import asyncio import functools import math import os import tempfile import traceback import uuid from typing import Any, Tuple import aiohttp import backoff import discord # An enum of two modes, TOP_P or TEMPERATURE import requests from PIL import Image from discord import File class Mode: TEMPERATURE = "temperature" TOP_P = "top_p" ALL_MODES = [TEMPERATURE, TOP_P] class Models: # Text models DAVINCI = "text-davinci-003" CURIE = "text-curie-001" BABBAGE = "text-babbage-001" ADA = "text-ada-001" # Code models CODE_DAVINCI = "code-davinci-002" CODE_CUSHMAN = "code-cushman-001" # Embedding models EMBEDDINGS = "text-embedding-ada-002" # Edit models EDIT = "text-davinci-edit-001" CODE_EDIT = "code-davinci-edit-001" # Model collections TEXT_MODELS = [DAVINCI, CURIE, BABBAGE, ADA, CODE_DAVINCI, CODE_CUSHMAN] EDIT_MODELS = [EDIT, CODE_EDIT] DEFAULT = DAVINCI LOW_USAGE_MODEL = CURIE class ImageSize: SMALL = "256x256" MEDIUM = "512x512" LARGE = "1024x1024" ALL_SIZES = [SMALL, MEDIUM, LARGE] class ModelLimits: MIN_TOKENS = 15 MAX_TOKENS = 4096 MIN_CONVERSATION_LENGTH = 1 MAX_CONVERSATION_LENGTH = 500 MIN_SUMMARIZE_THRESHOLD = 800 MAX_SUMMARIZE_THRESHOLD = 3500 MIN_NUM_IMAGES = 1 MAX_NUM_IMAGES = 4 MIN_NUM_STATIC_CONVERSATION_ITEMS = 5 MAX_NUM_STATIC_CONVERSATION_ITEMS = 20 MIN_NUM_CONVERSATION_LOOKBACK = 5 MAX_NUM_CONVERSATION_LOOKBACK = 15 MIN_TEMPERATURE = 0.0 MAX_TEMPERATURE = 2.0 MIN_TOP_P = 0.0 MAX_TOP_P = 1.0 MIN_PRESENCE_PENALTY = -2.0 MAX_PRESENCE_PENALTY = 2.0 MIN_FREQUENCY_PENALTY = -2.0 MAX_FREQUENCY_PENALTY = 2.0 MIN_BEST_OF = 1 MAX_BEST_OF = 3 MIN_PROMPT_MIN_LENGTH = 10 MAX_PROMPT_MIN_LENGTH = 4096 class Model: def __init__(self, usage_service): self._mode = Mode.TEMPERATURE self._temp = 0.8 # Higher value means more random, lower value means more likely to be a coherent sentence self._top_p = 0.95 # 1 is equivalent to greedy sampling, 0.1 means that the model will only consider the top 10% of the probability distribution self._max_tokens = 4000 # The maximum number of tokens the model can generate self._presence_penalty = ( 0 # Penalize new tokens based on whether they appear in the text so far ) # Penalize new tokens based on their existing frequency in the text so far. (Higher frequency = lower probability of being chosen.) self._frequency_penalty = 0 self._best_of = 1 # Number of responses to compare the loglikelihoods of self._prompt_min_length = 8 self._max_conversation_length = 100 self._model = Models.DEFAULT self._low_usage_mode = False self.usage_service = usage_service self.DAVINCI_ROLES = ["admin", "Admin", "GPT", "gpt"] self._image_size = ImageSize.MEDIUM self._num_images = 2 self._summarize_conversations = True self._summarize_threshold = 3000 self.model_max_tokens = 4024 self._welcome_message_enabled = True self._num_static_conversation_items = 10 self._num_conversation_lookback = 5 try: self.IMAGE_SAVE_PATH = os.environ["IMAGE_SAVE_PATH"] self.custom_image_path = True except Exception: self.IMAGE_SAVE_PATH = "dalleimages" # Try to make this folder called images/ in the local directory if it doesnt exist if not os.path.exists(self.IMAGE_SAVE_PATH): os.makedirs(self.IMAGE_SAVE_PATH) self.custom_image_path = False self._hidden_attributes = [ "usage_service", "DAVINCI_ROLES", "custom_image_path", "custom_web_root", "_hidden_attributes", "model_max_tokens", "openai_key", ] self.openai_key = os.getenv("OPENAI_TOKEN") # Use the @property and @setter decorators for all the self fields to provide value checking @property def num_static_conversation_items(self): return self._num_static_conversation_items @num_static_conversation_items.setter def num_static_conversation_items(self, value): value = int(value) if value < ModelLimits.MIN_NUM_STATIC_CONVERSATION_ITEMS: raise ValueError(f"Number of static conversation items must be >= {ModelLimits.MIN_NUM_STATIC_CONVERSATION_ITEMS}") if value > ModelLimits.MAX_NUM_STATIC_CONVERSATION_ITEMS: raise ValueError( f"Number of static conversation items must be <= {ModelLimits.MAX_NUM_STATIC_CONVERSATION_ITEMS}, this is to ensure reliability and reduce token wastage!" ) self._num_static_conversation_items = value @property def num_conversation_lookback(self): return self._num_conversation_lookback @num_conversation_lookback.setter def num_conversation_lookback(self, value): value = int(value) if value < ModelLimits.MIN_NUM_CONVERSATION_LOOKBACK: raise ValueError(f"Number of conversations to look back on must be >= {ModelLimits.MIN_NUM_CONVERSATION_LOOKBACK}") if value > ModelLimits.MAX_NUM_CONVERSATION_LOOKBACK: raise ValueError( f"Number of conversations to look back on must be <= {ModelLimits.MIN_NUM_CONVERSATION_LOOKBACK}, this is to ensure reliability and reduce token wastage!" ) self._num_conversation_lookback = value @property def welcome_message_enabled(self): return self._welcome_message_enabled @welcome_message_enabled.setter def welcome_message_enabled(self, value): if value.lower() == "true": self._welcome_message_enabled = True elif value.lower() == "false": self._welcome_message_enabled = False else: raise ValueError("Value must be either `true` or `false`!") @property def summarize_threshold(self): return self._summarize_threshold @summarize_threshold.setter def summarize_threshold(self, value): value = int(value) if value < ModelLimits.MIN_SUMMARIZE_THRESHOLD or value > ModelLimits.MAX_SUMMARIZE_THRESHOLD: raise ValueError( f"Summarize threshold should be a number between {ModelLimits.MIN_SUMMARIZE_THRESHOLD} and {ModelLimits.MAX_SUMMARIZE_THRESHOLD}!" ) self._summarize_threshold = value @property def summarize_conversations(self): return self._summarize_conversations @summarize_conversations.setter def summarize_conversations(self, value): # convert value string into boolean if value.lower() == "true": value = True elif value.lower() == "false": value = False else: raise ValueError("Value must be either `true` or `false`!") self._summarize_conversations = value @property def image_size(self): return self._image_size @image_size.setter def image_size(self, value): if value in ImageSize.ALL_SIZES: self._image_size = value else: raise ValueError( f"Image size must be one of the following: {ImageSize.ALL_SIZES}" ) @property def num_images(self): return self._num_images @num_images.setter def num_images(self, value): value = int(value) if value < ModelLimits.MIN_NUM_IMAGES or value > ModelLimits.MAX_NUM_IMAGES: raise ValueError(f"Number of images to generate should be a number between {ModelLimits.MIN_NUM_IMAGES} and {ModelLimits.MAX_NUM_IMAGES}!") self._num_images = value @property def low_usage_mode(self): return self._low_usage_mode @low_usage_mode.setter def low_usage_mode(self, value): # convert value string into boolean if value.lower() == "true": value = True elif value.lower() == "false": value = False else: raise ValueError("Value must be either `true` or `false`!") if value: self._model = Models.LOW_USAGE_MODEL self.max_tokens = 1900 self.model_max_tokens = 1000 else: self._model = Models.DEFAULT self.max_tokens = 4000 self.model_max_tokens = 4024 @property def model(self): return self._model @model.setter def model(self, model): if model not in Models.TEXT_MODELS: raise ValueError(f"Invalid model, must be one of: {Models.TEXT_MODELS}") self._model = model @property def max_conversation_length(self): return self._max_conversation_length @max_conversation_length.setter def max_conversation_length(self, value): value = int(value) if value < ModelLimits.MIN_CONVERSATION_LENGTH: raise ValueError(f"Max conversation length must be greater than {ModelLimits.MIN_CONVERSATION_LENGTH}") if value > ModelLimits.MAX_CONVERSATION_LENGTH: raise ValueError( f"Max conversation length must be less than {ModelLimits.MIN_CONVERSATION_LENGTH}, this will start using