app.py

import functools
import sys
import os
import traceback
import typing

from utils import set_seed, clear_torch_cache, save_generate_output, NullContext, KThread, wrapped_partial

SEED = 1236
set_seed(SEED)

os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'
from typing import Union
import numpy as np
import pandas as pd

import fire
import torch
from peft import PeftModel
from transformers import GenerationConfig, StoppingCriteriaList, AutoModel, TextIteratorStreamer
from accelerate import init_empty_weights, infer_auto_device_map

from prompter import Prompter

from finetune import get_loaders, example_data_points, generate_prompt, human, bot, inv_prompt_type_to_model_lower
from stopping import StoppingCriteriaSub

eval_extra_columns = ['prompt', 'response', 'score']


def main(
        load_8bit: bool = False,
        load_half: bool = True,
        infer_devices: bool = True,  # really if to "control" devices now
        base_model: str = '',
        tokenizer_base_model: str = '',
        lora_weights: str = "",
        gpu_id: int = 0,  # if infer_devices = True and gpu_id != -1

        prompt_type: Union[int, str] = None,
        # input to generation
        temperature: float = None,
        top_p: float = None,
        top_k: int = None,
        num_beams: int = None,
        repetition_penalty: float = None,
        num_return_sequences: int = None,
        do_sample: bool = None,
        max_new_tokens: int = None,
        min_new_tokens: int = None,
        early_stopping: Union[bool, str] = None,
        max_time: float = None,

        debug: bool = False,
        save_dir: str = None,
        share: bool = True,
        local_files_only: bool = False,
        resume_download: bool = True,
        use_auth_token: Union[str, bool] = False,  # True requires CLI did huggingface-cli login before running

        src_lang: str = "English",
        tgt_lang: str = "Russian",

        gradio: bool = True,
        gradio_avoid_processing_markdown: bool = False,
        chat: bool = True,
        chat_history: int = 4096,  # character length of chat context/history
        chat_context: bool = False,  # use default context if human_bot
        stream_output: bool = True,
        show_examples: bool = None,
        verbose: bool = False,
        h2ocolors: bool = True,
        height: int = 400,
        show_lora: bool = True,
        # set to True to load --base_model after client logs in,
        # to be able to free GPU memory when model is swapped
        login_mode_if_model0: bool = False,
        block_gradio_exit: bool = True,
        concurrency_count: int = 1,
        api_open: bool = False,  # don't let API skip queue
        allow_api: bool = True,
        input_lines: int = 1,

        sanitize_user_prompt: bool = True,
        sanitize_bot_response: bool = True,

        extra_model_options: typing.List[str] = [],
        extra_lora_options: typing.List[str] = [],

        score_model: str = 'OpenAssistant/reward-model-deberta-v3-large-v2',
        auto_score: bool = True,

        eval_sharegpt_prompts_only: int = 0,
        eval_sharegpt_prompts_only_seed: int = 1234,
        eval_sharegpt_as_output: bool = False,

        hard_stop_list: typing.List[str] = [],
):
    is_hf = bool(os.getenv("HUGGINGFACE_SPACES"))
    is_gpth2oai = bool(os.getenv("GPT_H2O_AI"))
    is_public = is_hf or is_gpth2oai  # multi-user case with fixed model and disclaimer
    is_low_mem = is_hf  # assumes run on 24GB consumer GPU
    admin_pass = os.getenv("ADMIN_PASS")
    # will sometimes appear in UI or sometimes actual generation, but maybe better than empty result
    # but becomes unrecoverable sometimes if raise, so just be silent for now
    raise_generate_gpu_exceptions = not is_public

    # allow set token directly
    use_auth_token = os.environ.get("HUGGINGFACE_API_TOKEN", use_auth_token)

    if is_public:
        input_lines = 1  # ensure set, for ease of use
        temperature = 0.4
        top_p = 0.85
        top_k = 70
        do_sample = True
        if is_low_mem:
            base_model = 'h2oai/h2ogpt-oasst1-512-12b'
            load_8bit = True
        else:
            base_model = 'h2oai/h2ogpt-oasst1-512-20b'
    if is_low_mem:
        load_8bit = True
    if is_hf:
        # must override share if in spaces
        share = False
    save_dir = os.getenv('SAVE_DIR', save_dir)
    score_model = os.getenv('SCORE_MODEL', score_model)
    if score_model == '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)))

    n_gpus = torch.cuda.device_count()

