Unsloth needs a gpu

app.py

@@ -5,7 +5,6 @@ import re, os, torch, cairosvg, lpips, clip, gradio as gr
 from io import BytesIO
 from pathlib import Path
 from PIL import Image
-from unsloth import FastLanguageModel
 from transformers import BitsAndBytesConfig, AutoTokenizer
 import gradio as gr
 import spaces
@@ -51,17 +50,24 @@ def fused_sim(a:Image.Image,b:Image.Image,α=.5):
     return α*clip_sim + (1-α)*lp_sim
 
 # ---------- load models once at startup ---------------------
-bnb_cfg = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_use_double_quant=True)
-print("Loading BASE …")
-base, tok = FastLanguageModel.from_pretrained(
-    BASE_MODEL, max_seq_length=2048,
-    load_in_4bit=True, quantization_config=bnb_cfg, device_map="auto")
-tok.pad_token = tok.eos_token
+@spaces.GPU
+def load_models():
+    from unsloth import FastLanguageModel
+    bnb_cfg = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_use_double_quant=True)
+    print("Loading BASE …")
+    base, tok = FastLanguageModel.from_pretrained(
+        BASE_MODEL, max_seq_length=2048,
+        load_in_4bit=True, quantization_config=bnb_cfg, device_map="auto")
+    tok.pad_token = tok.eos_token
+
+    print("Loading LoRA …")
+    lora, _ = FastLanguageModel.from_pretrained(
+        ADAPTER_DIR, max_seq_length=2048,
+        load_in_4bit=True, quantization_config=bnb_cfg, device_map="auto")
+    return base, tok, lora
+
 
-print("Loading LoRA …")
-lora, _ = FastLanguageModel.from_pretrained(
-    ADAPTER_DIR, max_seq_length=2048,
-    load_in_4bit=True, quantization_config=bnb_cfg, device_map="auto")
+base, tok, lora = load_models()
 
 def build_prompt(desc:str):
     msgs=[{"role":"system","content":"You are an SVG illustrator."},