File size: 9,888 Bytes
eb76b96
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
import json
import os
import shutil
import requests

import gradio as gr
from huggingface_hub import Repository
from text_generation import Client

from share_btn import community_icon_html, loading_icon_html, share_js, share_btn_css

HF_TOKEN = os.environ.get("HF_TOKEN", None)
API_URL = "https://api-inference.huggingface.co/models/bigcode/starcoder/"
API_URL_BASE ="https://api-inference.huggingface.co/models/bigcode/starcoderbase/"

FIM_PREFIX = "<fim_prefix>"
FIM_MIDDLE = "<fim_middle>"
FIM_SUFFIX = "<fim_suffix>"

FIM_INDICATOR = "<FILL_HERE>"

FORMATS = """## Model Formats

The model is pretrained on code and is formatted with special tokens in addition to the pure code data,\
such as prefixes specifying the source of the file or tokens separating code from a commit message.\
Use these templates to explore the model's capacities:

### 1. Prefixes 🏷️
For pure code files, use any combination of the following prefixes:

```
<reponame>REPONAME<filename>FILENAME<gh_stars>STARS\ncode<|endoftext|>
```
STARS can be one of: 0, 1-10, 10-100, 100-1000, 1000+

### 2. Commits 💾
The commits data is formatted as follows:

```
<commit_before>code<commit_msg>text<commit_after>code<|endoftext|>
```

### 3. Jupyter Notebooks 📓
The model is trained on Jupyter notebooks as Python scripts and structured formats like:

```
<start_jupyter><jupyter_text>text<jupyter_code>code<jupyter_output>output<jupyter_text>
```

### 4. Issues 🐛
We also trained on GitHub issues using the following formatting:
```
<issue_start><issue_comment>text<issue_comment>...<issue_closed>
```

### 5. Fill-in-the-middle 🧩
Fill in the middle requires rearranging the model inputs. The playground handles this for you - all you need is to specify where to fill:
```
code before<FILL_HERE>code after
```
"""

theme = gr.themes.Monochrome(
    primary_hue="indigo",
    secondary_hue="blue",
    neutral_hue="slate",
    radius_size=gr.themes.sizes.radius_sm,
    font=[
        gr.themes.GoogleFont("Open Sans"),
        "ui-sans-serif",
        "system-ui",
        "sans-serif",
    ],
)

client = Client(
    API_URL,
    headers={"Authorization": f"Bearer {HF_TOKEN}"},
)
client_base = Client(
    API_URL_BASE, headers={"Authorization": f"Bearer {HF_TOKEN}"},
)

def generate(
    prompt, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0, version="StarCoder",
):

    temperature = float(temperature)
    if temperature < 1e-2:
        temperature = 1e-2
    top_p = float(top_p)
    fim_mode = False

    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        do_sample=True,
        seed=42,
    )

    if FIM_INDICATOR in prompt:
        fim_mode = True
        try:
            prefix, suffix = prompt.split(FIM_INDICATOR)
        except:
            raise ValueError(f"Only one {FIM_INDICATOR} allowed in prompt!")
        prompt = f"{FIM_PREFIX}{prefix}{FIM_SUFFIX}{suffix}{FIM_MIDDLE}"

    if version == "StarCoder":
        stream = client.generate_stream(prompt, **generate_kwargs)
    else:
        stream = client_base.generate_stream(prompt, **generate_kwargs)

    if fim_mode:
        output = prefix
    else:
        output = prompt

    previous_token = ""
    for response in stream:
        if response.token.text == "<|endoftext|>":
            if fim_mode:
                output += suffix
            else:
                return output
        else:
            output += response.token.text
        previous_token = response.token.text
        yield output
    return output


examples = [
    "X_train, y_train, X_test, y_test = train_test_split(X, y, test_size=0.1)\n\n# Train a logistic regression model, predict the labels on the test set and compute the accuracy score",
    "// Returns every other value in the array as a new array.\nfunction everyOther(arr) {",
    "def alternating(list1, list2):\n   results = []\n   for i in range(min(len(list1), len(list2))):\n       results.append(list1[i])\n       results.append(list2[i])\n   if len(list1) > len(list2):\n       <FILL_HERE>\n   else:\n       results.extend(list2[i+1:])\n   return results",
]


