######################################################################################### # Title: Gradio Writing Assistant # Author: Andreas Fischer # Date: May 23th, 2024 # Last update: May 23th, 2024 ########################################################################################## #https://github.com/abetlen/llama-cpp-python/issues/306 #sudo apt install libclblast-dev #CMAKE_ARGS="-DLLAMA_CLBLAST=on" FORCE_CMAKE=1 pip install llama-cpp-python --force-reinstall --upgrade --no-cache-dir -v # Prepare resources #------------------- import torch import gc torch.cuda.empty_cache() gc.collect() # Chroma-DB #----------- import os import chromadb dbPath = "/home/af/Schreibtisch/Code/gradio/Chroma/db" onPrem = True if(os.path.exists(dbPath)) else False if(onPrem==False): dbPath="/home/user/app/db" #onPrem=True # uncomment to override automatic detection print(dbPath) #client = chromadb.Client() path=dbPath client = chromadb.PersistentClient(path=path) print(client.heartbeat()) print(client.get_version()) print(client.list_collections()) from chromadb.utils import embedding_functions default_ef = embedding_functions.DefaultEmbeddingFunction() #sentence_transformer_ef = embedding_functions.SentenceTransformerEmbeddingFunction(model_name="T-Systems-onsite/cross-en-de-roberta-sentence-transformer") #instructor_ef = embedding_functions.InstructorEmbeddingFunction(model_name="hkunlp/instructor-large", device="cuda") embeddingModel = embedding_functions.SentenceTransformerEmbeddingFunction(model_name="T-Systems-onsite/cross-en-de-roberta-sentence-transformer", device="cuda" if(onPrem) else "cpu") print(str(client.list_collections())) global collection dbName="writingStyleDB1" if("name="+dbName in str(client.list_collections())): client.delete_collection(name=dbName) # deletes collection if("name="+dbName in str(client.list_collections())): print(dbName+" found!") collection = client.get_collection(name=dbName, embedding_function=embeddingModel) #sentence_transformer_ef) else: #client.delete_collection(name=dbName) print(dbName+" created!") collection = client.create_collection( dbName, embedding_function=embeddingModel, metadata={"hnsw:space": "cosine"}) print("Database ready!") print(collection.count()) x=collection.get(include=[])["ids"] if(len(x)==0): x=collection.get(include=[])["ids"] collection.add( documents=["Ich möchte einen Blogbeitrag","Ich möchte einen Gliederungsvorschlag","Ich möchte einen Social Media Beitrag"], metadatas=[ {"prompt": "Bitte schreibe einen Blogbeitrag zur Anfrage des Users!"}, {"prompt": "Bitte entwerfe einen Gliederungsvorschlag zur Anfrage des Users!"}, {"prompt": "Bitte verfasse einen Beitrag für die professionelle social media Plattform LinkedIn zur Anfrage des Users!"}], ids=[str(len(x)+1),str(len(x)+2),str(len(x)+3)] ) RAGResults=collection.query( query_texts=["Dies ist ein Test"], n_results=1, #where={"source": "USER"} ) RAGResults["metadatas"][0][0]["prompt"] x=collection.get(where_document={"$contains":"Blogbeitrag"},include=["metadatas"])['metadatas'][0]['prompt'] # Model #------- onPrem=False myModel="mistralai/Mixtral-8x7B-Instruct-v0.1" if(onPrem==False): modelPath=myModel from huggingface_hub import InferenceClient import gradio as gr client = InferenceClient( model=modelPath, #token="hf_..." ) else: import os import requests import subprocess #modelPath="/home/af/gguf/models/c4ai-command-r-v01-Q4_0.gguf" #modelPath="/home/af/gguf/models/Discolm_german_7b_v1.Q4_0.gguf" modelPath="/home/af/gguf/models/Mixtral-8x7b-instruct-v0.1.Q4_0.gguf" if(os.path.exists(modelPath)==False): #url="https://huggingface.co/TheBloke/DiscoLM_German_7b_v1-GGUF/resolve/main/discolm_german_7b_v1.Q4_0.gguf?download=true" url="https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF/resolve/main/mixtral-8x7b-instruct-v0.1.Q4_0.gguf?download=true" response = requests.get(url) with open("./Mixtral-8x7b-instruct.gguf", mode="wb") as file: file.write(response.content) print("Model downloaded") modelPath="./Mixtral-8x7b-instruct.gguf" print(modelPath) n="20" if("Mixtral-8x7b-instruct" in modelPath): n="0" # mixtral seems to cause problems here... command = ["python3", "-m", "llama_cpp.server", "--model", modelPath, "--host", "0.0.0.0", "--port", "2600", "--n_threads", "8", "--n_gpu_layers", n] subprocess.Popen(command) print("Server ready!") # Check template #---------------- if(False): from transformers import AutoTokenizer #mod="mistralai/Mixtral-8x22B-Instruct-v0.1" #mod="mistralai/Mixtral-8x7b-instruct-v0.1" mod="VAGOsolutions/Llama-3-SauerkrautLM-8b-Instruct" tok=AutoTokenizer.from_pretrained(mod) #,token="hf_...") cha=[{"role":"system","content":"A"},{"role":"user","content":"B"},{"role":"assistant","content":"C"}] res=tok.apply_chat_template(cha) print(tok.decode(res)) cha=[{"role":"user","content":"U1"},{"role":"assistant","content":"A1"},{"role":"user","content":"U2"},{"role":"assistant","content":"A2"}] res=tok.apply_chat_template(cha) print(tok.decode(res)) # Gradio-GUI #------------ import re def extend_prompt(message="", history=None, system=None, RAGAddon=None, system2=None, zeichenlimit=None,historylimit=4, removeHTML=True): startOfString="" if zeichenlimit is None: zeichenlimit=1000000000 # :-) template0=" [INST]{system}\n [/INST] " template1=" [INST] {message} [/INST]" template2=" {response}" if("command-r" in modelPath): #https://huggingface.co/CohereForAI/c4ai-command-r-v01 ## <|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello, how are you?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|> template0="<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|> {system}<|END_OF_TURN_TOKEN|>" template1="<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{message}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>" template2="{response}<|END_OF_TURN_TOKEN|>" if("Gemma-" in modelPath): # https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1 template0="user{system}" template1="user{message}model" template2="{response}" if("Mixtral-8x22B-Instruct" in modelPath): # AutoTokenizer: [INST] U1[/INST] A1[INST] U2[/INST] A2 startOfString="" template0="[INST]{system}\n [/INST] " template1="[INST] {message}[/INST]" template2=" {response}" if("Mixtral-8x7b-instruct" in modelPath): # https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1 startOfString="" # AutoTokenzizer: [INST] U1 [/INST]A1 [INST] U2 [/INST]A2 template0=" [INST]{system}\n [/INST] " template1=" [INST] {message} [/INST]" template2=" {response}" if("Mistral-7B-Instruct" in modelPath): #https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2 startOfString="" template0="[INST]{system}\n [/INST]" template1="[INST] {message} [/INST]" template2=" {response}" if("Openchat-3.5" in modelPath): #https://huggingface.co/TheBloke/openchat-3.5-0106-GGUF template0="GPT4 Correct User: {system}<|end_of_turn|>GPT4 Correct Assistant: Okay.<|end_of_turn|>" template1="GPT4 Correct User: {message}<|end_of_turn|>GPT4 Correct Assistant: " template2="{response}<|end_of_turn|>" if(("Discolm_german_7b" in modelPath) or ("SauerkrautLM-7b-HerO" in modelPath)): #https://huggingface.co/VAGOsolutions/SauerkrautLM-7b-HerO template0="<|im_start|>system\n{system}<|im_end|>\n" template1="<|im_start|>user\n{message}<|im_end|>\n<|im_start|>assistant\n" template2="{response}<|im_end|>\n" if("Llama-3-SauerkrautLM-8b-Instruct" in modelPath): #https://huggingface.co/VAGOsolutions/SauerkrautLM-7b-HerO template0="<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|>" template1="<|start_header_id|>user<|end_header_id|>\n\n{message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n" template2="{response}<|eot_id|>\n" if("WizardLM-13B-V1.2" in modelPath): #https://huggingface.co/WizardLM/WizardLM-13B-V1.2 template0="{system} " # template1="USER: {message} ASSISTANT: " template2="{response}" if("Phi-2" in modelPath): #https://huggingface.co/TheBloke/phi-2-GGUF template0="Instruct: {system}\nOutput: Okay.\n" template1="Instruct: {message}\nOutput:" template2="{response}\n" prompt = "" if RAGAddon is not None: system += RAGAddon if system is not None: prompt += template0.format(system=system) #"" if history is not None: for user_message, bot_response in history[-historylimit:]: if user_message is None: user_message = "" if bot_response is None: bot_response = "" bot_response = re.sub("\n\n
((.|\n)*?)
