diff --git "a/generate.py" "b/generate.py" --- "a/generate.py" +++ "b/generate.py" @@ -1,27 +1,39 @@ import ast +import copy import functools import glob import inspect import queue -import shutil import sys import os import time import traceback +import types import typing import warnings from datetime import datetime import filelock +import requests import psutil +from requests import ConnectTimeout, JSONDecodeError +from urllib3.exceptions import ConnectTimeoutError, MaxRetryError, ConnectionError +from requests.exceptions import ConnectionError as ConnectionError2 +from requests.exceptions import ReadTimeout as ReadTimeout2 + +if os.path.dirname(os.path.abspath(__file__)) not in sys.path: + sys.path.append(os.path.dirname(os.path.abspath(__file__))) os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1' os.environ['BITSANDBYTES_NOWELCOME'] = '1' warnings.filterwarnings('ignore', category=UserWarning, message='TypedStorage is deprecated') +from enums import DocumentChoices, LangChainMode, no_lora_str, model_token_mapping, no_model_str, source_prefix, \ + source_postfix from loaders import get_loaders from utils import set_seed, clear_torch_cache, save_generate_output, NullContext, wrapped_partial, EThread, get_githash, \ - import_matplotlib, get_device, makedirs, get_kwargs + import_matplotlib, get_device, makedirs, get_kwargs, start_faulthandler, get_hf_server, FakeTokenizer, remove +start_faulthandler() import_matplotlib() SEED = 1236 @@ -31,17 +43,14 @@ from typing import Union import fire import torch -from peft import PeftModel from transformers import GenerationConfig, AutoModel, TextIteratorStreamer -from accelerate import init_empty_weights, infer_auto_device_map -from prompter import Prompter, inv_prompt_type_to_model_lower, non_hf_types +from prompter import Prompter, inv_prompt_type_to_model_lower, non_hf_types, PromptType, get_prompt, generate_prompt from stopping import get_stopping eval_extra_columns = ['prompt', 'response', 'score'] -langchain_modes = ['Disabled', 'ChatLLM', 'LLM', 'All', 'wiki', 'wiki_full', 'UserData', 'MyData', 'github h2oGPT', - 'DriverlessAI docs'] +langchain_modes = [x.value for x in list(LangChainMode)] scratch_base_dir = '/tmp/' @@ -56,8 +65,15 @@ def main( lora_weights: str = "", gpu_id: int = 0, compile_model: bool = True, - + use_cache: bool = None, + inference_server: str = "", prompt_type: Union[int, str] = None, + prompt_dict: typing.Dict = None, + + model_lock: typing.List[typing.Dict[str, str]] = None, + model_lock_columns: int = None, + fail_if_cannot_connect: bool = False, + # input to generation temperature: float = None, top_p: float = None, @@ -87,14 +103,13 @@ def main( cli: bool = False, cli_loop: bool = True, gradio: bool = True, - gradio_avoid_processing_markdown: bool = False, gradio_offline_level: int = 0, chat: bool = True, chat_context: bool = False, stream_output: bool = True, show_examples: bool = None, verbose: bool = False, - h2ocolors: bool = False, + h2ocolors: bool = True, height: int = 600, show_lora: bool = True, login_mode_if_model0: bool = False, @@ -103,16 +118,19 @@ def main( api_open: bool = False, allow_api: bool = True, input_lines: int = 1, + gradio_size: str = None, auth: typing.List[typing.Tuple[str, str]] = None, + max_max_time=None, + max_max_new_tokens=None, - sanitize_user_prompt: bool = True, - sanitize_bot_response: bool = True, + sanitize_user_prompt: bool = False, + sanitize_bot_response: bool = False, extra_model_options: typing.List[str] = [], extra_lora_options: typing.List[str] = [], + extra_server_options: typing.List[str] = [], score_model: str = 'OpenAssistant/reward-model-deberta-v3-large-v2', - auto_score: bool = True, eval_filename: str = None, eval_prompts_only_num: int = 0, @@ -120,8 +138,9 @@ def main( eval_as_output: bool = False, langchain_mode: str = 'Disabled', + force_langchain_evaluate: bool = False, visible_langchain_modes: list = ['UserData', 'MyData'], - document_choice: list = ['All'], + document_choice: list = [DocumentChoices.All_Relevant.name], user_path: str = None, detect_user_path_changes_every_query: bool = False, load_db_if_exists: bool = True, @@ -129,7 +148,7 @@ def main( db_type: str = 'chroma', use_openai_embedding: bool = False, use_openai_model: bool = False, - hf_embedding_model: str = "sentence-transformers/all-MiniLM-L6-v2", + hf_embedding_model: str = None, allow_upload_to_user_data: bool = True, allow_upload_to_my_data: bool = True, enable_url_upload: bool = True, @@ -137,7 +156,10 @@ def main( enable_sources_list: bool = True, chunk: bool = True, chunk_size: int = 512, - top_k_docs: int = 3, # FIXME: Can go back to 4 once https://github.com/h2oai/h2ogpt/issues/192 fixed + top_k_docs: int = None, + reverse_docs: bool = True, + auto_reduce_chunks: bool = True, + max_chunks: int = 100, n_jobs: int = -1, enable_captions: bool = True, captions_model: str = "Salesforce/blip-image-captioning-base", @@ -156,7 +178,31 @@ def main( :param lora_weights: LORA weights path/HF link :param gpu_id: if infer_devices, then use gpu_id for cuda device ID, or auto mode if gpu_id != -1 :param compile_model Whether to compile the model + :param use_cache: Whether to use caching in model (some models fail when multiple threads use) + :param inference_server: Consume base_model as type of model at this address + Address can be text-generation-server hosting that base_model + e.g. python generate.py --inference_server="http://192.168.1.46:6112" --base_model=h2oai/h2ogpt-oasst1-512-12b + Or Address can be "openai_chat" or "openai" for OpenAI API + e.g. python generate.py --inference_server="openai_chat" --base_model=gpt-3.5-turbo + e.g. python generate.py --inference_server="openai" --base_model=text-davinci-003 :param prompt_type: type of prompt, usually matched to fine-tuned model or plain for foundational model + :param prompt_dict: If prompt_type=custom, then expects (some) items returned by get_prompt(..., return_dict=True) + :param model_lock: Lock models to specific combinations, for ease of use and extending to many models + Only used if gradio = True + List of dicts, each dict has base_model, tokenizer_base_model, lora_weights, inference_server, prompt_type, and prompt_dict + If all models have same prompt_type, and prompt_dict, can still specify that once in CLI outside model_lock as default for dict + Can specify model_lock instead of those items on CLI + As with CLI itself, base_model can infer prompt_type and prompt_dict if in prompter.py. + Also, tokenizer_base_model and lora_weights are optional. + Also, inference_server is optional if loading model from local system. + All models provided will automatically appear in compare model mode + Model loading-unloading and related choices will be disabled. Model/lora/server adding will be disabled + :param model_lock_columns: How many columns to show if locking models (and so showing all at once) + If None, then defaults to up to 3 + if -1, then all goes into 1 row + Maximum value is 4 due to non-dynamic gradio rendering elements + :param fail_if_cannot_connect: if doing model locking (e.g. with many models), fail if True. Otherwise ignore. + Useful when many endpoints and want to just see what works, but still have to wait for timeout. :param temperature: generation temperature :param top_p: generation top_p :param top_k: generation top_k @@ -182,13 +228,13 @@ def main( :param cli: whether to use CLI (non-gradio) interface. :param cli_loop: whether to loop for CLI (False usually only for testing) :param gradio: whether to enable gradio, or to enable benchmark mode - :param gradio_avoid_processing_markdown: :param gradio_offline_level: > 0, then change fonts so full offline == 1 means backend won't need internet for fonts, but front-end UI might if font not cached == 2 means backend and frontend don't need internet to download any fonts. Note: Some things always disabled include HF telemetry, gradio telemetry, chromadb posthog that involve uploading. This option further disables google fonts for downloading, which is less intrusive than uploading, but still required in air-gapped case. The fonts don't look as nice as google fonts, but ensure full offline behavior. + Also set --share=False to avoid sharing a gradio live link. :param chat: whether to enable chat mode with chat history :param chat_context: whether to use extra helpful context if human_bot :param stream_output: whether to stream output from generate @@ -203,20 +249,25 @@ def main( :param api_open: If False, don't let API calls skip gradio queue :param allow_api: whether to allow API calls at all to gradio server :param input_lines: how many input lines to show for chat box (>1 forces shift-enter for submit, else enter is submit) + :param gradio_size: Overall size of text and spaces: "xsmall", "small", "medium", "large". + Small useful for many chatbots in model_lock mode :param auth: gradio auth for launcher in form [(user1, pass1), (user2, pass2), ...] e.g. --auth=[('jon','password')] with no spaces - :param sanitize_user_prompt: whether to remove profanity from user input - :param sanitize_bot_response: whether to remove profanity and repeat lines from bot output + :param max_max_time: Maximum max_time for gradio slider + :param max_max_new_tokens: Maximum max_new_tokens for gradio slider + :param sanitize_user_prompt: whether to remove profanity from user input (slows down input processing) + :param sanitize_bot_response: whether to remove profanity and repeat lines from bot output (about 2x slower generation for long streaming cases due to better_profanity being slow) :param extra_model_options: extra models to show in list in gradio :param extra_lora_options: extra LORA to show in list in gradio + :param extra_server_options: extra servers to show in list in gradio :param score_model: which model to score responses (None means no scoring) - :param auto_score: whether to automatically score responses :param eval_filename: json file to use for evaluation, if None is sharegpt :param eval_prompts_only_num: for no gradio benchmark, if using eval_filename prompts for eval instead of examples :param eval_prompts_only_seed: for no gradio benchmark, seed for eval_filename sampling :param eval_as_output: for no gradio benchmark, whether to test