Spaces:
Sleeping
Sleeping
File size: 7,521 Bytes
9255bb8 b1c5198 9255bb8 b1f91f1 9255bb8 41d4329 5a68917 9255bb8 ca38a58 cd7fced 4d0f760 c8f6eb0 cd7fced 4d0f760 cd7fced 4d0f760 cd7fced 4d0f760 cd7fced 4d0f760 cd7fced 4d0f760 cd7fced b033af5 4d0f760 b033af5 cd7fced 4d0f760 cd7fced 9255bb8 8c4290e 9255bb8 c8f6eb0 9255bb8 ca38a58 9255bb8 7d6a2e6 9255bb8 c8f6eb0 7eaa8e3 c8f6eb0 ca38a58 220e795 ca38a58 c8f6eb0 ca38a58 d0dc19e 5a68917 ca38a58 9255bb8 7d6a2e6 9255bb8 1ea7ac9 f18cc6e 7d6a2e6 1ea7ac9 7d6a2e6 9255bb8 ed3bcbc da6c800 d7c0687 9255bb8 ca38a58 83b80a1 cd7fced 83b80a1 9255bb8 b1f91f1 cd7fced 199759c cd7fced 199759c 9255bb8 199759c 9255bb8 c8f6eb0 9255bb8 7926189 9255bb8 4745d0f 549ab4b 9255bb8 c8f6eb0 cd7fced 9255bb8 |
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 |
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 = os.environ.get("API_URL")
with open("./HHH_prompt.txt", "r") as f:
HHH_PROMPT = f.read() + "\n\n"
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 in addition to the pure code data it is formatted with special tokens. E.g. prefixes specifying the source of the file or special tokens separating code from a commit message. See below:
### Chat mode
Chat mode prepends the [HHH prompt](https://gist.github.com/jareddk/2509330f8ef3d787fc5aaac67aab5f11#file-hhh_prompt-txt) from Anthropic to the request which conditions the model to be an assistant.
### Prefixes
Any combination of the three following prefixes can be found in pure code files:
```
<reponame>REPONAME<filename>FILENAME<gh_stars>STARS\ncode<|endoftext|>
```
STARS can be one of: 0, 1-10, 10-100, 100-1000, 1000+
### Commits
The commits data is formatted as follows:
```
<commit_before>code<commit_msg>text<commit_after>code<|endoftext|>
```
### Jupyter structure
Jupyter notebooks were both trained in form of Python scripts as well as the following structured format:
```
<start_jupyter><jupyter_text>text<jupyter_code>code<jupyter_output>output<jupyter_text>
```
### Issues
We also trained on GitHub issues using the following formatting:
```
<issue_start><issue_comment>text<issue_comment>...<issue_closed>
```
### Fill-in-the-middle
Fill in the middle requires rearranging the model inputs. The playground does 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}"},
)
def generate(prompt, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0, chat_mode=False):
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 chat_mode:
generate_kwargs.update({"stop_sequences": ["\nHuman", "\n-----"]})
if chat_mode and FIM_INDICATOR in prompt:
raise ValueError("Chat mode and FIM are mutually exclusive. Choose one or the other.")
if chat_mode:
chat_prompt = "Human: " + prompt + "\n\nAssistant:"
prompt = HHH_PROMPT + chat_prompt
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}"
stream = client.generate_stream(prompt, **generate_kwargs)
if fim_mode:
output = prefix
elif chat_mode:
output = chat_prompt
else:
output = prompt
previous_token = ""
for response in stream:
if fim_mode and response.token.text =="<|endoftext|>":
output += (suffix + "\n" + response.token.text)
elif chat_mode and response.token.text in ["Human", "-----"] and previous_token=="\n":
return output
else:
output += response.token.text
previous_token = response.token.text
yield output
return output
examples = [
"def print_hello_world():",
"def fibonacci(n):",
"class TransformerDecoder(nn.Module):",
"class ComplexNumbers:"
]
def process_example(args):
for x in generate(args):
pass
return x
css = ".generating {visibility: hidden}" + share_btn_css
with gr.Blocks(theme=theme, analytics_enabled=False, css=css) as demo:
with gr.Column():
gr.Markdown(
"""\
# BigCode - Playground
_Note:_ this is an internal playground - please do not share. The deployment can also change and thus the space not work as we continue development.\
"""
)
with gr.Row():
with gr.Column(scale=3):
instruction = gr.Textbox(placeholder="Enter your prompt here", label="Prompt", elem_id="q-input")
submit = gr.Button("Generate", variant="primary")
output = gr.Code(elem_id="q-output")
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)
with gr.Column(scale=1):
chat_mode = gr.Checkbox(
value=False,
label="Chat mode",
info="Uses Anthropic's HHH prompt to turn the model into an assistant."
)
temperature = gr.Slider(
label="Temperature",
value=0.2,
minimum=0.0,
maximum=2.0,
step=0.1,
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",
)
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",
)
submit.click(generate, inputs=[instruction, temperature, max_new_tokens, top_p, repetition_penalty, chat_mode], outputs=[output])
# instruction.submit(generate, inputs=[instruction, temperature, max_new_tokens, top_p, repetition_penalty, chat_mode], outputs=[output])
share_button.click(None, [], [], _js=share_js)
demo.queue(concurrency_count=16).launch(debug=True) |