from __future__ import annotations from typing import TYPE_CHECKING, Any, Callable, Dict, List, Tuple, Type import logging import json import os import datetime import hashlib import csv import requests import re import html import torch from torch import cuda, bfloat16 import sys import gc from pygments.lexers import guess_lexer, ClassNotFound import gradio as gr from pygments import highlight from pygments.lexers import guess_lexer,get_lexer_by_name from pygments.formatters import HtmlFormatter import transformers from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig import datasets from datasets import load_dataset import evaluate from transformers import LlamaForCausalLM, LlamaTokenizer from setfit import SetFitModel, SetFitTrainer def is_stop_word_or_prefix(s: str, stop_words: list) -> bool: for stop_word in stop_words: if s.endswith(stop_word): return True for i in range(1, len(stop_word)): if s.endswith(stop_word[:i]): return True return False def generate_prompt_with_history(text, history, tokenizer, max_length=2048): #prompt = "The following is a conversation between a human and an AI assistant named Baize (named after a mythical creature in Chinese folklore). Baize is an open-source AI assistant developed by UCSD and Sun Yat-Sen University. The human and the AI assistant take turns chatting. Human statements start with [|Human|] and AI assistant statements start with [|AI|]. The AI assistant always provides responses in as much detail as possible, and in Markdown format. The AI assistant always declines to engage with topics, questions and instructions related to unethical, controversial, or sensitive issues. Complete the transcript in exactly that format.\n[|Human|]Hello!\n[|AI|]Hi!" prompt = "Das folgende ist eine Unterhaltung zwischen einem Menschen und einem KI-Assistenten, der Baize genannt wird. Baize ist ein open-source KI-Assistent, der von UCSD entwickelt wurde. Der Mensch und der KI-Assistent chatten abwechselnd miteinander in deutsch. Die Antworten des KI Assistenten sind immer so ausführlich wie möglich und in Markdown Schreibweise und in deutscher Sprache. Wenn nötig übersetzt er sie ins Deutsche. Die Antworten des KI-Assistenten vermeiden Themen und Antworten zu unethischen, kontroversen oder sensiblen Themen. Die Antworten sind immer sehr höflich formuliert..\n[|Human|]Hallo!\n[|AI|]Hi!" history = ["\n[|Human|]{}\n[|AI|]{}".format(x[0],x[1]) for x in history] history.append("\n[|Human|]{}\n[|AI|]".format(text)) history_text = "" flag = False for x in history[::-1]: if tokenizer(prompt+history_text+x, return_tensors="pt")['input_ids'].size(-1) <= max_length: history_text = x + history_text flag = True else: break if flag: return prompt+history_text,tokenizer(prompt+history_text, return_tensors="pt") else: return None def load_tokenizer_and_model(base_model, load_8bit=False): if torch.cuda.is_available(): device = "cuda" else: device = "cpu" #tokenizer = AutoTokenizer.from_pretrained(base_model, use_fast = True, use_auth_token=True, bos_token='<|startoftext|>', eos_token='<|endoftext|>', pad_token='<|pad|>') tokenizer = AutoTokenizer.from_pretrained(base_model, use_fast = True, use_auth_token=True) tokenizer.pad_token = tokenizer.eos_token tokenizer.pad_token_id = tokenizer.eos_token_id if device == "cuda": model = AutoModelForCausalLM.from_pretrained( base_model, load_in_8bit=load_8bit, torch_dtype=torch.float16, device_map="auto", use_auth_token=True, ) else: model = AutoModelForCausalLM.from_pretrained( base_model, device_map={"": device}, low_cpu_mem_usage=True ) return tokenizer,model, device # hier werden aber Chat-Daten geladen!!!! def load_tokenizer_and_model_Baize(base_model, load_8bit=True): if torch.cuda.is_available(): device = "cuda" else: device = "cpu" tokenizer = LlamaTokenizer.from_pretrained(base_model, add_eos_token=True, use_auth_token=True) model = LlamaForCausalLM.from_pretrained(base_model, load_in_8bit=True, device_map="auto") return tokenizer,model, device def load_model(base_model, load_8bit=False): if torch.cuda.is_available(): device = "cuda" else: device = "cpu" if device == "cuda": model = AutoModelForCausalLM.from_pretrained( base_model, load_in_8bit=load_8bit, torch_dtype=torch.float16, device_map="auto", use_auth_token=True ) else: model = AutoModelForCausalLM.from_pretrained( base_model, device_map={"": device}, low_cpu_mem_usage=True, use_auth_token=True ) #if not load_8bit: #model.half() # seems to fix bugs for some users. model.eval() return model, device def load_tokenizer(base_model): if torch.cuda.is_available(): device = "cuda" else: device = "cpu" tokenizer = AutoTokenizer.from_pretrained(base_model, use_fast = True) return tokenizer # Greedy Search def greedy_search(input_ids: torch.Tensor, model: torch.nn.Module, tokenizer: transformers.PreTrainedTokenizer, stop_words: list, max_length: int, temperature: float = 1.0, top_p: float = 1.0, top_k: int = 25) -> Iterator[str]: generated_tokens = [] past_key_values = None current_length = 1 for i in range(max_length): with torch.no_grad(): if past_key_values is None: outputs = model(input_ids) else: outputs = model(input_ids[:, -1:], past_key_values=past_key_values) logits = outputs.logits[:, -1, :] past_key_values = outputs.past_key_values # apply temperature logits /= temperature probs = torch.softmax(logits, dim=-1) # apply top_p probs_sort, probs_idx = torch.sort(probs, dim=-1, descending=True) probs_sum = torch.cumsum(probs_sort, dim=-1) mask = probs_sum - probs_sort > top_p probs_sort[mask] = 0.0 # apply top_k #if top_k is not None: # probs_sort1, _ = torch.topk(probs_sort, top_k) # min_top_probs_sort = torch.min(probs_sort1, dim=-1, keepdim=True).values # probs_sort = torch.where(probs_sort < min_top_probs_sort, torch.full_like(probs_sort, float(0.0)), probs_sort) probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True)) next_token = torch.multinomial(probs_sort, num_samples=1) next_token = torch.gather(probs_idx, -1, next_token) input_ids = torch.cat((input_ids, next_token), dim=-1) generated_tokens.append(next_token[0].item()) text = tokenizer.decode(generated_tokens) yield text if any([x in text for x in stop_words]): del past_key_values del logits del probs del probs_sort del probs_idx del probs_sum gc.collect() return ######################################## #Predict def predict(model, tokenizer, device, text, history, top_p, temperature, max_length_tokens, max_context_length_tokens,): if text=="": return "Leer" try: model except: return [[text,"No Model Found"]] inputs = generate_prompt_with_history(text,history,tokenizer,max_length=max_context_length_tokens) if inputs is None: return "Too long" else: prompt,inputs=inputs begin_length = len(prompt) input_ids = inputs["input_ids"][:,-max_context_length_tokens:].to(device) torch.cuda.empty_cache() #torch.no_grad() bedeutet, dass für die betreffenden tensoren keine Ableitungen berechnet werden bei der backpropagation #hier soll das NN ja auch nicht geändert werden 8backprop ist nicht nötig), da es um interference-prompts geht! with torch.no_grad(): antwort=[[""],[""]] #die vergangenen prompts werden alle als Tupel in history abgelegt sortiert nach 'Human' und 'AI'- dass sind daher auch die stop-words, die den jeweils nächsten Eintrag kennzeichnen for x in greedy_search(input_ids,model,tokenizer,stop_words=["[|Human|]", "[|AI|]"],max_length=max_length_tokens,temperature=temperature,top_p=top_p): if is_stop_word_or_prefix(x,["[|Human|]", "[|AI|]"]) is False: if "[|Human|]" in x: x = x[:x.index("[|Human|]")].strip() if "[|AI|]" in x: x = x[:x.index("[|AI|]")].strip() x = x.strip() a, b= [[y[0],convert_to_markdown(y[1])] for y in history]+[[text, convert_to_markdown(x)]],history + [[text,x]] antwort = antwort + a del input_ids gc.collect() torch.cuda.empty_cache() try: return antwort except: pass #Funktion, die der trainer braucht, um das Training zu evaluieren - mit einer Metrik def compute_metrics(eval_pred): #Metrik berechnen, um das training messen zu können - wird es besser??? metric = evaluate.load("accuracy") #3 Arten von gegebener Metrik: f1 oder roc_auc oder accuracy logits, labels = eval_pred predictions = np.argmax(logits, axis=-1) #Call compute on metric to calculate the accuracy of your predictions. #Before passing your predictions to compute, you need to convert the predictions to logits (remember all Transformers models return logits): return metric.compute(predictions=predictions, references=labels) def compute_metrics2(p): pred, labels = p pred = np.argmax(pred, axis=1) accuracy = accuracy_score(y_true=labels, y_pred=pred) recall = recall_score(y_true=labels, y_pred=pred) precision = precision_score(y_true=labels, y_pred=pred) f1 = f1_score(y_true=labels, y_pred=pred) return {"accuracy": accuracy, "precision": precision, "recall": recall, "f1": f1} def convert_to_markdown(text): text = text.replace("$","$") def replace_leading_tabs_and_spaces(line): new_line = [] for char in line: if char == "\t": new_line.append(" ") elif char == " ": new_line.append(" ") else: break return "".join(new_line) + line[len(new_line):] markdown_text = "" lines = text.split("\n") in_code_block = False for line in lines: if in_code_block is False and line.startswith("```"): in_code_block = True markdown_text += f"{line}\n" elif in_code_block is True and line.startswith("```"): in_code_block = False markdown_text += f"{line}\n" elif in_code_block: markdown_text += f"{line}\n" else: line = replace_leading_tabs_and_spaces(line) line = re.sub(r"^(#)", r"\\\1", line) markdown_text += f"{line} \n" return markdown_text #Datasets encodieren - in train und val Sets class Dataset(torch.utils.data.Dataset): def __init__(self, encodings, labels=None): self.encodings = encodings self.labels = labels def __getitem__(self, idx): item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()} if self.labels: item["labels"] = torch.tensor(self.labels[idx]) return item def __len__(self): return len(self.encodings["input_ids"]) ####################################################### #Fine-Tuning ####################################################### #load Dataset def daten_laden(name): return load_dataset("alexkueck/tis", delimiter=";", column_names=["id", "text"]) #return load_dataset(name) return #Quantisation - tzo speed up training def bnb_config (load4Bit, double_quant): bnb_config = BitsAndBytesConfig( load_in_4bit= load4Bit, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=bfloat16, bnb_4bit_use_double_quant=double_quant, ) return bnb_config