eval_filename output itself :param langchain_mode: Data source to include. Choose "UserData" to only consume files from make_db.py. WARNING: wiki_full requires extra data processing via read_wiki_full.py and requires really good workstation to generate db, unless already present. + :param force_langchain_evaluate: Whether to force langchain LLM use even if not doing langchain, mostly for testing. :param user_path: user path to glob from to generate db for vector search, for 'UserData' langchain mode. If already have db, any new/changed files are added automatically if path set, does not have to be same path used for prior db sources :param detect_user_path_changes_every_query: whether to detect if any files changed or added every similarity search (by file hashes). @@ -234,6 +285,10 @@ def main( :param use_openai_embedding: Whether to use OpenAI embeddings for vector db :param use_openai_model: Whether to use OpenAI model for use with vector db :param hf_embedding_model: Which HF embedding model to use for vector db + Default is instructor-large with 768 parameters per embedding if have GPUs, else all-MiniLM-L6-v1 if no GPUs + Can also choose simpler model with 384 parameters per embedding: "sentence-transformers/all-MiniLM-L6-v2" + Can also choose even better embedding with 1024 parameters: 'hkunlp/instructor-xl' + We support automatically changing of embeddings for chroma, with a backup of db made if this is done :param allow_upload_to_user_data: Whether to allow file uploads to update shared vector db :param allow_upload_to_my_data: Whether to allow file uploads to update scratch vector db :param enable_url_upload: Whether to allow upload from URL @@ -242,12 +297,17 @@ def main( :param chunk: Whether to chunk data (True unless know data is already optimally chunked) :param chunk_size: Size of chunks, with typically top-4 passed to LLM, so neesd to be in context length :param top_k_docs: number of chunks to give LLM + :param reverse_docs: whether to reverse docs order so most relevant is closest to question. + Best choice for sufficiently smart model, and truncation occurs for oldest context, so best then too. + But smaller 6_9 models fail to use newest context and can get stuck on old information. + :param auto_reduce_chunks: Whether to automatically reduce top_k_docs to fit context given prompt + :param max_chunks: If top_k_docs=-1, maximum number of chunks to allow :param n_jobs: Number of processors to use when consuming documents (-1 = all, is default) :param enable_captions: Whether to support captions using BLIP for image files as documents, then preloads that model :param captions_model: Which model to use for captions. - captions_model: int = "Salesforce/blip-image-captioning-base", # continue capable + captions_model: str = "Salesforce/blip-image-captioning-base", # continue capable captions_model: str = "Salesforce/blip2-flan-t5-xl", # question/answer capable, 16GB state - captions_model: int = "Salesforce/blip2-flan-t5-xxl", # question/answer capable, 60GB state + captions_model: str = "Salesforce/blip2-flan-t5-xxl", # question/answer capable, 60GB state Note: opt-based blip2 are not permissive license due to opt and Meta license restrictions :param pre_load_caption_model: Whether to preload caption model, or load after forking parallel doc loader parallel loading disabled if preload and have images, to prevent deadlocking on cuda context @@ -256,8 +316,34 @@ def main( :param enable_ocr: Whether to support OCR on images :return: """ - is_hf = bool(os.getenv("HUGGINGFACE_SPACES")) - is_gpth2oai = bool(os.getenv("GPT_H2O_AI")) + if base_model is None: + base_model = '' + if tokenizer_base_model is None: + tokenizer_base_model = '' + if lora_weights is None: + lora_weights = '' + if inference_server is None: + inference_server = '' + + # listen to env if set + model_lock = os.getenv('model_lock', str(model_lock)) + model_lock = ast.literal_eval(model_lock) + + if model_lock: + assert gradio, "model_lock only supported for gradio=True" + if len(model_lock) > 1: + assert chat, "model_lock only works for multiple models for chat=True" + assert not cli, "model_lock only supported for cli=False" + assert not (not cli and not gradio), "model_lock only supported for eval (cli=gradio=False)" + assert not base_model, "Don't specify model_lock and base_model" + assert not tokenizer_base_model, "Don't specify model_lock and tokenizer_base_model" + assert not lora_weights, "Don't specify model_lock and lora_weights" + assert not inference_server, "Don't specify model_lock and inference_server" + # assert not prompt_type, "Don't specify model_lock and prompt_type" + # assert not prompt_dict, "Don't specify model_lock and prompt_dict" + + is_hf = bool(int(os.getenv("HUGGINGFACE_SPACES", '0'))) + is_gpth2oai = bool(int(os.getenv("GPT_H2O_AI", '0'))) is_public = is_hf or is_gpth2oai # multi-user case with fixed model and disclaimer if memory_restriction_level is None: memory_restriction_level = 2 if is_hf else 0 # 2 assumes run on 24GB consumer GPU @@ -270,9 +356,11 @@ def main( # allow set token directly use_auth_token = os.environ.get("HUGGINGFACE_API_TOKEN", use_auth_token) - allow_upload_to_user_data = bool(os.environ.get("allow_upload_to_user_data", allow_upload_to_user_data)) - allow_upload_to_my_data = bool(os.environ.get("allow_upload_to_my_data", allow_upload_to_my_data)) - height = os.environ.get("HEIGHT", height) + allow_upload_to_user_data = bool( + int(os.environ.get("allow_upload_to_user_data", str(int(allow_upload_to_user_data))))) + allow_upload_to_my_data = bool(int(os.environ.get("allow_upload_to_my_data", str(int(allow_upload_to_my_data))))) + height = int(os.environ.get("HEIGHT", height)) + h2ocolors = bool(int(os.getenv('h2ocolors', h2ocolors))) # allow enabling langchain via ENV # FIRST PLACE where LangChain referenced, but no imports related to it @@ -282,6 +370,12 @@ def main( if langchain_mode not in visible_langchain_modes and langchain_mode in langchain_modes: visible_langchain_modes += [langchain_mode] + # if specifically chose not to show My or User Data, disable upload, so gradio elements are simpler + if LangChainMode.MY_DATA.value not in visible_langchain_modes: + allow_upload_to_my_data = False + if LangChainMode.USER_DATA.value not in visible_langchain_modes: + allow_upload_to_user_data = False + if is_public: allow_upload_to_user_data = False input_lines = 1 # ensure set, for ease of use @@ -290,31 +384,57 @@ def main( top_k = 70 if top_k is None else top_k if is_hf: do_sample = True if do_sample is None else do_sample + top_k_docs = 3 if top_k_docs is None else top_k_docs else: # by default don't sample, too chatty do_sample = False if do_sample is None else do_sample + top_k_docs = 4 if top_k_docs is None else top_k_docs if memory_restriction_level == 2: - if not base_model: + if not base_model and not inference_server: base_model = 'h2oai/h2ogpt-oasst1-512-12b' # don't set load_8bit if passed base_model, doesn't always work so can't just override load_8bit = True load_4bit = False # FIXME - consider using 4-bit instead of 8-bit - else: - base_model = 'h2oai/h2ogpt-oasst1-512-20b' if not base_model else base_model + elif not inference_server: + top_k_docs = 10 if top_k_docs is None else top_k_docs if memory_restriction_level >= 2: load_8bit = True load_4bit = False # FIXME - consider using 4-bit instead of 8-bit + if hf_embedding_model is None: + hf_embedding_model = "sentence-transformers/all-MiniLM-L6-v2" + top_k_docs = 3 if top_k_docs is None else top_k_docs + if top_k_docs is None: + top_k_docs = 3 + if is_public: + if not max_time: + max_time = 60 * 2 + if not max_max_time: + max_max_time = max_time + if not max_new_tokens: + max_new_tokens = 256 + if not max_max_new_tokens: + max_max_new_tokens = 256 + else: + if not max_max_time: + max_max_time = 60 * 20 + if not max_max_new_tokens: + max_max_new_tokens = 512 if is_hf: # must override share if in spaces share = False + if not max_time: + max_time = 60 * 1 + if not max_max_time: + max_max_time = max_time + # HF accounted for later in get_max_max_new_tokens() save_dir = os.getenv('SAVE_DIR', save_dir) score_model = os.getenv('SCORE_MODEL', score_model) if score_model == 'None' or score_model is None: score_model = '' concurrency_count = int(os.getenv('CONCURRENCY_COUNT', concurrency_count)) - api_open = bool(int(os.getenv('API_OPEN', api_open))) - allow_api = bool(int(os.getenv('ALLOW_API', allow_api))) + api_open = bool(int(os.getenv('API_OPEN', str(int(api_open))))) + allow_api = bool(int(os.getenv('ALLOW_API', str(int(allow_api))))) n_gpus = torch.cuda.device_count() if torch.cuda.is_available else 0 if n_gpus == 0: @@ -326,10 +446,17 @@ def main( torch.backends.cudnn.benchmark = True torch.backends.cudnn.enabled = False torch.set_default_dtype(torch.float32) - if psutil.virtual_memory().available < 94 * 1024 ** 3: + if psutil.virtual_memory().available < 94 * 1024 ** 3 and not inference_server: # 12B uses ~94GB # 6.9B uses ~47GB base_model = 'h2oai/h2ogpt-oig-oasst1-512-6_9b' if not base_model else base_model + if hf_embedding_model is None: + # if no GPUs, use simpler embedding model to avoid cost in time + hf_embedding_model = "sentence-transformers/all-MiniLM-L6-v2" + else: + if hf_embedding_model is None: + # if still None, then set default + hf_embedding_model = 'hkunlp/instructor-large' # get defaults model_lower = base_model.lower() @@ -344,10 +471,13 @@ def main( if offload_folder: makedirs(offload_folder) + if user_path: + makedirs(user_path) placeholder_instruction, placeholder_input, \ stream_output, show_examples, \ - prompt_type, temperature, top_p, top_k, num_beams, \ + prompt_type, prompt_dict, \ + temperature, top_p, top_k, num_beams, \ max_new_tokens, min_new_tokens, early_stopping, max_time, \ repetition_penalty, num_return_sequences, \ do_sample, \ @@ -356,19 +486,23 @@ def main( task_info = \ get_generate_params(model_lower, chat, stream_output, show_examples, - prompt_type, temperature, top_p, top_k, num_beams, + prompt_type, prompt_dict, + temperature, top_p, top_k, num_beams, max_new_tokens, min_new_tokens, early_stopping, max_time, repetition_penalty, num_return_sequences, do_sample, top_k_docs, + chunk, + chunk_size, verbose, ) + git_hash = get_githash() locals_dict = locals() locals_print = '\n'.join(['%s: %s' % (k, v) for k, v in locals_dict.items()]) if verbose: print(f"Generating model with params:\n{locals_print}", flush=True) - print("Command: %s\nHash: %s" % (str(' '.join(sys.argv)), get_githash()), flush=True) + print("Command: %s\nHash: %s" % (str(' '.join(sys.argv)), git_hash), flush=True) if langchain_mode != "Disabled": # SECOND PLACE where LangChain referenced, but all imports are kept local so not required @@ -382,18 +516,22 @@ def main( for gpath1 in glob.glob(os.path.join(scratch_base_dir, 'db_dir_%s*' % langchain_mode1)): if os.path.isdir(gpath1): print("Removing old MyData: %s" % gpath1, flush=True) - shutil.rmtree(gpath1) + remove(gpath1) continue if langchain_mode1 in ['All']: # FIXME: All should be avoided until scans over each db, shouldn't be separate db continue persist_directory1 = 'db_dir_%s' % langchain_mode1 # single place, no special names for each case - db = prep_langchain(persist_directory1, - load_db_if_exists, - db_type, use_openai_embedding, - langchain_mode1, user_path, - hf_embedding_model, - kwargs_make_db=locals()) + try: + db = prep_langchain(persist_directory1, + load_db_if_exists, + db_type, use_openai_embedding, + langchain_mode1, user_path, + hf_embedding_model, + kwargs_make_db=locals()) + finally: + # in case updated embeddings or created new embeddings + clear_torch_cache() dbs[langchain_mode1] = db # remove None db's so can just rely upon k in dbs for if hav db dbs = {k: v for k, v in dbs.items() if v is not None} @@ -404,6 +542,10 @@ def main( assert 'gpt_langchain' not in sys.modules, "Dev bug, import of langchain when should not have" assert 'langchain' not in sys.modules, "Dev bug, import of langchain when should not have" + model_state_none = dict(model=None, tokenizer=None, device=None, + base_model=None, tokenizer_base_model=None, lora_weights=None, + inference_server=None, prompt_type=None, prompt_dict=None) + if cli: from cli import run_cli return run_cli(**get_kwargs(run_cli, exclude_names=['model_state0'], **locals())) @@ -415,20 +557,68 @@ def main( from gradio_runner import go_gradio # get default model - all_kwargs = locals().copy() - if all_kwargs.get('base_model') and not all_kwargs['login_mode_if_model0']: - model0, tokenizer0, device = get_model(reward_type=False, - **get_kwargs(get_model, exclude_names=['reward_type'], **all_kwargs)) - else: - # if empty model, then don't load anything, just get gradio up - model0, tokenizer0, device = None, None, None - model_state0 = [model0, tokenizer0, device, all_kwargs['base_model']] + model_states = [] + model_list = [dict(base_model=base_model, tokenizer_base_model=tokenizer_base_model, lora_weights=lora_weights, + inference_server=inference_server, prompt_type=prompt_type, prompt_dict=prompt_dict)] + model_list0 = copy.deepcopy(model_list) # just strings, safe to deepcopy + model_state0 = model_state_none.copy() + assert len(model_state_none) == len(model_state0) + if model_lock: + model_list = model_lock + for model_dict in reversed(model_list): + # do reverse, so first is default base_model etc., so some logic works in go_gradio() more easily + # handles defaults user didn't have to pass + model_dict['base_model'] = base_model = model_dict.get('base_model', '') + model_dict['tokenizer_base_model'] = tokenizer_base_model = model_dict.get('tokenizer_base_model', '') + model_dict['lora_weights'] = lora_weights = model_dict.get('lora_weights', '') + model_dict['inference_server'] = inference_server = model_dict.get('inference_server', '') + prompt_type = model_dict.get('prompt_type', model_list0[0]['prompt_type']) # don't use mutated value + # try to infer, ignore empty initial state leading to get_generate_params -> 'plain' + if model_dict.get('prompt_type') is None: + model_lower = base_model.lower() + if model_lower in inv_prompt_type_to_model_lower: + prompt_type = inv_prompt_type_to_model_lower[model_lower] + prompt_dict, error0 = get_prompt(prompt_type, '', + chat=False, context='', reduced=False, making_context=False, + return_dict=True) + model_dict['prompt_type'] = prompt_type + model_dict['prompt_dict'] = prompt_dict = model_dict.get('prompt_dict', prompt_dict) + all_kwargs = locals().copy() + if base_model and not login_mode_if_model0: + model0, tokenizer0, device = get_model(reward_type=False, + **get_kwargs(get_model, exclude_names=['reward_type'], + **all_kwargs)) + else: + # if empty model, then don't load anything, just get gradio up + model0, tokenizer0, device = None, None, None + if model0 is None: + if fail_if_cannot_connect: + raise RuntimeError("Could not connect, see logs") + # skip + if isinstance(model_lock, list): + model_lock.remove(model_dict) + continue + model_state_trial = dict(model=model0, tokenizer=tokenizer0, device=device) + model_state_trial.update(model_dict) + assert len(model_state_none) == len(model_state_trial) + print("Model %s" % model_dict, flush=True) + if model_lock: + # last in iteration will be first + model_states.insert(0, model_state_trial) + # fill model_state0 so go_gradio() easier, manage model_states separately + model_state0 = model_state_trial.copy() + else: + model_state0 = model_state_trial.copy() + assert len(model_state_none) == len(model_state0) # get score model + all_kwargs = locals().copy() smodel, stokenizer, sdevice = get_score_model(reward_type=True, **get_kwargs(get_score_model, exclude_names=['reward_type'], **all_kwargs)) - score_model_state0 = [smodel, stokenizer, sdevice, score_model] + score_model_state0 = dict(model=smodel, tokenizer=stokenizer, device=sdevice, + base_model=score_model, tokenizer_base_model='', lora_weights='', + inference_server='', prompt_type='', prompt_dict='') if enable_captions: if pre_load_caption_model: @@ -443,34 +633,33 @@ def main( go_gradio(**locals()) -def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type, - gpu_id=0, - use_auth_token=False, - trust_remote_code=True, - offload_folder=None, - triton_attn=False, - long_sequence=True, - ): - """ - Ensure model gets on correct device - :param base_model: - :param model_loader: - :param load_half: - :param model_kwargs: - :param reward_type: - :param gpu_id: - :param use_auth_token: - :param trust_remote_code: - :param offload_folder: - :param triton_attn: - :param long_sequence: - :return: - """ +def get_config(base_model, + use_auth_token=False, + trust_remote_code=True, + offload_folder=None, + triton_attn=False, + long_sequence=True, + return_model=False, + raise_exception=False, + ): + from accelerate import init_empty_weights with init_empty_weights(): from transformers import AutoConfig - config = AutoConfig.from_pretrained(base_model, use_auth_token=use_auth_token, - trust_remote_code=trust_remote_code, - offload_folder=offload_folder) + try: + config = AutoConfig.from_pretrained(base_model, use_auth_token=use_auth_token, + trust_remote_code=trust_remote_code, + offload_folder=offload_folder) + except OSError as e: + if raise_exception: + raise + if 'not a local folder and is not a valid model identifier listed on' in str( + e) or '404 Client Error' in str(e): + # e.g. llama, gpjt, etc. + # e.g. HF TGI but not model on HF or private etc. + # HF TGI server only should really require prompt_type, not HF model state + return None, None + else: + raise if triton_attn and 'mpt-' in base_model.lower(): config.attn_config['attn_impl'] = 'triton' if long_sequence: @@ -478,18 +667,36 @@ def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward config.update({"max_seq_len": 83968}) if 'mosaicml/mpt-7b-chat' in base_model.lower(): config.update({"max_seq_len": 4096}) - if issubclass(config.__class__, tuple(AutoModel._model_mapping.keys())): + if 'mpt-30b' in base_model.lower(): + config.update({"max_seq_len": 2 * 8192}) + if return_model and \ + issubclass(config.__class__, tuple(AutoModel._model_mapping.keys())): model = AutoModel.from_config( config, + trust_remote_code=trust_remote_code, ) else: # can't infer model = None + if 'falcon' in base_model.lower(): + config.use_cache = False + + return config, model + + +def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type, + config, model, + gpu_id=0, + ): + """ + Ensure model gets on correct device + """ if model is not None: # NOTE: Can specify max_memory={0: max_mem, 1: max_mem}, to shard model # NOTE: Some models require avoiding sharding some layers, # then would pass no_split_module_classes and give list of those layers. + from accelerate import infer_auto_device_map device_map = infer_auto_device_map( model, dtype=torch.float16 if load_half else torch.float32, @@ -541,12 +748,59 @@ def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward return model +def get_client_from_inference_server(inference_server, raise_connection_exception=False): + inference_server, headers = get_hf_server(inference_server) + # preload client since slow for gradio case especially + from gradio_utils.grclient import GradioClient + gr_client = None + hf_client = None + if headers is None: + try: + print("GR Client Begin: %s" % inference_server, flush=True) + # first do sanity check if alive, else gradio client takes too long by default + requests.get(inference_server, timeout=int(os.getenv('REQUEST_TIMEOUT', '30'))) + gr_client = GradioClient(inference_server) + print("GR Client End: %s" % inference_server, flush=True) + except (OSError, ValueError) as e: + # Occurs when wrong endpoint and should have been HF client, so don't hard raise, just move to HF + gr_client = None + print("GR Client Failed %s: %s" % (inference_server, str(e)), flush=True) + except (ConnectTimeoutError, ConnectTimeout, MaxRetryError, ConnectionError, ConnectionError2, + JSONDecodeError, ReadTimeout2, KeyError) as e: + t, v, tb = sys.exc_info() + ex = ''.join(traceback.format_exception(t, v, tb)) + print("GR Client Failed %s: %s" % (inference_server, str(ex)), flush=True) + if raise_connection_exception: + raise + + if gr_client is None: + res = None + from text_generation import Client as HFClient + print("HF Client Begin: %s" % inference_server) + try: + hf_client = HFClient(inference_server, headers=headers, timeout=int(os.getenv('REQUEST_TIMEOUT', '30'))) + # quick check valid TGI endpoint + res = hf_client.generate('What?', max_new_tokens=1) + hf_client = HFClient(inference_server, headers=headers, timeout=300) + except (ConnectTimeoutError, ConnectTimeout, MaxRetryError, ConnectionError, ConnectionError2, + JSONDecodeError, ReadTimeout2, KeyError) as e: + hf_client = None + t, v, tb = sys.exc_info() + ex = ''.join(traceback.format_exception(t, v, tb)) + print("HF Client Failed %s: %s" % (inference_server, str(ex))) + if raise_connection_exception: + raise + print("HF Client End: %s %s" % (inference_server, res)) + return inference_server, gr_client, hf_client + + def get_model( load_8bit: bool = False, load_4bit: bool = False, load_half: bool = True, infer_devices: bool = True, base_model: str = '', + inference_server: str = "", tokenizer_base_model: str = '', lora_weights: str = "", gpu_id: int = 0, @@ -570,6 +824,7 @@ def get_model( For non-LORA case, False will spread shards across multiple GPUs, but this can lead to cuda:x cuda:y mismatches So it is not the default :param base_model: name/path of base model + :param inference_server: whether base_model is hosted locally ('') or via http (url) :param tokenizer_base_model: name/path of tokenizer :param lora_weights: name/path :param gpu_id: which GPU (0..n_gpus-1) or allow all GPUs if relevant (-1) @@ -585,11 +840,120 @@ def get_model( """ if verbose: print("Get %s model" % base_model, flush=True) + + triton_attn = False + long_sequence = True + config_kwargs = dict(use_auth_token=use_auth_token, + trust_remote_code=trust_remote_code, + offload_folder=offload_folder, + triton_attn=triton_attn, + long_sequence=long_sequence) + config, _ = get_config(base_model, **config_kwargs, raise_exception=False) + + if base_model in non_hf_types: + assert config is None, "Expected config None for %s" % base_model + + llama_type_from_config = 'llama' in str(config).lower() + llama_type_from_name = "llama" in base_model.lower() + llama_type = llama_type_from_config or llama_type_from_name + if "xgen" in base_model.lower(): + llama_type = False + if llama_type: + if verbose: + print("Detected as llama type from" + " config (%s) or name (%s)" % (llama_type_from_config, llama_type_from_name), flush=True) + + model_loader, tokenizer_loader = get_loaders(model_name=base_model, reward_type=reward_type, llama_type=llama_type) + + tokenizer_kwargs = dict(local_files_only=local_files_only, + resume_download=resume_download, + use_auth_token=use_auth_token, + trust_remote_code=trust_remote_code, + offload_folder=offload_folder, + padding_side='left', + config=config, + ) + if not tokenizer_base_model: + tokenizer_base_model = base_model + + if config is not None and tokenizer_loader is not None and not isinstance(tokenizer_loader, str): + tokenizer = tokenizer_loader.from_pretrained(tokenizer_base_model, **tokenizer_kwargs) + # sets raw (no cushion) limit + set_model_max_len(config, tokenizer, verbose=False) + # if using fake tokenizer, not really accurate when lots of numbers, give a bit of buffer, else get: + # Generation Failed: Input validation error: `inputs` must have less than 2048 tokens. Given: 2233 + tokenizer.model_max_length = tokenizer.model_max_length - 50 + else: + tokenizer = FakeTokenizer() + + if isinstance(inference_server, str) and inference_server.startswith("http"): + inference_server, gr_client, hf_client = get_client_from_inference_server(inference_server) + client = gr_client or hf_client + # Don't return None, None for model, tokenizer so triggers + return client, tokenizer, 'http' + if isinstance(inference_server, str) and inference_server.startswith('openai'): + assert os.getenv('OPENAI_API_KEY'), "Set environment for OPENAI_API_KEY" + # Don't return None, None for model, tokenizer so triggers + # include small token cushion + tokenizer = FakeTokenizer(model_max_length=model_token_mapping[base_model] - 50) + return inference_server, tokenizer, inference_server + assert not inference_server, "Malformed inference_server=%s" % inference_server if base_model in non_hf_types: from gpt4all_llm import get_model_tokenizer_gpt4all model, tokenizer, device = get_model_tokenizer_gpt4all(base_model) return model, tokenizer, device + # get local torch-HF model + return get_hf_model(load_8bit=load_8bit, + load_4bit=load_4bit, + load_half=load_half, + infer_devices=infer_devices, + base_model=base_model, + tokenizer_base_model=tokenizer_base_model, + lora_weights=lora_weights, + gpu_id=gpu_id, + + reward_type=reward_type, + local_files_only=local_files_only, + resume_download=resume_download, + use_auth_token=use_auth_token, + trust_remote_code=trust_remote_code, + offload_folder=offload_folder, + compile_model=compile_model, + + llama_type=llama_type, + config_kwargs=config_kwargs, + tokenizer_kwargs=tokenizer_kwargs, + + verbose=verbose) + + +def get_hf_model(load_8bit: bool = False, + load_4bit: bool = False, + load_half: bool = True, + infer_devices: bool = True, + base_model: str = '', + tokenizer_base_model: str = '', + lora_weights: str = "", + gpu_id: int = 0, + + reward_type: bool = None, + local_files_only: bool = False, + resume_download: bool = True, + use_auth_token: Union[str, bool] = False, + trust_remote_code: bool = True, + offload_folder: str = None, + compile_model: bool = True, + + llama_type: bool = False, + config_kwargs=None, + tokenizer_kwargs=None, + + verbose: bool = False, + ): + assert config_kwargs is not None + assert tokenizer_kwargs is not None + if lora_weights is not None and lora_weights.strip(): if verbose: print("Get %s lora weights" % lora_weights, flush=True) @@ -604,30 +968,13 @@ def get_model( "Please choose a base model with --base_model (CLI) or load one from Models Tab (gradio)" ) - from transformers import AutoConfig - config = AutoConfig.from_pretrained(base_model, use_auth_token=use_auth_token, - trust_remote_code=trust_remote_code, - offload_folder=offload_folder) - llama_type_from_config = 'llama' in str(config).lower() - llama_type_from_name = "llama" in base_model.lower() - llama_type = llama_type_from_config or llama_type_from_name - if llama_type: - if verbose: - print("Detected as llama type from" - " config (%s) or name (%s)" % (llama_type_from_config, llama_type_from_name), flush=True) + model_loader, tokenizer_loader = get_loaders(model_name=base_model, reward_type=reward_type, llama_type=llama_type) - model_loader, tokenizer_loader = get_loaders(llama_type=llama_type, model_name=base_model, reward_type=reward_type) - if not tokenizer_base_model: - tokenizer_base_model = base_model + config, _ = get_config(base_model, return_model=False, raise_exception=True, **config_kwargs) if tokenizer_loader is not None and not isinstance(tokenizer_loader, str): tokenizer = tokenizer_loader.from_pretrained(tokenizer_base_model, - local_files_only=local_files_only, - resume_download=resume_download, - use_auth_token=use_auth_token, - trust_remote_code=trust_remote_code, - offload_folder=offload_folder, - ) + **tokenizer_kwargs) else: tokenizer = tokenizer_loader @@ -651,7 +998,7 @@ def get_model( load_in_4bit=load_4bit, device_map={"": 0} if (load_8bit or load_4bit) and device == 'cuda' else "auto", )) - if 'mpt-' in base_model.lower() and gpu_id >= 0: + if 'mpt-' in base_model.lower() and gpu_id is not None and gpu_id >= 0: model_kwargs.update(dict(device_map={"": gpu_id} if device == 'cuda' else "cpu")) if 'OpenAssistant/reward-model'.lower() in base_model.lower(): @@ -662,27 +1009,33 @@ def get_model( if not lora_weights: with torch.device(device): + if infer_devices: + config, model = get_config(base_model, return_model=True, raise_exception=True, **config_kwargs) model = get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type, + config, model, gpu_id=gpu_id, - use_auth_token=use_auth_token, - trust_remote_code=trust_remote_code, - offload_folder=offload_folder, ) else: + config, _ = get_config(base_model, **config_kwargs) if load_half and not (load_8bit or load_4bit): model = model_loader.from_pretrained( base_model, + config=config, **model_kwargs).half() else: model = model_loader.from_pretrained( base_model, + config=config, **model_kwargs) elif load_8bit or load_4bit: + config, _ = get_config(base_model, **config_kwargs) model = model_loader.from_pretrained( base_model, + config=config, **model_kwargs ) + from peft import PeftModel # loads cuda, so avoid in global scope model = PeftModel.from_pretrained( model, lora_weights, @@ -696,10 +1049,13 @@ def get_model( ) else: with torch.device(device): + config, _ = get_config(base_model, raise_exception=True, **config_kwargs) model = model_loader.from_pretrained( base_model, + config=config, **model_kwargs ) + from peft import PeftModel # loads cuda, so avoid in global scope model = PeftModel.from_pretrained( model, lora_weights, @@ -730,13 +1086,27 @@ def get_model( if torch.