    # get defaults
    model_lower = base_model.lower()
    if not gradio:
        # force, else not single response like want to look at
        stream_output = False
        # else prompt removal can mess up output
        chat = False

    placeholder_instruction, placeholder_input, \
    stream_output, show_examples, \
    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, \
    src_lang, tgt_lang, \
    examples, \
    task_info = \
        get_generate_params(model_lower, chat,
                            stream_output, show_examples,
                            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,
                            )

    if not gradio:
        if eval_sharegpt_prompts_only > 0:
            # override default examples with shareGPT ones for human-level eval purposes only
            eval_filename = 'ShareGPT_V3_unfiltered_cleaned_split_no_imsorry.json'
            if not os.path.isfile(eval_filename):
                os.system(
                    'wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/%s' % eval_filename)
            import json
            data = json.load(open(eval_filename, 'rt'))
            # focus on data that starts with human, else likely chopped from other data
            turn_start = 0  # odd in general
            data = [x for x in data if len(x['conversations']) > turn_start + 1 and
                    x['conversations'][turn_start]['from'] == 'human' and
                    x['conversations'][turn_start + 1]['from'] == 'gpt']
            np.random.seed(eval_sharegpt_prompts_only_seed)
            example1 = examples[-1]  # pick reference example
            examples = []
            responses = []
            for i in list(np.random.randint(0, len(data), size=eval_sharegpt_prompts_only)):
                assert data[i]['conversations'][turn_start]['from'] == 'human'
                instruction = data[i]['conversations'][turn_start]['value']
                assert data[i]['conversations'][turn_start + 1]['from'] == 'gpt'
                output = data[i]['conversations'][turn_start + 1]['value']
                examplenew = example1.copy()
                assert not chat, "No gradio must use chat=False, uses nochat instruct"
                examplenew[eval_func_param_names.index('instruction_nochat')] = instruction
                examplenew[eval_func_param_names.index('iinput_nochat')] = ''  # no input
                examplenew[eval_func_param_names.index('context')] = get_context(chat_context, prompt_type)
                examples.append(examplenew)
                responses.append(output)

        num_examples = len(examples)
        scoring_path = 'scoring'
        os.makedirs(scoring_path, exist_ok=True)
        if eval_sharegpt_as_output:
            used_base_model = 'gpt35'
            used_lora_weights = ''
        else:
            used_base_model = str(base_model.split('/')[-1])
            used_lora_weights = str(lora_weights.split('/')[-1])
        eval_filename = "df_scores_%s_%s_%s_%s_%s_%s.parquet" % (num_examples, eval_sharegpt_prompts_only,
                                                                 eval_sharegpt_prompts_only_seed,
                                                                 eval_sharegpt_as_output,
                                                                 used_base_model,
                                                                 used_lora_weights)
        eval_filename = os.path.join(scoring_path, eval_filename)

        # torch.device("cuda") leads to cuda:x cuda:y mismatches for multi-GPU consistently
        context_class = NullContext() if n_gpus > 1 else torch.device("cuda")

        with context_class:
            # ensure was set right above before examples generated
            assert not stream_output, "stream_output=True does not make sense with example loop"
            import time
            from functools import partial

            # get score model
            smodel, stokenizer, sdevice = get_score_model(**locals())

            if not eval_sharegpt_as_output:
                model, tokenizer, device = get_model(**locals())
                model_state = [model, tokenizer, device, base_model]
                fun = partial(evaluate, model_state, debug=debug, save_dir=save_dir, is_low_mem=is_low_mem,
                              raise_generate_gpu_exceptions=raise_generate_gpu_exceptions,
                              chat_context=chat_context,
                              concurrency_count=concurrency_count)
            else:
                assert eval_sharegpt_prompts_only > 0

                def get_response(*args, exi=0):
                    # assumes same ordering of examples and responses
                    yield responses[exi]

                fun = get_response
            t0 = time.time()
            score_dump = []

