Refactor the code to make it more modular

api-example.py

@@ -10,7 +10,6 @@ Optionally, you can also add the --share flag to generate a public gradio URL,
 allowing you to use the API remotely.
 
 '''
-
 import requests
 
 # Server address

convert-to-flexgen.py

@@ -3,13 +3,12 @@
 Converts a transformers model to a format compatible with flexgen.
 
 '''
-
 import argparse
 import os
-import numpy as np
 from pathlib import Path
 from sys import argv
 
+import numpy as np
 import torch
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM

convert-to-safetensors.py

@@ -10,7 +10,6 @@ Based on the original script by 81300:
 https://gist.github.com/81300/fe5b08bff1cba45296a829b9d6b0f303
 
 '''
-
 import argparse
 from pathlib import Path
 from sys import argv

modules/chat.py

@@ -0,0 +1,369 @@
+import io
+import json
+import re
+from datetime import datetime
+from pathlib import Path
+
+import modules.shared as shared
+from modules.extensions import apply_extensions
+from modules.html_generator import *
+from modules.prompt import encode
+from modules.prompt import generate_reply
+from modules.prompt import get_max_prompt_length
+
+history = {'internal': [], 'visible': []}
+character = None
+
+# This gets the new line characters right.
+def clean_chat_message(text):
+    text = text.replace('\n', '\n\n')
+    text = re.sub(r"\n{3,}", "\n\n", text)
+    text = text.strip()
+    return text
+
+def generate_chat_prompt(text, tokens, name1, name2, context, chat_prompt_size, impersonate=False):
+    text = clean_chat_message(text)
+    rows = [f"{context.strip()}\n"]
+    i = len(history['internal'])-1
+    count = 0
+
+    if shared.soft_prompt:
+        chat_prompt_size -= shared.soft_prompt_tensor.shape[1]
+    max_length = min(get_max_prompt_length(tokens), chat_prompt_size)
+
+    while i >= 0 and len(encode(''.join(rows), tokens)[0]) < max_length:
+        rows.insert(1, f"{name2}: {history['internal'][i][1].strip()}\n")
+        count += 1
+        if not (history['internal'][i][0] == '<|BEGIN-VISIBLE-CHAT|>'):
+            rows.insert(1, f"{name1}: {history['internal'][i][0].strip()}\n")
+            count += 1
+        i -= 1
+
+    if not impersonate:
+        rows.append(f"{name1}: {text}\n")
+        rows.append(apply_extensions(f"{name2}:", "bot_prefix"))
+        limit = 3
+    else:
+        rows.append(f"{name1}:")
+        limit = 2
+
+    while len(rows) > limit and len(encode(''.join(rows), tokens)[0]) >= max_length:
+        rows.pop(1)
+        rows.pop(1)
+
+    question = ''.join(rows)
+    return question
+
+def extract_message_from_reply(question, reply, current, other, check, extensions=False):
+    next_character_found = False
+    substring_found = False
+
+    previous_idx = [m.start() for m in re.finditer(f"(^|\n){re.escape(current)}:", question)]
+    idx = [m.start() for m in re.finditer(f"(^|\n){re.escape(current)}:", reply)]
+    idx = idx[len(previous_idx)-1]
+
+    if extensions:
+        reply = reply[idx + 1 + len(apply_extensions(f"{current}:", "bot_prefix")):]
+    else:
+        reply = reply[idx + 1 + len(f"{current}:"):]
+
+    if check:
+        reply = reply.split('\n')[0].strip()
+    else:
+        idx = reply.find(f"\n{other}:")
+        if idx != -1:
+            reply = reply[:idx]
+            next_character_found = True
+        reply = clean_chat_message(reply)
+
+        # Detect if something like "\nYo" is generated just before
+        # "\nYou:" is completed
+        tmp = f"\n{other}:"
+        for j in range(1, len(tmp)):
+            if reply[-j:] == tmp[:j]:
+                substring_found = True
+
+    return reply, next_character_found, substring_found
+
+def generate_chat_picture(picture, name1, name2):
+    text = f'*{name1} sends {name2} a picture that contains the following: "{bot_picture.caption_image(picture)}"*'
+    buffer = BytesIO()
+    picture.save(buffer, format="JPEG")
+    img_str = base64.b64encode(buffer.getvalue()).decode('utf-8')
+    visible_text = f'<img src="data:image/jpeg;base64,{img_str}">'
+    return text, visible_text
+
