Update app.py

app.py

@@ -1,14 +1,234 @@
 import spaces
-from transformers import pipeline
+print(5)
+import argparse
+# from dataclasses import dataclass
+from langchain.prompts import ChatPromptTemplate
+
+try:
+  from langchain_community.vectorstores import Chroma
+except:
+  from langchain_community.vectorstores import Chroma
+
+# from langchain.document_loaders import DirectoryLoader
+from langchain_community.document_loaders import DirectoryLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain.schema import Document
+# from langchain.embeddings import OpenAIEmbeddings
+#from langchain_openai import OpenAIEmbeddings
+from langchain_community.vectorstores import Chroma
+import openai
+from dotenv import load_dotenv
+import os
+import shutil
+import torch
+
+from transformers import AutoModel,AutoTokenizer
+model2 = AutoModel.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
+tokenizer2 = AutoTokenizer.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
+
+
+# this shoub be used when we can not use sentence_transformers (which reqiures transformers==4.39. we cannot use
+# this version since causes using large amount of RAm when loading falcon model)
+# a custom embedding
+#from sentence_transformers import SentenceTransformer
+from langchain_experimental.text_splitter import SemanticChunker
+from typing import List
+import re
+import warnings
+from typing import List
+
+import torch
+from langchain import PromptTemplate
+from langchain.chains import ConversationChain
+from langchain.chains.conversation.memory import ConversationBufferWindowMemory
+from langchain.llms import HuggingFacePipeline
+from langchain.schema import BaseOutputParser
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    StoppingCriteria,
+    StoppingCriteriaList,
+    pipeline,
+)
+
+warnings.filterwarnings("ignore", category=UserWarning)
+
+
+class MyEmbeddings:
+    def __init__(self):
+        #self.model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+        self.model=model2
+
+    def embed_documents(self, texts: List[str]) -> List[List[float]]:
+        inputs = tokenizer2(texts, padding=True, truncation=True, return_tensors="pt")
+
+        # Get the model outputs
+        with torch.no_grad():
+          outputs = self.model(**inputs)
+
+        # Mean pooling to get sentence embeddings
+        embeddings = outputs.last_hidden_state.mean(dim=1)
+        return [embeddings[i].tolist() for i, sentence in enumerate(texts)]
+    def embed_query(self, query: str) -> List[float]:
+        inputs = tokenizer2(query, padding=True, truncation=True, return_tensors="pt")
+
+        # Get the model outputs
+        with torch.no_grad():
+          outputs = self.model(**inputs)
+
+        # Mean pooling to get sentence embeddings
+        embeddings = outputs.last_hidden_state.mean(dim=1)
+        return embeddings[0].tolist() 
+
+
+embeddings = MyEmbeddings()
+
+splitter = SemanticChunker(embeddings)
+
+
+CHROMA_PATH = "chroma8"
+# call the chroma generated in a directory
+db = Chroma(persist_directory=CHROMA_PATH, embedding_function=embeddings)
+
+
+
+MODEL_NAME = "tiiuae/falcon-7b-instruct"
+
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_NAME, trust_remote_code=True, device_map="auto",offload_folder="offload"
+)
+model = model.eval()
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+print(f"Model device: {model.device}")
+
+
+generation_config = model.generation_config
+generation_config.temperature = 0
+generation_config.num_return_sequences = 1
+generation_config.max_new_tokens = 256
+generation_config.use_cache = False
+generation_config.repetition_penalty = 1.7
+generation_config.pad_token_id = tokenizer.eos_token_id
+generation_config.eos_token_id = tokenizer.eos_token_id
+generation_config
+
+
+prompt = """
+The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context.
+
+Current conversation:
+
+Human: Who is Dwight K Schrute?
+AI:
+""".strip()
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+input_ids = input_ids.to(model.device)
+
+
+
+class StopGenerationCriteria(StoppingCriteria):
+    def __init__(
+        self, tokens: List[List[str]], tokenizer: AutoTokenizer, device: torch.device
+    ):
+        stop_token_ids = [tokenizer.convert_tokens_to_ids(t) for t in tokens]
+        self.stop_token_ids = [
+            torch.tensor(x, dtype=torch.long, device=device) for x in stop_token_ids
+        ]
+
+    def __call__(
+        self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
+    ) -> bool:
+        for stop_ids in self.stop_token_ids:
+            if torch.eq(input_ids[0][-len(stop_ids) :], stop_ids).all():
+                return True
+        return False
+
+
+stop_tokens = [["Human", ":"], ["AI", ":"]]
+stopping_criteria = StoppingCriteriaList(
+    [StopGenerationCriteria(stop_tokens, tokenizer, model.device)]
+)
+
+generation_pipeline = pipeline(
+    model=model,
+    tokenizer=tokenizer,
+    return_full_text=True,
+    task="text-generation",
+    stopping_criteria=stopping_criteria,
+    generation_config=generation_config,
+)
+
+llm = HuggingFacePipeline(pipeline=generation_pipeline)
+
+
+class CleanupOutputParser(BaseOutputParser):
+    def parse(self, text: str) -> str:
+        user_pattern = r"\nUser"
+        text = re.sub(user_pattern, "", text)
+        human_pattern = r"\nHuman:"
+        text = re.sub(human_pattern, "", text)
+        ai_pattern = r"\nAI:"
+        return re.sub(ai_pattern, "", text).strip()
+
+    @property
+    def _type(self) -> str:
+        return "output_parser"
+
+
+template = """
+The following 
+Current conversation:
+
+{history}
+
+Human: {input}
+AI:""".strip()
+prompt = PromptTemplate(input_variables=["history", "input"], template=template)
+
+memory = ConversationBufferWindowMemory(
+    memory_key="history", k=6, return_only_outputs=True
+)
+
+chain = ConversationChain(
+    llm=llm,
+    memory=memory,
+    prompt=prompt,
+    output_parser=CleanupOutputParser(),
+    verbose=True,
+)
 
-# Load your model
-model = pipeline("text-generation", model="gpt2")
-model.to("cuda")
 
 @spaces.GPU
-def generate_text(prompt):
- return model(prompt)
+def get_llama_response(message: str, history: list) -> str:
+  query_text = message
+
+  results = db.similarity_search_with_relevance_scores(query_text, k=3)
+  if len(results) == 0 or results[0][1] < 0.5:
+      print(f"Unable to find matching results.")
+
+
+  context_text = "\n\n---\n\n".join([doc.page_content for doc, _score in results])
+  template = """
+  The following is a conversation between a human an AI. The AI acts exactly like Dwight K Schrute from the TV show The Office.
+  Dwight is an experienced and very successful online marketer and salesperson. He is witty, persuasive, uncensored, direct and practical.
+  Dwight helps with every marketing task is given to him. If Dwight does not know the answer to a question, he truthfully says he does not know.
+
+  Current conversation:
+  """
+  s="""
+  {history}
+  Human: {input}
+  AI:""".strip()
+
+
+  prompt = PromptTemplate(input_variables=["history", "input"], template=template+context_text+ s)
+
+  #print(template)
+  chain.prompt=prompt
+  res = chain(query_text)
+  return(res["response"])
 
 import gradio as gr
-iface = gr.Interface(fn=generate_text, inputs="text", outputs="text")
-iface.launch()
+
+gr.ChatInterface(get_llama_response).launch()