credits quick." ) self._max_conversation_length = value @property def mode(self): return self._mode @mode.setter def mode(self, value): if value not in Mode.ALL_MODES: raise ValueError(f"Mode must be one of: {Mode.ALL_MODES}") if value == Mode.TOP_P: self._top_p = 0.1 self._temp = 0.7 elif value == Mode.TEMPERATURE: self._top_p = 0.9 self._temp = 0.6 else: raise ValueError(f"Unknown mode: {value}") self._mode = value @property def temp(self): return self._temp @temp.setter def temp(self, value): value = float(value) if value < ModelLimits.MIN_TEMPERATURE or value > ModelLimits.MAX_TEMPERATURE: raise ValueError( f"Temperature must be between {ModelLimits.MIN_TEMPERATURE} and {ModelLimits.MAX_TEMPERATURE}, it is currently: {value}" ) self._temp = value @property def top_p(self): return self._top_p @top_p.setter def top_p(self, value): value = float(value) if value < ModelLimits.MIN_TOP_P or value > ModelLimits.MAX_TOP_P: raise ValueError( f"Top P must be between {ModelLimits.MIN_TOP_P} and {ModelLimits.MAX_TOP_P}, it is currently: {value}" ) self._top_p = value @property def max_tokens(self): return self._max_tokens @max_tokens.setter def max_tokens(self, value): value = int(value) if value < ModelLimits.MIN_TOKENS or value > ModelLimits.MAX_TOKENS: raise ValueError( f"Max tokens must be between {ModelLimits.MIN_TOKENS} and {ModelLimits.MAX_TOKENS}, it is currently: {value}" ) self._max_tokens = value @property def presence_penalty(self): return self._presence_penalty @presence_penalty.setter def presence_penalty(self, value): value = float(value) if value < ModelLimits.MIN_PRESENCE_PENALTY or value > ModelLimits.MAX_PRESENCE_PENALTY: raise ValueError( f"Presence penalty must be between {ModelLimits.MIN_PRESENCE_PENALTY} and {ModelLimits.MAX_PRESENCE_PENALTY}, it is currently: {value}" ) self._presence_penalty = value @property def frequency_penalty(self): return self._frequency_penalty @frequency_penalty.setter def frequency_penalty(self, value): value = float(value) if value < ModelLimits.MIN_FREQUENCY_PENALTY or value > ModelLimits.MAX_FREQUENCY_PENALTY: raise ValueError( f"Frequency penalty must be greater between {ModelLimits.MIN_FREQUENCY_PENALTY} and {ModelLimits.MAX_FREQUENCY_PENALTY}, it is currently: {value}" ) self._frequency_penalty = value @property def best_of(self): return self._best_of @best_of.setter def best_of(self, value): value = int(value) if value < ModelLimits.MIN_BEST_OF or value > ModelLimits.MAX_BEST_OF: raise ValueError( f"Best of must be between {ModelLimits.MIN_BEST_OF} and {ModelLimits.MAX_BEST_OF}, it is currently: {value}\nNote that increasing the value of this parameter will act as a multiplier on the number of tokens requested!" ) self._best_of = value @property def prompt_min_length(self): return self._prompt_min_length @prompt_min_length.setter def prompt_min_length(self, value): value = int(value) if value < ModelLimits.MIN_PROMPT_MIN_LENGTH or value > ModelLimits.MAX_PROMPT_MIN_LENGTH: raise ValueError( f"Minimal prompt length must be between {ModelLimits.MIN_PROMPT_MIN_LENGTH} and {ModelLimits.MAX_PROMPT_MIN_LENGTH}, it is currently: {value}" ) self._prompt_min_length = value def backoff_handler(details): print( f"Backing off {details['wait']:0.1f} seconds after {details['tries']} tries calling function {details['target']} | " f"{details['exception'].status}: {details['exception'].message}" ) async def valid_text_request(self, response): try: tokens_used = int(response["usage"]["total_tokens"]) await self.usage_service.update_usage(tokens_used) except Exception as e: raise ValueError( "The API returned an invalid response: " + str(response["error"]["message"]) ) from e @backoff.on_exception( backoff.expo, aiohttp.ClientResponseError, factor=3, base=5, max_tries=4, on_backoff=backoff_handler, ) async def send_embedding_request(self, text, custom_api_key=None): async