Create app_old.py
Browse files- app_old.py +674 -0
app_old.py
ADDED
@@ -0,0 +1,674 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
Hugging Face's logo Hugging Face
|
3 |
+
|
4 |
+
Models
|
5 |
+
Datasets
|
6 |
+
Spaces
|
7 |
+
Docs
|
8 |
+
Pricing
|
9 |
+
|
10 |
+
Hugging Face is way more fun with friends and colleagues! 🤗 Join an organization
|
11 |
+
Spaces:
|
12 |
+
hugo1234
|
13 |
+
/
|
14 |
+
galileo
|
15 |
+
private
|
16 |
+
App
|
17 |
+
Files
|
18 |
+
Community
|
19 |
+
Settings
|
20 |
+
galileo
|
21 |
+
/ app.py
|
22 |
+
hugo1234's picture
|
23 |
+
hugo1234
|
24 |
+
Update app.py
|
25 |
+
a43103a
|
26 |
+
about 16 hours ago
|
27 |
+
raw
|
28 |
+
history
|
29 |
+
blame
|
30 |
+
22.8 kB
|
31 |
+
import os
|
32 |
+
os.system('pip install bitsandbytes')
|
33 |
+
os.system('pip install -q datasets loralib sentencepiece accelerate')
|
34 |
+
# os.system('pip install -q git+https://github.com/zphang/transformers@c3dc391')
|
35 |
+
# os.system('pip install -q git+https://github.com/huggingface/transformers')
|
36 |
+
os.system('pip install -q git+https://github.com/mbehm/transformers')
|
37 |
+
os.system('pip install -q git+https://github.com/huggingface/peft.git')
|
38 |
+
# os.system('pip install gradio')
|
39 |
+
# os.system('pip install torch')
|
40 |
+
# os.system('pip install peft')
|
41 |
+
# os.system('pip install transformers')
|
42 |
+
os.system('pip install tenacity')
|
43 |
+
os.system('pip install scipy')
|
44 |
+
# os.system('pip install sentencepiece')
|
45 |
+
|
46 |
+
import re
|
47 |
+
import yaml
|
48 |
+
import gc
|
49 |
+
import copy
|
50 |
+
import time
|
51 |
+
from tenacity import RetryError
|
52 |
+
from tenacity import retry, stop_after_attempt, wait_fixed
|
53 |
+
import gradio as gr
|
54 |
+
# import torch
|
55 |
+
from peft import PeftModel
|
56 |
+
from transformers import (
|
57 |
+
LLaMATokenizer,
|
58 |
+
LlamaForCausalLM,
|
59 |
+
GenerationConfig,
|
60 |
+
AutoModelForCausalLM,
|
61 |
+
AutoModelForSeq2SeqLM,
|
62 |
+
AutoTokenizer,
|
63 |
+
LogitsProcessorList,
|
64 |
+
MinNewTokensLengthLogitsProcessor,
|
65 |
+
TemperatureLogitsWarper,
|
66 |
+
TopPLogitsWarper,
|
67 |
+
MinLengthLogitsProcessor
|
68 |
+
)
|
69 |
+
|
70 |
+
# assert torch.cuda.is_available(), "Change the runtime type to GPU"
|
71 |
+
|
72 |
+
# constants
|
73 |
+
num_of_characters_to_keep = 1000
|
74 |
+
|
75 |
+
# regex
|
76 |
+
html_tag_pattern = re.compile(r"<.*?>")
|
77 |
+
multi_line_pattern = re.compile(r"\n+")
|
78 |
+
multi_space_pattern = re.compile(r"( )")
|
79 |
+
multi_br_tag_pattern = re.compile(re.compile(r'<br>\s*(<br>\s*)*'))
|
80 |
+
|
81 |
+
# repl is short for replacement
|
82 |
+
repl_linebreak = "\n"
|
83 |
+
repl_empty_str = ""
|
84 |
+
|
85 |
+
TITLE = "Galileo"
|
86 |
+
|
87 |
+
ABSTRACT = """
|
88 |
+
Stambecco is a Italian Instruction-following model based on the [LLaMA](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) model. It comes in two versions: 7b and 13b parameters. It is trained on an Italian version of the [GPT-4-LLM](https://github.com/Instruction-Tuning-with-GPT-4/GPT-4-LLM) dataset, a dataset of `GPT-4` generated instruction-following data.
