Update Space
Browse files
app.py
CHANGED
@@ -1,64 +1,173 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
from huggingface_hub import InferenceClient
|
3 |
|
4 |
-
|
5 |
-
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
|
|
|
8 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
system_message,
|
14 |
-
max_tokens,
|
15 |
-
temperature,
|
16 |
-
top_p,
|
17 |
-
):
|
18 |
-
messages = [{"role": "system", "content": system_message}]
|
19 |
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
if val[1]:
|
24 |
-
messages.append({"role": "assistant", "content": val[1]})
|
25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
messages.append({"role": "user", "content": message})
|
27 |
|
28 |
-
|
|
|
29 |
|
30 |
-
|
|
|
|
|
|
|
31 |
messages,
|
32 |
max_tokens=max_tokens,
|
33 |
stream=True,
|
34 |
temperature=temperature,
|
35 |
top_p=top_p,
|
36 |
):
|
37 |
-
token =
|
38 |
-
|
39 |
response += token
|
40 |
-
yield response
|
41 |
|
|
|
|
|
|
|
|
|
42 |
|
|
|
|
|
|
|
43 |
"""
|
44 |
-
|
|
|
|
|
|
|
45 |
"""
|
46 |
demo = gr.ChatInterface(
|
47 |
-
respond,
|
48 |
additional_inputs=[
|
49 |
-
gr.Textbox(
|
50 |
-
|
51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
gr.Slider(
|
53 |
minimum=0.1,
|
54 |
maximum=1.0,
|
55 |
value=0.95,
|
56 |
step=0.05,
|
57 |
-
label="Top-p (nucleus sampling)"
|
58 |
),
|
59 |
],
|
60 |
)
|
61 |
|
62 |
-
|
63 |
if __name__ == "__main__":
|
64 |
-
demo.launch()
|
|
|
1 |
+
import sys
|
2 |
+
import onnxruntime as ort
|
3 |
+
import numpy as np
|
4 |
+
import string
|
5 |
+
|
6 |
+
# Transformers, HuggingFace Hub, and Gradio
|
7 |
+
from transformers import AutoTokenizer
|
8 |
import gradio as gr
|
9 |
from huggingface_hub import InferenceClient
|
10 |
|
11 |
+
# ------------------------------------------------
|
12 |
+
# Turn Detector Configuration
|
13 |
+
# ------------------------------------------------
|
14 |
+
HG_MODEL = "livekit/turn-detector" # or your HF model repo
|
15 |
+
ONNX_FILENAME = "model_quantized.onnx" # path to your ONNX file
|
16 |
+
MAX_HISTORY_TOKENS = 512
|
17 |
+
PUNCS = string.punctuation.replace("'", "")
|
18 |
+
|
19 |
+
# ------------------------------------------------
|
20 |
+
# Utility functions
|
21 |
+
# ------------------------------------------------
|
22 |
+
def softmax(logits: np.ndarray) -> np.ndarray:
|
23 |
+
exp_logits = np.exp(logits - np.max(logits))
|
24 |
+
return exp_logits / np.sum(exp_logits)
|
25 |
+
|
26 |
+
def normalize_text(text: str) -> str:
|
27 |
+
"""Lowercase, strip punctuation (except single quotes), and collapse whitespace."""
|
28 |
+
def strip_puncs(text_in):
|
29 |
+
return text_in.translate(str.maketrans("", "", PUNCS))
|
30 |
+
return " ".join(strip_puncs(text).lower().split())
|
31 |
+
|
32 |
+
def calculate_eou(chat_ctx, session, tokenizer) -> float:
|
33 |
+
"""
|
34 |
+
Given a conversation context (list of dicts with 'role' and 'content'),
|
35 |
+
returns the probability that the user is finished speaking.
|
36 |
+
"""
|
37 |
+
# Collect normalized messages from 'user' or 'assistant' roles
|
38 |
+
normalized_ctx = []
|
39 |
+
for msg in chat_ctx:
|
40 |
+
if msg["role"] in ("user", "assistant"):
|
41 |
+
content = normalize_text(msg["content"])
|
42 |
+
if content:
|
43 |
+
normalized_ctx.append(content)
|
44 |
+
|
45 |
+
# Join them into one input string
|
46 |
+
text = " ".join(normalized_ctx)
|
47 |
+
inputs = tokenizer(
|
48 |
+
text,
|
49 |
+
return_tensors="np",
|
50 |
+
truncation=True,
|
51 |
+
max_length=MAX_HISTORY_TOKENS,
|
52 |
+
)
|
53 |
+
|
54 |
+
input_ids = np.array(inputs["input_ids"], dtype=np.int64)
|
55 |
+
# Run inference
|
56 |
+
outputs = session.run(["logits"], {"input_ids": input_ids})
|
57 |
+
logits = outputs[0][0, -1, :]
|
58 |
+
|
59 |
+
# Softmax over logits
|
60 |
+
probs = softmax(logits)
|
61 |
+
# The ID for the <|im_end|> special token
|
62 |
+
eou_token_id = tokenizer.encode("<|im_end|>")[-1]
|
63 |
+
return probs[eou_token_id]
|
64 |
+
|
65 |
+
# ------------------------------------------------
|
66 |
+
# Load ONNX session & tokenizer once
|
67 |
+
# ------------------------------------------------
|
68 |
+
print("Loading ONNX model session...")