            import matplotlib.pyplot as plt

            for exi, ex in enumerate(examples):
                instruction = ex[eval_func_param_names.index('instruction_nochat')]
                iinput = ex[eval_func_param_names.index('iinput_nochat')]
                context = ex[eval_func_param_names.index('context')]
                clear_torch_cache()
                print("")
                print("START" + "=" * 100)
                print("Question: %s %s" % (instruction, ('input=%s' % iinput if iinput else '')))
                print("-" * 105)
                # fun yields as generator, so have to iterate over it
                # Also means likely do NOT want --stream_output=True, else would show all generations
                gener = fun(*tuple(ex), exi=exi) if eval_sharegpt_as_output else fun(*tuple(ex))
                for res in gener:
                    print(res)
                    if smodel:
                        score_with_prompt = False
                        if score_with_prompt:
                            data_point = dict(instruction=instruction, input=iinput, context=context)
                            prompter = Prompter(prompt_type, debug=debug, chat=chat, stream_output=stream_output)
                            prompt = prompter.generate_prompt(data_point)
                        else:
                            # just raw input and output
                            if eval_sharegpt_prompts_only > 0:
                                # only our own examples have this filled at moment
                                assert iinput in [None, ''], iinput  # should be no iinput
                            if not (chat_context and prompt_type == 'human_bot'):
                                assert context in [None, ''], context  # should be no context
                            prompt = instruction
                        cutoff_len = 768 if is_low_mem else 2048
                        inputs = stokenizer(prompt, res,
                                            return_tensors="pt",
                                            truncation=True,
                                            max_length=cutoff_len)
                        try:
                            score = torch.sigmoid(smodel(**inputs).logits[0]).cpu().detach().numpy()[0]
                        except torch.cuda.OutOfMemoryError as e:
                            print("GPU OOM 1: question: %s answer: %s exception: %s" % (prompt, res, str(e)), flush=True)
                            traceback.print_exc()
                            score = 0.0
                            clear_torch_cache()
                        except (Exception, RuntimeError) as e:
                            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):
                                print("GPU error: question: %s answer: %s exception: %s" % (prompt, res, str(e)),
                                      flush=True)
                                traceback.print_exc()
                                score = 0.0
                                clear_torch_cache()
                            else:
                                raise
                        print("SCORE %s: %s" % (exi, score), flush=True)
                        score_dump.append(ex + [prompt, res, score])
                        # dump every score in case abort
                        df_scores = pd.DataFrame(score_dump,
                                                 columns=eval_func_param_names + eval_extra_columns)
                        df_scores.to_parquet(eval_filename, index=False)
                        # plot histogram so far
                        plt.figure(figsize=(10, 10))
                        plt.hist(df_scores['score'], bins=20)
                        score_avg = np.mean(df_scores['score'])
                        score_median = np.median(df_scores['score'])
                        plt.title("Score avg: %s median: %s" % (score_avg, score_median))
                        plt.savefig(eval_filename.replace('.parquet', '.png'))
                        plt.close()

                print("END" + "=" * 102)
                print("")
                t2 = time.time()
                print("Time taken so far: %.4f about %.4g per example" % (t2 - t0, (t2 - t0) / (1 + exi)))
            t1 = time.time()
            print("Total time taken: %.4f about %.4g per example" % (t1 - t0, (t1 - t0) / num_examples))
        return eval_filename

    if gradio:
        # imported here so don't require gradio to run generate
        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(**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']]

        # get score model
        smodel, stokenizer, sdevice = get_score_model(**all_kwargs)
        score_model_state0 = [smodel, stokenizer, sdevice, score_model]

        go_gradio(**locals())


def get_device():
    if torch.cuda.is_available():
        device = "cuda"
    else:
        raise RuntimeError("only cuda supported")

    return device


def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type,
                       gpu_id=0,
                       use_auth_token=False):
    """
    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:
    :return:
    """
    with init_empty_weights():
        from transformers import AutoConfig
        config = AutoConfig.from_pretrained(base_model, use_auth_token=use_auth_token)
        model = AutoModel.from_config(
            config,
        )

    # 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.
    device_map = infer_auto_device_map(
        model,
        dtype=torch.float16 if load_half else torch.float32,
    )
    if hasattr(model, 'model'):
        device_map_model = infer_auto_device_map(
            model.model,
            dtype=torch.float16 if load_half else torch.float32,
        )
        device_map.update(device_map_model)
    print('device_map: %s' % device_map, flush=True)

    if gpu_id >= 0:
        # FIXME: If really distributes model, tend to get things like: ValueError: gpt_neox.embed_in.weight doesn't have any device set.
        # So avoid for now, just put on first GPU, unless score_model, put on last
        n_gpus = torch.cuda.device_count()
        if reward_type:
            device_map = {'': n_gpus - 1}
        else:
            device_map = {'': min(n_gpus - 1, gpu_id)}
    if gpu_id == -1:
        device_map = {'': 'cuda'}

    load_in_8bit = model_kwargs.get('load_in_8bit', False)
    model_kwargs['device_map'] = device_map

    if load_in_8bit or not load_half:
        model = model_loader.from_pretrained(
            base_model,
            **model_kwargs,
        )
    else:
        model = model_loader.from_pretrained(
            base_model,
            **model_kwargs,
        ).half()
    return model


def get_model(
        load_8bit: 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,
        compile: bool = True,
        **kwargs,
):
    """

    :param load_8bit: load model in 8-bit, not supported by all models
    :param load_half: load model in 16-bit
    :param infer_devices: Use torch infer of optimal placement of layers on devices (for non-lora case)
           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 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)
    :param reward_type: reward type model for sequence classification
    :param local_files_only: use local files instead of from HF
    :param resume_download: resume downloads from HF
    :param use_auth_token: assumes user did on CLI `huggingface-cli login` to access private repo
    :param compile: whether to compile torch model
    :param kwargs:
    :return:
    """
    print("Get %s model" % base_model, flush=True)
    if lora_weights is not None and lora_weights.strip():
        print("Get %s lora weights" % lora_weights, flush=True)
    device = get_device()

    if 'gpt2' in base_model.lower():
        # RuntimeError: where expected condition to be a boolean tensor, but got a tensor with dtype Half
        load_8bit = False

    assert base_model.strip(), (
        "Please choose a base model with --base_model (CLI) or in Models Tab (gradio)"
    )

    from transformers import AutoConfig
    config = AutoConfig.from_pretrained(base_model, use_auth_token=use_auth_token)
    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:
        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(llama_type=llama_type, model_name=base_model, reward_type=reward_type)
    if not tokenizer_base_model:
        tokenizer_base_model = base_model

