Spaces:
Running
Running
toaster61
commited on
Commit
•
021692e
1
Parent(s):
1391fc1
first real gradio commit
Browse files- Dockerfile +1 -1
- gradio_app.py +117 -0
- app.py → quart_app.py +1 -1
Dockerfile
CHANGED
@@ -29,4 +29,4 @@ RUN python3 -m pip install -U --no-cache-dir pip setuptools wheel
|
|
29 |
RUN pip install --no-cache-dir --upgrade -r /app/requirements.txt
|
30 |
|
31 |
# Now it's time to run Quart app using uvicorn! (It's faster, trust me.)
|
32 |
-
CMD ["
|
|
|
29 |
RUN pip install --no-cache-dir --upgrade -r /app/requirements.txt
|
30 |
|
31 |
# Now it's time to run Quart app using uvicorn! (It's faster, trust me.)
|
32 |
+
CMD ["python", "gradio_app.py"]
|
gradio_app.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Importing libraries
|
2 |
+
from transformers import M2M100Tokenizer, M2M100ForConditionalGeneration
|
3 |
+
from llama_cpp import Llama
|
4 |
+
import gradio as gr
|
5 |
+
import psutil
|
6 |
+
|
7 |
+
# Initing things
|
8 |
+
llm = Llama(model_path="./model.bin") # LLaMa model
|
9 |
+
llama_model_name = "TheBloke/Llama-2-13B-chat-GGUF"
|
10 |
+
translator_tokenizer = M2M100Tokenizer.from_pretrained( # tokenizer for translator
|
11 |
+
"facebook/m2m100_1.2B", cache_dir="translator/"
|
12 |
+
)
|
13 |
+
translator_model = M2M100ForConditionalGeneration.from_pretrained( # translator model
|
14 |
+
"facebook/m2m100_1.2B", cache_dir="translator/"
|
15 |
+
)
|
16 |
+
translator_model.eval()
|
17 |
+
|
18 |
+
# Preparing things to work
|
19 |
+
translator_tokenizer.src_lang = "en"
|
20 |
+
title = "llama.cpp API"
|
21 |
+
desc = '''<style>a:visited{color:black;}</style>
|
22 |
+
<h1>Hello, world!</h1>
|
23 |
+
This is showcase how to make own server with Llama2 model.<br>
|
24 |
+
I'm using here 7b model just for example. Also here's only CPU power.<br>
|
25 |
+
But you can use GPU power as well!<br>
|
26 |
+
<h1>How to GPU?</h1>
|
27 |
+
Change <code>`CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS`</code> in Dockerfile on <code>`CMAKE_ARGS="-DLLAMA_CUBLAS=on"`</code>. Also you can try <code>`DLLAMA_CLBLAST`</code>, <code>`DLLAMA_METAL`</code> or <code>`DLLAMA_METAL`</code>.<br>
|
28 |
+
Powered by <a href="https://github.com/abetlen/llama-cpp-python">llama-cpp-python</a>, <a href="https://quart.palletsprojects.com/">Quart</a> and <a href="https://www.uvicorn.org/">Uvicorn</a>.<br>
|
29 |
+
<h1>How to test it on own machine?</h1>
|
30 |
+
You can install Docker, build image and run it. I made <code>`run-docker.sh`</code> for ya. To stop container run <code>`docker ps`</code>, find name of container and run <code>`docker stop _dockerContainerName_`</code><br>
|
31 |
+
Or you can once follow steps in Dockerfile and try it on your machine, not in Docker.<br>
|
32 |
+
<br>''' + f"Memory used: {psutil.virtual_memory()[2]}<br>" + '''
|
33 |
+
<script>document.write("<b>URL of space:</b> "+window.location.href);</script>'''
|
34 |
+
|
35 |
+
# Loading prompt
|
36 |
+
with open('system.prompt', 'r', encoding='utf-8') as f:
|
37 |
+
prompt = f.read()
|
38 |
+
|
39 |
+
# this model was loaded from https://hf.co/models
|
40 |
+
model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")
|
41 |
+
tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M")
|
42 |
+
device = 0 if torch.cuda.is_available() else -1
|
43 |
+
LANGS = ["ace_Arab", "eng_Latn", "fra_Latn", "spa_Latn"]
|
44 |
+
|
45 |
+
def t1ranslate(text, src_lang, tgt_lang):
|
46 |
+
try:
|
47 |
+
maxTokens = data.get("max_tokens", 64)
|
48 |
+
if isinstance(data.get("system_prompt"), str):
|
49 |
+
userPrompt = data.get("system_prompt") + "\n\nUser: " + data['request'] + "\nAssistant: "
|
50 |
+
else:
|
51 |
+
userPrompt = prompt + "\n\nUser: " + data['request'] + "\nAssistant: "
|
52 |
+
except:
|
53 |
+
return {"error": "Not enough data", "output": "Oops! Error occured! If you're a developer, using this API, check 'error' key."}, 400
|
54 |
+
try:
|
55 |
+
output = llm(userPrompt, max_tokens=maxTokens, stop=["User:", "\n"], echo=False)
|
56 |
+