|
89 |
+
This demo is intended to show and evaluate the conversational capabilities of the model.
|
90 |
+
For more information, please visit [the project's website](https://github.com/mchl-labs/stambecco).
|
91 |
+
NOTE: Too long input (context, instruction) will not be allowed. Please keep context < 500 and instruction < 150
|
92 |
+
"""
|
93 |
+
|
94 |
+
BOTTOM_LINE = """
|
95 |
+
By default, this demo runs with streaming mode, but you can also run with dynamic batch generation model.
|
96 |
+
Stambecco is built on the same concept as Standford Alpaca project, but using LoRA it lets us train and inference on a smaller GPUs such as RTX4090 for 7B version. Also, we could build very small size of checkpoints on top of base models thanks to [🤗 transformers](https://huggingface.co/docs/transformers/index), [🤗 peft](https://github.com/huggingface/peft), and [bitsandbytes](https://github.com/TimDettmers/bitsandbytes/tree/main) libraries.
|
97 |
+
This demo currently runs 8Bit 7b version of the model.
|
98 |
+
"""
|
99 |
+
|
100 |
+
DEFAULT_EXAMPLES = {
|
101 |
+
"Typical Questions": [
|
102 |
+
{
|
103 |
+
"title": "Parlami di Giulio Cesare.",
|
104 |
+
"examples": [
|
105 |
+
["1", "Scrivi un articolo su Giulio Cesare"],
|
106 |
+
["2", "Davvero?"],
|
107 |
+
["3", "Quanto era ricco Giulio Cesare?"],
|
108 |
+
["4", "Chi è stato il suo successore?"],
|
109 |
+
]
|
110 |
+
},
|
111 |
+
{
|
112 |
+
"title": "Parigi",
|
113 |
+
"examples": [
|
114 |
+
["1", "Scrivi un tema sulla città di Parigi"],
|
115 |
+
["2", "Fai un elenco di 5 posti da visitare assolutamente"],
|
116 |
+
["3", "Quali eventi importanti della Storia sono avvenuti a Parigi?"],
|
117 |
+
["4", "Quale è il periodo migliore per visitare Parigi?"],
|
118 |
+
]
|
119 |
+
},
|
120 |
+
{
|
121 |
+
"title": "Scrivi un programma in Python che stampi i primi 10 numeri di Fibonacci",
|
122 |
+
"examples": [
|
123 |
+
["1", "Scrivi un programma in Python che stampi i primi 10 numeri di Fibonacci"],
|
124 |
+
["2", "Potresti spiegarmi come funziona il codice?"],
|
125 |
+
["3", "Cos'è la ricorsione?"],
|
126 |
+
]
|
127 |
+
}
|
128 |
+
],
|
129 |
+
}
|
130 |
+
|
131 |
+
SPECIAL_STRS = {
|
132 |
+
"continue": "continua",
|
133 |
+
"summarize": "Di cosa abbiamo discusso finora? Descrivi nella user's view."
|
134 |
+
}
|
135 |
+
|
136 |
+
PARENT_BLOCK_CSS = """
|
137 |
+
#col_container {
|
138 |
+
width: 95%;
|
139 |
+
margin-left: auto;
|
140 |
+
margin-right: auto;
|
141 |
+
}
|
142 |
+
#chatbot {
|
143 |
+
height: 500px;
|
144 |
+
overflow: auto;
|
145 |
+
}
|
146 |
+
"""
|
147 |
+
|
148 |
+
def load_model(
|
149 |
+
base="decapoda-research/llama-7b-hf",
|
150 |
+
finetuned="mchl-labs/stambecco-7b-plus",
|
151 |
+
):
|
152 |
+
tokenizer = LLaMATokenizer.from_pretrained(base)
|
153 |
+
tokenizer.pad_token_id = 0
|
154 |
+
tokenizer.padding_side = "left"
|
155 |
+
|
156 |
+
model = LlamaForCausalLM.from_pretrained(
|
157 |
+
base,
|
158 |
+
load_in_8bit=True,
|
159 |
+
device_map="from_pretrained",
|
160 |
+
# load_in_8bit_fp32_cpu_offload=True
|
161 |
+
)
|
162 |
+
# model = PeftModel.from_pretrained(model, finetuned, device_map={'': 0})
|
163 |
+
|
164 |
+
model = PeftModel.from_pretrained(model, finetuned)
|
165 |
+
return model, tokenizer
|
166 |
+
|
167 |
+
def get_generation_config(path):
|
168 |
+
with open(path, 'rb') as f:
|
169 |
+
generation_config = yaml.safe_load(f.read())
|
170 |
+
|
171 |
+
return GenerationConfig(**generation_config["generation_config"])
|
172 |
+
|
173 |
+
def generate_prompt(prompt, histories, ctx=None, partial=False):
|
174 |
+
convs = f"""Di seguito è riportata una cronologia delle istruzioni che descrivono le tasks, abbinate a un input che fornisce ulteriore contesto. Scrivi una risposta che completi adeguatamente la richiesta ricordando la cronologia della conversazione.
