nizar-sayad
commited on
Commit
•
56fd19b
1
Parent(s):
0f2976b
add custom handler
Browse files- handler.py +32 -0
- requirements.txt +3 -0
handler.py
ADDED
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Any
|
2 |
+
from transformers import AutoModelForMultipleChoice, AutoTokenizer
|
3 |
+
import torch
|
4 |
+
|
5 |
+
|
6 |
+
|
7 |
+
class EndpointHandler:
|
8 |
+
def __init__(self, path=""):
|
9 |
+
# load model and processor from path
|
10 |
+
self.model = AutoModelForMultipleChoice.from_pretrained(path)
|
11 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path)
|
12 |
+
|
13 |
+
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
14 |
+
"""
|
15 |
+
Args:
|
16 |
+
data (:dict:):
|
17 |
+
The payload with the text prompt.
|
18 |
+
"""
|
19 |
+
# process input
|
20 |
+
input = data.pop("input", data)
|
21 |
+
gen_outputs_no_input_decoded = data.pop("gen_outputs_no_input_decoded", None)
|
22 |
+
|
23 |
+
rank_inputs = self.tokenizer([[input, gen_output_no_input_decoded] for gen_output_no_input_decoded in gen_outputs_no_input_decoded],
|
24 |
+
return_tensors="pt",
|
25 |
+
padding=True)
|
26 |
+
|
27 |
+
rank_labels = torch.tensor(0).unsqueeze(0)
|
28 |
+
|
29 |
+
rank_outputs = self.model(**{k: v.unsqueeze(0) for k, v in rank_inputs.items()}, labels=rank_labels)
|
30 |
+
|
31 |
+
rank_predictions = torch.nn.functional.softmax(rank_outputs.logits, dim=-1)[0].tolist()
|
32 |
+
return {"rank_predictions": rank_predictions}
|
requirements.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
accelerate
|
2 |
+
bitsandbytes
|
3 |
+
transformers
|