dwolfe66/text-generation-webui-space

1import gc2import re3import time4 5import numpy as np6import torch7import transformers8 9import modules.shared as shared10from modules.callbacks import (Iteratorize, Stream,11                               _SentinelTokenStoppingCriteria)12from modules.extensions import apply_extensions13from modules.html_generator import generate_4chan_html, generate_basic_html14from modules.models import local_rank15 16 17def get_max_prompt_length(tokens):18    max_length = 2048-tokens19    if shared.soft_prompt:20        max_length -= shared.soft_prompt_tensor.shape[1]21    return max_length22 23def encode(prompt, tokens_to_generate=0, add_special_tokens=True):24    if shared.is_RWKV:25        input_ids = shared.tokenizer.encode(str(prompt))26        input_ids = np.array(input_ids).reshape(1, len(input_ids))27        return input_ids28    else:29        input_ids = shared.tokenizer.encode(str(prompt), return_tensors='pt', truncation=True, max_length=get_max_prompt_length(tokens_to_generate), add_special_tokens=add_special_tokens)30        if shared.args.cpu:31            return input_ids32        elif shared.args.flexgen:33            return input_ids.numpy()34        elif shared.args.deepspeed:35            return input_ids.to(device=local_rank)36        else:37            return input_ids.cuda()38 39def decode(output_ids):40    # Open Assistant relies on special tokens like <|endoftext|>41    if re.match('oasst-*', shared.model_name.lower()):42        return shared.tokenizer.decode(output_ids, skip_special_tokens=False)43    else:44        reply = shared.tokenizer.decode(output_ids, skip_special_tokens=True)45        reply = reply.replace(r'<|endoftext|>', '')46        return reply47 48def generate_softprompt_input_tensors(input_ids):49    inputs_embeds = shared.model.transformer.wte(input_ids)50    inputs_embeds = torch.cat((shared.soft_prompt_tensor, inputs_embeds), dim=1)51    filler_input_ids = torch.zeros((1, inputs_embeds.shape[1]), dtype=input_ids.dtype).to(shared.model.device)52    #filler_input_ids += shared.model.config.bos_token_id # setting dummy input_ids to bos tokens53    return inputs_embeds, filler_input_ids54 55# Removes empty replies from gpt4chan outputs56def fix_gpt4chan(s):57    for i in range(10):58        s = re.sub("--- [0-9]*\n>>[0-9]*\n---", "---", s)59        s = re.sub("--- [0-9]*\n *\n---", "---", s)60        s = re.sub("--- [0-9]*\n\n\n---", "---", s)61    return s62 63# Fix the LaTeX equations in galactica64def fix_galactica(s):65    s = s.replace(r'\[', r'$')66    s = s.replace(r'\]', r'$')67    s = s.replace(r'\(', r'$')68    s = s.replace(r'\)', r'$')69    s = s.replace(r'$$', r'$')70    s = re.sub(r'\n', r'\n\n', s)71    s = re.sub(r"\n{3,}", "\n\n", s)72    return s73 74def formatted_outputs(reply, model_name):75    if not (shared.args.chat or shared.args.cai_chat):76        if model_name.lower().startswith('galactica'):77            reply = fix_galactica(reply)78            return reply, reply, generate_basic_html(reply)79        elif model_name.lower().startswith(('gpt4chan', 'gpt-4chan', '4chan')):80            reply = fix_gpt4chan(reply)81            return reply, 'Only applicable for GALACTICA models.', generate_4chan_html(reply)82        else:83            return reply, 'Only applicable for GALACTICA models.', generate_basic_html(reply)84    else:85        return reply86 87def clear_torch_cache():88    gc.collect()89    if not shared.args.cpu:90        torch.cuda.empty_cache()91 92def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=None, stopping_string=None):93    clear_torch_cache()94    t0 = time.time()95 96    # These models are not part of Hugging Face, so we handle them97    # separately and terminate the function call earlier98    if shared.is_RWKV:99        try:100            if shared.args.no_stream:101                reply = shared.model.generate(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k)102                yield formatted_outputs(reply, shared.model_name)103            else:104                yield formatted_outputs(question, shared.model_name)105                # RWKV has proper streaming, which is very nice.106                # No need to generate 8 tokens at a time.107                for reply in shared.model.generate_with_streaming(context=question, token_count=max_new_tokens, temperature=temperature, top_p=top_p, top_k=top_k):108                    yield formatted_outputs(reply, shared.model_name)109        finally:110            t1 = time.time()111            output = encode(reply)[0]112            input_ids = encode(question)113            print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(input_ids[0]))/(t1-t0):.2f} tokens/s, {len(output)-len(input_ids[0])} tokens)")114            return115 116    original_question = question117    if not (shared.args.chat or shared.args.cai_chat):118        question = apply_extensions(question, "input")119    if shared.args.verbose:120        print(f"\n\n{question}\n--------------------\n")121 122    input_ids = encode(question, max_new_tokens)123    original_input_ids = input_ids124    output = input_ids[0]125    cuda = "" if (shared.args.cpu or shared.args.deepspeed or shared.args.flexgen) else ".cuda()"126    eos_token_ids = [shared.tokenizer.eos_token_id] if shared.tokenizer.eos_token_id is not None else []127    