+def stop_everything_event():
+    global stop_everything
+    stop_everything = True
+
+def chatbot_wrapper(text, tokens, do_sample, max_new_tokens, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, picture=None):
+    global stop_everything
+    stop_everything = False
+
+    if 'pygmalion' in shared.model_name.lower():
+        name1 = "You"
+
+    if shared.args.picture and picture is not None:
+        text, visible_text = generate_chat_picture(picture, name1, name2)
+    else:
+        visible_text = text
+        if shared.args.chat:
+            visible_text = visible_text.replace('\n', '<br>')
+
+    text = apply_extensions(text, "input")
+    question = generate_chat_prompt(text, tokens, name1, name2, context, chat_prompt_size)
+    eos_token = '\n' if check else None
+    first = True
+    for reply in generate_reply(question, tokens, do_sample, max_new_tokens, 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=eos_token, stopping_string=f"\n{name1}:"):
+        reply, next_character_found, substring_found = extract_message_from_reply(question, reply, name2, name1, check, extensions=True)
+        visible_reply = apply_extensions(reply, "output")
+        if shared.args.chat:
+            visible_reply = visible_reply.replace('\n', '<br>')
+
+        # We need this global variable to handle the Stop event,
+        # otherwise gradio gets confused
+        if stop_everything:
+            return history['visible']
+
+        if first:
+            first = False
+            history['internal'].append(['', ''])
+            history['visible'].append(['', ''])
+
+        history['internal'][-1] = [text, reply]
+        history['visible'][-1] = [visible_text, visible_reply]
+        if not substring_found:
+            yield history['visible']
+        if next_character_found:
+            break
+    yield history['visible']
+
+def impersonate_wrapper(text, tokens, do_sample, max_new_tokens, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, picture=None):
+    if 'pygmalion' in shared.model_name.lower():
+        name1 = "You"
+
+    question = generate_chat_prompt(text, tokens, name1, name2, context, chat_prompt_size, impersonate=True)
+    eos_token = '\n' if check else None
+    for reply in generate_reply(question, tokens, do_sample, max_new_tokens, 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=eos_token, stopping_string=f"\n{name2}:"):
+        reply, next_character_found, substring_found = extract_message_from_reply(question, reply, name1, name2, check, extensions=False)
+        if not substring_found:
+            yield reply
+        if next_character_found:
+            break
+    yield reply
+
+def cai_chatbot_wrapper(text, tokens, do_sample, max_new_tokens, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, picture=None):
+    for _history in chatbot_wrapper(text, tokens, do_sample, max_new_tokens, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, picture):
+        yield generate_chat_html(_history, name1, name2, character)
+
+def regenerate_wrapper(text, tokens, do_sample, max_new_tokens, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, picture=None):
+    if character is not None and len(history['visible']) == 1:
+        if shared.args.cai_chat:
+            yield generate_chat_html(history['visible'], name1, name2, character)
+        else:
+            yield history['visible']
+    else:
+        last_visible = history['visible'].pop()
+        last_internal = history['internal'].pop()
+
+        for _history in chatbot_wrapper(last_internal[0], tokens, do_sample, max_new_tokens, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, picture):
+            if shared.args.cai_chat:
+                history['visible'][-1] = [last_visible[0], _history[-1][1]]
+                yield generate_chat_html(history['visible'], name1, name2, character)
+            else:
+                history['visible'][-1] = (last_visible[0], _history[-1][1])
+                yield history['visible']
+
+def remove_last_message(name1, name2):
+    if not history['internal'][-1][0] == '<|BEGIN-VISIBLE-CHAT|>':
