Create handler.py
Browse files- handler.py +106 -0
handler.py
ADDED
@@ -0,0 +1,106 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import AutoProcessor, AutoModelForVision2Seq, GenerationConfig
|
3 |
+
from transformers.image_utils import load_image
|
4 |
+
|
5 |
+
from typing import Any, Dict
|
6 |
+
|
7 |
+
import base64
|
8 |
+
import re
|
9 |
+
from copy import deepcopy
|
10 |
+
|
11 |
+
|
12 |
+
def is_base64(s: str) -> bool:
|
13 |
+
try:
|
14 |
+
return base64.b64encode(base64.b64decode(s)).decode() == s
|
15 |
+
except Exception:
|
16 |
+
return False
|
17 |
+
|
18 |
+
|
19 |
+
def is_url(s: str) -> bool:
|
20 |
+
url_pattern = re.compile(r"https?://(?:[-\w.]|(?:%[\da-fA-F]{2}))+")
|
21 |
+
return bool(url_pattern.match(s))
|
22 |
+
|
23 |
+
|
24 |
+
class EndpointHandler:
|
25 |
+
def __init__(
|
26 |
+
self,
|
27 |
+
model_dir: str = "HuggingFaceTB/SmolVLM-Instruct",
|
28 |
+
**kwargs: Any, # type: ignore
|
29 |
+
) -> None:
|
30 |
+
self.processor = AutoProcessor.from_pretrained(model_dir)
|
31 |
+
self.model = AutoModelForVision2Seq.from_pretrained(
|
32 |
+
model_dir,
|
33 |
+
torch_dtype=torch.bfloat16,
|
34 |
+
_attn_implementation="flash_attention_2",
|
35 |
+
device_map="auto",
|
36 |
+
).eval()
|
37 |
+
self.generation_config = GenerationConfig.from_pretrained(model_dir)
|
38 |
+
|
39 |
+
def __call__(self, data: Dict[str, Any]) -> Any:
|
40 |
+
if "inputs" not in data:
|
41 |
+
raise ValueError(
|
42 |
+
"The request body must contain a key 'inputs' with a list of inputs."
|
43 |
+
)
|
44 |
+
|
45 |
+
if not isinstance(data["inputs"], list):
|
46 |
+
raise ValueError(
|
47 |
+
"The request inputs must be a list of dictionaries with the keys 'text' and 'images', being a"
|
48 |
+
" string with the prompt and a list with the image URLs or base64 encodings, respectively; and"
|
49 |
+
" optionally including the key 'generation_parameters' key too."
|
50 |
+
)
|
51 |
+
|
52 |
+
predictions = []
|
53 |
+
for input in data["inputs"]:
|
54 |
+
if "text" not in input:
|
55 |
+
raise ValueError(
|
56 |
+
"The request input body must contain the key 'text' with the prompt to use."
|
57 |
+
)
|
58 |
+
|
59 |
+
if "images" not in input or (
|
60 |
+
not isinstance(input["images"], list)
|
61 |
+
and all(isinstance(i, str) for i in input["images"])
|
62 |
+
):
|
63 |
+
raise ValueError(
|
64 |
+
"The request input body must contain the key 'images' with a list of strings,"
|
65 |
+
" where each string corresponds to an image on either base64 encoding, or provided"
|
66 |
+
" as a valid URL (needs to be publicly accessible and contain a valid image)."
|
67 |
+
)
|
68 |
+
|
69 |
+
images = []
|
70 |
+
for image in input["images"]:
|
71 |
+
try:
|
72 |
+
images.append(load_image(image))
|
73 |
+
except Exception as e:
|
74 |
+
raise ValueError(
|
75 |
+
f"Provided {image=} is not valid, please make sure that's either a base64 encoding"
|
76 |
+
f" of a valid image, or a publicly accesible URL to a valid image.\nFailed with {e=}."
|
77 |
+
)
|
78 |
+
|
79 |
+
generation_config = deepcopy(self.generation_config)
|
80 |
+
generation_config.update(**input.get("generation_parameters", {}))
|
81 |
+
|
82 |
+
messages = [
|
83 |
+
{
|
84 |
+
"role": "user",
|
85 |
+
"content": [{"type": "image"} for _ in images]
|
86 |
+
+ [{"type": "text", "text": input["text"]}],
|
87 |
+
},
|
88 |
+
]
|
89 |
+
prompt = self.processor.apply_chat_template(
|
90 |
+
messages, add_generation_prompt=True
|
91 |
+
)
|
92 |
+
processed_inputs = self.processor(
|
93 |
+
text=prompt, images=images, return_tensors="pt"
|
94 |
+
)
|
95 |
+
|
96 |
+
with torch.no_grad(), torch.autocast(self.model.device):
|
97 |
+
generated_ids = self.model.generate(
|
98 |
+
**processed_inputs, **generation_config
|
99 |
+
)
|
100 |
+
generated_texts = self.processor.batch_decode(
|
101 |
+
generated_ids,
|
102 |
+
skip_special_tokens=True,
|
103 |
+
)
|
104 |
+
predictions.append(generated_texts[0])
|
105 |
+
|
106 |
+
return {"predictions": predictions}
|