    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,
                                                     )
    else:
        tokenizer = tokenizer_loader

    if isinstance(tokenizer, str):
        # already a pipeline, tokenizer_loader is string for task
        model = model_loader(tokenizer,
                             model=base_model,
                             device=0 if device == "cuda" else -1,
                             torch_dtype=torch.float16)
    else:
        assert device == "cuda", "Unsupported device %s" % device
        model_kwargs = dict(local_files_only=local_files_only,
                            torch_dtype=torch.float16,
                            resume_download=resume_download,
                            use_auth_token=use_auth_token)
        if 'mbart-' not in base_model.lower():
            model_kwargs.update(dict(load_in_8bit=load_8bit,
                                     device_map={"": 0} if load_8bit else "auto",
                                     ))
        if 'OpenAssistant/reward-model'.lower() in base_model.lower():
            # could put on other GPUs
            model_kwargs['device_map'] = {"": 0}
            model_kwargs.pop('torch_dtype', None)

        if not lora_weights:
            with torch.device("cuda"):
                if infer_devices:
                    model = get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type,
                                               gpu_id=gpu_id, use_auth_token=use_auth_token)
                else:
                    if load_half and not load_8bit:
                        model = model_loader.from_pretrained(
                            base_model,
                            **model_kwargs).half()
                    else:
                        model = model_loader.from_pretrained(
                            base_model,
                            **model_kwargs)
        elif load_8bit:
            model = model_loader.from_pretrained(
                base_model,
                **model_kwargs
            )
            model = PeftModel.from_pretrained(
                model,
                lora_weights,
                torch_dtype=torch.float16,
                local_files_only=local_files_only,
                resume_download=resume_download,
                use_auth_token=use_auth_token,
                device_map={"": 0},  # seems to be required
            )
        else:
            with torch.device("cuda"):
                model = model_loader.from_pretrained(
                    base_model,
                    **model_kwargs
                )
                model = PeftModel.from_pretrained(
                    model,
                    lora_weights,
                    torch_dtype=torch.float16,
                    local_files_only=local_files_only,
                    resume_download=resume_download,
                    use_auth_token=use_auth_token,
                    device_map="auto",
                )
                if load_half:
                    model.half()

    # unwind broken decapoda-research config
    if llama_type:
        model.config.pad_token_id = tokenizer.pad_token_id = 0  # unk
        model.config.bos_token_id = 1
        model.config.eos_token_id = 2
    if 'gpt2' in base_model.lower():
        # add special tokens that otherwise all share the same id
        tokenizer.add_special_tokens({'bos_token': '<bos>',
                                      'eos_token': '<eos>',
                                      'pad_token': '<pad>'})

    if not isinstance(tokenizer, str):
        model.eval()
        if torch.__version__ >= "2" and sys.platform != "win32" and compile:
            model = torch.compile(model)

    return model, tokenizer, device


def get_score_model(**kwargs):
    # score model
    if kwargs.get('score_model') is not None and kwargs.get('score_model').strip():
        score_all_kwargs = kwargs.copy()
        score_all_kwargs['load_8bit'] = False
        score_all_kwargs['load_half'] = False
        score_all_kwargs['base_model'] = kwargs.get('score_model').strip()
        score_all_kwargs['tokenizer_base_model'] = ''
        score_all_kwargs['lora_weights'] = ''
        score_all_kwargs['llama_type'] = False
        score_all_kwargs['compile'] = False
        smodel, stokenizer, sdevice = get_model(**score_all_kwargs)
    else:
        smodel, stokenizer, sdevice = None, None, None
    return smodel, stokenizer, sdevice


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',
                         'instruction_nochat',
                         'iinput_nochat',
                         ]


def evaluate(
        model_state,
        # START NOTE: Examples must have same order of parameters
        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,
        instruction_nochat,
        iinput_nochat,
        # END NOTE: Examples must have same order of parameters
        src_lang=None,
        tgt_lang=None,
        debug=False,
        concurrency_count=None,
        save_dir=None,
        hard_stop_list=None,
        sanitize_bot_response=True,
        model_state0=None,
        is_low_mem=None,
        raise_generate_gpu_exceptions=None,
        chat_context=None,
):
    # ensure passed these
    assert concurrency_count is not None
    assert is_low_mem is not None
    assert raise_generate_gpu_exceptions is not None
    assert chat_context is not None

    if debug:
        locals_dict = locals().copy()
        locals_dict.pop('model_state', None)
        locals_dict.pop('model_state0', None)
        print(locals_dict)

    no_model_msg = "Please choose a base model with --base_model (CLI) or in Models Tab (gradio).\nThen start New Conversation"