+        last = history['visible'].pop()
+        history['internal'].pop()
+    else:
+        last = ['', '']
+    if shared.args.cai_chat:
+        return generate_chat_html(history['visible'], name1, name2, character), last[0]
+    else:
+        return history['visible'], last[0]
+
+def send_last_reply_to_input():
+    if len(history['internal']) > 0:
+        return history['internal'][-1][1]
+    else:
+        return ''
+
+def replace_last_reply(text, name1, name2):
+    if len(history['visible']) > 0:
+        if shared.args.cai_chat:
+            history['visible'][-1][1] = text
+        else:
+            history['visible'][-1] = (history['visible'][-1][0], text)
+        history['internal'][-1][1] = apply_extensions(text, "input")
+
+    if shared.args.cai_chat:
+        return generate_chat_html(history['visible'], name1, name2, character)
+    else:
+        return history['visible']
+
+def clear_html():
+    return generate_chat_html([], "", "", character)
+
+def clear_chat_log(_character, name1, name2):
+    global history
+    if _character != 'None':
+        for i in range(len(history['internal'])):
+            if '<|BEGIN-VISIBLE-CHAT|>' in history['internal'][i][0]:
+                history['visible'] = [['', history['internal'][i][1]]]
+                history['internal'] = history['internal'][:i+1]
+                break
+    else:
+        history['internal'] = []
+        history['visible'] = []
+    if shared.args.cai_chat:
+        return generate_chat_html(history['visible'], name1, name2, character)
+    else:
+        return history['visible'] 
+
+def redraw_html(name1, name2):
+    global history
+    return generate_chat_html(history['visible'], name1, name2, character)
+
+def tokenize_dialogue(dialogue, name1, name2):
+    _history = []
+
+    dialogue = re.sub('<START>', '', dialogue)
+    dialogue = re.sub('<start>', '', dialogue)
+    dialogue = re.sub('(\n|^)[Aa]non:', '\\1You:', dialogue)
+    dialogue = re.sub('(\n|^)\[CHARACTER\]:', f'\\g<1>{name2}:', dialogue)
+    idx = [m.start() for m in re.finditer(f"(^|\n)({re.escape(name1)}|{re.escape(name2)}):", dialogue)]
+    if len(idx) == 0:
+        return _history
+
+    messages = []
+    for i in range(len(idx)-1):
+        messages.append(dialogue[idx[i]:idx[i+1]].strip())
+    messages.append(dialogue[idx[-1]:].strip())
+
+    entry = ['', '']
+    for i in messages:
+        if i.startswith(f'{name1}:'):
+            entry[0] = i[len(f'{name1}:'):].strip()
+        elif i.startswith(f'{name2}:'):
+            entry[1] = i[len(f'{name2}:'):].strip()
+            if not (len(entry[0]) == 0 and len(entry[1]) == 0):
+                _history.append(entry)
+            entry = ['', '']
+
+    print(f"\033[1;32;1m\nDialogue tokenized to:\033[0;37;0m\n", end='')
+    for row in _history:
+        for column in row:
+            print("\n")
+            for line in column.strip().split('\n'):
+                print("|  "+line+"\n")
+            print("|\n")
+        print("------------------------------")
+
+    return _history
+
+def save_history(timestamp=True):
+    if timestamp:
+        fname = f"{character or ''}{'_' if character else ''}{datetime.now().strftime('%Y%m%d-%H%M%S')}.json"
+    else:
+        fname = f"{character or ''}{'_' if character else ''}persistent.json"
+    if not Path('logs').exists():
+        Path('logs').mkdir()
+    with open(Path(f'logs/{fname}'), 'w') as f:
+        f.write(json.dumps({'data': history['internal'], 'data_visible': history['visible']}, indent=2))
+    return Path(f'logs/{fname}')
+
+def load_history(file, name1, name2):
+    global history
+    file = file.decode('utf-8')
+    try:
+        j = json.loads(file)
+        if 'data' in j:
+            history['internal'] = j['data']
+            if 'data_visible' in j:
+                history['visible'] = j['data_visible']
+            else:
+                history['visible'] = copy.deepcopy(history['internal'])
+        # Compatibility with Pygmalion AI's official web UI
+        elif 'chat' in j:
