revert to origin

config/infer.yaml

@@ -2,6 +2,7 @@ device: 'cuda'
 
 vicuna:
   model_path: '/home/user/app/vicuna-7b'
+  # model_path: '/mnt/petrelfs/wangyiqin/vid_cap/ChatVID_huggingface/vicuna-7b'
   device: 'cuda'
   num_gpus: 'auto'
   max_gpu_memory: '24Gib'
@@ -11,6 +12,7 @@ vicuna:
   max_new_tokens: 512
   debug: False
   output_path: '/home/user/app/vicuna_out.json'
+  # output_path: '/mnt/petrelfs/wangyiqin/vid_cap/ChatVID_huggingface/vicuna_out.json'
 
 vid2seq:
   enable: False

model/fastchat/serve/compression.py

@@ -28,7 +28,7 @@ class CLinear(nn.Module):
     def __init__(self, weight, bias, device):
         super().__init__()
 
-        self.weight = compress(weight.detach().to(device), default_compression_config)
+        self.weight = compress(weight.data.to(device), default_compression_config)
         self.bias = bias
 
     def forward(self, input: Tensor) -> Tensor:

model/vision/ImageCaptioner.py

@@ -15,6 +15,11 @@ class ImageCaptioner:
         self.model = Blip2ForConditionalGeneration.from_pretrained(
             "/home/user/app/pretrained_models/blip2-opt-2.7b",
             torch_dtype=self.data_type, device_map="auto")
+        # self.processor = Blip2Processor.from_pretrained(
+        #     "/mnt/petrelfs/wangyiqin/vid_cap/ChatVID_huggingface/pretrained_models/blip2-opt-2.7b")
+        # self.model = Blip2ForConditionalGeneration.from_pretrained(
+        #     "/mnt/petrelfs/wangyiqin/vid_cap/ChatVID_huggingface/pretrained_models/blip2-opt-2.7b",
+        #     torch_dtype=self.data_type, device_map="auto")
 
     def __call__(self, imgs):
         inputs = self.processor(