__version__ >= "2" and sys.platform != "win32" and compile_model: model = torch.compile(model) - if hasattr(config, 'max_position_embeddings') and isinstance(config.max_position_embeddings, int): + set_model_max_len(config, tokenizer, verbose=False, reward_type=reward_type) + + return model, tokenizer, device + + +def set_model_max_len(config, tokenizer, verbose=False, reward_type=False): + if reward_type: + # limit deberta, else uses too much memory and not worth response score + tokenizer.model_max_length = 512 + if hasattr(config, 'max_seq_len') and isinstance(config.max_seq_len, int): + tokenizer.model_max_length = config.max_seq_len + elif hasattr(config, 'max_position_embeddings') and isinstance(config.max_position_embeddings, int): # help automatically limit inputs to generate tokenizer.model_max_length = config.max_position_embeddings else: + if verbose: + print("Could not determine model_max_length, setting to 2048", flush=True) + tokenizer.model_max_length = 2048 + # for bug in HF transformers + if tokenizer.model_max_length > 100000000: tokenizer.model_max_length = 2048 - - return model, tokenizer, device def pop_unused_model_kwargs(model_kwargs): @@ -758,6 +1128,7 @@ def get_score_model(score_model: str = None, load_half: bool = True, infer_devices: bool = True, base_model: str = '', + inference_server: str = '', tokenizer_base_model: str = '', lora_weights: str = "", gpu_id: int = 0, @@ -779,6 +1150,7 @@ def get_score_model(score_model: str = None, base_model = score_model.strip() tokenizer_base_model = '' lora_weights = '' + inference_server = '' llama_type = False compile_model = False smodel, stokenizer, sdevice = get_model(reward_type=True, @@ -788,30 +1160,155 @@ def get_score_model(score_model: str = None, return smodel, stokenizer, sdevice +no_default_param_names = [ + 'instruction', + 'iinput', + 'context', + 'instruction_nochat', + 'iinput_nochat', +] + +gen_hyper = ['temperature', + 'top_p', + 'top_k', + 'num_beams', + 'max_new_tokens', + 'min_new_tokens', + 'early_stopping', + 'max_time', + 'repetition_penalty', + 'num_return_sequences', + 'do_sample', + ] + eval_func_param_names = ['instruction', 'iinput', 'context', 'stream_output', 'prompt_type', - 'temperature', - 'top_p', - 'top_k', - 'num_beams', - 'max_new_tokens', - 'min_new_tokens', - 'early_stopping', - 'max_time', - 'repetition_penalty', - 'num_return_sequences', - 'do_sample', - 'chat', + 'prompt_dict'] + \ + gen_hyper + \ + ['chat', 'instruction_nochat', 'iinput_nochat', 'langchain_mode', 'top_k_docs', + 'chunk', + 'chunk_size', 'document_choice', ] +# form evaluate defaults for submit_nochat_api +eval_func_param_names_defaults = eval_func_param_names.copy() +for k in no_default_param_names: + if k in eval_func_param_names_defaults: + eval_func_param_names_defaults.remove(k) + + +def evaluate_from_str( + model_state, + my_db_state, + # START NOTE: Examples must have same order of parameters + user_kwargs, + # END NOTE: Examples must have same order of parameters + default_kwargs=None, + src_lang=None, + tgt_lang=None, + debug=False, + concurrency_count=None, + save_dir=None, + sanitize_bot_response=False, + model_state0=None, + memory_restriction_level=None, + max_max_new_tokens=None, + is_public=None, + max_max_time=None, + raise_generate_gpu_exceptions=None, + chat_context=None, + lora_weights=None, + load_db_if_exists=True, + dbs=None, + user_path=None, + detect_user_path_changes_every_query=None, + use_openai_embedding=None, + use_openai_model=None, + hf_embedding_model=None, + db_type=None, + n_jobs=None, + first_para=None, + text_limit=None, + verbose=False, + cli=False, + reverse_docs=True, + use_cache=None, + auto_reduce_chunks=None, + max_chunks=None, + model_lock=None, + force_langchain_evaluate=None, + model_state_none=None, +): + if isinstance(user_kwargs, str): + user_kwargs = ast.literal_eval(user_kwargs) + # only used for submit_nochat_api + user_kwargs['chat'] = False + if 'stream_output' not in user_kwargs: + user_kwargs['stream_output'] = False + if 'langchain_mode' not in user_kwargs: + # if user doesn't specify, then assume disabled, not use default + user_kwargs['langchain_mode'] = 'Disabled' + + assert set(list(default_kwargs.keys())) == set(eval_func_param_names) + # correct ordering. Note some things may not be in default_kwargs, so can't be default of user_kwargs.get() + args_list = [user_kwargs[k] if k in user_kwargs else default_kwargs[k] for k in eval_func_param_names] + + ret = evaluate( + model_state, + my_db_state, + # START NOTE: Examples must have same order of parameters + *tuple(args_list), + # END NOTE: Examples must have same order of parameters + src_lang=src_lang, + tgt_lang=tgt_lang, + debug=debug, + concurrency_count=concurrency_count, + save_dir=save_dir, + sanitize_bot_response=sanitize_bot_response, + model_state0=model_state0, + memory_restriction_level=memory_restriction_level, + max_max_new_tokens=max_max_new_tokens, + is_public=is_public, + max_max_time=max_max_time, + raise_generate_gpu_exceptions=raise_generate_gpu_exceptions, + chat_context=chat_context, + lora_weights=lora_weights, + load_db_if_exists=load_db_if_exists, + dbs=dbs, + user_path=user_path, + detect_user_path_changes_every_query=detect_user_path_changes_every_query, + use_openai_embedding=use_openai_embedding, + use_openai_model=use_openai_model, + hf_embedding_model=hf_embedding_model, + db_type=db_type, + n_jobs=n_jobs, + first_para=first_para, + text_limit=text_limit, + verbose=verbose, + cli=cli, + reverse_docs=reverse_docs, + use_cache=use_cache, + auto_reduce_chunks=auto_reduce_chunks, + max_chunks=max_chunks, + model_lock=model_lock, + force_langchain_evaluate=force_langchain_evaluate, + model_state_none=model_state_none, + ) + try: + for ret1 in ret: + yield ret1 + finally: + # clear before return, in finally in case GPU OOM exception + clear_torch_cache() + def evaluate( model_state, @@ -822,6 +1319,7 @@ def evaluate( context, stream_output, prompt_type, + prompt_dict, temperature, top_p, top_k, @@ -838,6 +1336,8 @@ def evaluate( iinput_nochat, langchain_mode, top_k_docs, + chunk, + chunk_size, document_choice, # END NOTE: Examples must have same order of parameters src_lang=None, @@ -845,9 +1345,12 @@ def evaluate( debug=False, concurrency_count=None, save_dir=None, - sanitize_bot_response=True, + sanitize_bot_response=False, model_state0=None, memory_restriction_level=None, + max_max_new_tokens=None, + is_public=None, + max_max_time=None, raise_generate_gpu_exceptions=None, chat_context=None, lora_weights=None, @@ -858,14 +1361,19 @@ def evaluate( use_openai_embedding=None, use_openai_model=None, hf_embedding_model=None, - chunk=None, - chunk_size=None, db_type=None, n_jobs=None, first_para=None, text_limit=None, verbose=False, cli=False, + reverse_docs=True, + use_cache=None, + auto_reduce_chunks=None, + max_chunks=None, + model_lock=None, + force_langchain_evaluate=None, + model_state_none=None, ): # ensure passed these assert concurrency_count is not None @@ -875,10 +1383,10 @@ def evaluate( assert use_openai_embedding is not None assert use_openai_model is not None assert hf_embedding_model is not None - assert chunk is not None - assert chunk_size is not None assert db_type is not None assert top_k_docs is not None and isinstance(top_k_docs, int) + assert chunk is not None and isinstance(chunk, bool) + assert chunk_size is not None and isinstance(chunk_size, int) assert n_jobs is not None assert first_para is not None @@ -886,29 +1394,58 @@ def evaluate( locals_dict = locals().copy() locals_dict.pop('model_state', None) locals_dict.pop('model_state0', None) + locals_dict.pop('model_states', None) print(locals_dict) - no_model_msg = "Please choose a base model with --base_model (CLI) or load in Models Tab (gradio).\nThen start New Conversation" + no_model_msg = "Please choose a base model with --base_model (CLI) or load in Models Tab (gradio).\n" \ + "Then start New Conversation" + if model_state is None: + model_state = model_state_none.copy() if model_state0 is None: # e.g. for no gradio case, set dummy value, else should be set - model_state0 = [None, None, None, None] - - if model_state is not None and len(model_state) == 4 and not isinstance(model_state[0], str): - # try to free-up original model (i.e. list was passed as reference) - if model_state0 is not None and model_state0[0] is not None: - model_state0[0].cpu() - model_state0[0] = None - # try to free-up original tokenizer (i.e. list was passed as reference) - if model_state0 is not None and model_state0[1] is not None: - model_state0[1] = None - clear_torch_cache() - model, tokenizer, device, base_model = model_state - elif model_state0 is not None and len(model_state0) == 4 and model_state0[0] is not None: - assert isinstance(model_state[0], str) - model, tokenizer, device, base_model = model_state0 + model_state0 = model_state_none.copy() + + # model_state['model] is only 'model' if should use model_state0 + # model could also be None + have_model_lock = model_lock is not None + have_fresh_model = model_state['model'] not in [None, 'model', no_model_str] + # for gradio UI control, expect model_state and model_state0 to match, so if have_model_lock=True, then should have_fresh_model=True + # but gradio API control will only use nochat api etc. and won't use fresh model, so can't assert in general + # if have_model_lock: + # assert have_fresh_model, "Expected model_state and model_state0 to match if have_model_lock" + have_cli_model = model_state0['model'] not in [None, 'model', no_model_str] + + if have_fresh_model: + # USE FRESH MODEL + if not have_model_lock: + # model_state0 is just one of model_state if model_lock, so don't nuke + # try to free-up original model (i.e. list was passed as reference) + if model_state0['model'] and hasattr(model_state0['model'], 'cpu'): + model_state0['model'].cpu() + model_state0['model'] = None + # try to free-up original tokenizer (i.e. list was passed as reference) + if model_state0['tokenizer']: + model_state0['tokenizer'] = None + clear_torch_cache() + chosen_model_state = model_state + elif have_cli_model: + # USE MODEL SETUP AT CLI + assert isinstance(model_state['model'], str) # expect no fresh model + chosen_model_state = model_state0 else: raise AssertionError(no_model_msg) + # get variables + model = chosen_model_state['model'] + tokenizer = chosen_model_state['tokenizer'] + device = chosen_model_state['device'] + base_model = chosen_model_state['base_model'] + tokenizer_base_model = chosen_model_state['tokenizer_base_model'] + lora_weights = chosen_model_state['lora_weights'] + inference_server = chosen_model_state['inference_server'] + # prefer use input from API over model state + prompt_type = prompt_type or chosen_model_state['prompt_type'] + prompt_dict = prompt_dict or chosen_model_state['prompt_dict'] if base_model is None: raise AssertionError(no_model_msg) @@ -922,11 +1459,49 @@ def evaluate( instruction = instruction_nochat iinput = iinput_nochat + # in some cases, like lean nochat API, don't want to force sending prompt_type, allow default choice + model_lower = base_model.lower() + if not prompt_type and model_lower in inv_prompt_type_to_model_lower: + prompt_type = inv_prompt_type_to_model_lower[model_lower] + if verbose: + print("Auto-selecting prompt_type=%s for %s" % (prompt_type, model_lower), flush=True) + assert prompt_type is not None, "prompt_type was None" + + # Control generation hyperparameters + # adjust for bad inputs, e.g. in case also come from API that doesn't get constrained by gradio sliders + # below is for TGI server, not required for HF transformers + # limits are chosen similar to gradio_runner.py sliders/numbers + top_p = min(max(1e-3, top_p), 1.0 - 1e-3) + top_k = min(max(1, int(top_k)), 100) + temperature = min(max(0.01, temperature), 