-
Notifications
You must be signed in to change notification settings - Fork 0
/
translate.py
178 lines (151 loc) · 5.35 KB
/
translate.py
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
import json
from argparse import ArgumentParser
from pathlib import Path
from typing import Annotated, Dict
import pycountry
import webvtt
from exllamav2 import (
ExLlamaV2,
ExLlamaV2Cache,
ExLlamaV2Cache_Q4,
ExLlamaV2Cache_Q6,
ExLlamaV2Cache_Q8,
ExLlamaV2Config,
ExLlamaV2Tokenizer,
)
from exllamav2.generator import (
ExLlamaV2DynamicGenerator,
ExLlamaV2DynamicJob,
ExLlamaV2Sampler,
)
from exllamav2.generator.filters import ExLlamaV2PrefixFilter
from jinja2 import Template
from lmformatenforcer import JsonSchemaParser
from lmformatenforcer.integrations.exllamav2 import ExLlamaV2TokenEnforcerFilter
from pydantic import Field, RootModel, StringConstraints
from rich import print
parser = ArgumentParser()
parser.add_argument("-m", "--model", type=Path, required=True)
parser.add_argument("-c", "--cache_bits", type=int, default=16, choices=(4, 6, 8, 16))
parser.add_argument("-s", "--seq_len", type=int, default=8192)
parser.add_argument("-l", "--line_len", type=int, default=128)
parser.add_argument("-f", "--lang_from", type=str, default="Chinese")
parser.add_argument("-t", "--lang_to", type=str, default="English")
parser.add_argument("-i", "--input", type=Path, default=".")
args = parser.parse_args()
lang_from = args.lang_from.capitalize().strip()
lang_to = args.lang_to.capitalize().strip()
lang_code = pycountry.languages.get(name=lang_to)
lang_code = lang_code.alpha_2 if lang_code else lang_to.lower()[:2]
inputs = (
[i for i in args.input.glob("*.*") if i.suffix in (".srt", ".vtt")]
if args.input.is_dir()
else [args.input] if args.input.is_file() else []
)
config = ExLlamaV2Config(str(args.model))
config.fasttensors = True
init_len = args.seq_len or config.max_seq_len
tokenizer = ExLlamaV2Tokenizer(config)
chat_template = tokenizer.tokenizer_config_dict["chat_template"]
template = Template(chat_template)
settings = ExLlamaV2Sampler.Settings()
settings.greedy()
caches = {
4: lambda m: ExLlamaV2Cache_Q4(m, lazy=True),
6: lambda m: ExLlamaV2Cache_Q6(m, lazy=True),
8: lambda m: ExLlamaV2Cache_Q8(m, lazy=True),
16: lambda m: ExLlamaV2Cache(m, lazy=True),
}
for input in inputs:
subs = webvtt.from_srt(input) if input.suffix == ".srt" else webvtt.read(input)
subs_len = len(subs)
class Translation(RootModel):
root: Annotated[
Dict[
Annotated[int, Field(ge=0, lt=subs_len)],
Annotated[
str,
StringConstraints(
min_length=1,
max_length=args.line_len,
strip_whitespace=True,
),
],
],
subs_len,
]
schema = JsonSchemaParser(Translation.schema())
lines = {index: line.text for index, line in enumerate(subs)}
lines = json.dumps(lines, ensure_ascii=False, separators=(",", ":"))
instruction = (
f"Translate each line from {lang_from} to {lang_to}. "
"Consider the meaning of all lines when translating. "
f"Keep the line length under {args.line_len} characters. "
"Return the translation in JSON format."
)
messages = [
{"role": "system", "content": instruction},
{"role": "user", "content": lines},
]
try:
prompt = template.render(
messages=messages,
bos_token="",
add_generation_prompt=True,
)
except:
prompt = template.render(
messages=[{"role": "user", "content": f"{instruction}\n{lines}"}],
bos_token="",
add_generation_prompt=True,
)
input_ids = tokenizer.encode(prompt, add_bos=True, encode_special_tokens=True)
input_len = input_ids.shape[-1]
config.max_seq_len = int(max(init_len, subs_len * args.line_len / 4) // 256 * 256)
model = ExLlamaV2(config)
cache = caches[args.cache_bits](model)
model.load_autosplit(cache, progress=True)
generator = ExLlamaV2DynamicGenerator(model, cache, tokenizer)
filters = [
ExLlamaV2TokenEnforcerFilter(schema, tokenizer),
ExLlamaV2PrefixFilter(model, tokenizer, "{"),
]
job = ExLlamaV2DynamicJob(
input_ids=input_ids,
gen_settings=settings,
filters=filters,
filter_prefer_eos=True,
max_new_tokens=config.max_seq_len - input_len,
stop_conditions=[tokenizer.eos_token_id],
)
generator.enqueue(job)
eos = False
chunks = []
print(
f'Input: "{input.name}"\n'
f"Lines: {subs_len}\n"
f"Tokens: {input_len}\n"
f"Sequence: {config.max_seq_len}\n"
)
while not eos:
for result in generator.iterate():
if result["stage"] == "streaming":
chunk = result.get("text", "")
print(chunk, end="", flush=True)
chunks.append(chunk)
eos = result["eos"]
print("\n")
result = "".join(chunks)
result = json.loads(result)
result = Translation(**result).dict()
for key, value in result.items():
subs[key].text = value
output_name = input.name.replace("".join(input.suffixes), "")
output_name = f"{output_name}.{lang_code}"
output = input.parent / f"{output_name}.vtt"
index = 1
while output.exists():
output = input.parent / f"{output_name}.{index}.vtt"
index += 1
subs.save(output)
model.unload()