Adding initial eval code

app.py

@@ -53,6 +53,7 @@ def fused_sim(a:Image.Image,b:Image.Image,α=.5):
 bnb_cfg = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_use_double_quant=True)
 
 # ---------- load models once at startup ---------------------
+base = None
 @spaces.GPU
 def load_models():
     from unsloth import FastLanguageModel
@@ -85,21 +86,7 @@ def build_prompt(desc:str):
 @spaces.GPU
 @torch.no_grad()
 def draw(model, desc:str):
-    # ensure_models()
-    from unsloth import FastLanguageModel
-    global base, tok, lora
-    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")
+    ensure_models()
     prompt = build_prompt(desc)
     ids = tok(prompt, return_tensors="pt").to(DEVICE)
     out = model.generate(**ids, max_new_tokens=MAX_NEW,
@@ -111,6 +98,7 @@ def draw(model, desc:str):
 
 # ---------- gradio interface --------------------------------
 def compare(desc):
+    ensure_models()
     img_base, svg_base = draw(base, desc)
     img_lora, svg_lora = draw(lora, desc)
     # sim = (fused_sim(img_lora, img_base) if img_base and img_lora else float("nan"))