|
175 |
+
"""
|
176 |
+
|
177 |
+
if ctx is not None:
|
178 |
+
convs = f"""### Input: {ctx}
|
179 |
+
"""
|
180 |
+
|
181 |
+
sub_convs = ""
|
182 |
+
start_idx = 0
|
183 |
+
|
184 |
+
for idx, history in enumerate(histories):
|
185 |
+
history_prompt = history[0]
|
186 |
+
history_response = history[1]
|
187 |
+
if history_response == "✅ Riepilogo della conversazione effettuato e impostato come contesto" or history_prompt == SPECIAL_STRS["summarize"]:
|
188 |
+
start_idx = idx
|
189 |
+
|
190 |
+
# drop the previous conversations if user has summarized
|
191 |
+
for history in histories[start_idx if start_idx == 0 else start_idx+1:]:
|
192 |
+
history_prompt = history[0]
|
193 |
+
history_response = history[1]
|
194 |
+
|
195 |
+
history_response = history_response.replace("<br>", "\n")
|
196 |
+
history_response = re.sub(
|
197 |
+
html_tag_pattern, repl_empty_str, history_response
|
198 |
+
)
|
199 |
+
|
200 |
+
sub_convs = sub_convs + f"""### Istruzione: {history_prompt}
|
201 |
+
### Risposta: {history_response}
|
202 |
+
"""
|
203 |
+
|
204 |
+
sub_convs = sub_convs + f"""### Istruzione: {prompt}
|
205 |
+
### Risposta:"""
|
206 |
+
|
207 |
+
convs = convs + sub_convs
|
208 |
+
return sub_convs if partial else convs, len(sub_convs)
|
209 |
+
|
210 |
+
def common_post_process(original_str):
|
211 |
+
original_str = re.sub(
|
212 |
+
multi_line_pattern, repl_linebreak, original_str
|
213 |
+
)
|
214 |
+
return original_str
|
215 |
+
|
216 |
+
def post_process_stream(bot_response):
|
217 |
+
# sometimes model spits out text containing
|
218 |
+
# "### Risposta:" and "### Istruzione: -> in this case, we want to stop generating
|
219 |
+
if "### Risposta:" in bot_response or "### Input:" in bot_response:
|
220 |
+
bot_response = bot_response.replace("### Risposta:", '').replace("### Input:", '').strip()
|
221 |
+
return bot_response, True
|
222 |
+
|
223 |
+
return common_post_process(bot_response), False
|
224 |
+
|
225 |
+
def post_process_batch(bot_response):
|
226 |
+
bot_response = bot_response.split("### Risposta:")[-1].strip()
|
227 |
+
return common_post_process(bot_response)
|
228 |
+
|
229 |
+
def post_processes_batch(bot_responses):
|
230 |
+
return [post_process_batch(r) for r in bot_responses]
|
231 |
+
|
232 |
+
def get_output_batch(
|
233 |
+
model, tokenizer, prompts, generation_config
|
234 |
+
):
|
235 |
+
if len(prompts) == 1:
|
236 |
+
encoding = tokenizer(prompts, return_tensors="pt")
|
237 |
+
input_ids = encoding["input_ids"].cuda()
|
238 |
+
generated_id = model.generate(
|
239 |
+
input_ids=input_ids,
|
240 |
+
generation_config=generation_config,
|
241 |
+
max_new_tokens=256
|
242 |
+
)
|
243 |
+
|
244 |
+
decoded = tokenizer.batch_decode(generated_id)
|
245 |
+
del input_ids, generated_id
|
246 |
+
torch.cuda.empty_cache()
|
247 |
+
return decoded
|
248 |
+
else:
|
249 |
+
encodings = tokenizer(prompts, padding=True, return_tensors="pt").to('cuda')
|
250 |
+
generated_ids = model.generate(
|
251 |
+
**encodings,
|
252 |
+
generation_config=generation_config,
|
253 |
+
max_new_tokens=256
|
254 |
+
)
|
255 |
+
|
256 |
+
decoded = tokenizer.batch_decode(generated_ids)
|
257 |
+
del encodings, generated_ids
|
258 |
+
torch.cuda.empty_cache()
|
259 |
+
return decoded
|
260 |
+
|
261 |
+
|
262 |
+
# StreamModel is borrowed from basaran project
|
263 |
+
# please find more info about it -> https://github.com/hyperonym/basaran
|
264 |
+
class StreamModel:
|
265 |
+
"""StreamModel wraps around a language model to provide stream decoding."""
