Spaces:
Paused
Paused
| import os | |
| # from dotenv import load_dotenv | |
| from chromadb.config import Settings | |
| # https://python.langchain.com/en/latest/modules/indexes/document_loaders/examples/excel.html?highlight=xlsx#microsoft-excel | |
| from langchain.document_loaders import CSVLoader, PDFMinerLoader, TextLoader, UnstructuredExcelLoader, Docx2txtLoader | |
| from langchain.document_loaders import UnstructuredFileLoader | |
| # load_dotenv() | |
| ROOT_DIRECTORY = os.path.dirname(os.path.realpath(__file__)) | |
| PATH_NAME_SOURCE_DIRECTORY = "SOURCE_DOCUMENTS" | |
| SHOW_SOURCES=True | |
| # Define the folder for storing database | |
| SOURCE_DIRECTORY = f"{ROOT_DIRECTORY}/{PATH_NAME_SOURCE_DIRECTORY}" | |
| PERSIST_DIRECTORY = f"{ROOT_DIRECTORY}/DB" | |
| MODELS_PATH = "./models" | |
| # Can be changed to a specific number | |
| INGEST_THREADS = os.cpu_count() or 8 | |
| # Define the Chroma settings | |
| CHROMA_SETTINGS = Settings( | |
| anonymized_telemetry=False, | |
| is_persistent=True, | |
| ) | |
| # Context Window and Max New Tokens | |
| CONTEXT_WINDOW_SIZE = 4096 | |
| MAX_NEW_TOKENS = 1024 # int(CONTEXT_WINDOW_SIZE/4) | |
| #### If you get a "not enough space in the buffer" error, you should reduce the values below, start with half of the original values and keep halving the value until the error stops appearing | |
| N_GPU_LAYERS = 100 # Llama-2-70B has 83 layers | |
| N_BATCH = CONTEXT_WINDOW_SIZE | |
| ### From experimenting with the Llama-2-7B-Chat-GGML model on 8GB VRAM, these values work: | |
| # N_GPU_LAYERS = 20 | |
| # N_BATCH = 512 | |
| # https://python.langchain.com/en/latest/_modules/langchain/document_loaders/excel.html#UnstructuredExcelLoader | |
| DOCUMENT_MAP = { | |
| ".txt": TextLoader, | |
| ".md": TextLoader, | |
| ".py": TextLoader, | |
| ".pdf": PDFMinerLoader, | |
| # ".pdf": UnstructuredFileLoader, | |
| ".csv": CSVLoader, | |
| ".xls": UnstructuredExcelLoader, | |
| ".xlsx": UnstructuredExcelLoader, | |
| ".docx": Docx2txtLoader, | |
| ".doc": Docx2txtLoader, | |
| } | |
| # Default Instructor Model | |
| EMBEDDING_MODEL_NAME = "hkunlp/instructor-large" # Uses 1.5 GB of VRAM (High Accuracy with lower VRAM usage) | |
| #### | |
| #### OTHER EMBEDDING MODEL OPTIONS | |
| #### | |
| # EMBEDDING_MODEL_NAME = "hkunlp/instructor-xl" # Uses 5 GB of VRAM (Most Accurate of all models) | |
| # EMBEDDING_MODEL_NAME = "intfloat/e5-large-v2" # Uses 1.5 GB of VRAM (A little less accurate than instructor-large) | |
| # EMBEDDING_MODEL_NAME = "intfloat/e5-base-v2" # Uses 0.5 GB of VRAM (A good model for lower VRAM GPUs) | |
| # EMBEDDING_MODEL_NAME = "all-MiniLM-L6-v2" # Uses 0.2 GB of VRAM (Less accurate but fastest - only requires 150mb of vram) | |
| #### | |
| #### MULTILINGUAL EMBEDDING MODELS | |
| #### | |
| # EMBEDDING_MODEL_NAME = "intfloat/multilingual-e5-large" # Uses 2.5 GB of VRAM | |
| # EMBEDDING_MODEL_NAME = "intfloat/multilingual-e5-base" # Uses 1.2 GB of VRAM | |
| #### SELECT AN OPEN SOURCE LLM (LARGE LANGUAGE MODEL) | |
| # Select the Model ID and model_basename | |
| # load the LLM for generating Natural Language responses | |
| #### GPU VRAM Memory required for LLM Models (ONLY) by Billion Parameter value (B Model) | |
| #### Does not include VRAM used by Embedding Models - which use an additional 2GB-7GB of VRAM depending on the model. | |
| #### | |
| #### (B Model) (float32) (float16) (GPTQ 8bit) (GPTQ 4bit) | |
| #### 7b 28 GB 14 GB 7 GB - 9 GB 3.5 GB - 5 GB | |
| #### 13b 52 GB 26 GB 13 GB - 15 GB 6.5 GB - 8 GB | |