text = output["choices"][0]["text"]
|
57 |
+
# i allowed only certain languages:
|
58 |
+
# russian (ru), ukranian (uk), chinese (zh)
|
59 |
+
if isinstance(data.get("target_lang"), str) and data.get("target_lang").lower() in ["ru", "uk", "zh"]:
|
60 |
+
encoded_input = translator_tokenizer(output, return_tensors="pt")
|
61 |
+
generated_tokens = translator_model.generate(
|
62 |
+
**encoded_input, forced_bos_token_id=translator_tokenizer.get_lang_id(data.get("target_lang"))
|
63 |
+
)
|
64 |
+
translated_text = translator_tokenizer.batch_decode(
|
65 |
+
generated_tokens, skip_special_tokens=True
|
66 |
+
)[0]
|
67 |
+
return {"output": text, "translated_output": translated_text}
|
68 |
+
|
69 |
+
return {"output": text}
|
70 |
+
except Exception as e:
|
71 |
+
print(e)
|
72 |
+
return {"error": str(e), "output": "Oops! Internal server error. Check the logs. If you're a developer, using this API, check 'error' key."}, 500
|
73 |
+
|
74 |
+
def translate(request: str, max_tokens: int = 256, language: str = "en", custom_prompt: str = None):
|
75 |
+
try:
|
76 |
+
maxTokens = max_tokens if 16 <= max_tokens <= 256 else 64
|
77 |
+
if isinstance(custom_prompt, str):
|
78 |
+
userPrompt = custom_prompt + "\n\nUser: " + request + "\nAssistant: "
|
79 |
+
else:
|
80 |
+
userPrompt = prompt + "\n\nUser: " + request + "\nAssistant: "
|
81 |
+
except:
|
82 |
+
return "Not enough data! Check that you passed all needed data."
|
83 |
+
|
84 |
+
try:
|
85 |
+
output = llm(userPrompt, max_tokens=maxTokens, stop=["User:", "\n"], echo=False)
|
86 |
+
text = output["choices"][0]["text"]
|
87 |
+
# i allowed only certain languages (its not discrimination, its just other popular language on my opinion!!!):
|
88 |
+
# russian (ru), ukranian (uk), chinese (zh)
|
89 |
+
if language in ["ru", "uk", "zh"]:
|
90 |
+
encoded_input = translator_tokenizer(output, return_tensors="pt")
|
91 |
+
generated_tokens = translator_model.generate(
|
92 |
+
**encoded_input, forced_bos_token_id=translator_tokenizer.get_lang_id(language)
|
93 |
+
)
|
94 |
+
translated_text = translator_tokenizer.batch_decode(
|
95 |
+
generated_tokens, skip_special_tokens=True
|
96 |
+
)[0]
|
97 |
+
return translated_text
|
98 |
+
return text
|
99 |
+
except Exception as e:
|
100 |
+
print(e)
|
101 |
+
return "Oops! Internal server error. Check the logs of space/instance."
|
102 |
+
|
103 |
+
|
104 |
+
demo = gr.Interface(
|
105 |
+
fn=translate,
|
106 |
+
inputs=[
|
107 |
+
gr.components.Textbox(label="Input"),
|
108 |
+
gr.components.Number(value=256),
|
109 |
+
gr.components.Dropdown(label="Target Language", value="en", choices=["en", "ru", "uk", "zh"]),
|
110 |
+
gr.components.Textbox(label="Custom system prompt"),
|
111 |
+
],
|
112 |
+
outputs=["text"],
|
113 |
+
title=title,
|
114 |
+
description=desc
|
115 |
+
)
|
116 |
+
demo.queue()
|
117 |
+
demo.launch()
|
app.py → quart_app.py
RENAMED
@@ -68,5 +68,5 @@ Powered by <a href="https://github.com/abetlen/llama-cpp-python">llama-cpp-pytho
|
|
68 |
<h1>How to test it on own machine?</h1>
|
69 |
You can install Docker, build image and run it. I made <code>`run-docker.sh`</code> for ya. To stop container run <code>`docker ps`</code>, find name of container and run <code>`docker stop _dockerContainerName_`</code><br>
|
70 |
Or you can once follow steps in Dockerfile and try it on your machine, not in Docker.<br>
|
71 |
-
<br>''' + f"Memory
|
72 |
<script>document.write("<b>URL of space:</b> "+window.location.href);</script>'''
|
|
|
68 |
<h1>How to test it on own machine?</h1>
|
69 |
You can install Docker, build image and run it. I made <code>`run-docker.sh`</code> for ya. To stop container run <code>`docker ps`</code>, find name of container and run <code>`docker stop _dockerContainerName_`</code><br>
|
70 |
Or you can once follow steps in Dockerfile and try it on your machine, not in Docker.<br>
|
71 |
+
<br>''' + f"Memory used: {psutil.virtual_memory()[2]}<br>" + '''
|
72 |
<script>document.write("<b>URL of space:</b> "+window.location.href);</script>'''
|