ZeroGPU: force cuda placement via env; use transformers 5.x dtype arg

app.py

@@ -38,7 +38,13 @@ LAYOUT_MODELS = ["PP-DocLayoutV2", "PP-DocLayout_plus-L", "PP-DocLayout-L"]
 RENDER_DPI = 150          # PDF rasterization DPI
 MAX_SIDE = 2200           # cap the long side of a page (px)
 
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+# On ZeroGPU the GPU is not attached at import time, so torch.cuda.is_available()
+# is False here; the `spaces` package intercepts `.to("cuda")` at module scope and
+# streams the weights in on the first @spaces.GPU call. So trust the ZeroGPU env
+# var, not is_available(), when deciding to place weights on cuda.
+ON_ZEROGPU = bool(os.environ.get("SPACES_ZERO_GPU"))
+USE_CUDA = ON_ZEROGPU or torch.cuda.is_available()
+DEVICE = "cuda" if USE_CUDA else "cpu"
 
 # Prompt per PaddleOCR-VL task.
 PROMPTS = {
@@ -84,11 +90,18 @@ def _load_vlm():
     from transformers import AutoModelForCausalLM, AutoProcessor
 
     print("[vlm] loading", MODEL_PATH, flush=True)
-    model = AutoModelForCausalLM.from_pretrained(
-        MODEL_PATH,
-        trust_remote_code=True,
-        torch_dtype=torch.bfloat16,
-    ).eval()
+    try:
+        model = AutoModelForCausalLM.from_pretrained(
+            MODEL_PATH,
+            trust_remote_code=True,
+            dtype=torch.bfloat16,  # transformers 5.x name (was torch_dtype)
+        ).eval()
+    except TypeError:
+        model = AutoModelForCausalLM.from_pretrained(
+            MODEL_PATH,
+            trust_remote_code=True,
+            torch_dtype=torch.bfloat16,
+        ).eval()
     if DEVICE == "cuda":
         model = model.to("cuda")
     processor = AutoProcessor.from_pretrained(MODEL_PATH, trust_remote_code=True)