first commit
Browse files- app.py +144 -0
- requirements.txt +4 -0
app.py
ADDED
@@ -0,0 +1,144 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from threading import Thread
|
3 |
+
from open_lm.hf import *
|
4 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
|
5 |
+
import torch
|
6 |
+
from gradio.layouts import Accordion
|
7 |
+
|
8 |
+
# Define model options
|
9 |
+
MODEL_OPTIONS = {
|
10 |
+
"TRI-ML/DCLM-1B": "TRI-ML/DCLM-1B",
|
11 |
+
"Apple DCLM-Baseline-7B": "apple/DCLM-Baseline-7B"
|
12 |
+
}
|
13 |
+
|
14 |
+
# Global variables for model and tokenizer
|
15 |
+
current_model = None
|
16 |
+
current_tokenizer = None
|
17 |
+
|
18 |
+
def load_model(model_name):
|
19 |
+
global current_model, current_tokenizer
|
20 |
+
current_tokenizer = AutoTokenizer.from_pretrained(MODEL_OPTIONS[model_name])
|
21 |
+
current_model = AutoModelForCausalLM.from_pretrained(MODEL_OPTIONS[model_name])
|
22 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
23 |
+
current_model = current_model.to(device)
|
24 |
+
return f"Loaded model: {model_name}"
|
25 |
+
|
26 |
+
def generate(
|
27 |
+
prompt, model_choice, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0,
|
28 |
+
):
|
29 |
+
global current_model, current_tokenizer
|
30 |
+
|
31 |
+
if current_model is None or current_tokenizer is None:
|
32 |
+
return "Please load a model first."
|
33 |
+
|
34 |
+
temperature = float(temperature)
|
35 |
+
if temperature < 1e-2:
|
36 |
+
temperature = 1e-2
|
37 |
+
top_p = float(top_p)
|
38 |
+
|
39 |
+
inputs = current_tokenizer(prompt, return_tensors="pt").to(current_model.device)
|
40 |
+
|
41 |
+
generate_kwargs = dict(
|
42 |
+
**inputs,
|
43 |
+
max_new_tokens=max_new_tokens,
|
44 |
+
temperature=temperature,
|
45 |
+
top_p=top_p,
|
46 |
+
repetition_penalty=repetition_penalty,
|
47 |
+
do_sample=True,
|
48 |
+
pad_token_id=current_tokenizer.eos_token_id
|
49 |
+
)
|
50 |
+
|
51 |
+
streamer = TextIteratorStreamer(current_tokenizer, skip_prompt=True, skip_special_tokens=True)
|
52 |
+
generate_kwargs["streamer"] = streamer
|
53 |
+
|
54 |
+
thread = Thread(target=current_model.generate, kwargs=generate_kwargs)
|
55 |
+
thread.start()
|
56 |
+
|
57 |
+
# Write the prompt in blue
|
58 |
+
output = "<span style='color: blue;'>" + prompt + "</span>"
|
59 |
+
for new_text in streamer:
|
60 |
+
if isinstance(new_text, torch.Tensor):
|
61 |
+
new_text = current_tokenizer.decode(new_text)
|
62 |
+
output += new_text
|
63 |
+
yield output
|
64 |
+
|
65 |
+
thread.join()
|
66 |
+
return output
|
67 |
+
|
68 |
+
additional_inputs=[
|
69 |
+
gr.Slider(
|
70 |
+
label="Temperature",
|
71 |
+
value=0.9,
|
72 |
+
minimum=0.0,
|
73 |
+
maximum=1.0,
|
74 |
+
step=0.05,
|
75 |
+
interactive=True,
|
76 |
+
info="Higher values produce more diverse outputs",
|
77 |
+
),
|
78 |
+
gr.Slider(
|
79 |
+
label="Max new tokens",
|
80 |
+
value=256,
|
81 |
+
minimum=0,
|
82 |
+
maximum=1048,
|
83 |
+
step=64,
|
84 |
+
interactive=True,
|
85 |
+
info="The maximum numbers of new tokens",
|
86 |
+
),
|
87 |
+
gr.Slider(
|
88 |
+
label="Top-p (nucleus sampling)",
|
89 |
+
value=0.90,
|
90 |
+
minimum=0.0,
|
91 |
+
maximum=1,
|
92 |
+
step=0.05,
|
93 |
+
interactive=True,
|
94 |
+
info="Higher values sample more low-probability tokens",
|
95 |
+
),
|
96 |
+
gr.Slider(
|
97 |
+
label="Repetition penalty",
|
98 |
+
value=1.2,
|
99 |
+
minimum=1.0,
|
100 |
+
maximum=2.0,
|
101 |
+
step=0.05,
|
102 |
+
interactive=True,
|
103 |
+
info="Penalize repeated tokens",
|
104 |
+
)
|
105 |
+
]
|
106 |
+
|
107 |
+
with gr.Blocks() as demo:
|
108 |
+
gr.Markdown(
|
109 |
+
"""
|
110 |
+
# DCLM Text Completion Demo
|
111 |
+
This demo allows you to generate text using a DCLM model.
|
112 |
+
These models are trained to predict the next word in a sequence of text, and can be used to generate text completions, they are not chatbots.
|
113 |
+
|
114 |
+
First select a model from the dropdown and click "Load Model".
|
115 |
+
Then enter some text in the text box and click "Generate" to see the model's completion.
|
116 |
+
"""
|
117 |
+
)
|
118 |
+
|
119 |
+
|
120 |
+
with gr.Row():
|
121 |
+
model_dropdown = gr.Dropdown(choices=list(MODEL_OPTIONS.keys()), label="Select Model")
|
122 |
+
|
123 |
+
model_dropdown.select(
|
124 |
+
load_model,
|
125 |
+
inputs=[model_dropdown],
|
126 |
+
outputs=[gr.Textbox(label="Model Status")]
|
127 |
+
)
|
128 |
+
|
129 |
+
text_input = gr.Textbox(lines=3, label="Input Text")
|
130 |
+
text_output = gr.HTML(label="Generated Text")
|
131 |
+
|
132 |
+
generate_button = gr.Button("Generate")
|
133 |
+
|
134 |
+
generate_button.click(
|
135 |
+
generate,
|
136 |
+
inputs=[text_input, model_dropdown, *additional_inputs],
|
137 |
+
outputs=[text_output]
|
138 |
+
)
|
139 |
+
with Accordion(label="Advanced Options", open=False):
|
140 |
+
for input_component in additional_inputs:
|
141 |
+
if not input_component.is_rendered:
|
142 |
+
input_component.render()
|
143 |
+
|
144 |
+
demo.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
git+https://github.com/mlfoundations/open_lm.git
|
2 |
+
gradio
|
3 |
+
transformers
|
4 |
+
torch
|