Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,86 +1,134 @@
|
|
1 |
import streamlit as st
|
2 |
from huggingface_hub import login
|
3 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM
|
4 |
import torch
|
|
|
5 |
import os
|
6 |
|
7 |
def initialize_model():
|
8 |
-
"""Initialize the model and tokenizer"""
|
9 |
# Log in to Hugging Face
|
10 |
token = os.environ.get("hf")
|
11 |
-
|
|
|
12 |
|
13 |
-
#
|
14 |
-
model_id = "
|
15 |
-
|
16 |
-
|
17 |
-
# Configure INT8 quantization
|
18 |
-
bnb_config = BitsAndBytesConfig(
|
19 |
-
load_in_8bit=True,
|
20 |
-
llm_int8_enable_fp32_cpu_offload=True
|
21 |
-
)
|
22 |
-
|
23 |
-
# Load tokenizer and model
|
24 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
|
31 |
# Ensure padding token is defined
|
32 |
if tokenizer.pad_token is None:
|
33 |
tokenizer.pad_token = tokenizer.eos_token
|
34 |
|
35 |
-
return model, tokenizer
|
36 |
|
37 |
def format_conversation(conversation_history):
|
38 |
"""Format the conversation history into a single string."""
|
39 |
formatted = ""
|
40 |
for turn in conversation_history:
|
41 |
-
formatted += f"
|
42 |
return formatted.strip()
|
43 |
|
44 |
-
def generate_response(model, tokenizer,
|
45 |
"""Generate model response"""
|
46 |
# Format the entire conversation context
|
47 |
context = format_conversation(conversation_history[:-1])
|
48 |
if context:
|
49 |
-
full_prompt = f"{context}\
|
50 |
else:
|
51 |
-
full_prompt = f"
|
52 |
|
53 |
# Tokenize input
|
54 |
-
inputs = tokenizer(full_prompt, return_tensors="pt", padding=True, truncation=True)
|
|
|
|
|
|
|
|
|
55 |
|
56 |
# Calculate max new tokens
|
57 |
input_length = inputs["input_ids"].shape[1]
|
58 |
-
max_model_length =
|
59 |
-
max_new_tokens = min(
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
|
81 |
def main():
|
82 |
-
st.set_page_config(
|
83 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
|
85 |
# Initialize session state for chat history
|
86 |
if "chat_history" not in st.session_state:
|
@@ -89,10 +137,14 @@ def main():
|
|
89 |
# Initialize model (only once)
|
90 |
if "model" not in st.session_state:
|
91 |
with st.spinner("Loading the model... This might take a minute..."):
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
|
|
|
|
|
|
|
|
96 |
|
97 |
# Display chat messages
|
98 |
for message in st.session_state.chat_history:
|
@@ -116,7 +168,6 @@ def main():
|
|
116 |
response = generate_response(
|
117 |
st.session_state.model,
|
118 |
st.session_state.tokenizer,
|
119 |
-
st.session_state.device,
|
120 |
prompt,
|
121 |
st.session_state.chat_history
|
122 |
)
|
@@ -128,10 +179,20 @@ def main():
|
|
128 |
if len(st.session_state.chat_history) > 5:
|
129 |
st.session_state.chat_history = st.session_state.chat_history[-5:]
|
130 |
|
131 |
-
#
|
132 |
-
|
133 |
-
st.
|
134 |
-
st.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
|
136 |
if __name__ == "__main__":
|
137 |
main()
|
|
|
1 |
import streamlit as st
|
2 |
from huggingface_hub import login
|
3 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
4 |
import torch
|
5 |
+
from transformers import BitsAndBytesConfig
|
6 |
import os
|
7 |
|
8 |
def initialize_model():
|
9 |
+
"""Initialize the model and tokenizer with CPU support"""
|
10 |
# Log in to Hugging Face
|
11 |
token = os.environ.get("hf")
|
12 |
+
if token:
|
13 |
+
login(token)
|
14 |
|
15 |
+
# Use a smaller model that's more CPU-friendly
|
16 |
+
model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0" # Much smaller model
|
17 |
+
|
18 |
+
# Load tokenizer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
20 |
+
|
21 |
+
# Configure 4-bit quantization for CPU
|
22 |
+
try:
|
23 |
+
# First try with bitsandbytes 4-bit quantization
|
24 |
+
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
|
25 |
+
|
26 |
+
compute_dtype = getattr(torch, "float16")
|
27 |
+
|
28 |
+
bnb_config = BitsAndBytesConfig(
|
29 |
+
load_in_4bit=True,
|
30 |
+
bnb_4bit_quant_type="nf4",
|
31 |
+
bnb_4bit_compute_dtype=compute_dtype,
|
32 |
+
bnb_4bit_use_double_quant=False,
|
33 |
+
)
|
34 |
+
|
35 |
+
model = AutoModelForCausalLM.from_pretrained(
|
36 |
+
model_id,
|
37 |
+
quantization_config=bnb_config,
|
38 |
+
device_map="auto",
|
39 |
+
trust_remote_code=True
|
40 |
+
)
|
41 |
+
except:
|
42 |
+
# Fallback to CPU without quantization
|
43 |
+
print("Falling back to CPU without quantization")
|
44 |
+
model = AutoModelForCausalLM.from_pretrained(
|
45 |
+
model_id,
|
46 |
+
device_map="cpu",
|
47 |
+
trust_remote_code=True,
|
48 |
+
low_cpu_mem_usage=True
|
49 |
+
)
|
50 |
|
51 |
# Ensure padding token is defined
|
52 |
if tokenizer.pad_token is None:
|
53 |
tokenizer.pad_token = tokenizer.eos_token
|
54 |
|
55 |
+
return model, tokenizer
|
56 |
|
57 |
def format_conversation(conversation_history):
|
58 |
"""Format the conversation history into a single string."""