|
266 |
+
|
267 |
+
def __init__(self, model, tokenizer):
|
268 |
+
super().__init__()
|
269 |
+
self.model = model
|
270 |
+
self.tokenizer = tokenizer
|
271 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
272 |
+
|
273 |
+
self.processor = LogitsProcessorList()
|
274 |
+
self.processor.append(TemperatureLogitsWarper(0.9))
|
275 |
+
self.processor.append(TopPLogitsWarper(0.75))
|
276 |
+
|
277 |
+
|
278 |
+
def __call__(
|
279 |
+
self,
|
280 |
+
prompt,
|
281 |
+
min_tokens=0,
|
282 |
+
max_tokens=16,
|
283 |
+
temperature=1.0,
|
284 |
+
top_p=1.0,
|
285 |
+
n=1,
|
286 |
+
logprobs=0,
|
287 |
+
):
|
288 |
+
"""Create a completion stream for the provided prompt."""
|
289 |
+
input_ids = self.tokenize(prompt)
|
290 |
+
logprobs = max(logprobs, 0)
|
291 |
+
|
292 |
+
# bigger than 1
|
293 |
+
chunk_size = 2
|
294 |
+
chunk_count = 0
|
295 |
+
|
296 |
+
# Generate completion tokens.
|
297 |
+
final_tokens = torch.empty(0)
|
298 |
+
|
299 |
+
for tokens in self.generate(
|
300 |
+
input_ids[None, :].repeat(n, 1),
|
301 |
+
logprobs=logprobs,
|
302 |
+
min_new_tokens=min_tokens,
|
303 |
+
max_new_tokens=max_tokens,
|
304 |
+
temperature=temperature,
|
305 |
+
top_p=top_p,
|
306 |
+
):
|
307 |
+
if chunk_count < chunk_size:
|
308 |
+
chunk_count = chunk_count + 1
|
309 |
+
|
310 |
+
final_tokens = torch.cat((final_tokens, tokens.to("cpu")))
|
311 |
+
|
312 |
+
if chunk_count == chunk_size-1:
|
313 |
+
chunk_count = 0
|
314 |
+
yield self.tokenizer.decode(final_tokens, skip_special_tokens=True)
|
315 |
+
|
316 |
+
if chunk_count > 0:
|
317 |
+
yield self.tokenizer.decode(final_tokens, skip_special_tokens=True)
|
318 |
+
|
319 |
+
del final_tokens, input_ids
|
320 |
+
if self.device == "cuda":
|
321 |
+
torch.cuda.empty_cache()
|
322 |
+
|
323 |
+
def _infer(self, model_fn, **kwargs):
|
324 |
+
with torch.inference_mode():
|
325 |
+
return model_fn(**kwargs)
|
326 |
+
|
327 |
+
def tokenize(self, text):
|
328 |
+
"""Tokenize a string into a tensor of token IDs."""
|
329 |
+
batch = self.tokenizer.encode(text, return_tensors="pt")
|
330 |
+
return batch[0].to(self.device)
|
331 |
+
|
332 |
+
def generate(self, input_ids, logprobs=0, **kwargs):
|
333 |
+
"""Generate a stream of predicted tokens using the language model."""