def process_example(args):
    for x in generate(args):
        pass
    return x


css = ".generating {visibility: hidden}"

monospace_css = """
#q-input textarea {
    font-family: monospace, 'Consolas', Courier, monospace;
}
"""


css += share_btn_css + monospace_css + ".gradio-container {color: black}"


description = """
<div style="text-align: center;">
    <h1> 💫 StarCoder<span style='color: #e6b800;'> - </span>Code Completion Playground 🪐</h1>
    <p>This is a demo to generate code with <a href="https://huggingface.co/bigcode/starcoder" style='color: #e6b800;'>StarCoder</a>, a 15B parameter model for code generation in 86 programming languages.
    <b>This is not an instruction model</b>. For instruction and chatting you can chat with a prompted version of the model at <a href="https://huggingface.co/chat/?model=bigcode/starcoder">hf.co/chat</a></p>
</div>
"""
disclaimer = """⚠️<b>Any use or sharing of this demo constitues your acceptance of the BigCode [OpenRAIL-M](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement) License Agreement and the use restrictions included within.</b>\
 <br>**Intended Use**: this app and its [supporting model](https://huggingface.co/bigcode) are provided for demonstration purposes; not to serve as replacement for human expertise. For more details on the model's limitations in terms of factuality and biases, see the [model card.](hf.co/bigcode)"""

with gr.Blocks(theme=theme, analytics_enabled=False, css=css) as demo:
    with gr.Column():
        gr.Markdown(description)
        with gr.Row():
            with gr.Column():
                instruction = gr.Textbox(
                    placeholder="Enter your code here",
                    label="Code",
                    elem_id="q-input",
                )
                submit = gr.Button("Generate", variant="primary")
                output = gr.Code(elem_id="q-output", lines=30)
                with gr.Row():
                    with gr.Column():
                        with gr.Accordion("Advanced settings", open=False):
                            with gr.Row():
                                column_1, column_2 = gr.Column(), gr.Column()
                                with column_1:
                                    temperature = gr.Slider(
                                        label="Temperature",
                                        value=0.2,
                                        minimum=0.0,
                                        maximum=1.0,
                                        step=0.05,
                                        interactive=True,
                                        info="Higher values produce more diverse outputs",
                                    )
                                    max_new_tokens = gr.Slider(
                                        label="Max new tokens",
                                        value=256,
                                        minimum=0,
                                        maximum=8192,
                                        step=64,
                                        interactive=True,
                                        info="The maximum numbers of new tokens",
                                    )
                                with column_2:
                                    top_p = gr.Slider(
                                        label="Top-p (nucleus sampling)",
                                        value=0.90,
                                        minimum=0.0,
                                        maximum=1,
                                        step=0.05,
                                        interactive=True,
                                        info="Higher values sample more low-probability tokens",
                                    )
                                    repetition_penalty = gr.Slider(
                                        label="Repetition penalty",
                                        value=1.2,
                                        minimum=1.0,
                                        maximum=2.0,
                                        step=0.05,
                                        interactive=True,
                                        info="Penalize repeated tokens",
                                    )
                    with gr.Column():
                        version = gr.Dropdown(
                                    ["StarCoderBase", "StarCoder"],
                                    value="StarCoder",
                                    label="Version",
                                    info="",
                                    )
                gr.Markdown(disclaimer)
                with gr.Group(elem_id="share-btn-container"):
                    community_icon = gr.HTML(community_icon_html, visible=True)
                    loading_icon = gr.HTML(loading_icon_html, visible=True)
                    share_button = gr.Button(
                        "Share to community", elem_id="share-btn", visible=True
                    )
                gr.Examples(
                    examples=examples,
                    inputs=[instruction],
                    cache_examples=False,
                    fn=process_example,
                    outputs=[output],
                )
                gr.Markdown(FORMATS)

    submit.click(
        generate,
        inputs=[instruction, temperature, max_new_tokens, top_p, repetition_penalty, version],
        outputs=[output],
    )
    share_button.click(None, [], [], _js=share_js)
demo.queue(concurrency_count=16).launch(debug=True)