Delete mm_builder.py
Browse files- mm_builder.py +0 -33
mm_builder.py
DELETED
@@ -1,33 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
from modeling_mmalaya import MMAlayaMPTForCausalLM
|
3 |
-
from transformers import AutoTokenizer
|
4 |
-
from mm_utils import DEFAULT_IMAGE_TOKEN
|
5 |
-
|
6 |
-
|
7 |
-
def load_pretrained_model(model_path, device_map="auto", device="cuda"):
|
8 |
-
kwargs = {"device_map": device_map}
|
9 |
-
if device != "cuda":
|
10 |
-
kwargs['device_map'] = {"": device}
|
11 |
-
kwargs['torch_dtype'] = torch.bfloat16
|
12 |
-
|
13 |
-
print('******** load mpt model from here kwargs ', kwargs)
|
14 |
-
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
15 |
-
model = MMAlayaMPTForCausalLM.from_pretrained(
|
16 |
-
model_path,
|
17 |
-
low_cpu_mem_usage=True,
|
18 |
-
**kwargs
|
19 |
-
)
|
20 |
-
|
21 |
-
tokenizer.add_tokens([DEFAULT_IMAGE_TOKEN], special_tokens=True)
|
22 |
-
model.resize_token_embeddings(len(tokenizer))
|
23 |
-
vision_tower = model.get_vision_tower()
|
24 |
-
vision_tower.to(device=device, dtype=torch.float16)
|
25 |
-
image_processor = vision_tower.image_processor
|
26 |
-
|
27 |
-
if hasattr(model.config, "max_sequence_length"):
|
28 |
-
context_len = model.config.max_sequence_length
|
29 |
-
else:
|
30 |
-
context_len = 2048
|
31 |
-
|
32 |
-
return tokenizer, model, image_processor, context_len
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|