    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
    else:
        raise AssertionError(no_model_msg)

    if base_model is None:
        raise AssertionError(no_model_msg)

    assert base_model.strip(), no_model_msg
    assert model, "Model is missing"
    assert tokenizer, "Tokenizer is missing"

    # choose chat or non-chat mode
    if not chat:
        instruction = instruction_nochat
        iinput = iinput_nochat

    if not context:
        # get hidden context if have one
        context = get_context(chat_context, prompt_type)

    data_point = dict(context=context, instruction=instruction, input=iinput)
    prompter = Prompter(prompt_type, debug=debug, chat=chat, stream_output=stream_output)
    prompt = prompter.generate_prompt(data_point)

    if hard_stop_list is None:
        # acts like undo on user entry and bot response
        hard_stop_list = []

    if isinstance(tokenizer, str):
        # pipeline
        if tokenizer == "summarization":
            key = 'summary_text'
        else:
            raise RuntimeError("No such task type %s" % tokenizer)
        # NOTE: uses max_length only
        yield model(prompt, max_length=max_new_tokens)[0][key]

    if 'mbart-' in base_model.lower():
        assert src_lang is not None
        tokenizer.src_lang = languages_covered()[src_lang]

    if chat:
        # override, ignore user change
        num_return_sequences = 1
    if prompt_type in ['human_bot', 'instruct_vicuna', 'instruct_with_end']:
        if prompt_type == 'human_bot':
            # encounters = [prompt.count(human) + 1, prompt.count(bot) + 1]
            # stopping only starts once output is beyond prompt
            # 1 human is enough to trigger, but need 2 bots, because very first view back will be bot we added
            stop_words = [human, bot, '\n' + human, '\n' + bot]
            encounters = [1, 2]
        elif prompt_type == 'instruct_vicuna':
            # even below is not enough, generic strings and many ways to encode
            stop_words = [
                '### Human:',
                """
### Human:""",
                """
### Human:
""",
                '### Assistant:',
                """
### Assistant:""",
                """
### Assistant:
""",
            ]
            encounters = [1, 2]
        else:
            # some instruct prompts have this as end, doesn't hurt to stop on it since not common otherwise
            stop_words = ['### End']
            encounters = [1]
        stop_words_ids = [
            tokenizer(stop_word, return_tensors='pt')['input_ids'].squeeze() for stop_word in stop_words]
        # handle single token case
        stop_words_ids = [x if len(x.shape) > 0 else torch.tensor([x]) for x in stop_words_ids]
        stop_words_ids = [x for x in stop_words_ids if x.shape[0] > 0]
        # avoid padding in front of tokens
        if tokenizer.pad_token:
            stop_words_ids = [x[1:] if x[0] == tokenizer.pad_token_id and len(x) > 1 else x for x in stop_words_ids]
        # handle fake \n added
        stop_words_ids = [x[1:] if y[0] == '\n' else x for x, y in zip(stop_words_ids, stop_words)]
        # build stopper
        stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids, encounters=encounters)])
    else:
        stopping_criteria = StoppingCriteriaList()

    # help to avoid errors like:
    # RuntimeError: The size of tensor a (2048) must match the size of tensor b (2049) at non-singleton dimension 3
    # RuntimeError: expected scalar type Half but found Float
    # with - 256
    max_length_tokenize = 768 - 256 if is_low_mem else 2048 - 256
    cutoff_len = max_length_tokenize * 4  # if reaches limit, then can't generate new tokens
    output_smallest = 30 * 4
    prompt = prompt[-cutoff_len - output_smallest:]
    inputs = tokenizer(prompt,
                       return_tensors="pt",
                       truncation=True,
                       max_length=max_length_tokenize)
    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)
    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,
    )

    gen_kwargs = dict(input_ids=input_ids,
                      generation_config=generation_config,
                      return_dict_in_generate=True,
                      output_scores=True,
                      max_new_tokens=max_new_tokens,  # prompt + new
                      min_new_tokens=min_new_tokens,  # prompt + new
                      early_stopping=early_stopping,  # False, True, "never"
                      max_time=max_time,
                      stopping_criteria=stopping_criteria,
                      )
    if 'gpt2' in base_model.lower():
        gen_kwargs.update(dict(bos_token_id=tokenizer.bos_token_id, pad_token_id=tokenizer.eos_token_id))
    elif 'mbart-' in base_model.lower():
        assert tgt_lang is not None
        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))

