Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,4 +1,119 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
import gradio as gr
|
4 |
def qna_chatbot(message, history):
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
from transformers import (
|
4 |
+
AutoTokenizer,
|
5 |
+
AutoModelForCausalLM,
|
6 |
+
BitsAndBytesConfig,
|
7 |
+
pipeline
|
8 |
+
)
|
9 |
|
10 |
+
from transformers import BitsAndBytesConfig
|
11 |
+
from langchain.embeddings.huggingface import HuggingFaceEmbeddings
|
12 |
+
from langchain.vectorstores import FAISS
|
13 |
+
|
14 |
+
from langchain.prompts import PromptTemplate
|
15 |
+
from langchain.schema.runnable import RunnablePassthrough
|
16 |
+
from langchain.llms import HuggingFacePipeline
|
17 |
+
from langchain.chains import LLMChain
|
18 |
+
import transformers
|
19 |
+
model_name='mistralai/Mistral-7B-Instruct-v0.1'
|
20 |
+
|
21 |
+
model_config = transformers.AutoConfig.from_pretrained(
|
22 |
+
model_name,
|
23 |
+
)
|
24 |
+
|
25 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
26 |
+
tokenizer.pad_token = tokenizer.eos_token
|
27 |
+
tokenizer.padding_side = "right"
|
28 |
+
#################################################################
|
29 |
+
# bitsandbytes parameters
|
30 |
+
#################################################################
|
31 |
+
|
32 |
+
# Activate 4-bit precision base model loading
|
33 |
+
use_4bit = True
|
34 |
+
|
35 |
+
# Compute dtype for 4-bit base models
|
36 |
+
bnb_4bit_compute_dtype = "float16"
|
37 |
+
|
38 |
+
# Quantization type (fp4 or nf4)
|
39 |
+
bnb_4bit_quant_type = "nf4"
|
40 |
+
|
41 |
+
# Activate nested quantization for 4-bit base models (double quantization)
|
42 |
+
use_nested_quant = False
|
43 |
+
#################################################################
|
44 |
+
# Set up quantization config
|
45 |
+
#################################################################
|
46 |
+
compute_dtype = getattr(torch, bnb_4bit_compute_dtype)
|
47 |
+
|
48 |
+
bnb_config = BitsAndBytesConfig(
|
49 |
+
load_in_4bit=use_4bit,
|
50 |
+
bnb_4bit_quant_type=bnb_4bit_quant_type,
|
51 |
+
bnb_4bit_compute_dtype=compute_dtype,
|
52 |
+
bnb_4bit_use_double_quant=use_nested_quant,
|
53 |
+
)
|
54 |
+
#############################################################
|
55 |
+
# Load pre-trained config
|
56 |
+
#################################################################
|
57 |
+
model = AutoModelForCausalLM.from_pretrained(
|
58 |
+
model_name,
|
59 |
+
quantization_config=bnb_config,
|
60 |
+
)
|
61 |
+
# Connect query to FAISS index using a retriever
|
62 |
+
retriever = db.as_retriever(
|
63 |
+
search_type="mmr",
|
64 |
+
search_kwargs={'k': 1}
|
65 |
+
)
|
66 |
+
from langchain.llms import HuggingFacePipeline
|
67 |
+
from langchain.prompts import PromptTemplate
|
68 |
+
from langchain.embeddings.huggingface import HuggingFaceEmbeddings
|
69 |
+
|
70 |
+
text_generation_pipeline = transformers.pipeline(
|
71 |
+
model=model,
|
72 |
+
tokenizer=tokenizer,
|
73 |
+
task="text-generation",
|
74 |
+
|
75 |
+
temperature=0.02,
|
76 |
+
repetition_penalty=1.1,
|
77 |
+
return_full_text=True,
|
78 |
+
max_new_tokens=512,
|
79 |
+
)
|
80 |
+
|
81 |
+
prompt_template = """
|
82 |
+
### [INST]
|
83 |
+
Instruction: You are a Q&A assistant. Your goal is to answer questions as accurately as possible based on the instructions and context provided without using prior knowledge.You answer in FRENCH
|
84 |
+
Analyse carefully the context and provide a direct answer based on the context.
|
85 |
+
Answer in french only
|
86 |
+
{context}
|
87 |
+
Vous devez répondre aux questions en français.
|
88 |
+
|
89 |
+
### QUESTION:
|
90 |
+
{question}
|
91 |
+
[/INST]
|
92 |
+
Answer in french only
|
93 |
+
Vous devez répondre aux questions en français.
|
94 |
+
|
95 |
+
"""
|
96 |
+
|
97 |
+
mistral_llm = HuggingFacePipeline(pipeline=text_generation_pipeline)
|
98 |
+
|
99 |
+
# Create prompt from prompt template
|
100 |
+
prompt = PromptTemplate(
|
101 |
+
input_variables=["question"],
|
102 |
+
template=prompt_template,
|
103 |
+
)
|
104 |
+
|
105 |
+
# Create llm chain
|
106 |
+
llm_chain = LLMChain(llm=mistral_llm, prompt=prompt)
|
107 |
+
from langchain.chains import RetrievalQA
|
108 |
+
|
109 |
+
|
110 |
+
retriever.search_kwargs = {'k':1}
|
111 |
+
qa = RetrievalQA.from_chain_type(
|
112 |
+
llm=mistral_llm,
|
113 |
+
chain_type="stuff",
|
114 |
+
retriever=retriever,
|
115 |
+
chain_type_kwargs={"prompt": prompt},
|
116 |
+
)
|
117 |
|
118 |
import gradio as gr
|
119 |
def qna_chatbot(message, history):
|