zaddyzaddy
commited on
Commit
•
8fe825e
1
Parent(s):
01f6258
Upload 2 files
Browse files- handler.py +73 -0
- requirements.txt +10 -0
handler.py
ADDED
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from unsloth import FastLanguageModel
|
2 |
+
from typing import Dict, List, Any
|
3 |
+
import torch
|
4 |
+
|
5 |
+
class EndpointHandler:
|
6 |
+
def __init__(self, path=""):
|
7 |
+
max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
|
8 |
+
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
|
9 |
+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
|
10 |
+
|
11 |
+
self.model, self.tokenizer = FastLanguageModel.from_pretrained(
|
12 |
+
model_name=path,
|
13 |
+
max_seq_length=max_seq_length,
|
14 |
+
dtype=dtype,
|
15 |
+
load_in_4bit=load_in_4bit,
|
16 |
+
# token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
|
17 |
+
)
|
18 |
+
|
19 |
+
self.alpaca_prompt = """
|
20 |
+
### Instruction:
|
21 |
+
{}
|
22 |
+
|
23 |
+
### Input:
|
24 |
+
{}
|
25 |
+
|
26 |
+
### Response:
|
27 |
+
"""
|
28 |
+
|
29 |
+
self.EOS_TOKEN = self.tokenizer.eos_token
|
30 |
+
|
31 |
+
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
32 |
+
"""
|
33 |
+
data args:
|
34 |
+
inputs (:obj: `str`)
|
35 |
+
date (:obj: `str`)
|
36 |
+
Return:
|
37 |
+
A :obj:`list` | `dict`: will be serialized and returned
|
38 |
+
"""
|
39 |
+
|
40 |
+
data = data.pop("inputs", data)
|
41 |
+
input_text = data.get("input_text", "")
|
42 |
+
lex_diversity = data.get("lex_diversity", 80)
|
43 |
+
order_diversity = data.get("order_diversity", 20)
|
44 |
+
repetition_penalty = data.get("repetition_penalty", 1.0)
|
45 |
+
use_cache = data.get("use_cache", False)
|
46 |
+
max_length = data.get("max_length", 128)
|
47 |
+
|
48 |
+
prediction = self.paraphrase(
|
49 |
+
input_text,
|
50 |
+
lex_diversity,
|
51 |
+
order_diversity,
|
52 |
+
repetition_penalty=repetition_penalty,
|
53 |
+
use_cache=use_cache,
|
54 |
+
max_length=max_length
|
55 |
+
)
|
56 |
+
|
57 |
+
prediction = {'prediction': prediction}
|
58 |
+
return prediction
|
59 |
+
|
60 |
+
def paraphrase(self, input_text, lex_diversity, order_diversity, repetition_penalty, use_cache, max_length, **kwargs):
|
61 |
+
FastLanguageModel.for_inference(self.model) # Enable native 2x faster inference
|
62 |
+
inputs = self.tokenizer(
|
63 |
+
[
|
64 |
+
self.alpaca_prompt.format(
|
65 |
+
"You are an AI assistant, capable of paraphrasing any text to a human-like version of the text. Human writing often exhibits bursts and lulls, with a mix of long and short sentences", # instruction
|
66 |
+
f"lexical = {lex_diversity}, order = {order_diversity} {input_text}",
|
67 |
+
"", # output - leave this blank for generation!
|
68 |
+
)
|
69 |
+
], return_tensors="pt").to("cuda")
|
70 |
+
|
71 |
+
outputs = self.model.generate(**inputs, max_new_tokens=max_length, use_cache=False, repetition_penalty=repetition_penalty)
|
72 |
+
output_text = self.tokenizer.batch_decode(outputs)
|
73 |
+
return output_text
|
requirements.txt
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git
|
2 |
+
packaging
|
3 |
+
ninja
|
4 |
+
einops
|
5 |
+
flash-attn
|
6 |
+
xformers
|
7 |
+
trl
|
8 |
+
peft
|
9 |
+
accelerate
|
10 |
+
bitsandbytes
|