|
334 |
+
|
335 |
+
# Store the original batch size and input length.
|
336 |
+
batch_size = input_ids.shape[0]
|
337 |
+
input_length = input_ids.shape[-1]
|
338 |
+
|
339 |
+
# Separate model arguments from generation config.
|
340 |
+
config = self.model.generation_config
|
341 |
+
config = copy.deepcopy(config)
|
342 |
+
kwargs = config.update(**kwargs)
|
343 |
+
kwargs["output_attentions"] = False
|
344 |
+
kwargs["output_hidden_states"] = False
|
345 |
+
kwargs["use_cache"] = True
|
346 |
+
|
347 |
+
# Collect special token IDs.
|
348 |
+
pad_token_id = config.pad_token_id
|
349 |
+
bos_token_id = config.bos_token_id
|
350 |
+
eos_token_id = config.eos_token_id
|
351 |
+
if isinstance(eos_token_id, int):
|
352 |
+
eos_token_id = [eos_token_id]
|
353 |
+
if pad_token_id is None and eos_token_id is not None:
|
354 |
+
pad_token_id = eos_token_id[0]
|
355 |
+
|
356 |
+
# Generate from eos if no input is specified.
|
357 |
+
if input_length == 0:
|
358 |
+
input_ids = input_ids.new_ones((batch_size, 1)).long()
|
359 |
+
if eos_token_id is not None:
|
360 |
+
input_ids = input_ids * eos_token_id[0]
|
361 |
+
input_length = 1
|
362 |
+
|
363 |
+
# Keep track of which sequences are already finished.
|
364 |
+
unfinished = input_ids.new_ones(batch_size)
|
365 |
+
|
366 |
+
# Start auto-regressive generation.
|
367 |
+
while True:
|
368 |
+
inputs = self.model.prepare_inputs_for_generation(
|
369 |
+
input_ids, **kwargs
|
370 |
+
) # noqa: E501
|
371 |
+
|
372 |
+
outputs = self._infer(
|
373 |
+
self.model,
|
374 |
+
**inputs,
|
375 |
+
# return_dict=True,
|
376 |
+
output_attentions=False,
|
377 |
+
output_hidden_states=False,
|
378 |
+
)
|
379 |
+
|
380 |
+
# Pre-process the probability distribution of the next tokens.
|
381 |
+
logits = outputs.logits[:, -1, :]
|
382 |
+
with torch.inference_mode():
|
383 |
+
logits = self.processor(input_ids, logits)
|
384 |
+
probs = torch.nn.functional.softmax(logits, dim=-1)
|
385 |
+
|
386 |
+
# Select deterministic or stochastic decoding strategy.
|
387 |
+
if (config.top_p is not None and config.top_p <= 0) or (
|
388 |
+
config.temperature is not None and config.temperature <= 0
|
389 |
+
):
|
390 |
+
tokens = torch.argmax(probs, dim=-1)[:, None]
|
391 |
+
else:
|
392 |
+
tokens = torch.multinomial(probs, num_samples=1)
|
393 |
+
|
394 |
+
tokens = tokens.squeeze(1)
|
395 |
+
|
396 |
+
# Finished sequences should have their next token be a padding.
|
397 |
+
if pad_token_id is not None:
|
398 |
+
tokens = tokens * unfinished + pad_token_id * (1 - unfinished)
|
399 |
+
|
400 |
+
# Append selected tokens to the inputs.
|
401 |
+
input_ids = torch.cat([input_ids, tokens[:, None]], dim=-1)
|
402 |
+
|
403 |
+
# Mark sequences with eos tokens as finished.
|
404 |
+
if eos_token_id is not None:
|
405 |
+
not_eos = sum(tokens != i for i in eos_token_id)
|
406 |
+
unfinished = unfinished.mul(not_eos.long())
|
407 |
+
|
408 |
+
# Set status to -1 if exceeded the max length.
|
409 |
+
status = unfinished.clone()
|
410 |
+
if input_ids.shape[-1] - input_length >= config.max_new_tokens:
