load_models.py

import torch

import logging

from auto_gptq import AutoGPTQForCausalLM
from huggingface_hub import hf_hub_download
from langchain.llms import LlamaCpp, HuggingFacePipeline

from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler

from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    LlamaForCausalLM,
    LlamaTokenizer,
    GenerationConfig,
    pipeline,
)


torch.set_grad_enabled(False)


from constants import CONTEXT_WINDOW_SIZE, MAX_NEW_TOKENS, N_GPU_LAYERS, N_BATCH, MODELS_PATH


def load_quantized_model_gguf_ggml(model_id, model_basename, device_type, logging, stream = False, callbacks = []):
    """
    Load a GGUF/GGML quantized model using LlamaCpp.

    This function attempts to load a GGUF/GGML quantized model using the LlamaCpp library.
    If the model is of type GGML, and newer version of LLAMA-CPP is used which does not support GGML,
    it logs a message indicating that LLAMA-CPP has dropped support for GGML.

    Parameters:
    - model_id (str): The identifier for the model on HuggingFace Hub.
    - model_basename (str): The base name of the model file.
    - device_type (str): The type of device where the model will run, e.g., 'mps', 'cuda', etc.
    - logging (logging.Logger): Logger instance for logging messages.

    Returns:
    - LlamaCpp: An instance of the LlamaCpp model if successful, otherwise None.

    Notes:
    - The function uses the `hf_hub_download` function to download the model from the HuggingFace Hub.
    - The number of GPU layers is set based on the device type.
    """

    try:
        logging.info("Using Llamacpp for GGUF/GGML quantized models")
        model_path = hf_hub_download(
            repo_id=model_id,
            filename=model_basename,
            resume_download=True,
            cache_dir=MODELS_PATH,
        )
        kwargs = {
            "model_path": model_path,
            "n_ctx": CONTEXT_WINDOW_SIZE,
            "max_tokens": MAX_NEW_TOKENS,
             # set this based on your GPU & CPU RAM
        }
        if device_type.lower() == "mps":
            kwargs["n_gpu_layers"] = 1
        if device_type.lower() == "cuda":
            kwargs["n_gpu_layers"] = N_GPU_LAYERS
            kwargs["n_batch"] = N_BATCH  # set this based on your GPU

        # kwargs["stream"] = stream

        # if stream == True:
        #     kwargs["callbacks"] = callbacks

        return LlamaCpp(**kwargs)
    except:
        if "ggml" in model_basename:
            logging.INFO("If you were using GGML model, LLAMA-CPP Dropped Support, Use GGUF Instead")
        return None


def load_quantized_model_qptq(model_id, model_basename, device_type, logging):
    """
    Load a GPTQ quantized model using AutoGPTQForCausalLM.

    This function loads a quantized model that ends with GPTQ and may have variations
    of .no-act.order or .safetensors in their HuggingFace repo.

    Parameters:
    - model_id (str): The identifier for the model on HuggingFace Hub.
    - model_basename (str): The base name of the model file.
    - device_type (str): The type of device where the model will run.
    - logging (logging.Logger): Logger instance for logging messages.

    Returns:
    - model (AutoGPTQForCausalLM): The loaded quantized model.
    - tokenizer (AutoTokenizer): The tokenizer associated with the model.

    Notes:
    - The function checks for the ".safetensors" ending in the model_basename and removes it if present.
    """

    # The code supports all huggingface models that ends with GPTQ and have some variation
    # of .no-act.order or .safetensors in their HF repo.
    logging.info("Using AutoGPTQForCausalLM for quantized models")

    if ".safetensors" in model_basename:
        # Remove the ".safetensors" ending if present
        model_basename = model_basename.replace(".safetensors", "")

    tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
    logging.info("Tokenizer loaded")

    model = AutoGPTQForCausalLM.from_quantized(
        model_id,
        model_basename=model_basename,
        use_safetensors=True,
        trust_remote_code=True,
        device_map="auto",
        use_triton=False,
        quantize_config=None,
    )

    return model, tokenizer


def load_full_model(model_id, model_basename, device_type, logging):
    """
    Load a full model using either LlamaTokenizer or AutoModelForCausalLM.