2.0) + # FIXME: https://github.com/h2oai/h2ogpt/issues/106 + num_beams = 1 if stream_output else num_beams # See max_beams in gradio_runner + max_max_new_tokens = get_max_max_new_tokens(chosen_model_state, + memory_restriction_level=memory_restriction_level, + max_new_tokens=max_new_tokens, + max_max_new_tokens=max_max_new_tokens) + model_max_length = get_model_max_length(chosen_model_state) + max_new_tokens = min(max(1, int(max_new_tokens)), max_max_new_tokens) + min_new_tokens = min(max(0, int(min_new_tokens)), max_new_tokens) + max_time = min(max(0, max_time), max_max_time) + repetition_penalty = min(max(0.01, repetition_penalty), 3.0) + num_return_sequences = 1 if chat else min(max(1, int(num_return_sequences)), 10) + min_top_k_docs, max_top_k_docs, label_top_k_docs = get_minmax_top_k_docs(is_public) + top_k_docs = min(max(min_top_k_docs, int(top_k_docs)), max_top_k_docs) + chunk_size = min(max(128, int(chunk_size)), 2048) if not context: # get hidden context if have one context = get_context(chat_context, prompt_type) - prompter = Prompter(prompt_type, debug=debug, chat=chat, stream_output=stream_output) + # restrict instruction, typically what has large input + from h2oai_pipeline import H2OTextGenerationPipeline + instruction, num_prompt_tokens1 = H2OTextGenerationPipeline.limit_prompt(instruction, tokenizer) + context, num_prompt_tokens2 = H2OTextGenerationPipeline.limit_prompt(context, tokenizer) + iinput, num_prompt_tokens3 = H2OTextGenerationPipeline.limit_prompt(iinput, tokenizer) + num_prompt_tokens = (num_prompt_tokens1 or 0) + (num_prompt_tokens2 or 0) + (num_prompt_tokens3 or 0) + + # get prompt + prompter = Prompter(prompt_type, prompt_dict, debug=debug, chat=chat, stream_output=stream_output) data_point = dict(context=context, instruction=instruction, input=iinput) prompt = prompter.generate_prompt(data_point) @@ -938,20 +1513,36 @@ def evaluate( db1 = dbs[langchain_mode] else: db1 = None - if langchain_mode not in [False, 'Disabled', 'ChatLLM', 'LLM'] and db1 is not None or base_model in non_hf_types: + do_langchain_path = langchain_mode not in [False, 'Disabled', 'ChatLLM', 'LLM'] and \ + db1 is not None or \ + base_model in non_hf_types or \ + force_langchain_evaluate + if do_langchain_path: query = instruction if not iinput else "%s\n%s" % (instruction, iinput) outr = "" # use smaller cut_distanct for wiki_full since so many matches could be obtained, and often irrelevant unless close from gpt_langchain import run_qa_db + gen_hyper_langchain = dict(do_sample=do_sample, + temperature=temperature, + repetition_penalty=repetition_penalty, + top_k=top_k, + top_p=top_p, + num_beams=num_beams, + min_new_tokens=min_new_tokens, + max_new_tokens=max_new_tokens, + early_stopping=early_stopping, + max_time=max_time, + num_return_sequences=num_return_sequences, + ) for r in run_qa_db(query=query, model_name=base_model, model=model, tokenizer=tokenizer, + inference_server=inference_server, stream_output=stream_output, prompter=prompter, load_db_if_exists=load_db_if_exists, db=db1, user_path=user_path, detect_user_path_changes_every_query=detect_user_path_changes_every_query, - max_new_tokens=max_new_tokens, cut_distanct=1.1 if langchain_mode in ['wiki_full'] else 1.64, # FIXME, too arbitrary use_openai_embedding=use_openai_embedding, use_openai_model=use_openai_model, @@ -963,20 +1554,33 @@ def evaluate( langchain_mode=langchain_mode, document_choice=document_choice, db_type=db_type, - k=top_k_docs, - temperature=temperature, - repetition_penalty=repetition_penalty, - top_k=top_k, - top_p=top_p, + top_k_docs=top_k_docs, + + **gen_hyper_langchain, + prompt_type=prompt_type, + prompt_dict=prompt_dict, n_jobs=n_jobs, verbose=verbose, cli=cli, + sanitize_bot_response=sanitize_bot_response, + reverse_docs=reverse_docs, + + lora_weights=lora_weights, + + auto_reduce_chunks=auto_reduce_chunks, + max_chunks=max_chunks, ): outr, extra = r # doesn't accumulate, new answer every yield, so only save that full answer yield dict(response=outr, sources=extra) if save_dir: - save_generate_output(output=outr, base_model=base_model, save_dir=save_dir) + extra_dict = gen_hyper_langchain.copy() + extra_dict.update(prompt_type=prompt_type, inference_server=inference_server, + langchain_mode=langchain_mode, document_choice=document_choice, + num_prompt_tokens=num_prompt_tokens) + save_generate_output(prompt=query, output=outr, base_model=base_model, save_dir=save_dir, + where_from='run_qa_db', + extra_dict=extra_dict) if verbose: print( 'Post-Generate Langchain: %s decoded_output: %s' % (str(datetime.now()), len(outr) if outr else -1), @@ -985,8 +1589,270 @@ def evaluate( # if got no response (e.g. not showing sources and got no sources, # so nothing to give to LLM), then slip through and ask LLM # Or if llama/gptj, then just return since they had no response and can't go down below code path + # clear before return, since .then() never done if from API + clear_torch_cache() return + if inference_server.startswith('openai') or inference_server.startswith('http'): + if inference_server.startswith('openai'): + import openai + where_from = "openai_client" + + openai.api_key = os.getenv("OPENAI_API_KEY") + stop_sequences = list(set(prompter.terminate_response + [prompter.PreResponse])) + # OpenAI will complain if ask for too many new tokens, takes it as min in some sense, wrongly so. + max_new_tokens_openai = min(max_new_tokens, model_max_length - num_prompt_tokens) + gen_server_kwargs = dict(temperature=temperature if do_sample else 0, + max_tokens=max_new_tokens_openai, + top_p=top_p if do_sample else 1, + frequency_penalty=0, + n=num_return_sequences, + presence_penalty=1.07 - repetition_penalty + 0.6, # so good default + ) + if inference_server == 'openai': + response = openai.Completion.create( + model=base_model, + prompt=prompt, + **gen_server_kwargs, + stop=stop_sequences, + stream=stream_output, + ) + if not stream_output: + text = response['choices'][0]['text'] + yield dict(response=prompter.get_response(prompt + text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources='') + else: + collected_events = [] + text = '' + for event in response: + collected_events.append(event) # save the event response + event_text = event['choices'][0]['text'] # extract the text + text += event_text # append the text + yield dict(response=prompter.get_response(prompt + text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources='') + elif inference_server == 'openai_chat': + response = openai.ChatCompletion.create( + model=base_model, + messages=[ + {"role": "system", "content": "You are a helpful assistant."}, + {'role': 'user', + 'content': prompt, + } + ], + stream=stream_output, + **gen_server_kwargs, + ) + if not stream_output: + text = response["choices"][0]["message"]["content"] + yield dict(response=prompter.get_response(prompt + text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources='') + else: + text = "" + for chunk in response: + delta = chunk["choices"][0]["delta"] + if 'content' in delta: + text += delta['content'] + yield dict(response=prompter.get_response(prompt + text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources='') + else: + raise RuntimeError("No such OpenAI mode: %s" % inference_server) + elif inference_server.startswith('http'): + inference_server, headers = get_hf_server(inference_server) + from gradio_utils.grclient import GradioClient + from text_generation import Client as HFClient + if isinstance(model, GradioClient): + gr_client = model + hf_client = None + elif isinstance(model, HFClient): + gr_client = None + hf_client = model + else: + inference_server, gr_client, hf_client = get_client_from_inference_server(inference_server) + + # quick sanity check to avoid long timeouts, just see if can reach server + requests.get(inference_server, timeout=int(os.getenv('REQUEST_TIMEOUT_FAST', '10'))) + + if gr_client is not None: + # Note: h2oGPT gradio server could handle input token size issues for prompt, + # but best to handle here so send less data to server + + chat_client = False + where_from = "gr_client" + client_langchain_mode = 'Disabled' + gen_server_kwargs = dict(temperature=temperature, + top_p=top_p, + top_k=top_k, + num_beams=num_beams, + max_new_tokens=max_new_tokens, + min_new_tokens=min_new_tokens, + early_stopping=early_stopping, + max_time=max_time, + repetition_penalty=repetition_penalty, + num_return_sequences=num_return_sequences, + do_sample=do_sample, + chat=chat_client, + ) + # account for gradio into gradio that handles prompting, avoid duplicating prompter prompt injection + if prompt_type in [None, '', PromptType.plain.name, PromptType.plain.value, + str(PromptType.plain.value)]: + # if our prompt is plain, assume either correct or gradio server knows different prompt type, + # so pass empty prompt_Type + gr_prompt_type = '' + gr_prompt_dict = '' + gr_prompt = prompt # already prepared prompt + gr_context = '' + gr_iinput = '' + else: + # if already have prompt_type that is not plain, None, or '', then already applied some prompting + # But assume server can handle prompting, and need to avoid double-up. + # Also assume server can do better job of using stopping.py to stop early, so avoid local prompting, let server handle + # So avoid "prompt" and let gradio server reconstruct from prompt_type we passed + # Note it's ok that prompter.get_response() has prompt+text, prompt=prompt passed, + # because just means extra processing and removal of prompt, but that has no human-bot prompting doesn't matter + # since those won't appear + gr_context = context + gr_prompt = instruction + gr_iinput = iinput + gr_prompt_type = prompt_type + gr_prompt_dict = prompt_dict + client_kwargs = dict(instruction=gr_prompt if chat_client else '', # only for chat=True + iinput=gr_iinput, # only for chat=True + context=gr_context, + # streaming output is supported, loops over and outputs each generation in streaming mode + # but leave stream_output=False for simple input/output mode + stream_output=stream_output, + + **gen_server_kwargs, + + prompt_type=gr_prompt_type, + prompt_dict=gr_prompt_dict, + + instruction_nochat=gr_prompt if not chat_client else '', + iinput_nochat=gr_iinput, # only for chat=False + langchain_mode=client_langchain_mode, + top_k_docs=top_k_docs, + chunk=chunk, + chunk_size=chunk_size, + document_choice=[DocumentChoices.All_Relevant.name], + ) + api_name = '/submit_nochat_api' # NOTE: like submit_nochat but stable API for string dict passing + if not stream_output: + res = gr_client.predict(str(dict(client_kwargs)), api_name=api_name) + res_dict = ast.literal_eval(res) + text = res_dict['response'] + sources = res_dict['sources'] + yield dict(response=prompter.get_response(prompt + text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources=sources) + else: + job = gr_client.submit(str(dict(client_kwargs)), api_name=api_name) + text = '' + sources = '' + res_dict = dict(response=text, sources=sources) + while not job.done(): + outputs_list = job.communicator.job.outputs + if outputs_list: + res = job.communicator.job.outputs[-1] + res_dict = ast.literal_eval(res) + text = res_dict['response'] + sources = res_dict['sources'] + if gr_prompt_type == 'plain': + # then gradio server passes back full prompt + text + prompt_and_text = text + else: + prompt_and_text = prompt + text + yield dict(response=prompter.get_response(prompt_and_text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources=sources) + time.sleep(0.01) + # ensure get last output to avoid race + res_all = job.outputs() + if len(res_all) > 0: + res = res_all[-1] + res_dict = ast.literal_eval(res) + text = res_dict['response'] + sources = res_dict['sources'] + else: + # go with old text if last call didn't work + e = job.future._exception + if e is not None: + stre = str(e) + strex = ''.join(traceback.format_tb(e.__traceback__)) + else: + stre = '' + strex = '' + + print("Bad final response: %s %s %s %s %s: %s %s" % (base_model, inference_server, + res_all, prompt, text, stre, strex), + flush=True) + if gr_prompt_type == 'plain': + # then gradio server passes back full prompt + text + prompt_and_text = text + else: + prompt_and_text = prompt + text + yield dict(response=prompter.get_response(prompt_and_text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources=sources) + elif hf_client: + # HF inference server needs control over input tokens + where_from = "hf_client" + + # prompt must include all human-bot like tokens, already added by prompt + # https://github.com/huggingface/text-generation-inference/tree/main/clients/python#types + stop_sequences = list(set(prompter.terminate_response + [prompter.PreResponse])) + gen_server_kwargs = dict(do_sample=do_sample, + max_new_tokens=max_new_tokens, + # best_of=None, + repetition_penalty=repetition_penalty, + return_full_text=True, + seed=SEED, + stop_sequences=stop_sequences, + temperature=temperature, + top_k=top_k, + top_p=top_p, + # truncate=False, # behaves oddly + # typical_p=top_p, + # watermark=False, + # decoder_input_details=False, + ) + # work-around for timeout at constructor time, will be issue if multi-threading, + # so just do something reasonable or max_time if larger + # lower bound because client is re-used if multi-threading + hf_client.timeout = max(300, max_time) + if not stream_output: + text = hf_client.generate(prompt, **gen_server_kwargs).generated_text + yield dict(response=prompter.get_response(text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources='') + else: + text = "" + for response in hf_client.generate_stream(prompt, **gen_server_kwargs): + if not response.token.special: + # stop_sequences + text_chunk = response.token.text + text += text_chunk + yield dict(response=prompter.get_response(prompt + text, prompt=prompt, + sanitize_bot_response=sanitize_bot_response), + sources='') + else: + raise RuntimeError("Failed to get client: %s" % inference_server) + else: + raise RuntimeError("No such inference_server %s" % inference_server) + + if save_dir and text: + # save prompt + new text + extra_dict = gen_server_kwargs.copy() + extra_dict.update(dict(inference_server=inference_server, num_prompt_tokens=num_prompt_tokens)) + save_generate_output(prompt=prompt, output=text, base_model=base_model, save_dir=save_dir, + where_from=where_from, extra_dict=extra_dict) + return + else: + assert not inference_server, "inferene_server=%s not supported" % inference_server + if isinstance(tokenizer, str): # pipeline if tokenizer == "summarization": @@ -1000,36 +1866,37 @@ def evaluate( assert src_lang is not None tokenizer.src_lang = languages_covered()[src_lang] - if chat: - # override, ignore user change - num_return_sequences = 1 - stopping_criteria = get_stopping(prompt_type, tokenizer, device) - _, _, max_length_tokenize, max_prompt_length = get_cutoffs(memory_restriction_level, model_max_length=tokenizer.model_max_length) - prompt = prompt[-max_prompt_length:] - inputs = tokenizer(prompt, - return_tensors="pt", - truncation=True, - max_length=max_length_tokenize) - if inputs['input_ids'].shape[1] >= max_length_tokenize - 1: - print("Cutting off input: %s %s" % (inputs['input_ids'].shape[1], max_length_tokenize), flush=True) + stopping_criteria = get_stopping(prompt_type, prompt_dict, tokenizer, device, + model_max_length=tokenizer.model_max_length) + + inputs = tokenizer(prompt, return_tensors="pt") if debug and len(inputs["input_ids"]) > 0: print('input_ids length', len(inputs["input_ids"][0]), flush=True) input_ids = inputs["input_ids"].to(device) # CRITICAL LIMIT else will fail max_max_tokens = tokenizer.model_max_length - max_input_tokens = max_max_tokens - max_new_tokens + max_input_tokens = max_max_tokens - min_new_tokens + # NOTE: Don't limit up front due to max_new_tokens, let go up to max or reach max_max_tokens in stopping.py input_ids = input_ids[:, -max_input_tokens:] - generation_config = GenerationConfig( - temperature=float(temperature), - top_p=float(top_p), - top_k=top_k, - num_beams=num_beams, - do_sample=do_sample, - repetition_penalty=float(repetition_penalty), - num_return_sequences=num_return_sequences, - renormalize_logits=True, - remove_invalid_values=True, - ) + # required for falcon if multiple threads or asyncio accesses to model during generation + if use_cache is None: + use_cache = False if 'falcon' in base_model else True + gen_config_kwargs = dict(temperature=float(temperature), + top_p=float(top_p), + top_k=top_k, + num_beams=num_beams, + do_sample=do_sample, + repetition_penalty=float(repetition_penalty), + num_return_sequences=num_return_sequences, + renormalize_logits=True, + remove_invalid_values=True, + use_cache=use_cache, + ) + token_ids = ['eos_token_id', 'pad_token_id', 'bos_token_id', 'cls_token_id', 'sep_token_id'] + for token_id in token_ids: + if hasattr(tokenizer, token_id) and getattr(tokenizer, token_id) is not None: + gen_config_kwargs.update({token_id: getattr(tokenizer, token_id)}) + generation_config = GenerationConfig(**gen_config_kwargs) gen_kwargs = dict(input_ids=input_ids, generation_config=generation_config, @@ -1048,7 +1915,10 @@ def evaluate( tgt_lang = languages_covered()[tgt_lang] gen_kwargs.update(dict(forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang])) else: - gen_kwargs.update(dict(pad_token_id=tokenizer.eos_token_id)) + token_ids = ['eos_token_id', 'bos_token_id', 'pad_token_id'] + for token_id in token_ids: + if hasattr(tokenizer, token_id) and getattr(tokenizer, token_id) is not None: + gen_kwargs.update({token_id: getattr(tokenizer, token_id)}) decoder_kwargs = dict(skip_special_tokens=True, clean_up_tokenization_spaces=True) @@ -1064,7 +1934,8 @@ def evaluate( ) with torch.no_grad(): - context_class_cast = NullContext if device == 'cpu' or lora_weights else torch.autocast + have_lora_weights = lora_weights not in [no_lora_str, '', None] + context_class_cast = NullContext if device == 'cpu' or have_lora_weights else torch.autocast with context_class_cast(device): # protection for gradio not keeping track of closed users, # else hit bitsandbytes lack of thread safety: @@ -1127,6 +1998,8 @@ def evaluate( raise thread.exc raise finally: + # clear before return, since .then() never done if from API + clear_torch_cache() # in case no exception and didn't join with thread yet, then join if not thread.exc: thread.join() @@ -1135,14 +2008,21 @@ def evaluate( raise thread.exc decoded_output = outputs else: - outputs = model.generate(**gen_kwargs) + try: + outputs = model.generate(**gen_kwargs) + finally: + clear_torch_cache() # has to be here for API submit_nochat_api since.then() not called outputs = [decoder(s) for s in outputs.sequences] yield dict(response=prompter.get_response(outputs, prompt=inputs_decoded, sanitize_bot_response=sanitize_bot_response), sources='') if outputs and len(outputs) >= 1: decoded_output = prompt + outputs[0] if save_dir and decoded_output: - save_generate_output(output=decoded_output, base_model=base_model, save_dir=save_dir) + extra_dict = gen_config_kwargs.copy() + extra_dict.update(dict(num_prompt_tokens=num_prompt_tokens)) + save_generate_output(prompt=prompt, output=decoded_output, base_model=base_model, save_dir=save_dir, + where_from="evaluate_%s" % str(stream_output), + extra_dict=gen_config_kwargs) if verbose: print('Post-Generate: %s decoded_output: %s' % ( str(datetime.now()), len(decoded_output) if decoded_output else -1), flush=True) @@ -1161,6 +2041,7 @@ def get_cutoffs(memory_restriction_level, for_context=False, model_max_length=20 if memory_restriction_level > 0: max_length_tokenize = 768 - 256 if memory_restriction_level <= 2 else 512 - 256 else: + # at least give room for 1 paragraph output max_length_tokenize = model_max_length - 256 cutoff_len = max_length_tokenize * 4 # if reaches limit, then can't generate new tokens output_smallest = 30 * 4 @@ -1251,18 +2132,22 @@ def generate_with_exceptions(func, *args, prompt='', inputs_decoded='', raise_ge def get_generate_params(model_lower, chat, stream_output, show_examples, - prompt_type, temperature, top_p, top_k, num_beams, + prompt_type, prompt_dict, + temperature, top_p, top_k, num_beams, max_new_tokens, min_new_tokens, early_stopping, max_time, repetition_penalty, num_return_sequences, - do_sample, k, verbose): + do_sample, + top_k_docs, chunk, chunk_size, + verbose): use_defaults = False use_default_examples = True examples = [] - task_info = f"{prompt_type}" + task_info = 'LLM' if model_lower: print(f"Using Model {model_lower}", flush=True) else: - print("No