with aiohttp.ClientSession(raise_for_status=True) as session: payload = { "model": Models.EMBEDDINGS, "input": text, } headers = { "Content-Type": "application/json", "Authorization": f"Bearer {self.openai_key if not custom_api_key else custom_api_key}", } async with session.post( "https://api.openai.com/v1/embeddings", json=payload, headers=headers ) as resp: response = await resp.json() try: return response["data"][0]["embedding"] except Exception: print(response) traceback.print_exc() return @backoff.on_exception( backoff.expo, aiohttp.ClientResponseError, factor=3, base=5, max_tries=6, on_backoff=backoff_handler, ) async def send_edit_request( self, instruction, text=None, temp_override=None, top_p_override=None, codex=False, custom_api_key=None, ): # Validate that all the parameters are in a good state before we send the request if len(instruction) < self.prompt_min_length: raise ValueError( "Instruction must be greater than 8 characters, it is currently " + str(len(instruction)) ) print( f"The text about to be edited is [{text}] with instructions [{instruction}] codex [{codex}]" ) print(f"Overrides -> temp:{temp_override}, top_p:{top_p_override}") async with aiohttp.ClientSession(raise_for_status=True) as session: payload = { "model": Models.EDIT if codex is False else Models.CODE_EDIT, "input": "" if text is None else text, "instruction": instruction, "temperature": self.temp if temp_override is None else temp_override, "top_p": self.top_p if top_p_override is None else top_p_override, } headers = { "Content-Type": "application/json", "Authorization": f"Bearer {self.openai_key if not custom_api_key else custom_api_key}", } async with session.post( "https://api.openai.com/v1/edits", json=payload, headers=headers ) as resp: response = await resp.json() await self.valid_text_request(response) return response @backoff.on_exception( backoff.expo, aiohttp.ClientResponseError, factor=3, base=5, max_tries=6, on_backoff=backoff_handler, ) async def send_moderations_request(self, text): # Use aiohttp to send the above request: async with aiohttp.ClientSession(raise_for_status=True) as session: headers = { "Content-Type": "application/json", "Authorization": f"Bearer {self.openai_key}", } payload = {"input": text} async with session.post( "https://api.openai.com/v1/moderations", headers=headers, json=payload, ) as response: return await response.json() @backoff.on_exception( backoff.expo, aiohttp.ClientResponseError, factor=3, base=5, max_tries=4, on_backoff=backoff_handler, ) async def send_summary_request(self, prompt, custom_api_key=None): """ Sends a summary request to the OpenAI API """ summary_request_text = [] summary_request_text.append( "The following is a conversation instruction set and a conversation between two people, a , and GPTie." " Firstly, determine the 's name from the conversation history, then summarize the conversation." " Do not summarize the instructions for GPTie, only the conversation. Summarize the conversation in a detailed fashion. If mentioned" " their name, be sure to mention it in the summary. Pay close attention to things the has told you, such as personal details." ) summary_request_text.append(prompt + "\nDetailed summary of conversation: \n") summary_request_text = "".join(summary_request_text) tokens = self.usage_service.count_tokens(summary_request_text) async with aiohttp.ClientSession(raise_for_status=True) as session: payload = { "model": Models.DAVINCI, "prompt": summary_request_text, "temperature": 0.5, "top_p": 1, "max_tokens": self.max_tokens - tokens, "presence_penalty": self.presence_penalty, "frequency_penalty": self.frequency_penalty, "best_of": self.best_of, } headers = { "Content-Type": "application/json", "Authorization": f"Bearer {self.openai_key if not custom_api_key else custom_api_key}", } async with session.post( "https://api.openai.com/v1/completions", json=payload, headers=headers ) as resp: response = await resp.json() await self.valid_text_request(response) # print(response["choices"][0]["text"]) return response @backoff.on_exception( backoff.expo, aiohttp.ClientResponseError, factor=3, base=5, max_tries=4, on_backoff=backoff_handler, ) async def send_request( self, prompt, tokens, temp_override=None, top_p_override=None, best_of_override=None, frequency_penalty_override=None, presence_penalty_override=None, max_tokens_override=None, model=None, stop=None, custom_api_key=None, ) -> ( Tuple[dict, bool] ): # The response, and a boolean indicating whether or not the context limit was reached. # Validate that all the parameters are in a good state before we send the request if len(prompt) < self.prompt_min_length: raise ValueError( f"Prompt must be greater than {self.prompt_min_length} characters, it is currently: {len(prompt)} characters" ) print(f"The prompt about to be sent is {prompt}") print( f"Overrides -> temp:{temp_override}, top_p:{top_p_override} frequency:{frequency_penalty_override}, presence:{presence_penalty_override}" ) async with aiohttp.ClientSession(raise_for_status=True) as session: payload = { "model": self.model if model is None else model, "prompt": prompt, "stop": "" if stop is None else stop, "temperature": self.temp if temp_override is None else temp_override, "top_p": self.top_p if top_p_override is None else top_p_override, "max_tokens": self.max_tokens - tokens if not max_tokens_override else max_tokens_override, "presence_penalty": self.presence_penalty if presence_penalty_override is None else presence_penalty_override, "frequency_penalty": self.frequency_penalty if frequency_penalty_override is None else frequency_penalty_override, "best_of": self.best_of if not best_of_override else best_of_override, } headers = { "Authorization": f"Bearer {self.openai_key if not custom_api_key else custom_api_key}" } async with session.post( "https://api.openai.com/v1/completions", json=payload, headers=headers ) as resp: response = await resp.json() # print(f"Payload -> {payload}") # Parse the total tokens used for this request and response pair from the response await self.valid_text_request(response) print(f"Response -> {response}") return response @staticmethod async def send_test_request(api_key): async with aiohttp.ClientSession() as session: payload = { "model": Models.LOW_USAGE_MODEL, "prompt": "test.", "temperature": 1, "top_p": 1, "max_tokens": 10, } headers = {"Authorization": f"Bearer {api_key}"} async with session.post( "https://api.openai.com/v1/completions", json=payload, headers=headers ) as resp: response = await resp.json() try: int(response["usage"]["total_tokens"]) except: raise ValueError(str(response["error"]["message"])) return response @backoff.on_exception( backoff.expo, aiohttp.ClientResponseError, factor=3, base=5, max_tries=4, on_backoff=backoff_handler, ) async def send_image_request( self, ctx, prompt, vary=None, custom_api_key=None ) -> tuple[File, list[Any]]: # Validate that all the parameters are in a good state before we send the request words = len(prompt.split(" ")) if words < 3 or words > 75: raise ValueError( "Prompt must be greater than 3 words and less than 75, it is currently " + str(words) ) # print("The prompt about to be sent is " + prompt) await self.usage_service.update_usage_image(self.image_size) response = None if not vary: payload = {"prompt": prompt, "n": self.num_images, "size": self.image_size} headers = { "Content-Type": "application/json", "Authorization": f"Bearer {self.openai_key if not custom_api_key else custom_api_key}", } async with aiohttp.ClientSession(raise_for_status=True) as session: async with session.post( "https://api.openai.com/v1/images/generations", json=payload, headers=headers, ) as resp: response = await resp.json() else: async with aiohttp.ClientSession(raise_for_status=True) as