|
411 |
+
status = 0 - status
|
412 |
+
|
413 |
+
# Yield predictions and status.
|
414 |
+
yield tokens
|
415 |
+
|
416 |
+
# Stop when finished or exceeded the max length.
|
417 |
+
if status.max() <= 0:
|
418 |
+
break
|
419 |
+
|
420 |
+
generation_config = get_generation_config(
|
421 |
+
"./generation_config_default.yaml"
|
422 |
+
)
|
423 |
+
|
424 |
+
model, tokenizer = load_model(
|
425 |
+
# base="decapoda-research/llama-13b-hf",
|
426 |
+
# finetuned="mchl-labs/stambecco-13b-plus",
|
427 |
+
)
|
428 |
+
|
429 |
+
stream_model = StreamModel(model, tokenizer)
|
430 |
+
|
431 |
+
def chat_stream(
|
432 |
+
context,
|
433 |
+
instruction,
|
434 |
+
state_chatbot,
|
435 |
+
):
|
436 |
+
if len(context) > 1000 or len(instruction) > 300:
|
437 |
+
raise gr.Error("Context or prompt is too long!")
|
438 |
+
|
439 |
+
bot_summarized_response = ''
|
440 |
+
# user input should be appropriately formatted (don't be confused by the function name)
|
441 |
+
instruction_display = instruction
|
442 |
+
instruction_prompt, conv_length = generate_prompt(instruction, state_chatbot, context)
|
443 |
+
|
444 |
+
if conv_length > num_of_characters_to_keep:
|
445 |
+
instruction_prompt = generate_prompt(SPECIAL_STRS["summarize"], state_chatbot, context, partial=True)[0]
|
446 |
+
|
447 |
+
state_chatbot = state_chatbot + [
|
448 |
+
(
|
449 |
+
None,
|
450 |
+
"![](https://s2.gifyu.com/images/icons8-loading-circle.gif) Conversazione troppo lunga, sto riassumendo..."
|
451 |
+
)
|
452 |
+
]
|
453 |
+
yield (state_chatbot, state_chatbot, context)
|
454 |
+
|
455 |
+
bot_summarized_response = get_output_batch(
|
456 |
+
model, tokenizer, [instruction_prompt], generation_config
|
457 |
+
)[0]
|
458 |
+
bot_summarized_response = bot_summarized_response.split("### Risposta:")[-1].strip()
|
459 |
+
|
460 |
+
state_chatbot[-1] = (
|
461 |
+
None,
|
462 |
+
"✅ Riepilogo della conversazione effettuato e impostato come contesto"
|
463 |
+
)
|
464 |
+
print(f"bot_summarized_response: {bot_summarized_response}")
|
465 |
+
yield (state_chatbot, state_chatbot, f"{context}. {bot_summarized_response}".strip())
|
466 |
+
|
467 |
+
instruction_prompt = generate_prompt(instruction, state_chatbot, f"{context} {bot_summarized_response}")[0]
|
468 |
+
|
469 |
+
bot_response = stream_model(
|
470 |
+
instruction_prompt,
|
471 |
+
max_tokens=256,
|
472 |
+
temperature=1,
|
473 |
+
top_p=0.9
|
474 |
+
)
|
475 |
+
|
476 |
+
instruction_display = None if instruction_display == SPECIAL_STRS["continue"] else instruction_display
|
477 |
+
state_chatbot = state_chatbot + [(instruction_display, None)]
|
478 |
+
yield (state_chatbot, state_chatbot, f"{context}. {bot_summarized_response}".strip())
|
479 |
+
|
480 |
+
prev_index = 0
|
481 |
+
agg_tokens = ""
|
482 |
+
cutoff_idx = 0
|
483 |
+
for tokens in bot_response:
|
484 |
+
tokens = tokens.strip()
|
485 |
+
cur_token = tokens[prev_index:]
|
486 |
+
|
487 |
+
if "#" in cur_token and agg_tokens == "":
|
488 |
+
cutoff_idx = tokens.find("#")
|
489 |
+
agg_tokens = tokens[cutoff_idx:]
|
490 |
+
|
491 |
+
if agg_tokens != "":
|
492 |
+
if len(agg_tokens) < len("### Istruzione:") :
|
493 |
+
agg_tokens = agg_tokens + cur_token
|
494 |
+
elif len(agg_tokens) >= len("### Istruzione:"):
|
495 |
+
if tokens.find("### Istruzione:") > -1:
|
496 |
+
