetensors = get_safetensors_metadata(model_info.id)
        return round(sum(safetensors.parameter_count.values()) / 1e6)
    except Exception as e:
        pass

    filenames = [sib.rfilename for sib in model_info.siblings]
    if "pytorch_model.bin" in filenames:
        url = hf_hub_url(model_info.id, filename="pytorch_model.bin")
        meta = get_hf_file_metadata(url)
        bytes_per_param = KNOWN_BYTES_PER_PARAM.get(model_info.id, 4)
        return round(meta.size / bytes_per_param / 1e6)
    
    if "pytorch_model.bin.index.json" in filenames:
        index_path = hf_hub_download(model_info.id, filename="pytorch_model.bin.index.json")
        """
        {
        "metadata": {
            "total_size": 28272820224
        },....
        """
        size = json.load(open(index_path))
        bytes_per_param = KNOWN_BYTES_PER_PARAM.get(model_info.id, 4)
        if ("metadata" in size) and ("total_size" in size["metadata"]):
            return round(size["metadata"]["total_size"] / bytes_per_param / 1e6)

    return None