    decoder = functools.partial(tokenizer.decode,
                                skip_special_tokens=True,
                                clean_up_tokenization_spaces=True,
                                )
    decoder_raw = functools.partial(tokenizer.decode,
                                    skip_special_tokens=False,
                                    clean_up_tokenization_spaces=True,
                                    )

    with torch.no_grad():
        # decoded tokenized prompt can deviate from prompt due to special characters
        inputs_decoded = decoder(input_ids[0])
        inputs_decoded_raw = decoder_raw(input_ids[0])
        if inputs_decoded == prompt:
            # normal
            pass
        elif inputs_decoded.lstrip() == prompt.lstrip():
            # sometimes extra space in front, make prompt same for prompt removal
            prompt = inputs_decoded
        elif inputs_decoded_raw == prompt:
            # some models specify special tokens that are part of normal prompt, so can't skip them
            inputs_decoded_raw = inputs_decoded
            decoder = decoder_raw
        else:
            print("WARNING: Special characters in prompt", flush=True)
        if stream_output:
            #skip_prompt = prompt_type != 'plain'
            skip_prompt = False
            streamer = TextIteratorStreamer(tokenizer, skip_prompt=skip_prompt)
            gen_kwargs.update(dict(streamer=streamer))
            if debug:
                KThread.show_threads()
            target_func = generate_with_exceptions
            if concurrency_count == 1:
                # otherwise can't do this
                KThread.kill_threads(target_func.__name__, debug=debug)
            target = wrapped_partial(generate_with_exceptions, model.generate, prompt, inputs_decoded,
                                     raise_generate_gpu_exceptions, **gen_kwargs)
            thread = KThread(target=target)
            thread.start()
            outputs = ""
            for new_text in streamer:
                outputs += new_text
                yield prompter.get_response(outputs, prompt=inputs_decoded,
                                            sanitize_bot_response=sanitize_bot_response)
        else:
            outputs = model.generate(**gen_kwargs)
            outputs = [decoder(s) for s in outputs.sequences]
            yield prompter.get_response(outputs, prompt=inputs_decoded,
                                        sanitize_bot_response=sanitize_bot_response)
        if save_dir and outputs and len(outputs) >= 1:
            decoded_output = prompt + outputs[0]
            save_generate_output(output=decoded_output, base_model=base_model, save_dir=save_dir)


def generate_with_exceptions(func, prompt, inputs_decoded, raise_generate_gpu_exceptions, **kwargs):
    try:
        func(**kwargs)
    except torch.cuda.OutOfMemoryError as e:
        print("GPU OOM 2: prompt: %s inputs_decoded: %s exception: %s" % (prompt, inputs_decoded, str(e)),
              flush=True)
        if kwargs['input_ids'] is not None:
            kwargs['input_ids'].cpu()
        kwargs['input_ids'] = None
        traceback.print_exc()
        clear_torch_cache()
        return
    except (Exception, RuntimeError) as e:
        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) or \
                'mat1 and mat2 shapes cannot be multiplied' in str(e):
            print(
                "GPU Error: prompt: %s inputs_decoded: %s exception: %s" % (prompt, inputs_decoded, str(e)),
                flush=True)
            traceback.print_exc()
            clear_torch_cache()
            if raise_generate_gpu_exceptions:
                raise
            return
        else:
            clear_torch_cache()
            raise


def get_generate_params(model_lower, chat,
                        stream_output, show_examples,
                        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):
    use_defaults = False
    use_default_examples = True
    examples = []
    task_info = f"{prompt_type}"
    if model_lower:
        print(f"Using Model {model_lower}", flush=True)
    else:
        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
    max_time_defaults = 60 * 3
    max_time = max_time if max_time is not None else max_time_defaults

    if not prompt_type and model_lower in inv_prompt_type_to_model_lower:
        prompt_type = inv_prompt_type_to_model_lower[model_lower]

    # examples at first don't include chat, instruction_nochat, iinput_nochat, added at end
    if show_examples is None:
        if chat:
            show_examples = False
        else:
            show_examples = True

    summarize_example1 = """Jeff: Can I train a ? Transformers model on Amazon SageMaker? 
Philipp: Sure you can use the new Hugging Face Deep Learning Container. 
Jeff: ok.
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"""