| #### 32b 130 GB 65 GB 32.5 GB - 35 GB 16.25 GB - 19 GB | |
| #### 65b 260.8 GB 130.4 GB 65.2 GB - 67 GB 32.6 GB - - 35 GB | |
| # MODEL_ID = "TheBloke/Llama-2-7B-Chat-GGML" | |
| # MODEL_BASENAME = "llama-2-7b-chat.ggmlv3.q4_0.bin" | |
| #### | |
| #### (FOR GGUF MODELS) | |
| #### | |
| # MODEL_ID = "TheBloke/Llama-2-13b-Chat-GGUF" | |
| # MODEL_BASENAME = "llama-2-13b-chat.Q4_K_M.gguf" | |
| # MODEL_ID = "TheBloke/Llama-2-7b-Chat-GGUF" | |
| # MODEL_BASENAME = "llama-2-7b-chat.Q4_K_M.gguf" | |
| MODEL_ID = "TheBloke/Mistral-7B-Instruct-v0.1-GGUF" | |
| MODEL_BASENAME = "mistral-7b-instruct-v0.1.Q8_0.gguf" | |
| # MODEL_ID = "TheBloke/Llama-2-70b-Chat-GGUF" | |
| # MODEL_BASENAME = "llama-2-70b-chat.Q4_K_M.gguf" | |
| #### | |
| #### (FOR HF MODELS) | |
| #### | |
| # MODEL_ID = "NousResearch/Llama-2-7b-chat-hf" | |
| # MODEL_BASENAME = None | |
| # 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) Select a llm model based on your GPU and VRAM GB. Does not include Embedding Models VRAM usage. | |
| #### | |
| ##### 48GB VRAM Graphics Cards (RTX 6000, RTX A6000 and other 48GB VRAM GPUs) ##### | |
| ### 65b GPTQ LLM Models for 48GB GPUs (*** With best embedding model: hkunlp/instructor-xl ***) | |
| # MODEL_ID = "TheBloke/guanaco-65B-GPTQ" | |
| # MODEL_BASENAME = "model.safetensors" | |
| # MODEL_ID = "TheBloke/Airoboros-65B-GPT4-2.0-GPTQ" | |
| # MODEL_BASENAME = "model.safetensors" | |
| # MODEL_ID = "TheBloke/gpt4-alpaca-lora_mlp-65B-GPTQ" | |
| # MODEL_BASENAME = "model.safetensors" | |
| # MODEL_ID = "TheBloke/Upstage-Llama1-65B-Instruct-GPTQ" | |
| # MODEL_BASENAME = "model.safetensors" | |
| ##### 24GB VRAM Graphics Cards (RTX 3090 - RTX 4090 (35% Faster) - RTX A5000 - RTX A5500) ##### | |
| ### 13b GPTQ Models for 24GB GPUs (*** With best embedding model: hkunlp/instructor-xl ***) | |
| # MODEL_ID = "TheBloke/Wizard-Vicuna-13B-Uncensored-GPTQ" | |
| # MODEL_BASENAME = "Wizard-Vicuna-13B-Uncensored-GPTQ-4bit-128g.compat.no-act-order.safetensors" | |
| # MODEL_ID = "TheBloke/vicuna-13B-v1.5-GPTQ" | |
| # MODEL_BASENAME = "model.safetensors" | |
| # MODEL_ID = "TheBloke/Nous-Hermes-13B-GPTQ" | |
| # MODEL_BASENAME = "nous-hermes-13b-GPTQ-4bit-128g.no-act.order" | |
| # MODEL_ID = "TheBloke/WizardLM-13B-V1.2-GPTQ" | |
| # MODEL_BASENAME = "gptq_model-4bit-128g.safetensors | |
| ### 30b GPTQ Models for 24GB GPUs (*** Requires using intfloat/e5-base-v2 instead of hkunlp/instructor-large as embedding model ***) | |
| # MODEL_ID = "TheBloke/Wizard-Vicuna-30B-Uncensored-GPTQ" | |
| # MODEL_BASENAME = "Wizard-Vicuna-30B-Uncensored-GPTQ-4bit--1g.act.order.safetensors" | |
| # MODEL_ID = "TheBloke/WizardLM-30B-Uncensored-GPTQ" | |
| # MODEL_BASENAME = "WizardLM-30B-Uncensored-GPTQ-4bit.act-order.safetensors" | |
| ##### 8-10GB VRAM Graphics Cards (RTX 3080 - RTX 3080 Ti - RTX 3070 Ti - 3060 Ti - RTX 2000 Series, Quadro RTX 4000, 5000, 6000) ##### | |
| ### (*** Requires using intfloat/e5-small-v2 instead of hkunlp/instructor-large as embedding model ***) | |
| ### 7b GPTQ Models for 8GB GPUs | |
| # MODEL_ID = "TheBloke/Wizard-Vicuna-7B-Uncensored-GPTQ" | |
| # MODEL_BASENAME = "Wizard-Vicuna-7B-Uncensored-GPTQ-4bit-128g.no-act.order.safetensors" | |
| # MODEL_ID = "TheBloke/WizardLM-7B-uncensored-GPTQ" | |
| # MODEL_BASENAME = "WizardLM-7B-uncensored-GPTQ-4bit-128g.compat.no-act-order.safetensors" | |
| # MODEL_ID = "TheBloke/wizardLM-7B-GPTQ" | |
| # MODEL_BASENAME = "wizardLM-7B-GPTQ-4bit.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" | |