|
59 |
formatted = ""
|
60 |
for turn in conversation_history:
|
61 |
+
formatted += f"Human: {turn['user']}\nAssistant: {turn['assistant']}\n"
|
62 |
return formatted.strip()
|
63 |
|
64 |
+
def generate_response(model, tokenizer, prompt, conversation_history):
|
65 |
"""Generate model response"""
|
66 |
# Format the entire conversation context
|
67 |
context = format_conversation(conversation_history[:-1])
|
68 |
if context:
|
69 |
+
full_prompt = f"{context}\nHuman: {prompt}"
|
70 |
else:
|
71 |
+
full_prompt = f"Human: {prompt}"
|
72 |
|
73 |
# Tokenize input
|
74 |
+
inputs = tokenizer(full_prompt, return_tensors="pt", padding=True, truncation=True)
|
75 |
+
|
76 |
+
# Move inputs to the same device as the model
|
77 |
+
device = next(model.parameters()).device
|
78 |
+
inputs = {k: v.to(device) for k, v in inputs.items()}
|
79 |
|
80 |
# Calculate max new tokens
|
81 |
input_length = inputs["input_ids"].shape[1]
|
82 |
+
max_model_length = 1024 # Reduced context window for memory efficiency
|
83 |
+
max_new_tokens = min(150, max_model_length - input_length)
|
84 |
+
|
85 |
+
try:
|
86 |
+
# Generate response with lower temperature for faster generation
|
87 |
+
outputs = model.generate(
|
88 |
+
inputs["input_ids"],
|
89 |
+
attention_mask=inputs["attention_mask"],
|
90 |
+
max_new_tokens=max_new_tokens,
|
91 |
+
temperature=0.5, # Lower temperature for faster, more focused responses
|
92 |
+
top_p=0.9,
|
93 |
+
pad_token_id=tokenizer.pad_token_id,
|
94 |
+
do_sample=True,
|
95 |
+
min_length=10, # Reduced minimum length
|
96 |
+
no_repeat_ngram_size=3
|
97 |
+
)
|
98 |
+
|
99 |
+
# Decode response
|
100 |
+
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
101 |
+
response_parts = response.split("Human: ")
|
102 |
+
model_response = response_parts[-1].split("Assistant: ")[-1].strip()
|
103 |
+
|
104 |
+
return model_response
|
105 |
+
except RuntimeError as e:
|
106 |
+
if "out of memory" in str(e):
|
107 |
+
torch.cuda.empty_cache()
|
108 |
+
return "I apologize, but I ran out of memory. Please try a shorter message or clear the chat history."
|
109 |
+
else:
|
110 |
+
return f"An error occurred: {str(e)}"
|
111 |
|
112 |
def main():
|
113 |
+
st.set_page_config(
|
114 |
+
page_title="LLM Chat Interface",
|
115 |
+
page_icon="🤖",
|
116 |
+
layout="wide"
|
117 |
+
)
|
118 |
+
|
119 |
+
# Add CSS to make the chat interface more compact
|
120 |
+
st.markdown("""
|
121 |
+
<style>
|
122 |
+
.stChat {
|
123 |
+
padding-top: 0rem;
|
124 |
+
}
|
125 |
+
.stChatMessage {
|
126 |
+
padding: 0.5rem;
|
127 |
+
}
|
128 |
+
</style>
|
129 |
+
""", unsafe_allow_html=True)
|
130 |
+
|
131 |
+
st.title("Welcome to LowCode No Code Demo")
|
132 |
|
133 |
# Initialize session state for chat history
|
134 |
if "chat_history" not in st.session_state:
|
|
|
137 |
# Initialize model (only once)
|
138 |
if "model" not in st.session_state:
|
139 |
with st.spinner("Loading the model... This might take a minute..."):
|
140 |
+
try:
|
141 |
+
model, tokenizer = initialize_model()
|
142 |
+
st.session_state.model = model
|
143 |
+
st.session_state.tokenizer = tokenizer
|
144 |
+
st.success("Model loaded successfully!")
|
145 |
+
except Exception as e:
|
146 |
+
st.error(f"Error loading model: {str(e)}")
|
147 |
+
return
|
148 |
|
149 |
# Display chat messages
|
150 |
for message in st.session_state.chat_history:
|
|
|
168 |
response = generate_response(
|
169 |
st.session_state.model,
|
170 |
st.session_state.tokenizer,
|
|
|
171 |
prompt,
|
172 |
st.session_state.chat_history
|
173 |
)
|
|
|
179 |
if len(st.session_state.chat_history) > 5:
|
180 |
st.session_state.chat_history = st.session_state.chat_history[-5:]
|
181 |
|
182 |
+
# Sidebar controls
|
183 |
+
with st.sidebar:
|
184 |
+
st.title("Controls")
|
185 |
+
if st.button("Clear Chat"):
|
186 |
+
st.session_state.chat_history = []
|
187 |
+
st.rerun()
|
188 |
+
|
189 |
+
st.markdown("---")
|
190 |
+
st.markdown("""
|
191 |
+
### Model Info
|
192 |
+
- Using TinyLlama 1.1B Chat
|
193 |
+
- Optimized for CPU usage
|
194 |
+
- Context window: 1024 tokens
|
195 |
+
""")
|
196 |
|
197 |
if __name__ == "__main__":
|
198 |
main()
|