VascoBartolo
commited on
Commit
·
776aee4
1
Parent(s):
4d14f79
add custom handler
Browse files- __pycache__/handler.cpython-39.pyc +0 -0
- handler.py +43 -0
- model-00002-of-00003.safetensors +0 -3
- model-00003-of-00003.safetensors +0 -3
__pycache__/handler.cpython-39.pyc
ADDED
Binary file (1.73 kB). View file
|
|
handler.py
ADDED
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Any
|
2 |
+
import torch
|
3 |
+
from transformers import LlavaNextVideoProcessor, LlavaNextVideoForConditionalGeneration
|
4 |
+
|
5 |
+
class EndpointHandler:
|
6 |
+
def __init__(self, path=""):
|
7 |
+
# Adjust the loading process to handle potential configuration mismatches
|
8 |
+
self.model = LlavaNextVideoForConditionalGeneration.from_pretrained(
|
9 |
+
path
|
10 |
+
)
|
11 |
+
self.processor = LlavaNextVideoProcessor.from_pretrained(path)
|
12 |
+
|
13 |
+
# Ensure the model is in evaluation mode
|
14 |
+
self.model.eval()
|
15 |
+
|
16 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
17 |
+
"""
|
18 |
+
Args:
|
19 |
+
data (Dict): Contains the input data including "clip" and "prompt".
|
20 |
+
|
21 |
+
Returns:
|
22 |
+
List[Dict[str, Any]]: The generated text from the model.
|
23 |
+
"""
|
24 |
+
# Extract inputs from the data dictionary
|
25 |
+
clip = data.get("clip")
|
26 |
+
prompt = data.get("prompt")
|
27 |
+
|
28 |
+
if clip is None or prompt is None:
|
29 |
+
return [{"error": "Missing 'clip' or 'prompt' in input data"}]
|
30 |
+
|
31 |
+
# Prepare the inputs for the model
|
32 |
+
inputs_video = self.processor(text=prompt, videos=clip, padding=True, return_tensors="pt").to(self.model.device)
|
33 |
+
|
34 |
+
# Generate output from the model
|
35 |
+
generate_kwargs = {"max_new_tokens": 512, "do_sample": True, "top_p": 0.9}
|
36 |
+
output = self.model.generate(**inputs_video, **generate_kwargs)
|
37 |
+
generated_text = self.processor.batch_decode(output, skip_special_tokens=True)
|
38 |
+
|
39 |
+
# Extract the relevant part of the assistant's answer
|
40 |
+
assistant_answer_start = generated_text[0].find("ASSISTANT:") + len("ASSISTANT:")
|
41 |
+
assistant_answer = generated_text[0][assistant_answer_start:].strip()
|
42 |
+
|
43 |
+
return [{"generated_text": assistant_answer}]
|
model-00002-of-00003.safetensors
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:d33895934eef0fd0121b24eb475248232d3d2a9e847e7c957814d9cb4ca4e8e4
|
3 |
-
size 4957878440
|
|
|
|
|
|
|
|
model-00003-of-00003.safetensors
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:06fdf4f4afa75b24b6467da6c4bc10a4d609e0aae8ea648a24b43eff0cba5742
|
3 |
-
size 4176137408
|
|
|
|
|
|
|
|