fix: add offload folder for Hugging Face dispatch

model/load_model.py

@@ -5,34 +5,33 @@ import torch
 from PIL import ImageDraw
 from transformers import PaliGemmaProcessor, PaliGemmaForConditionalGeneration
 
-# --- Cấu hình model ---
+# ==== CẤU HÌNH ====
 pretrained_model_id = "google/paligemma2-3b-pt-224"
 finetuned_model_id = "pyimagesearch/brain-tumor-od-finetuned-paligemma2"
 token = os.environ.get("HUGGINGFACE_TOKEN")
 
-# --- Tạo thư mục offload nếu chưa tồn tại ---
-offload_folder = "./offload"
-os.makedirs(offload_folder, exist_ok=True)
+# ==== TẠO OFFLOAD FOLDER (fix lỗi Hugging Face không tự tạo) ====
+os.makedirs("./offload", exist_ok=True)
 
-# --- Load processor & model ---
+# ==== LOAD MODEL & PROCESSOR ====
 processor = PaliGemmaProcessor.from_pretrained(pretrained_model_id, token=token)
 model = PaliGemmaForConditionalGeneration.from_pretrained(
     finetuned_model_id,
     device_map="auto",
     low_cpu_mem_usage=True,
-    offload_folder=offload_folder,
+    offload_folder="./offload",  # ✅ Bắt buộc để tránh lỗi dispatch_model
     token=token
 )
 model.eval()
 
-# --- Xóa bộ nhớ GPU/CPU ---
+# ==== DỌN BỘ NHỚ (nếu có GPU) ====
 def clear_memory():
     gc.collect()
     if torch.cuda.is_available():
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()
 
-# --- Tách toạ độ từ output của model ---
+# ==== PARSE OUTPUT RA BOUNDING BOX ====
 def parse_multiple_locations(decoded_output):
     loc_pattern = r"<loc(\d{4})><loc(\d{4})><loc(\d{4})><loc(\d{4})>\s+([^;]+)"
     matches = re.findall(loc_pattern, decoded_output)
@@ -49,17 +48,17 @@ def parse_multiple_locations(decoded_output):
         })
     return coords_and_labels
 
-# --- Vẽ bounding boxes lên ảnh ---
+# ==== VẼ BOX LÊN ẢNH ====
 def draw_boxes(image, coords_and_labels):
     draw = ImageDraw.Draw(image)
     width, height = image.size
     for obj in coords_and_labels:
-        y1, x1, y2, x2 = [int(v * s) for v, s in zip(obj['bbox'], [height, width, height, width])]
+        y1, x1, y2, x2 = obj['bbox'][0]*height, obj['bbox'][1]*width, obj['bbox'][2]*height, obj['bbox'][3]*width
         draw.rectangle([x1, y1, x2, y2], outline="red", width=3)
         draw.text((x1, y1), obj['label'], fill="red")
     return image
 
-# --- Hàm xử lý chính để inference ---
+# ==== HÀM CHÍNH GỌI MODEL ====
 def process_image(image, prompt="detect yes"):
     if not prompt.startswith("<image>"):
         prompt = "<image>" + prompt