+            history['internal'] = [':'.join(x.split(':')[1:]).strip() for x in j['chat']]
+            if len(j['chat']) > 0 and j['chat'][0].startswith(f'{name2}:'):
+                history['internal'] = [['<|BEGIN-VISIBLE-CHAT|>', history['internal'][0]]] + [[history['internal'][i], history['internal'][i+1]] for i in range(1, len(history['internal'])-1, 2)]
+                history['visible'] = copy.deepcopy(history['internal'])
+                history['visible'][0][0] = ''
+            else:
+                history['internal'] = [[history['internal'][i], history['internal'][i+1]] for i in range(0, len(history['internal'])-1, 2)]
+                history['visible'] = copy.deepcopy(history['internal'])
+    except:
+        history['internal'] = tokenize_dialogue(file, name1, name2)
+        history['visible'] = copy.deepcopy(history['internal'])
+
+def load_character(_character, name1, name2):
+    global history, character
+    context = ""
+    history['internal'] = []
+    history['visible'] = []
+    if _character != 'None':
+        character = _character
+        data = json.loads(open(Path(f'characters/{_character}.json'), 'r').read())
+        name2 = data['char_name']
+        if 'char_persona' in data and data['char_persona'] != '':
+            context += f"{data['char_name']}'s Persona: {data['char_persona']}\n"
+        if 'world_scenario' in data and data['world_scenario'] != '':
+            context += f"Scenario: {data['world_scenario']}\n"
+        context = f"{context.strip()}\n<START>\n"
+        if 'example_dialogue' in data and data['example_dialogue'] != '':
+            history['internal'] = tokenize_dialogue(data['example_dialogue'], name1, name2)
+        if 'char_greeting' in data and len(data['char_greeting'].strip()) > 0:
+            history['internal'] += [['<|BEGIN-VISIBLE-CHAT|>', data['char_greeting']]]
+            history['visible'] += [['', apply_extensions(data['char_greeting'], "output")]]
+        else:
+            history['internal'] += [['<|BEGIN-VISIBLE-CHAT|>', "Hello there!"]]
+            history['visible'] += [['', "Hello there!"]]
+    else:
+        character = None
+        context = settings['context_pygmalion']
+        name2 = settings['name2_pygmalion']
+
+    if Path(f'logs/{character}_persistent.json').exists():
+        load_history(open(Path(f'logs/{character}_persistent.json'), 'rb').read(), name1, name2)
+
+    if shared.args.cai_chat:
+        return name2, context, generate_chat_html(history['visible'], name1, name2, character)
+    else:
+        return name2, context, history['visible']
+
+def upload_character(json_file, img, tavern=False):
+    json_file = json_file if type(json_file) == str else json_file.decode('utf-8')
+    data = json.loads(json_file)
+    outfile_name = data["char_name"]
+    i = 1
+    while Path(f'characters/{outfile_name}.json').exists():
+        outfile_name = f'{data["char_name"]}_{i:03d}'
+        i += 1
+    if tavern:
+        outfile_name = f'TavernAI-{outfile_name}'
+    with open(Path(f'characters/{outfile_name}.json'), 'w') as f:
+        f.write(json_file)
+    if img is not None:
+        img = Image.open(io.BytesIO(img))
+        img.save(Path(f'characters/{outfile_name}.png'))
+    print(f'New character saved to "characters/{outfile_name}.json".')
+    return outfile_name
+
+def upload_tavern_character(img, name1, name2):
+    _img = Image.open(io.BytesIO(img))
+    _img.getexif()
+    decoded_string = base64.b64decode(_img.info['chara'])
+    _json = json.loads(decoded_string)
+    _json = {"char_name": _json['name'], "char_persona": _json['description'], "char_greeting": _json["first_mes"], "example_dialogue": _json['mes_example'], "world_scenario": _json['scenario']}
+    _json['example_dialogue'] = _json['example_dialogue'].replace('{{user}}', name1).replace('{{char}}', _json['char_name'])
+    return upload_character(json.dumps(_json), img, tavern=True)
+
+def upload_your_profile_picture(img):
+    img = Image.open(io.BytesIO(img))
+    img.save(Path(f'img_me.png'))
+    print(f'Profile picture saved to "img_me.png"')

