oleksandrfluxon
commited on
Commit
•
44092ff
1
Parent(s):
e7b6e5f
Create pipeline.py
Browse files- pipeline.py +81 -0
pipeline.py
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import cuda
|
2 |
+
import transformers
|
3 |
+
from accelerate import dispatch_model, infer_auto_device_map
|
4 |
+
from accelerate.utils import get_balanced_memory
|
5 |
+
from transformers import BitsAndBytesConfig, StoppingCriteria, StoppingCriteriaList
|
6 |
+
from typing import Dict, List, Any
|
7 |
+
|
8 |
+
# define custom stopping criteria object
|
9 |
+
class StopOnTokens(StoppingCriteria):
|
10 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
11 |
+
for stop_ids in stop_token_ids:
|
12 |
+
if torch.eq(input_ids[0][-len(stop_ids):], stop_ids).all():
|
13 |
+
return True
|
14 |
+
return False
|
15 |
+
|
16 |
+
class PreTrainedPipeline():
|
17 |
+
def __init__(self, path=""):
|
18 |
+
path = "oleksandrfluxon/mpt-7b-instruct-evaluate"
|
19 |
+
print("===> path", path)
|
20 |
+
|
21 |
+
device = f'cuda:{cuda.current_device()}' if cuda.is_available() else 'cpu'
|
22 |
+
print("===> device", device)
|
23 |
+
|
24 |
+
model = transformers.AutoModelForCausalLM.from_pretrained(
|
25 |
+
'oleksandrfluxon/mpt-7b-instruct-evaluate',
|
26 |
+
trust_remote_code=True,
|
27 |
+
load_in_8bit=True, # this requires the `bitsandbytes` library
|
28 |
+
max_seq_len=8192,
|
29 |
+
init_device=device
|
30 |
+
)
|
31 |
+
model.eval()
|
32 |
+
#model.to(device)
|
33 |
+
print(f"===> Model loaded on {device}")
|
34 |
+
|
35 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained("mosaicml/mpt-7b")
|
36 |
+
|
37 |
+
# we create a list of stopping criteria
|
38 |
+
stop_token_ids = [
|
39 |
+
tokenizer.convert_tokens_to_ids(x) for x in [
|
40 |
+
['Human', ':'], ['AI', ':']
|
41 |
+
]
|
42 |
+
]
|
43 |
+
stop_token_ids = [torch.LongTensor(x).to(device) for x in stop_token_ids]
|
44 |
+
print("===> stop_token_ids", stop_token_ids)
|
45 |
+
|
46 |
+
stopping_criteria = StoppingCriteriaList([StopOnTokens()])
|
47 |
+
|
48 |
+
self.pipeline = transformers.pipeline(
|
49 |
+
model=model, tokenizer=tokenizer,
|
50 |
+
return_full_text=True, # langchain expects the full text
|
51 |
+
task='text-generation',
|
52 |
+
# we pass model parameters here too
|
53 |
+
stopping_criteria=stopping_criteria, # without this model rambles during chat
|
54 |
+
temperature=0.1, # 'randomness' of outputs, 0.0 is the min and 1.0 the max
|
55 |
+
top_p=0.15, # select from top tokens whose probability add up to 15%
|
56 |
+
top_k=0, # select from top 0 tokens (because zero, relies on top_p)
|
57 |
+
max_new_tokens=128, # mex number of tokens to generate in the output
|
58 |
+
repetition_penalty=1.1 # without this output begins repeating
|
59 |
+
)
|
60 |
+
|
61 |
+
print("===> init finished")
|
62 |
+
|
63 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
64 |
+
"""
|
65 |
+
data args:
|
66 |
+
inputs (:obj: `str`)
|
67 |
+
parameters (:obj: `str`)
|
68 |
+
Return:
|
69 |
+
A :obj:`str`: todo
|
70 |
+
"""
|
71 |
+
# get inputs
|
72 |
+
inputs = data.pop("inputs",data)
|
73 |
+
parameters = data.pop("parameters", {})
|
74 |
+
date = data.pop("date", None)
|
75 |
+
print("===> inputs", inputs)
|
76 |
+
print("===> parameters", parameters)
|
77 |
+
|
78 |
+
result = self.pipeline(inputs, **parameters)
|
79 |
+
print("===> result", result)
|
80 |
+
|
81 |
+
return result
|