Adding initial eval code

app.py

@@ -50,26 +50,31 @@ def fused_sim(a:Image.Image,b:Image.Image,α=.5):
     lp_sim = 1 - _LP(ta,tb,normalize=True).item()
     return α*clip_sim + (1-α)*lp_sim
 
+bnb_cfg = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_use_double_quant=True)
+
 # ---------- load models once at startup ---------------------
 @spaces.GPU
 def load_models():
     from unsloth import FastLanguageModel
     global base, tok, lora
-    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")
-    print("✔ models loaded")
+    if base is None:
+        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")
+        print("✔ models loaded")
 
+@spaces.GPU
+def ensure_models():
+    load_models()
+    return True           # small, pickle-able sentinel
 
-load_models()
 
 def build_prompt(desc:str):
     msgs=[{"role":"system","content":"You are an SVG illustrator."},
@@ -80,6 +85,7 @@ def build_prompt(desc:str):
 @spaces.GPU
 @torch.no_grad()
 def draw(model, desc:str):
+    ensure_models()
     prompt = build_prompt(desc)
     ids = tok(prompt, return_tensors="pt").to(DEVICE)
     out = model.generate(**ids, max_new_tokens=MAX_NEW,