modules/extensions.py

@@ -0,0 +1,41 @@
+import modules.shared as shared
+
+import extensions
+
+extension_state = {}
+available_extensions = []
+
+def apply_extensions(text, typ):
+    for ext in sorted(extension_state, key=lambda x : extension_state[x][1]):
+        if extension_state[ext][0] == True:
+            ext_string = f"extensions.{ext}.script"
+            if typ == "input" and hasattr(eval(ext_string), "input_modifier"):
+                text = eval(f"{ext_string}.input_modifier(text)")
+            elif typ == "output" and hasattr(eval(ext_string), "output_modifier"):
+                text = eval(f"{ext_string}.output_modifier(text)")
+            elif typ == "bot_prefix" and hasattr(eval(ext_string), "bot_prefix_modifier"):
+                text = eval(f"{ext_string}.bot_prefix_modifier(text)")
+    return text
+
+def update_extensions_parameters(*kwargs):
+    i = 0
+    for ext in sorted(extension_state, key=lambda x : extension_state[x][1]):
+        if extension_state[ext][0] == True:
+            params = eval(f"extensions.{ext}.script.params")
+            for param in params:
+                if len(kwargs) >= i+1:
+                    params[param] = eval(f"kwargs[{i}]")
+                    i += 1
+
+def load_extensions():
+    global extension_state
+    for i,ext in enumerate(shared.args.extensions.split(',')):
+        if ext in available_extensions:
+            print(f'Loading the extension "{ext}"... ', end='')
+            ext_string = f"extensions.{ext}.script"
+            exec(f"import {ext_string}")
+            extension_state[ext] = [True, i]
+            print(f'Ok.')
+
+def get_params(name):
+    return eval(f"extensions.{name}.script.params")

modules/html_generator.py

@@ -3,9 +3,7 @@
 This is a library for formatting GPT-4chan and chat outputs as nice HTML.
 