processed_response, _ = post_process_stream(tokens[:tokens.find("### Istruzione:")].strip())
|
497 |
+
|
498 |
+
state_chatbot[-1] = (
|
499 |
+
instruction_display,
|
500 |
+
processed_response
|
501 |
+
)
|
502 |
+
yield (state_chatbot, state_chatbot, f"{context} {bot_summarized_response}".strip())
|
503 |
+
break
|
504 |
+
else:
|
505 |
+
agg_tokens = ""
|
506 |
+
cutoff_idx = 0
|
507 |
+
|
508 |
+
if agg_tokens == "":
|
509 |
+
processed_response, to_exit = post_process_stream(tokens)
|
510 |
+
state_chatbot[-1] = (instruction_display, processed_response)
|
511 |
+
yield (state_chatbot, state_chatbot, f"{context} {bot_summarized_response}".strip())
|
512 |
+
|
513 |
+
if to_exit:
|
514 |
+
break
|
515 |
+
|
516 |
+
prev_index = len(tokens)
|
517 |
+
|
518 |
+
yield (
|
519 |
+
state_chatbot,
|
520 |
+
state_chatbot,
|
521 |
+
f"{context} {bot_summarized_response}".strip()
|
522 |
+
)
|
523 |
+
|
524 |
+
|
525 |
+
def chat_batch(
|
526 |
+
contexts,
|
527 |
+
instructions,
|
528 |
+
state_chatbots,
|
529 |
+
):
|
530 |
+
state_results = []
|
531 |
+
ctx_results = []
|
532 |
+
|
533 |
+
instruct_prompts = [
|
534 |
+
generate_prompt(instruct, histories, ctx)
|
535 |
+
for ctx, instruct, histories in zip(contexts, instructions, state_chatbots)
|
536 |
+
]
|
537 |
+
|
538 |
+
bot_responses = get_output_batch(
|
539 |
+
model, tokenizer, instruct_prompts, generation_config
|
540 |
+
)
|
541 |
+
bot_responses = post_processes_batch(bot_responses)
|
542 |
+
|
543 |
+
for ctx, instruction, bot_response, state_chatbot in zip(contexts, instructions, bot_responses, state_chatbots):
|
544 |
+
new_state_chatbot = state_chatbot + [('' if instruction == SPECIAL_STRS["continue"] else instruction, bot_response)]
|
545 |
+
ctx_results.append(gr.Textbox.update(value=bot_response) if instruction == SPECIAL_STRS["summarize"] else ctx)
|
546 |
+
state_results.append(new_state_chatbot)
|
547 |
+
|
548 |
+
return (state_results, state_results, ctx_results)
|
549 |
+
|
550 |
+
def reset_textbox():
|
551 |
+
return gr.Textbox.update(value='')
|
552 |
+
|
553 |
+
def reset_everything(
|
554 |
+
context_txtbox,
|
555 |
+
instruction_txtbox,
|
556 |
+
state_chatbot):
|
557 |
+
|
558 |
+
state_chatbot = []
|
559 |
+
|
560 |
+
return (
|
561 |
+
state_chatbot,
|
562 |
+
state_chatbot,
|
563 |
+
gr.Textbox.update(value=''),
|
564 |
+
gr.Textbox.update(value=''),
|
565 |
+
)
|
566 |
+
|
567 |
+
with gr.Blocks(css=PARENT_BLOCK_CSS) as demo:
|
568 |
+
state_chatbot = gr.State([])
|
569 |
+
|
570 |
+
with gr.Column(elem_id='col_container'):
|
571 |
+
gr.Markdown(f"## {TITLE}\n\n\n{ABSTRACT}")
|
572 |
+
|
573 |
+
with gr.Accordion("Context Setting", open=False):
|
574 |
+
context_txtbox = gr.Textbox(placeholder="Surrounding information to AI", label="Enter Context")
|
575 |
+
hidden_txtbox = gr.Textbox(placeholder="", label="Order", visible=False)
|
576 |
+
|
577 |
+
chatbot = gr.Chatbot(elem_id='chatbot', label="Stambecco")
|
578 |
+
instruction_txtbox = gr.Textbox(placeholder="What do you want to say to AI?", label="Instruction")
|
579 |
+
with gr.Row():
|
580 |
+
cancel_btn = gr.Button(value="Cancel")
|
581 |
+
reset_btn = gr.Button(value="Reset")
|
582 |
+
|
583 |
+
with gr.Accordion("Helper Buttons", open=False):
|
584 |
+
gr.Markdown(f"`Continue` lets AI to complete the previous incomplete answers. `Summarize` lets AI to summarize the conversations so far.")