    This function loads a full model based on the specified device type.
    If the device type is 'mps' or 'cpu', it uses LlamaTokenizer and LlamaForCausalLM.
    Otherwise, it uses AutoModelForCausalLM.

    Parameters:
    - model_id (str): The identifier for the model on HuggingFace Hub.
    - model_basename (str): The base name of the model file.
    - device_type (str): The type of device where the model will run.
    - logging (logging.Logger): Logger instance for logging messages.

    Returns:
    - model (Union[LlamaForCausalLM, AutoModelForCausalLM]): The loaded model.
    - tokenizer (Union[LlamaTokenizer, AutoTokenizer]): The tokenizer associated with the model.

    Notes:
    - The function uses the `from_pretrained` method to load both the model and the tokenizer.
    - Additional settings are provided for NVIDIA GPUs, such as loading in 4-bit and setting the compute dtype.
    """

    if device_type.lower() in ["mps", "cpu"]:
        logging.info("Using LlamaTokenizer")
        tokenizer = LlamaTokenizer.from_pretrained(model_id, cache_dir="./models/")
        model = LlamaForCausalLM.from_pretrained(model_id, cache_dir="./models/")
    else:
        logging.info("Using AutoModelForCausalLM for full models")
        tokenizer = AutoTokenizer.from_pretrained(model_id, cache_dir="./models/")
        logging.info("Tokenizer loaded")
        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            device_map="auto",
            torch_dtype=torch.float16,
            low_cpu_mem_usage=True,
            cache_dir=MODELS_PATH,
            # trust_remote_code=True, # set these if you are using NVIDIA GPU
            # load_in_4bit=True,
            # bnb_4bit_quant_type="nf4",
            # bnb_4bit_compute_dtype=torch.float16,
            max_memory={0: "15GB"} # Uncomment this line with you encounter CUDA out of memory errors
        )
        model.tie_weights()
    return model, tokenizer


def load_model(device_type, model_id, model_basename=None, LOGGING=logging, stream=False, callbacks = []):
    """
    Select a model for text generation using the HuggingFace library.
    If you are running this for the first time, it will download a model for you.
    subsequent runs will use the model from the disk.

    Args:
        device_type (str): Type of device to use, e.g., "cuda" for GPU or "cpu" for CPU.
        model_id (str): Identifier of the model to load from HuggingFace's model hub.
        model_basename (str, optional): Basename of the model if using quantized models.
            Defaults to None.

    Returns:
        HuggingFacePipeline: A pipeline object for text generation using the loaded model.

    Raises:
        ValueError: If an unsupported model or device type is provided.
    """

    logging.info(f"Loading Model: {model_id}, on: {device_type}")
    logging.info("This action can take a few minutes!")

    if model_basename is not None:
        if ".gguf" in model_basename.lower():
            llm = load_quantized_model_gguf_ggml(model_id, model_basename, device_type, LOGGING, stream, callbacks)
            return llm
        elif ".ggml" in model_basename.lower():
            model, tokenizer = load_quantized_model_gguf_ggml(model_id, model_basename, device_type, LOGGING)
        else:
            model, tokenizer = load_quantized_model_qptq(model_id, model_basename, device_type, LOGGING)
    else:
        model, tokenizer = load_full_model(model_id, model_basename, device_type, LOGGING)

    # Load configuration from the model to avoid warnings
    generation_config = GenerationConfig.from_pretrained(model_id)
    # see here for details:
    # https://huggingface.co/docs/transformers/
    # main_classes/text_generation#transformers.GenerationConfig.from_pretrained.returns

    # Create a pipeline for text generation

    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        max_length=MAX_NEW_TOKENS,
        temperature=0.2,
        # top_p=0.95,
        repetition_penalty=1.15,
        generation_config=generation_config,
    )

    local_llm = HuggingFacePipeline(pipeline=pipe)
    logging.info("Local LLM Loaded")

    return local_llm