 '''
-
 import base64
-import copy
 import os
 import re
 from io import BytesIO

modules/prompt.py

@@ -0,0 +1,174 @@
+import time
+
+import modules.shared as shared
+import torch
+import transformers
+from modules.extensions import apply_extensions
+from modules.html_generator import *
+from modules.stopping_criteria import _SentinelTokenStoppingCriteria
+from tqdm import tqdm
+
+
+def get_max_prompt_length(tokens):
+    max_length = 2048-tokens
+    if shared.soft_prompt:
+        max_length -= shared.soft_prompt_tensor.shape[1]
+    return max_length
+
+def encode(prompt, tokens_to_generate=0, add_special_tokens=True):
+    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)
+    if shared.args.cpu or shared.args.flexgen:
+        return input_ids
+    elif shared.args.deepspeed:
+        return input_ids.to(device=local_rank)
+    else:
+        return input_ids.cuda()
+
+def decode(output_ids):
+    reply = shared.tokenizer.decode(output_ids, skip_special_tokens=True)
+    reply = reply.replace(r'<|endoftext|>', '')
+    return reply
+
+def generate_softprompt_input_tensors(input_ids):
+    inputs_embeds = shared.model.transformer.wte(input_ids)
+    inputs_embeds = torch.cat((shared.soft_prompt_tensor, inputs_embeds), dim=1)
+    filler_input_ids = torch.zeros((1, inputs_embeds.shape[1]), dtype=input_ids.dtype).to(shared.model.device)
+    filler_input_ids += shared.model.config.bos_token_id # setting dummy input_ids to bos tokens
+    return inputs_embeds, filler_input_ids
+
+# Removes empty replies from gpt4chan outputs
+def fix_gpt4chan(s):
+    for i in range(10):
+        s = re.sub("--- [0-9]*\n>>[0-9]*\n---", "---", s)
+        s = re.sub("--- [0-9]*\n *\n---", "---", s)
+        s = re.sub("--- [0-9]*\n\n\n---", "---", s)
+    return s
+
+# Fix the LaTeX equations in galactica
+def fix_galactica(s):
+    s = s.replace(r'\[', r'$')
+    s = s.replace(r'\]', r'$')
+    s = s.replace(r'\(', r'$')
+    s = s.replace(r'\)', r'$')
+    s = s.replace(r'$$', r'$')
+    s = re.sub(r'\n', r'\n\n', s)
+    s = re.sub(r"\n{3,}", "\n\n", s)
+    return s
+
+def formatted_outputs(reply, model_name):
+    if not (shared.args.chat or shared.args.cai_chat):
+        if shared.model_name.lower().startswith('galactica'):
+            reply = fix_galactica(reply)
+            return reply, reply, generate_basic_html(reply)
+        elif shared.model_name.lower().startswith(('gpt4chan', 'gpt-4chan', '4chan')):
+            reply = fix_gpt4chan(reply)
+            return reply, 'Only applicable for GALACTICA models.', generate_4chan_html(reply)
+        else:
+            return reply, 'Only applicable for GALACTICA models.', generate_basic_html(reply)
+    else:
+        return reply
+
+def generate_reply(question, tokens, do_sample, max_new_tokens, 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):
+    original_question = question
+    if not (shared.args.chat or shared.args.cai_chat):
+        question = apply_extensions(question, "input")
+    if shared.args.verbose:
+        print(f"\n\n{question}\n--------------------\n")
+
+    input_ids = encode(question, tokens)
+    cuda = "" if (shared.args.cpu or shared.args.deepspeed or shared.args.flexgen) else ".cuda()"
+    if not shared.args.flexgen:
+        n = shared.tokenizer.eos_token_id if eos_token is None else shared.tokenizer.encode(eos_token, return_tensors='pt')[0][-1]
+    else:
+        n = shared.tokenizer(eos_token).input_ids[0] if eos_token else None
+
+    if stopping_string is not None:
+        # The stopping_criteria code below was copied from
+        # https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py
+        t = encode(stopping_string, 0, add_special_tokens=False)
+        stopping_criteria_list = transformers.StoppingCriteriaList([
+            _SentinelTokenStoppingCriteria(
+                sentinel_token_ids=t,
+                starting_idx=len(input_ids[0])
+            )
+        ])
+    else:
+        stopping_criteria_list = None
+
+    if not shared.args.flexgen:
+        generate_params = [
+            f"eos_token_id={n}",
+            f"stopping_criteria=stopping_criteria_list",
+            f"do_sample={do_sample}",
+            f"temperature={temperature}",
+            f"top_p={top_p}",
+            f"typical_p={typical_p}",
+            f"repetition_penalty={repetition_penalty}",
+            f"top_k={top_k}",
+            f"min_length={min_length if shared.args.no_stream else 0}",
+            f"no_repeat_ngram_size={no_repeat_ngram_size}",
+            f"num_beams={num_beams}",
+            f"penalty_alpha={penalty_alpha}",
+            f"length_penalty={length_penalty}",
+            f"early_stopping={early_stopping}",
+        ]
+    else:
+        generate_params = [
+            f"do_sample={do_sample}",
+            f"temperature={temperature}",
+            f"stop={n}",
+        ]
+
+    if shared.args.deepspeed:
+        generate_params.append("synced_gpus=True")
+    if shared.args.no_stream:
+        generate_params.append(f"max_new_tokens=tokens")
+    else:
+        generate_params.append(f"max_new_tokens=8")
+
+    if shared.soft_prompt:
+        inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)
+        generate_params.insert(0, "inputs_embeds=inputs_embeds")
+        generate_params.insert(0, "filler_input_ids")
+    else:
+        generate_params.insert(0, "input_ids")
+
+    # Generate the entire reply at once
+    if shared.args.no_stream:
+        t0 = time.time()
+        with torch.no_grad():
+            output = eval(f"shared.model.generate({', '.join(generate_params)}){cuda}")[0]
+        if shared.soft_prompt:
+            output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
+
+        reply = decode(output)
+        if not (shared.args.chat or shared.args.cai_chat):
+            reply = original_question + apply_extensions(reply[len(question):], "output")
+        yield formatted_outputs(reply, shared.model_name)
+
+        t1 = time.time()
+        print(f"Output generated in {(t1-t0):.2f} seconds ({(len(output)-len(input_ids[0]))/(t1-t0)/8:.2f} it/s, {len(output)-len(input_ids[0])} tokens)")
+
+    # Generate the reply 8 tokens at a time
+    else:
+        yield formatted_outputs(original_question, shared.model_name)
+        for i in tqdm(range(tokens//8+1)):
+            with torch.no_grad():
+                output = eval(f"shared.model.generate({', '.join(generate_params)}){cuda}")[0]
+            if shared.soft_prompt:
+                output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:]))
+
+            reply = decode(output)
+            if not (shared.args.chat or shared.args.cai_chat):
+                reply = original_question + apply_extensions(reply[len(question):], "output")
+            yield formatted_outputs(reply, shared.model_name)
+
+            if not shared.args.flexgen:
+                input_ids = torch.reshape(output, (1, output.shape[0]))
+            else:
+                input_ids = np.reshape(output, (1, output.shape[0]))
+            if shared.soft_prompt:
+                inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids)
+
+            if output[-1] == n:
+                break