|
585 |
+
continue_txtbox = gr.Textbox(value=SPECIAL_STRS["continue"], visible=False)
|
586 |
+
summrize_txtbox = gr.Textbox(value=SPECIAL_STRS["summarize"], visible=False)
|
587 |
+
|
588 |
+
continue_btn = gr.Button(value="Continue")
|
589 |
+
summarize_btn = gr.Button(value="Summarize")
|
590 |
+
|
591 |
+
gr.Markdown("#### Examples")
|
592 |
+
for _, (category, examples) in enumerate(DEFAULT_EXAMPLES.items()):
|
593 |
+
with gr.Accordion(category, open=False):
|
594 |
+
if category == "Identity":
|
595 |
+
for item in examples:
|
596 |
+
with gr.Accordion(item["title"], open=False):
|
597 |
+
gr.Examples(
|
598 |
+
examples=item["examples"],
|
599 |
+
inputs=[
|
600 |
+
hidden_txtbox, context_txtbox, instruction_txtbox
|
601 |
+
],
|
602 |
+
label=None
|
603 |
+
)
|
604 |
+
else:
|
605 |
+
for item in examples:
|
606 |
+
with gr.Accordion(item["title"], open=False):
|
607 |
+
gr.Examples(
|
608 |
+
examples=item["examples"],
|
609 |
+
inputs=[
|
610 |
+
hidden_txtbox, instruction_txtbox
|
611 |
+
],
|
612 |
+
label=None
|
613 |
+
)
|
614 |
+
|
615 |
+
gr.Markdown(f"{BOTTOM_LINE}")
|
616 |
+
|
617 |
+
|
618 |
+
send_event = instruction_txtbox.submit(
|
619 |
+
chat_stream,
|
620 |
+
[context_txtbox, instruction_txtbox, state_chatbot],
|
621 |
+
[state_chatbot, chatbot, context_txtbox],
|
622 |
+
)
|
623 |
+
reset_event = instruction_txtbox.submit(
|
624 |
+
reset_textbox,
|
625 |
+
[],
|
626 |
+
[instruction_txtbox],
|
627 |
+
)
|
628 |
+
|
629 |
+
continue_event = continue_btn.click(
|
630 |
+
chat_stream,
|
631 |
+
[context_txtbox, continue_txtbox, state_chatbot],
|
632 |
+
[state_chatbot, chatbot, context_txtbox],
|
633 |
+
)
|
634 |
+
reset_continue_event = continue_btn.click(
|
635 |
+
reset_textbox,
|
636 |
+
[],
|
637 |
+
[instruction_txtbox],
|
638 |
+
)
|
639 |
+
|
640 |
+
summarize_event = summarize_btn.click(
|
641 |
+
chat_stream,
|
642 |
+
[context_txtbox, summrize_txtbox, state_chatbot],
|
643 |
+
[state_chatbot, chatbot, context_txtbox],
|
644 |
+
)
|
645 |
+
summarize_reset_event = summarize_btn.click(
|
646 |
+
reset_textbox,
|
647 |
+
[],
|
648 |
+
[instruction_txtbox],
|
649 |
+
)
|
650 |
+
|
651 |
+
cancel_btn.click(
|
652 |
+
None, None, None,
|
653 |
+
cancels=[
|
654 |
+
send_event, continue_event, summarize_event
|
655 |
+
]
|
656 |
+
)
|
657 |
+
|
658 |
+
reset_btn.click(
|
659 |
+
reset_everything,
|
660 |
+
[context_txtbox, instruction_txtbox, state_chatbot],
|
661 |
+
[state_chatbot, chatbot, context_txtbox, instruction_txtbox],
|
662 |
+
cancels=[
|
663 |
+
send_event, continue_event, summarize_event
|
664 |
+
]
|
665 |
+
)
|
666 |
+
|
667 |
+
demo.queue(
|
668 |
+
concurrency_count=1,
|
669 |
+
max_size=100,
|
670 |
+
).launch(
|
671 |
+
max_threads=5,
|
672 |
+
server_name="0.0.0.0",
|
673 |
+
share=True
|
674 |
+
)
|