    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]]
        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?"
        placeholder_input = ""
        use_defaults = True
        use_default_examples = True
        task_info = "Multi-Task: Q/A, translation, Chain-of-Thought, Logical Reasoning, Summarization, etc.  Best to use task prefix as trained on, e.g. `translate English to German: ` (space after colon)"
    elif 'mbart-' in model_lower:
        placeholder_instruction = "The girl has long hair."
        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]]
    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]]
        task_info = "Auto-complete phrase, code, etc."
        use_defaults = True
    else:
        if chat:
            placeholder_instruction = "Enter a question or imperative."
        else:
            placeholder_instruction = "Give detailed answer for whether Einstein or Newton is smarter."
        placeholder_input = ""
        if model_lower:
            prompt_type = prompt_type or 'human_bot'
        else:
            prompt_type = ''
        examples += [[summarize_example1, 'Summarize' if prompt_type not in ['plain', 'instruct_simple'] else '', "",
                      stream_output, prompt_type or 'plain', 0.1, 0.75, 40, 4, 256, 0, False, max_time_defaults, 1.0, 1,
                      False]]
        task_info = "No task"
        if prompt_type == 'instruct':
            task_info = "Answer question or follow imperative as instruction with optionally input."
        elif prompt_type == 'plain':
            task_info = "Auto-complete phrase, code, etc."
        elif prompt_type == 'human_bot':
            if chat:
                task_info = "Chat (Shift-Enter to give question/imperative, input concatenated with instruction)"
            else:
                task_info = "Ask question/imperative (input concatenated with instruction)"

    # revert to plain if still nothing
    prompt_type = prompt_type or 'plain'
    if use_defaults:
        temperature = 1.0 if temperature is None else temperature
        top_p = 1.0 if top_p is None else top_p
        top_k = 40 if top_k is None else top_k
        num_beams = num_beams or 1
        max_new_tokens = max_new_tokens or 128
        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
    else:
        temperature = 0.4 if temperature is None else temperature
        top_p = 0.85 if top_p is None else top_p
        top_k = 70 if top_k is None else top_k
        if chat:
            num_beams = num_beams or 1
        else:
            num_beams = num_beams or 4
        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 = True 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,
                   early_stopping, max_time, repetition_penalty, num_return_sequences, do_sample]

    if use_default_examples:
        examples += [
            ["Translate English to French", "Good morning"] + params_list,
            ["Give detailed answer for whether Einstein or Newton is smarter.", ''] + params_list,
            ["Explain in detailed list, all the best practices for coding in python.", ''] + params_list,
            [
                "Create a markdown table with 3 rows for the primary colors, and 2 columns, with color name and hex codes.",
                ''] + params_list,
            ['Translate to German:  My name is Arthur', ''] + params_list,
            ["Please answer to the following question. Who is going to be the next Ballon d'or?", ''] + params_list,
            ['Can Geoffrey Hinton have a conversation with George Washington? Give the rationale before answering.',
             ''] + params_list,
            ['Please answer the following question. What is the boiling point of Nitrogen?', ''] + params_list,
            ['Answer the following yes/no question. Can you write a whole Haiku in a single tweet?', ''] + params_list,
            ["Simplify the following expression: (False or False and True). Explain your answer.", ''] + params_list,
            [
                "Premise: At my age you will probably have learnt one lesson. Hypothesis:  It's not certain how many lessons you'll learn by your thirties. Does the premise entail the hypothesis?",
                ''] + params_list,
            ['The square root of x is the cube root of y. What is y to the power of 2, if x = 4?', ''] + params_list,
            [
                'Answer the following question by reasoning step by step.  The cafeteria had 23 apples. If they used 20 for lunch, and bought 6 more, how many apple do they have?',
                ''] + params_list,
            ["""def area_of_rectangle(a: float, b: float):
    \"\"\"Return the area of the rectangle.\"\"\"""", ''] + params_list,
            ["""# a function in native python:
def mean(a):
    return sum(a)/len(a)

# the same function using numpy:
import numpy as np
def mean(a):""", ''] + params_list,
            ["""X = np.random.randn(100, 100)
y = np.random.randint(0, 1, 100)

# fit random forest classifier with 20 estimators""", ''] + params_list,
        ]

    src_lang = "English"
    tgt_lang = "Russian"

    # move to correct position
    for example in examples:
        example += [chat, '', '']
        # adjust examples if non-chat mode
        if not chat:
            example[eval_func_param_names.index('instruction_nochat')] = example[
                eval_func_param_names.index('instruction')]
            example[eval_func_param_names.index('instruction')] = ''

            example[eval_func_param_names.index('iinput_nochat')] = example[eval_func_param_names.index('iinput')]
            example[eval_func_param_names.index('iinput')] = ''

    return placeholder_instruction, placeholder_input, \
           stream_output, show_examples, \
           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, \
           src_lang, tgt_lang, \
           examples, \
           task_info


