set HF_TOKEN

app.py

@@ -24,7 +24,7 @@ import os
 import torch
 
 device = 'cuda'
-model_path = os.getenv("MODEL_PATH", "omni-research/Tarsier-7b")
+model_path = os.getenv("MODEL_PATH", "omni-research/Tarsier2-7b")
 max_n_frames = int(os.getenv("MAX_N_FRAMES", 8))
 debug = False
 

dataset/processor.py

@@ -20,6 +20,8 @@ import re
 
 from .utils import sample_image, sample_video, sample_gif, get_visual_type
 
+HF_TOKEN = os.environ.get('HF_TOKEN', '')
+
 ext2sampler = {
     'image': sample_image,
     'gif': sample_gif,
@@ -83,6 +85,7 @@ class Processor(object):
             model_name_or_path,
             padding_side='left',
             trust_remote_code=True,
+            token=HF_TOKEN,
         )
         self.processor = CustomImageProcessor(sub_processor)
         self.tokenizer = sub_processor.tokenizer

tools/utils.py

@@ -15,6 +15,9 @@ from models.modeling_tarsier import TarsierForConditionalGeneration, LlavaConfig
 from dataset.processor import Processor
 import torch
 import base64
+import os
+
+HF_TOKEN = os.environ.get('HF_TOKEN', '')
 
 class Color:
 
@@ -52,13 +55,15 @@ def load_model_and_processor(model_name_or_path, max_n_frames=8):
     model_config = LlavaConfig.from_pretrained(
         model_name_or_path,
         trust_remote_code=True,
+        token=HF_TOKEN,
     )
     model = TarsierForConditionalGeneration.from_pretrained(
         model_name_or_path,
         config=model_config,
         device_map='auto',
         torch_dtype=torch.float16,
-        trust_remote_code=True
+        trust_remote_code=True,
+        token=HF_TOKEN,
     )
     model.eval()
     return model, processor