modules/shared.py

@@ -0,0 +1,39 @@
+import argparse
+
+global tokenizer
+
+model = None
+tokenizer = None
+model_name = ""
+soft_prompt_tensor = None
+soft_prompt = False
+stop_everything = False
+
+parser = argparse.ArgumentParser(formatter_class=lambda prog: argparse.HelpFormatter(prog,max_help_position=54))
+parser.add_argument('--model', type=str, help='Name of the model to load by default.')
+parser.add_argument('--notebook', action='store_true', help='Launch the web UI in notebook mode, where the output is written to the same text box as the input.')
+parser.add_argument('--chat', action='store_true', help='Launch the web UI in chat mode.')
+parser.add_argument('--cai-chat', action='store_true', help='Launch the web UI in chat mode with a style similar to Character.AI\'s. If the file img_bot.png or img_bot.jpg exists in the same folder as server.py, this image will be used as the bot\'s profile picture. Similarly, img_me.png or img_me.jpg will be used as your profile picture.')
+parser.add_argument('--picture', action='store_true', help='Adds an ability to send pictures in chat UI modes. Captions are generated by BLIP.')
+parser.add_argument('--cpu', action='store_true', help='Use the CPU to generate text.')
+parser.add_argument('--load-in-8bit', action='store_true', help='Load the model with 8-bit precision.')
+parser.add_argument('--bf16', action='store_true', help='Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU.')
+parser.add_argument('--auto-devices', action='store_true', help='Automatically split the model across the available GPU(s) and CPU.')
+parser.add_argument('--disk', action='store_true', help='If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk.')
+parser.add_argument('--disk-cache-dir', type=str, default="cache", help='Directory to save the disk cache to. Defaults to "cache".')
+parser.add_argument('--gpu-memory', type=int, help='Maximum GPU memory in GiB to allocate. This is useful if you get out of memory errors while trying to generate text. Must be an integer number.')
+parser.add_argument('--cpu-memory', type=int, help='Maximum CPU memory in GiB to allocate for offloaded weights. Must be an integer number. Defaults to 99.')
+parser.add_argument('--flexgen', action='store_true', help='Enable the use of FlexGen offloading.')
+parser.add_argument('--percent', nargs="+", type=int, default=[0, 100, 100, 0, 100, 0], help='FlexGen: allocation percentages. Must be 6 numbers separated by spaces (default: 0, 100, 100, 0, 100, 0).')
+parser.add_argument("--compress-weight", action="store_true", help="FlexGen: activate weight compression.")
+parser.add_argument('--deepspeed', action='store_true', help='Enable the use of DeepSpeed ZeRO-3 for inference via the Transformers integration.')
+parser.add_argument('--nvme-offload-dir', type=str, help='DeepSpeed: Directory to use for ZeRO-3 NVME offloading.')
+parser.add_argument('--local_rank', type=int, default=0, help='DeepSpeed: Optional argument for distributed setups.')
+parser.add_argument('--no-stream', action='store_true', help='Don\'t stream the text output in real time. This improves the text generation performance.')
+parser.add_argument('--settings', type=str, help='Load the default interface settings from this json file. See settings-template.json for an example.')
+parser.add_argument('--extensions', type=str, help='The list of extensions to load. If you want to load more than one extension, write the names separated by commas and between quotation marks, "like,this".')
+parser.add_argument('--listen', action='store_true', help='Make the web UI reachable from your local network.')
+parser.add_argument('--listen-port', type=int, help='The listening port that the server will use.')
+parser.add_argument('--share', action='store_true', help='Create a public URL. This is useful for running the web UI on Google Colab or similar.')
+parser.add_argument('--verbose', action='store_true', help='Print the prompts to the terminal.')
+args = parser.parse_args()

modules/stopping_criteria.py

@@ -4,7 +4,6 @@ This code was copied from
 https://github.com/PygmalionAI/gradio-ui/
 
 '''
-
 import torch
 import transformers