session: data = aiohttp.FormData() data.add_field("n", str(self.num_images)) data.add_field("size", self.image_size) with open(vary, "rb") as f: data.add_field( "image", f, filename="file.png", content_type="image/png" ) async with session.post( "https://api.openai.com/v1/images/variations", headers={ "Authorization": f"Bearer {self.openai_key if not custom_api_key else custom_api_key}", }, data=data, ) as resp: response = await resp.json() print(response) image_urls = [] for result in response["data"]: image_urls.append(result["url"]) # For each image url, open it as an image object using PIL images = await asyncio.get_running_loop().run_in_executor( None, lambda: [ Image.open(requests.get(url, stream=True, timeout=10).raw) for url in image_urls ], ) # Save all the images with a random name to self.IMAGE_SAVE_PATH image_names = [f"{uuid.uuid4()}.png" for _ in range(len(images))] for image, name in zip(images, image_names): await asyncio.get_running_loop().run_in_executor( None, image.save, f"{self.IMAGE_SAVE_PATH}/{name}" ) # Update image_urls to include the local path to these new images image_urls = [f"{self.IMAGE_SAVE_PATH}/{name}" for name in image_names] widths, heights = zip(*(i.size for i in images)) # Calculate the number of rows and columns needed for the grid num_rows = num_cols = int(math.ceil(math.sqrt(len(images)))) # If there are only 2 images, set the number of rows to 1 if len(images) == 2: num_rows = 1 # Calculate the size of the combined image width = max(widths) * num_cols height = max(heights) * num_rows # Create a transparent image with the same size as the images transparent = await asyncio.get_running_loop().run_in_executor( None, lambda: Image.new("RGBA", (max(widths), max(heights))) ) # Create a new image with the calculated size new_im = await asyncio.get_running_loop().run_in_executor( None, lambda: Image.new("RGBA", (width, height)) ) # Paste the images and transparent segments into the grid x_offset = y_offset = 0 for im in images: await asyncio.get_running_loop().run_in_executor( None, new_im.paste, im, (x_offset, y_offset) ) x_offset += im.size[0] if x_offset >= width: x_offset = 0 y_offset += im.size[1] # Fill the remaining cells with transparent segments while y_offset < height: while x_offset < width: await asyncio.get_running_loop().run_in_executor( None, new_im.paste, transparent, (x_offset, y_offset) ) x_offset += transparent.size[0] x_offset = 0 y_offset += transparent.size[1] # Save the new_im to a temporary file and return it as a discord.File temp_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False) await asyncio.get_running_loop().run_in_executor( None, new_im.save, temp_file.name ) # Print the filesize of new_im, in mega bytes image_size = os.path.getsize(temp_file.name) / 1048576 if ctx.guild is None: guild_file_limit = 8 else: guild_file_limit = ctx.guild.filesize_limit / 1048576 # If the image size is greater than 8MB, we can't return this to the user, so we will need to downscale the # image and try again safety_counter = 0 while image_size > guild_file_limit: safety_counter += 1 if safety_counter >= 3: break print( f"Image size is {image_size}MB, which is too large for this server {guild_file_limit}MB. Downscaling and trying again" ) # We want to do this resizing asynchronously, so that it doesn't block the main thread during the resize. # We can use the asyncio.run_in_executor method to do this new_im = await asyncio.get_running_loop().run_in_executor( None, functools.partial( new_im.resize, (int(new_im.width / 1.05), int(new_im.height / 1.05)) ), ) temp_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False) await asyncio.get_running_loop().run_in_executor( None, new_im.save, temp_file.name ) image_size = os.path.getsize(temp_file.name) / 1000000 print(f"New image size is {image_size}MB") return (discord.File(temp_file.name), image_urls)