","", bot_response) # remove RAG-compontents if removeHTML==True: bot_response = re.sub("<(.*?)>","\n", bot_response) # remove HTML-components in general (may cause bugs with markdown-rendering) if user_message is not None: prompt += template1.format(message=user_message[:zeichenlimit]) if bot_response is not None: prompt += template2.format(response=bot_response[:zeichenlimit]) if message is not None: prompt += template1.format(message=message[:zeichenlimit]) if system2 is not None: prompt += system2 return startOfString+prompt import gradio as gr import requests import json from datetime import datetime import os import re def response(message, history,customSysPrompt, genre, hfToken): if((onPrem==False) & (hfToken.startswith("hf_"))): # use HF-hub with custom token if token is provided from huggingface_hub import InferenceClient import gradio as gr client = InferenceClient( model=myModel, token=hfToken ) removeHTML=True system=customSysPrompt # system-prompt can be changed in the UI (usually defaults to something like the following system-prompt) if(system==""): system="Du bist wissenschaftlicher Mitarbeiter an einem Forschungsinstitut und zuständig für die Wissenschaftskommunikation." message=message.replace("[INST]","") message=message.replace("[/INST]","") message=message.replace("","") message=re.sub("<[|](im_start|im_end|end_of_turn)[|]>", '', message) x=collection.get(include=[])["ids"] rag=None # RAG is turned off until history gets too long historylimit=2 if(genre==""): # use RAG to define genre if there is none RAGResults=collection.query(query_texts=[message], n_results=1) genre=str(RAGResults['documents'][0][0]) # determine genre based on best-matching db-entry rag="\n\n"+collection.get(where_document={"$contains":genre},include=["metadatas"])['metadatas'][0]['prompt'] # genre-specific addendum to system prompt (rag) if(len(history)>0): rag=rag+"\nFalls der User Rückfragen oder Änderungsvorschläge zu deinem Entwurf hat, gehe darauf ein." # add dialog-specific addendum to rag system2=None # system2 can be used as fictive first words of the AI, which are not displayed or stored prompt=extend_prompt( message, # current message of the user history, # complete history system, # system prompt rag, # RAG-component added to the system prompt system2, # fictive first words of the AI (neither displayed nor stored) historylimit=historylimit,# number of past messages to consider for response to current message removeHTML=removeHTML # remove HTML-components from History (to prevent bugs with Markdown) ) if(True): print("\n\nMESSAGE:"+str(message)) print("\n\nHISTORY:"+str(history)) print("\n\nSYSTEM:"+str(system)) print("\n\nRAG:"+str(rag)) print("\n\nSYSTEM2:"+str(system2)) print("\n\n*** Prompt:\n"+prompt+"\n***\n\n") ## Request response from model #------------------------------ print("AI running on prem!" if(onPrem) else "AI running HFHub!") if(onPrem==False): temperature=float(0.9) max_new_tokens=1000 top_p=0.95 repetition_penalty=1.0 if temperature < 1e-2: temperature = 1e-2 top_p = float(top_p) generate_kwargs = dict( temperature=temperature, max_new_tokens=max_new_tokens, top_p=top_p, repetition_penalty=repetition_penalty, do_sample=True, seed=42, ) stream = client.text_generation(prompt, **generate_kwargs, stream=True, details=True, return_full_text=False) response = "" #print("User: "+message+"\nAI: ") for text in stream: part=text.token.text #print(part, end="", flush=True) response += part if removeHTML==True: response = re.sub("<(.*?)>","\n", response) # remove HTML-components in general (may cause bugs with markdown-rendering) yield response if(onPrem==True): # url="https://afischer1985-wizardlm-13b-v1-2-q4-0-gguf.hf.space/v1/completions" url="http://0.0.0.0:2600/v1/completions" body={"prompt":prompt,"max_tokens":None, "echo":"False","stream":"True"} # e.g. Mixtral-Instruct if("Discolm_german_7b" in modelPath): body.update({"stop": ["<|im_end|>"]}) # fix stop-token of DiscoLM if("Gemma-" in modelPath): body.update({"stop": ["<|im_end|>",""]}) # fix stop-token of Gemma response="" #+"("+myType+")\n" buffer="" #print("URL: "+url) #print("User: "+message+"\nAI: ") for text in requests.post(url, json=body, stream=True): #-H 'accept: application/json' -H 'Content-Type: application/json' if buffer is None: buffer="" buffer=str("".join(buffer)) # print("*** Raw String: "+str(text)+"\n***\n") text=text.decode('utf-8') if((text.startswith(": ping -")==False) & (len(text.strip("\n\r"))>0)): buffer=buffer+str(text) # print("\n*** Buffer: "+str(buffer)+"\n***\n") buffer=buffer.split('"finish_reason": null}]}') if(len(buffer)==1): buffer="".join(buffer) pass if(len(buffer)==2): part=buffer[0]+'"finish_reason": null}]}' if(part.lstrip('\n\r').startswith("data: ")): part=part.lstrip('\n\r').replace("data: ", "") try: part = str(json.loads(part)["choices"][0]["text"]) #print(part, end="", flush=True) response=response+part buffer="" # reset buffer except Exception as e: print("Exception:"+str(e)) pass if removeHTML==True: response = re.sub("<(.*?)>","\n", response) # remove HTML-components in general (may cause bugs with markdown-rendering) yield response history.append((message, response)) # add current dialog to history val=None gr.ChatInterface( response, chatbot=gr.Chatbot(value=val, render_markdown=True), title="KI Schreibassistenz (on prem)" if onPrem else "KI Schreibassistenz (HFHub)", additional_inputs=[ gr.Textbox( value="Du bist wissenschaftlicher Mitarbeiter an einem Forschungsinstitut und zuständig für die Wissenschaftskommunikation.", label="System Prompt"), gr.Dropdown( ["Blogbeitrag","Gliederungsvorschlag","Social Media Beitrag",""], value="Blogbeitrag", label="Genre"), gr.Textbox( value="", label="HF_token"), ] ).queue().launch(share=True) #False, server_name="0.0.0.0", server_port=7864) print("Interface up and running!")