viethoangtranduong
commited on
Commit
·
a679b46
1
Parent(s):
958469b
Create handler.py
Browse files- handler.py +38 -0
handler.py
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from typing import Dict, List, Any
|
3 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
4 |
+
|
5 |
+
MAX_TOKENS_IN_BATCH = 4_000
|
6 |
+
DEFAULT_MAX_NEW_TOKENS = 10
|
7 |
+
|
8 |
+
|
9 |
+
class EndpointHandler():
|
10 |
+
def __init__(self, path: str = ""):
|
11 |
+
assert torch.cuda.device_count() >= 4, f"Only found access to {torch.cuda.device_count()} GPUs"
|
12 |
+
|
13 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path)
|
14 |
+
self.model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.float16)
|
15 |
+
self.model = self.model.to('cuda:0')
|
16 |
+
|
17 |
+
self.model.parallelize()
|
18 |
+
|
19 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
20 |
+
"""
|
21 |
+
Args:
|
22 |
+
data (:obj:):
|
23 |
+
includes the input data and the parameters for the inference.
|
24 |
+
Return:
|
25 |
+
A :obj:`list`:. The list contains the answer and scores of the inference inputs
|
26 |
+
"""
|
27 |
+
|
28 |
+
prompts = [f"<human>: {prompt}\n<bot>:" for prompt in data["inputs"]]
|
29 |
+
|
30 |
+
|
31 |
+
inputs = tokenizer(prompts, padding=True, return_tensors='pt').to(model.device)
|
32 |
+
input_length = inputs.input_ids.shape[1]
|
33 |
+
outputs = model.generate(
|
34 |
+
**inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.7, top_k=50
|
35 |
+
)
|
36 |
+
output_strs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
37 |
+
|
38 |
+
return output_strs
|