model defined yet", flush=True) + if verbose: + print("No model defined yet", flush=True) min_new_tokens = min_new_tokens if min_new_tokens is not None else 0 early_stopping = early_stopping if early_stopping is not None else False @@ -1288,15 +2173,13 @@ Jeff: and how can I get started? Jeff: where can I find documentation? Philipp: ok, ok you can find everything here. https://huggingface.co/blog/the-partnership-amazon-sagemaker-and-hugging-face""" + use_placeholder_instruction_as_example = False if 'bart-large-cnn-samsum' in model_lower or 'flan-t5-base-samsum' in model_lower: placeholder_instruction = summarize_example1 placeholder_input = "" use_defaults = True use_default_examples = False - examples += [ - [placeholder_instruction, "", "", stream_output, 'plain', 1.0, 1.0, 50, 1, 128, 0, False, max_time_defaults, - 1.0, 1, - False]] + use_placeholder_instruction_as_example = True task_info = "Summarization" elif 't5-' in model_lower or 't5' == model_lower or 'flan-' in model_lower: placeholder_instruction = "The square root of x is the cube root of y. What is y to the power of 2, if x = 4?" @@ -1309,29 +2192,25 @@ Philipp: ok, ok you can find everything here. https://huggingface.co/blog/the-pa placeholder_input = "" use_defaults = True use_default_examples = False - examples += [ - [placeholder_instruction, "", "", stream_output, 'plain', 1.0, 1.0, 50, 1, 128, 0, False, max_time_defaults, - 1.0, 1, - False]] + use_placeholder_instruction_as_example = True elif 'gpt2' in model_lower: placeholder_instruction = "The sky is" placeholder_input = "" prompt_type = prompt_type or 'plain' use_default_examples = True # some will be odd "continuations" but can be ok - examples += [ - [placeholder_instruction, "", "", stream_output, 'plain', 1.0, 1.0, 50, 1, 128, 0, False, max_time_defaults, - 1.0, 1, - False]] + use_placeholder_instruction_as_example = True task_info = "Auto-complete phrase, code, etc." use_defaults = True else: if chat: - placeholder_instruction = "Enter a question or imperative." + placeholder_instruction = "" else: placeholder_instruction = "Give detailed answer for whether Einstein or Newton is smarter." placeholder_input = "" - if model_lower: - # default is plain, because might relly upon trust_remote_code to handle prompting + if model_lower in inv_prompt_type_to_model_lower: + prompt_type = inv_prompt_type_to_model_lower[model_lower] + elif model_lower: + # default is plain, because might rely upon trust_remote_code to handle prompting prompt_type = prompt_type or 'plain' else: prompt_type = '' @@ -1361,18 +2240,22 @@ Philipp: ok, ok you can find everything here. https://huggingface.co/blog/the-pa temperature = 0.1 if temperature is None else temperature top_p = 0.75 if top_p is None else top_p top_k = 40 if top_k is None else top_k - if chat: - num_beams = num_beams or 1 - else: - num_beams = num_beams or 4 + num_beams = num_beams or 1 max_new_tokens = max_new_tokens or 256 repetition_penalty = repetition_penalty or 1.07 num_return_sequences = min(num_beams, num_return_sequences or 1) do_sample = False if do_sample is None else do_sample # doesn't include chat, instruction_nochat, iinput_nochat, added later - params_list = ["", stream_output, prompt_type, temperature, top_p, top_k, num_beams, max_new_tokens, min_new_tokens, + params_list = ["", + stream_output, + prompt_type, prompt_dict, + temperature, top_p, top_k, num_beams, + max_new_tokens, min_new_tokens, early_stopping, max_time, repetition_penalty, num_return_sequences, do_sample] + if use_placeholder_instruction_as_example: + examples += [[placeholder_instruction, ''] + params_list] + if use_default_examples: examples += [ ["Translate English to French", "Good morning"] + params_list, @@ -1410,14 +2293,15 @@ y = np.random.randint(0, 1, 100) # fit random forest classifier with 20 estimators""", ''] + params_list, ] # add summary example - examples += [[summarize_example1, 'Summarize' if prompt_type not in ['plain', 'instruct_simple'] else ''] + params_list] + examples += [ + [summarize_example1, 'Summarize' if prompt_type not in ['plain', 'instruct_simple'] else ''] + params_list] src_lang = "English" tgt_lang = "Russian" # move to correct position for example in examples: - example += [chat, '', '', 'Disabled', k, ['All']] + example += [chat, '', '', 'Disabled', top_k_docs, chunk, chunk_size, [DocumentChoices.All_Relevant.name]] # adjust examples if non-chat mode if not chat: example[eval_func_param_names.index('instruction_nochat')] = example[ @@ -1429,9 +2313,19 @@ y = np.random.randint(0, 1, 100) assert len(example) == len(eval_func_param_names), "Wrong example: %s %s" % ( len(example), len(eval_func_param_names)) + if prompt_type == PromptType.custom.name and not prompt_dict: + raise ValueError("Unexpected to get non-empty prompt_dict=%s for prompt_type=%s" % (prompt_dict, prompt_type)) + + # get prompt_dict from prompt_type, so user can see in UI etc., or for custom do nothing except check format + prompt_dict, error0 = get_prompt(prompt_type, prompt_dict, + chat=False, context='', reduced=False, making_context=False, return_dict=True) + if error0: + raise RuntimeError("Prompt wrong: %s" % error0) + return placeholder_instruction, placeholder_input, \ stream_output, show_examples, \ - prompt_type, temperature, top_p, top_k, num_beams, \ + prompt_type, prompt_dict, \ + temperature, top_p, top_k, num_beams, \ max_new_tokens, min_new_tokens, early_stopping, max_time, \ repetition_penalty, num_return_sequences, \ do_sample, \ @@ -1477,7 +2371,8 @@ def score_qa(smodel, stokenizer, max_length_tokenize, question, answer, cutoff_l if 'Expected all tensors to be on the same device' in str(e) or \ 'expected scalar type Half but found Float' in str(e) or \ 'probability tensor contains either' in str(e) or \ - 'cublasLt ran into an error!' in str(e): + 'cublasLt ran into an error!' in str(e) or \ + 'device-side assert triggered' in str(e): print("GPU Error: question: %s answer: %s exception: %s" % (question, answer, str(e)), flush=True) traceback.print_exc() @@ -1491,12 +2386,7 @@ def score_qa(smodel, stokenizer, max_length_tokenize, question, answer, cutoff_l def check_locals(**kwargs): # ensure everything in evaluate is here - can_skip_because_locally_generated = [ # evaluate - 'instruction', - 'iinput', - 'context', - 'instruction_nochat', - 'iinput_nochat', + can_skip_because_locally_generated = no_default_param_names + [ # get_model: 'reward_type' ] @@ -1515,7 +2405,101 @@ def check_locals(**kwargs): assert k in kwargs, "Missing %s" % k -if __name__ == "__main__": +def get_model_max_length(model_state): + if not isinstance(model_state['tokenizer'], (str, types.NoneType)): + return model_state['tokenizer'].model_max_length + else: + return 2048 + + +def get_max_max_new_tokens(model_state, **kwargs): + if not isinstance(model_state['tokenizer'], (str, types.NoneType)): + max_max_new_tokens = model_state['tokenizer'].model_max_length + else: + max_max_new_tokens = None + + if kwargs['max_max_new_tokens'] is not None and max_max_new_tokens is not None: + return min(max_max_new_tokens, kwargs['max_max_new_tokens']) + elif kwargs['max_max_new_tokens'] is not None: + return kwargs['max_max_new_tokens'] + elif kwargs['memory_restriction_level'] == 1: + return 768 + elif kwargs['memory_restriction_level'] == 2: + return 512 + elif kwargs['memory_restriction_level'] >= 3: + return 256 + else: + # FIXME: Need to update after new model loaded, so user can control with slider + return 2048 + + +def get_minmax_top_k_docs(is_public): + if is_public: + min_top_k_docs = 1 + max_top_k_docs = 3 + label_top_k_docs = "Number of document chunks" + else: + min_top_k_docs = -1 + max_top_k_docs = 100 + label_top_k_docs = "Number of document chunks (-1 = auto fill model context)" + return min_top_k_docs, max_top_k_docs, label_top_k_docs + + +def history_to_context(history, langchain_mode1, prompt_type1, prompt_dict1, chat1, model_max_length1, + memory_restriction_level1, keep_sources_in_context1): + """ + consumes all history up to (but not including) latest history item that is presumed to be an [instruction, None] pair + :param history: + :param langchain_mode1: + :param prompt_type1: + :param prompt_dict1: + :param chat1: + :param model_max_length1: + :param memory_restriction_level1: + :param keep_sources_in_context1: + :return: + """ + # ensure output will be unique to models + _, _, _, max_prompt_length = get_cutoffs(memory_restriction_level1, + for_context=True, model_max_length=model_max_length1) + context1 = '' + if max_prompt_length is not None and langchain_mode1 not in ['LLM']: + context1 = '' + # - 1 below because current instruction already in history from user() + for histi in range(0, len(history) - 1): + data_point = dict(instruction=history[histi][0], input='', output=history[histi][1]) + prompt, pre_response, terminate_response, chat_sep, chat_turn_sep = generate_prompt(data_point, + prompt_type1, + prompt_dict1, + chat1, + reduced=True, + making_context=True) + # md -> back to text, maybe not super important if model trained enough + if not keep_sources_in_context1 and langchain_mode1 != 'Disabled' and prompt.find(source_prefix) >= 0: + # FIXME: This is relatively slow even for small amount of text, like 0.3s each history item + import re + prompt = re.sub(f'{re.escape(source_prefix)}.*?{re.escape(source_postfix)}', '', prompt, + flags=re.DOTALL) + if prompt.endswith('\n

'): + prompt = prompt[:-4] + prompt = prompt.replace('
', chat_turn_sep) + if not prompt.endswith(chat_turn_sep): + prompt += chat_turn_sep + # most recent first, add older if can + # only include desired chat history + if len(prompt + context1) > max_prompt_length: + break + context1 += prompt + + _, pre_response, terminate_response, chat_sep, chat_turn_sep = generate_prompt({}, prompt_type1, prompt_dict1, + chat1, reduced=True, + making_context=True) + if context1 and not context1.endswith(chat_turn_sep): + context1 += chat_turn_sep # ensure if terminates abruptly, then human continues on next line + return context1 + + +def entrypoint_main(): """ Examples: @@ -1546,3 +2530,7 @@ if __name__ == "__main__": python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b """ fire.Fire(main) + + +if __name__ == "__main__": + entrypoint_main()