def languages_covered():
    # https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt#languages-covered
    covered = """Arabic (ar_AR), Czech (cs_CZ), German (de_DE), English (en_XX), Spanish (es_XX), Estonian (et_EE), Finnish (fi_FI), French (fr_XX), Gujarati (gu_IN), Hindi (hi_IN), Italian (it_IT), Japanese (ja_XX), Kazakh (kk_KZ), Korean (ko_KR), Lithuanian (lt_LT), Latvian (lv_LV), Burmese (my_MM), Nepali (ne_NP), Dutch (nl_XX), Romanian (ro_RO), Russian (ru_RU), Sinhala (si_LK), Turkish (tr_TR), Vietnamese (vi_VN), Chinese (zh_CN), Afrikaans (af_ZA), Azerbaijani (az_AZ), Bengali (bn_IN), Persian (fa_IR), Hebrew (he_IL), Croatian (hr_HR), Indonesian (id_ID), Georgian (ka_GE), Khmer (km_KH), Macedonian (mk_MK), Malayalam (ml_IN), Mongolian (mn_MN), Marathi (mr_IN), Polish (pl_PL), Pashto (ps_AF), Portuguese (pt_XX), Swedish (sv_SE), Swahili (sw_KE), Tamil (ta_IN), Telugu (te_IN), Thai (th_TH), Tagalog (tl_XX), Ukrainian (uk_UA), Urdu (ur_PK), Xhosa (xh_ZA), Galician (gl_ES), Slovene (sl_SI)"""
    covered = covered.split(', ')
    covered = {x.split(' ')[0]: x.split(' ')[1].replace(')', '').replace('(', '') for x in covered}
    return covered


def get_context(chat_context, prompt_type):
    if chat_context and prompt_type == 'human_bot':
        context0 = """<bot>: I am an intelligent, helpful, truthful, and fair assistant named h2oGPT, who will give accurate, balanced, and reliable responses.  I will not respond with I don't know or I don't understand.
<human>: I am a human person seeking useful assistance and request all questions be answered completely, and typically expect detailed responses.  Give answers in numbered list format if several distinct but related items are being listed."""
    else:
        context0 = ''
    return context0


def test_test_prompt(prompt_type='instruct', data_point=0):
    example_data_point = example_data_points[data_point]
    example_data_point.pop('output', None)
    return generate_prompt(example_data_point, prompt_type, False, False)


def score_qa(smodel, stokenizer, max_length_tokenize, question, answer, cutoff_len):
    question = question[-cutoff_len:]
    answer = answer[-cutoff_len:]

    inputs = stokenizer(question, answer,
                        return_tensors="pt",
                        truncation=True,
                        max_length=max_length_tokenize).to(smodel.device)
    try:
        score = torch.sigmoid(smodel(**inputs).logits[0]).cpu().detach().numpy()[0]
    except torch.cuda.OutOfMemoryError as e:
        print("GPU OOM 3: question: %s answer: %s exception: %s" % (question, answer, str(e)), flush=True)
        del inputs
        traceback.print_exc()
        clear_torch_cache()
        return 'Response Score: GPU OOM'
    except (Exception, RuntimeError) as e:
        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):
            print("GPU Error: question: %s answer: %s exception: %s" % (question, answer, str(e)),
                  flush=True)
            traceback.print_exc()
            clear_torch_cache()
            return 'Response Score: GPU Error'
        else:
            raise
    os.environ['TOKENIZERS_PARALLELISM'] = 'true'
    return score


if __name__ == "__main__":
    print("""
    WORLD_SIZE=4 CUDA_VISIBLE_DEVICES="0,1,2,3" torchrun --nproc_per_node=4 --master_port=1234 generate.py --base_model='EleutherAI/gpt-j-6B' --lora_weights=lora-alpaca_6B
    python generate.py --base_model='EleutherAI/gpt-j-6B' --lora_weights='lora-alpaca_6B'
    python generate.py --base_model='EleutherAI/gpt-neox-20b' --lora_weights='lora-alpaca_20B'
    
    # generate without lora weights, no prompt
    python generate.py --base_model='EleutherAI/gpt-neox-20b' --prompt_type='plain'
    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='dai_faq'

    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='dai_faq' --lora_weights='lora_20B_daifaq'
    # OpenChatKit settings:
    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='human_bot --debug=True --num_beams=1 --temperature=0.6 --top_k=40 --top_p=1.0

    python generate.py --base_model='distilgpt2' --prompt_type='plain' --debug=True --num_beams=1 --temperature=0.6 --top_k=40 --top_p=1.0 --share=False
    python generate.py --base_model='t5-large' --prompt_type='simple_instruct'
    python generate.py --base_model='philschmid/bart-large-cnn-samsum'
    python generate.py --base_model='philschmid/flan-t5-base-samsum'
    python generate.py --base_model='facebook/mbart-large-50-many-to-many-mmt'

    python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='human_bot' --lora_weights='GPT-NeoXT-Chat-Base-20B.merged.json.8_epochs.57b2892c53df5b8cefac45f84d019cace803ef26.28'

    must have 4*48GB GPU and run without 8bit in order for sharding to work with infer_devices=False
    can also pass --prompt_type='human_bot' and model can somewhat handle instructions without being instruct tuned
    python generate.py --base_model=decapoda-research/llama-65b-hf --load_8bit=False --infer_devices=False --prompt_type='human_bot'

    python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6.9b

    """, flush=True)
    fire.Fire(main)