Update app.py
Browse files
app.py
CHANGED
@@ -1,84 +1,68 @@
|
|
1 |
-
|
2 |
-
|
|
|
|
|
3 |
import torch
|
4 |
import gradio as gr
|
5 |
-
import
|
6 |
-
from textwrap import wrap
|
7 |
|
|
|
|
|
8 |
|
9 |
-
|
10 |
-
|
11 |
-
|
|
|
12 |
|
13 |
-
|
14 |
-
|
15 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
16 |
|
17 |
-
|
18 |
-
|
19 |
-
wrapped_lines = [textwrap.fill(line, width=width) for line in lines]
|
20 |
-
wrapped_text = '\n'.join(wrapped_lines)
|
21 |
-
return wrapped_text
|
22 |
|
23 |
-
|
|
|
|
|
|
|
|
|
24 |
|
25 |
-
|
|
|
|
|
|
|
26 |
|
27 |
-
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
-
|
31 |
-
**model_inputs,
|
32 |
-
max_length=max_length,
|
33 |
-
use_cache=True,
|
34 |
-
early_stopping=True,
|
35 |
-
bos_token_id=model.config.bos_token_id,
|
36 |
-
eos_token_id=model.config.eos_token_id,
|
37 |
-
pad_token_id=model.config.eos_token_id,
|
38 |
-
temperature=0.1,
|
39 |
-
do_sample=False
|
40 |
-
)
|
41 |
|
42 |
-
|
43 |
-
response_text = tokenizer.decode(output[0], skip_special_tokens=True)
|
44 |
|
45 |
-
|
46 |
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
tokenizer.padding_side = 'left'
|
51 |
-
peft_config = PeftConfig.from_pretrained("Tonic/mistralmed")
|
52 |
-
peft_model = MistralForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
|
53 |
-
peft_model = PeftModel.from_pretrained(peft_model, "Tonic/mistralmed")
|
54 |
-
peft_model = peft_model.to(torch.bfloat16)
|
55 |
-
peft_model = peft_model.to(device)
|
56 |
-
|
57 |
-
class ChatBot:
|
58 |
-
def __init__(self):
|
59 |
-
self.history = []
|
60 |
-
|
61 |
-
class ChatBot:
|
62 |
-
def __init__(self):
|
63 |
-
self.history = []
|
64 |
-
|
65 |
-
def predict(self, user_input, system_prompt="You are an expert medical analyst:"):
|
66 |
-
formatted_input = f"<s>[INST]{system_prompt} {user_input}[/INST]"
|
67 |
-
user_input_ids = tokenizer.encode(formatted_input, return_tensors="pt")
|
68 |
-
user_input_ids = user_input_ids.to(device)
|
69 |
-
response = peft_model.generate(input_ids=user_input_ids, max_length=256, pad_token_id=tokenizer.eos_token_id)
|
70 |
-
response_text = tokenizer.decode(response[0], skip_special_tokens=True)
|
71 |
-
|
72 |
-
return response_text
|
73 |
-
|
74 |
-
bot = ChatBot()
|
75 |
|
76 |
iface = gr.Interface(
|
77 |
-
fn=
|
78 |
title=title,
|
79 |
description=description,
|
80 |
-
|
81 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
outputs="text",
|
83 |
theme="ParityError/Anime"
|
84 |
)
|
|
|
1 |
+
import os
|
2 |
+
import math
|
3 |
+
import transformers
|
4 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
5 |
import torch
|
6 |
import gradio as gr
|
7 |
+
import sentencepiece
|
|
|
8 |
|
9 |
+
title = "Welcome to Tonic's ๐๐ณOrca-2-13B (in 8bit)!"
|
10 |
+
description = "You can use [๐๐ณmicrosoft/Orca-2-13b](https://huggingface.co/microsoft/Orca-2-13b) via API using Gradio by scrolling down and clicking Use 'Via API' or privately by [cloning this space on huggingface](https://huggingface.co/spaces/Tonic1/TonicsOrca2?duplicate=true) . [Join my active builders' server on discord](https://discord.gg/VqTxc76K3u). Big thanks to the HuggingFace Organisation for the Community Grant."
|
11 |
|
12 |
+
# os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:50'
|
13 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
14 |
+
model_name = "microsoft/Orca-2-13b"
|
15 |
+
# offload_folder = './model_weights'
|
16 |
|
17 |
+
# if not os.path.exists(offload_folder):
|
18 |
+
# os.makedirs(offload_folder)
|
|
|
19 |
|
20 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
|
21 |
+
model = transformers.AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True)
|
|
|
|
|
|
|
22 |
|
23 |
+
class OrcaChatBot:
|
24 |
+
def __init__(self, model, tokenizer, system_message="You are Orca, an AI language model created by Microsoft. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior."):
|
25 |
+
self.model = model
|
26 |
+
self.tokenizer = tokenizer
|
27 |
+
self.system_message = system_message
|
28 |
|
29 |
+
def predict(self, user_message, temperature=0.4, max_new_tokens=70, top_p=0.99, repetition_penalty=1.9):
|
30 |
+
prompt = f"<|im_start|>system\n{self.system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
|
31 |
+
inputs = self.tokenizer(prompt, return_tensors='pt', add_special_tokens=False)
|
32 |
+
input_ids = inputs["input_ids"].to(self.model.device)
|
33 |
|
34 |
+
output_ids = self.model.generate(
|
35 |
+
input_ids,
|
36 |
+
max_length=input_ids.shape[1] + max_new_tokens,
|
37 |
+
temperature=temperature,
|
38 |
+
top_p=top_p,
|
39 |
+
repetition_penalty=repetition_penalty,
|
40 |
+
pad_token_id=self.tokenizer.eos_token_id,
|
41 |
+
do_sample=True
|
42 |
+
)
|
43 |
|
44 |
+
response = self.tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
+
return response
|
|
|
47 |
|
48 |
+
Orca_bot = OrcaChatBot(model, tokenizer)
|
49 |
|
50 |
+
def gradio_predict(user_message, system_message, max_new_tokens, temperature, top_p, repetition_penalty):
|
51 |
+
full_message = f"{system_message}\n{user_message}" if system_message else user_message
|
52 |
+
return Orca_bot.predict(full_message, temperature, max_new_tokens, top_p, repetition_penalty)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
|
54 |
iface = gr.Interface(
|
55 |
+
fn=gradio_predict,
|
56 |
title=title,
|
57 |
description=description,
|
58 |
+
inputs=[
|
59 |
+
gr.Textbox(label="Your Message", type="text", lines=3),
|
60 |
+
gr.Textbox(label="Introduce a Character Here or Set a Scene (system prompt)", type="text", lines=2),
|
61 |
+
gr.Slider(label="Max new tokens", value=125, minimum=25, maximum=256, step=1),
|
62 |
+
gr.Slider(label="Temperature", value=0.1, minimum=0.05, maximum=1.0, step=0.05),
|
63 |
+
gr.Slider(label="Top-p (nucleus sampling)", value=0.90, minimum=0.01, maximum=0.99, step=0.05),
|
64 |
+
gr.Slider(label="Repetition penalty", value=1.9, minimum=1.0, maximum=2.0, step=0.05)
|
65 |
+
],
|
66 |
outputs="text",
|
67 |
theme="ParityError/Anime"
|
68 |
)
|