if eos_token is not None:128        eos_token_ids.append(int(encode(eos_token)[0][-1]))129    stopping_criteria_list = transformers.StoppingCriteriaList()130    if stopping_string is not None:131        # Copied from https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py132        t = encode(stopping_string, 0, add_special_tokens=False)133        stopping_criteria_list.append(_SentinelTokenStoppingCriteria(sentinel_token_ids=t, starting_idx=len(input_ids[0])))134 135    if not shared.args.flexgen:136        generate_params = [137            f"max_new_tokens=max_new_tokens",138            f"eos_token_id={eos_token_ids}",139            f"stopping_criteria=stopping_criteria_list",140            f"do_sample={do_sample}",141            f"temperature={temperature}",142            f"top_p={top_p}",143            f"typical_p={typical_p}",144            f"repetition_penalty={repetition_penalty}",145            f"top_k={top_k}",146            f"min_length={min_length if shared.args.no_stream else 0}",147            f"no_repeat_ngram_size={no_repeat_ngram_size}",148            f"num_beams={num_beams}",149            f"penalty_alpha={penalty_alpha}",150            f"length_penalty={length_penalty}",151            f"early_stopping={early_stopping}",152        ]153    else:154        generate_params = [155            f"max_new_tokens={max_new_tokens if shared.args.no_stream else 8}",156            f"do_sample={do_sample}",157            f"temperature={temperature}",158            f"stop={eos_token_ids[-1]}",159        ]160    if shared.args.deepspeed:161        generate_params.append("synced_gpus=True")162    if shared.soft_prompt:163        inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)164        generate_params.insert(0, "inputs_embeds=inputs_embeds")165        generate_params.insert(0, "inputs=filler_input_ids")166    else:167        generate_params.insert(0, "inputs=input_ids")168 169    try:170        # Generate the entire reply at once.171        if shared.args.no_stream:172            with torch.no_grad():173                output = eval(f"shared.model.generate({', '.join(generate_params)}){cuda}")[0]174            if shared.soft_prompt:175                output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))176 177            reply = decode(output)178            if not (shared.args.chat or shared.args.cai_chat):179                reply = original_question + apply_extensions(reply[len(question):], "output")180 181            yield formatted_outputs(reply, shared.model_name)182 183        # Stream the reply 1 token at a time.184        # This is based on the trick of using 'stopping_criteria' to create an iterator.185        elif not shared.args.flexgen:186 187            def generate_with_callback(callback=None, **kwargs):188                kwargs['stopping_criteria'].append(Stream(callback_func=callback))189                clear_torch_cache()190                with torch.no_grad():191                    shared.model.generate(**kwargs)192 193            def generate_with_streaming(**kwargs):194                return Iteratorize(generate_with_callback, kwargs, callback=None)195 196            yield formatted_outputs(original_question, shared.model_name)197            with eval(f"generate_with_streaming({', '.join(generate_params)})") as generator:198                for output in generator:199                    if shared.soft_prompt:200                        output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))201                    reply = decode(output)202 203                    if not (shared.args.chat or shared.args.cai_chat):204                        reply = original_question + apply_extensions(reply[len(question):], "output")205 206                    if output[-1] in eos_token_ids:207                        break208                    yield formatted_outputs(reply, shared.model_name)209 210                yield formatted_outputs(reply, shared.model_name)211 212        # Stream the output naively for FlexGen since it doesn't support 'stopping_criteria'213        else:214            for i in range(max_new_tokens//8+1):215                clear_torch_cache()216                with torch.no_grad():217                    output = eval(f"shared.model.generate({', '.join(generate_params)})")[0]218                if shared.soft_prompt:219                    output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))220                reply = decode(output)221 222                if not (shared.args.chat or shared.args.cai_chat):223                    reply = original_question + apply_extensions(reply[len(question):], "output")224 225                if np.count_nonzero(np.isin(input_ids[0], eos_token_ids)) < np.count_nonzero(np.isin(output, eos_token_ids)):226                    break227                yield formatted_outputs(reply, shared.model_name)228 229                input_ids = np.reshape(output, (1, output.shape[0]))230                if shared.soft_prompt:231                    inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)232 233            yield formatted_outputs(reply, shared.model_name)234 235    finally:236        t1 = time.time()237        print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(original_input_ids[0]))/(t1-t0):.2f} tokens/s, {len(output)-len(original_input_ids[0])} tokens)")238        return239