Update xgen-mm-vid-inference-script.py

xgen-mm-vid-inference-script.py

@@ -50,23 +50,11 @@ def generate(messages, images):
     image_sizes = [image.size for image in images]
     # Similar operation in model_worker.py
 
-    if cfg.vision_encoder_config.image_aspect_ratio == "anyres":
-        image_list = [
-            image_processor([img], image_aspect_ratio="anyres")["pixel_values"].to(
-                model.device, dtype=torch.float16
-            )
-            for img in images
-        ]
-
-        inputs = {"pixel_values": [image_list]}
-    else:
-        image_tensor = [image_processor([img])["pixel_values"].to(model.device, dtype=torch.float16) for img in images]
-
-        for i in range(0, 8):
-            image_tensor[i] = torch.zeros([1, 1, 1, 3, 384, 384], device=model.device, dtype=torch.float16)
-        image_tensor = torch.stack(image_tensor, dim=1)
-        image_tensor = image_tensor.squeeze(2)
-        inputs = {"pixel_values": image_tensor}
+    image_tensor = [image_processor([img])["pixel_values"].to(model.device, dtype=torch.float16) for img in images]
+
+    image_tensor = torch.stack(image_tensor, dim=1)
+    image_tensor = image_tensor.squeeze(2)
+    inputs = {"pixel_values": image_tensor}
 
     full_conv = "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.<|end|>\n"
     for msg in messages:
@@ -123,12 +111,6 @@ import torch
 
 your_checkpoint_path = ""
 sd = torch.load(your_checkpoint_path)
-
-sd = sd["model_state_dict"]
-for k, v in list(sd.items()):
-    sd["vlm." + k] = v
-    del sd[k]
-
 model.load_state_dict(sd)
 
 # %%