Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,61 +1,80 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
import spaces
|
3 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
from langchain.embeddings import HuggingFaceInstructEmbeddings
|
5 |
from langchain.vectorstores import Chroma
|
6 |
from langchain.prompts import PromptTemplate
|
7 |
from langchain.chains import RetrievalQA
|
8 |
-
from langchain.llms import HuggingFacePipeline
|
9 |
-
from huggingface_hub import InferenceClient
|
10 |
|
11 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
def initialize_model():
|
13 |
-
import
|
14 |
-
from
|
15 |
-
AutoTokenizer,
|
16 |
-
TextStreamer,
|
17 |
-
pipeline,
|
18 |
-
BitsAndBytesConfig,
|
19 |
-
AutoModelForCausalLM
|
20 |
-
)
|
21 |
|
22 |
model_id = "meta-llama/Llama-3.2-3B-Instruct"
|
23 |
token = os.environ.get("HF_TOKEN")
|
24 |
|
25 |
-
bnb_config = BitsAndBytesConfig(
|
26 |
-
load_in_4bit=True,
|
27 |
-
bnb_4bit_use_double_quant=True,
|
28 |
-
bnb_4bit_quant_type="nf4",
|
29 |
-
bnb_4bit_compute_dtype=torch.bfloat16
|
30 |
-
)
|
31 |
-
|
32 |
tokenizer = AutoTokenizer.from_pretrained(model_id, token=token)
|
33 |
model = AutoModelForCausalLM.from_pretrained(
|
34 |
model_id,
|
35 |
token=token,
|
36 |
-
quantization_config=bnb_config
|
37 |
)
|
38 |
|
|
|
|
|
|
|
39 |
return model, tokenizer
|
40 |
|
41 |
-
|
42 |
-
embeddings = HuggingFaceInstructEmbeddings(
|
43 |
-
model_name="hkunlp/instructor-base",
|
44 |
-
model_kwargs={"device": "cpu"}
|
45 |
-
)
|
46 |
-
|
47 |
-
db = Chroma(
|
48 |
-
persist_directory="db",
|
49 |
-
embedding_function=embeddings
|
50 |
-
)
|
51 |
-
|
52 |
-
@spaces.GPU(duration=30)
|
53 |
def respond(message, history, system_message, max_tokens, temperature, top_p):
|
54 |
try:
|
55 |
-
# Initialize model components inside
|
56 |
model, tokenizer = initialize_model()
|
57 |
-
|
58 |
|
|
|
59 |
text_pipeline = pipeline(
|
60 |
"text-generation",
|
61 |
model=model,
|
|
|
|
|
|
|
1 |
import os
|
2 |
+
import spaces
|
3 |
+
import gradio as gr
|
4 |
+
import torch
|
5 |
+
torch.jit.script = lambda f: f # Avoid script error in lambda
|
6 |
+
|
7 |
+
# Initialize non-GPU components first
|
8 |
from langchain.embeddings import HuggingFaceInstructEmbeddings
|
9 |
from langchain.vectorstores import Chroma
|
10 |
from langchain.prompts import PromptTemplate
|
11 |
from langchain.chains import RetrievalQA
|
|
|
|
|
12 |
|
13 |
+
# System prompts
|
14 |
+
DEFAULT_SYSTEM_PROMPT = """
|
15 |
+
Based on the information in this document provided in context, answer the question as accurately as possible in 1 or 2 lines. If the information is not in the context,
|
16 |
+
respond with "I don't know" or a similar acknowledgment that the answer is not available.
|
17 |
+
""".strip()
|
18 |
+
|
19 |
+
SYSTEM_PROMPT = "Use the following pieces of context to answer the question at the end. Do not provide commentary or elaboration more than 1 or 2 lines.?"
|
20 |
+
|
21 |
+
def generate_prompt(prompt: str, system_prompt: str = DEFAULT_SYSTEM_PROMPT) -> str:
|
22 |
+
return f"""
|
23 |
+
[INST] <<SYS>>
|
24 |
+
{system_prompt}
|
25 |
+
<</SYS>>
|
26 |
+
|
27 |
+
{prompt} [/INST]
|
28 |
+
""".strip()
|
29 |
+
|
30 |
+
template = generate_prompt(
|
31 |
+
"""
|
32 |
+
{context}
|
33 |
+
|
34 |
+
Question: {question}
|
35 |
+
""",
|
36 |
+
system_prompt=SYSTEM_PROMPT,
|
37 |
+
)
|
38 |
+
|
39 |
+
prompt_template = PromptTemplate(template=template, input_variables=["context", "question"])
|
40 |
+
|
41 |
+
# Initialize database and embeddings
|
42 |
+
embeddings = HuggingFaceInstructEmbeddings(
|
43 |
+
model_name="hkunlp/instructor-base",
|
44 |
+
model_kwargs={"device": "cpu"}
|
45 |
+
)
|
46 |
+
|
47 |
+
db = Chroma(
|
48 |
+
persist_directory="db",
|
49 |
+
embedding_function=embeddings
|
50 |
+
)
|
51 |
+
|
52 |
def initialize_model():
|
53 |
+
from transformers import AutoTokenizer, TextStreamer, pipeline, AutoModelForCausalLM
|
54 |
+
from langchain.llms import HuggingFacePipeline
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
|
56 |
model_id = "meta-llama/Llama-3.2-3B-Instruct"
|
57 |
token = os.environ.get("HF_TOKEN")
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
tokenizer = AutoTokenizer.from_pretrained(model_id, token=token)
|
60 |
model = AutoModelForCausalLM.from_pretrained(
|
61 |
model_id,
|
62 |
token=token,
|
|
|
63 |
)
|
64 |
|
65 |
+
if torch.cuda.is_available():
|
66 |
+
model = model.to("cuda")
|
67 |
+
|
68 |
return model, tokenizer
|
69 |
|
70 |
+
@spaces.GPU
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
def respond(message, history, system_message, max_tokens, temperature, top_p):
|
72 |
try:
|
73 |
+
# Initialize model components inside GPU context
|
74 |
model, tokenizer = initialize_model()
|
75 |
+
from transformers import TextStreamer, pipeline
|
76 |
|
77 |
+
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
|
78 |
text_pipeline = pipeline(
|
79 |
"text-generation",
|
80 |
model=model,
|