|
69 |
+
onnx_session = ort.InferenceSession(ONNX_FILENAME, providers=["CPUExecutionProvider"])
|
70 |
+
|
71 |
+
print("Loading tokenizer...")
|
72 |
+
turn_detector_tokenizer = AutoTokenizer.from_pretrained(HG_MODEL)
|
73 |
+
|
74 |
+
# ------------------------------------------------
|
75 |
+
# HF InferenceClient for text generation (example)
|
76 |
+
# ------------------------------------------------
|
77 |
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
|
78 |
+
# Adjust above to any other endpoint that suits your use case.
|
79 |
|
80 |
+
# ------------------------------------------------
|
81 |
+
# Gradio Chat Handler
|
82 |
+
# ------------------------------------------------
|
83 |
+
def respond(message, history, system_message, max_tokens, temperature, top_p):
|
84 |
+
"""
|
85 |
+
This function is called on each new user message in the ChatInterface.
|
86 |
+
- 'message' is the new user input
|
87 |
+
- 'history' is a list of (user, assistant) tuples
|
88 |
+
- 'system_message' is from the system Textbox
|
89 |
+
- max_tokens, temperature, top_p come from the Sliders
|
90 |
+
"""
|
91 |
|
92 |
+
# 1) Build a list of messages in the OpenAI-style format:
|
93 |
+
# [{'role': 'system', 'content': ...},
|
94 |
+
# {'role': 'user', 'content': ...}, ...]
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
96 |
+
messages = []
|
97 |
+
if system_message.strip():
|
98 |
+
messages.append({"role": "system", "content": system_message})
|
|
|
|
|
99 |
|
100 |
+
# history is a list of tuples: [(user1, assistant1), (user2, assistant2), ...]
|
101 |
+
for user_text, assistant_text in history:
|
102 |
+
if user_text:
|
103 |
+
messages.append({"role": "user", "content": user_text})
|
104 |
+
if assistant_text:
|
105 |
+
messages.append({"role": "assistant", "content": assistant_text})
|
106 |
+
|
107 |
+
# Append the new user message
|
108 |
messages.append({"role": "user", "content": message})
|
109 |
|
110 |
+
# 2) Calculate EOU probability on the entire conversation
|
111 |
+
eou_prob = calculate_eou(messages, onnx_session, turn_detector_tokenizer)
|
112 |
|
113 |
+
# 3) Generate the assistant response from your HF model.
|
114 |
+
# (This code streams token-by-token.)
|
115 |
+
response = ""
|
116 |
+
for resp_chunk in client.chat_completion(
|
117 |
messages,
|
118 |
max_tokens=max_tokens,
|
119 |
stream=True,
|
120 |
temperature=temperature,
|
121 |
top_p=top_p,
|
122 |
):
|
123 |
+
token = resp_chunk.choices[0].delta.get("content", "")
|
|
|
124 |
response += token
|
|
|
125 |
|
126 |
+
# If you want to display the partial response with the EOU probability
|
127 |
+
# appended at the bottom, you can do so each step. For cleanliness,
|
128 |
+
# we'll do it in-line as a bracketed note at the end.
|
129 |
+
yield response + f"\n\n[EOU Probability: {eou_prob:.4f}]"
|
130 |
|
131 |
+
# ------------------------------------------------
|
132 |
+
# Gradio ChatInterface
|
133 |
+
# ------------------------------------------------
|
134 |
"""
|
135 |
+
This ChatInterface will have:
|
136 |
+
- A chat box
|
137 |
+
- A system message textbox
|
138 |
+
- 3 sliders for max_tokens, temperature, and top_p
|
139 |
"""
|
140 |
demo = gr.ChatInterface(
|
141 |
+
fn=respond,
|
142 |
additional_inputs=[
|
143 |
+
gr.Textbox(
|
144 |
+
value="You are a friendly Chatbot.",
|
145 |
+
label="System message",
|
146 |
+
lines=2
|
147 |
+
),
|
148 |
+
gr.Slider(
|
149 |
+
minimum=1,
|
150 |
+
maximum=2048,
|
151 |
+
value=512,
|
152 |
+
step=1,
|
153 |
+
label="Max new tokens"
|
154 |
+
),
|
155 |
+
gr.Slider(
|
156 |
+
minimum=0.1,
|
157 |
+
maximum=4.0,
|
158 |
+
value=0.7,
|
159 |
+
step=0.1,
|
160 |
+
label="Temperature"
|
161 |
+
),
|
162 |
gr.Slider(
|
163 |
minimum=0.1,
|
164 |
maximum=1.0,
|
165 |
value=0.95,
|
166 |
step=0.05,
|
167 |
+
label="Top-p (nucleus sampling)"
|
168 |
),
|
169 |
],
|
170 |
)
|
171 |
|
|
|
172 |
if __name__ == "__main__":
|
173 |
+
demo.launch()
|