import logging import click import torch from auto_gptq import AutoGPTQForCausalLM from huggingface_hub import hf_hub_download from langchain.chains import RetrievalQA from langchain.embeddings import HuggingFaceInstructEmbeddings from langchain.llms import HuggingFacePipeline, LlamaCpp # from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler from langchain.vectorstores import Chroma from transformers import ( AutoModelForCausalLM, AutoTokenizer, GenerationConfig, LlamaForCausalLM, LlamaTokenizer, pipeline, ) from constants import CHROMA_SETTINGS, EMBEDDING_MODEL_NAME, PERSIST_DIRECTORY def load_model(device_type, model_id, model_basename=None): """ 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 ".ggml" in model_basename: logging.info("Using Llamacpp for GGML quantized models") model_path = hf_hub_download(repo_id=model_id, filename=model_basename) max_ctx_size = 2048 kwargs = { "model_path": model_path, "n_ctx": max_ctx_size, "max_tokens": max_ctx_size, } if device_type.lower() == "mps": kwargs["n_gpu_layers"] = 1000 if device_type.lower() == "cuda": kwargs["n_gpu_layers"] = 1000 kwargs["n_batch"] = max_ctx_size return LlamaCpp(**kwargs) else: # 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="cuda:0", use_triton=False, quantize_config=None, ) elif ( device_type.lower() == "cuda" ): # The code supports all huggingface models that ends with -HF or which have a .bin # file in their HF repo. logging.info("Using AutoModelForCausalLM for full models") tokenizer = AutoTokenizer.from_pretrained(model_id) logging.info("Tokenizer loaded") model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto", torch_dtype=torch.float16, low_cpu_mem_usage=True, trust_remote_code=True, # max_memory={0: "15GB"} # Uncomment this line with you encounter CUDA out of memory errors ) model.tie_weights() else: logging.info("Using LlamaTokenizer") tokenizer = LlamaTokenizer.from_pretrained(model_id) model = LlamaForCausalLM.from_pretrained(model_id) # 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=2048, temperature=0, 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 # # chose device typ to run on as well as to show source documents. # @click.command() # @click.option( # "--device_type", # default="cuda" if torch.cuda.is_available() else "cpu", # type=click.Choice( # [ # "cpu", # "cuda", # "ipu", # "xpu", # "mkldnn", # "opengl", # "opencl", # "ideep", # "hip", # "ve", # "fpga", # "ort", # "xla", # "lazy", # "vulkan", # "mps", # "meta", # "hpu", # "mtia", # ], # ), # help="Device to run on. (Default is cuda)", # ) # @click.option( # "--show_sources", # "-s", # is_flag=True, # help="Show sources along with answers (Default is False)", # ) def main(device_type, strQuery): """ This function implements the information retrieval task. 1. Loads an embedding model, can be HuggingFaceInstructEmbeddings or HuggingFaceEmbeddings 2. Loads the existing vectorestore that was created by inget.py 3. Loads the local LLM using load_model function - You can now set different LLMs. 4. Setup the Question Answer retreival chain. 5. Question answers. """ logging.info(f"Running on: {device_type}") #logging.info(f"Display Source Documents set to: {show_sources}") embeddings = HuggingFaceInstructEmbeddings(model_name=EMBEDDING_MODEL_NAME, model_kwargs={"device": device_type}) # uncomment the following line if you used HuggingFaceEmbeddings in the ingest.py # embeddings = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL_NAME) # load the vectorstore db = Chroma( persist_directory=PERSIST_DIRECTORY, embedding_function=embeddings, client_settings=CHROMA_SETTINGS, ) retriever = db.as_retriever() # load the LLM for generating Natural Language responses # for HF models # model_id = "TheBloke/vicuna-7B-1.1-HF" # model_basename = None # model_id = "TheBloke/Wizard-Vicuna-7B-Uncensored-HF" # model_id = "TheBloke/guanaco-7B-HF" # model_id = 'NousResearch/Nous-Hermes-13b' # Requires ~ 23GB VRAM. Using STransformers # alongside will 100% create OOM on 24GB cards. # llm = load_model(device_type, model_id=model_id) # for GPTQ (quantized) models # model_id = "TheBloke/Nous-Hermes-13B-GPTQ" # model_basename = "nous-hermes-13b-GPTQ-4bit-128g.no-act.order" # model_id = "TheBloke/WizardLM-30B-Uncensored-GPTQ" # model_basename = "WizardLM-30B-Uncensored-GPTQ-4bit.act-order.safetensors" # Requires # ~21GB VRAM. Using STransformers alongside can potentially create OOM on 24GB cards. # model_id = "TheBloke/wizardLM-7B-GPTQ" # model_basename = "wizardLM-7B-GPTQ-4bit.compat.no-act-order.safetensors" # model_id = "TheBloke/WizardLM-7B-uncensored-GPTQ" # model_basename = "WizardLM-7B-uncensored-GPTQ-4bit-128g.compat.no-act-order.safetensors" # for GGML (quantized cpu+gpu+mps) models - check if they support llama.cpp # model_id = "TheBloke/wizard-vicuna-13B-GGML" # model_basename = "wizard-vicuna-13B.ggmlv3.q4_0.bin" # model_basename = "wizard-vicuna-13B.ggmlv3.q6_K.bin" # model_basename = "wizard-vicuna-13B.ggmlv3.q2_K.bin" # model_id = "TheBloke/orca_mini_3B-GGML" # model_basename = "orca-mini-3b.ggmlv3.q4_0.bin" model_id = "TheBloke/Llama-2-7B-Chat-GGML" model_basename = "llama-2-7b-chat.ggmlv3.q4_0.bin" llm = load_model(device_type, model_id=model_id, model_basename=model_basename) qa = RetrievalQA.from_chain_type(llm=llm, chain_type="stuff", retriever=retriever, return_source_documents=True) # Interactive questions and answers query = strQuery # Get the answer from the chain res = qa(query) answer, docs = res["result"], res["source_documents"] return(answer) if __name__ == "__main__": logging.basicConfig( format="%(asctime)s - %(levelname)s - %(filename)s:%(lineno)s - %(message)s", level=logging.INFO ) main()