Defer model loading into first GPU call , fixes ZeroGPU adapter load crash

app.py

@@ -13,18 +13,23 @@ LANG_META = {
     "JavaScript": {"repo": "AmareshHebbar/leetcode-javascript-qwen25-coder-7b", "icon": "🟨", "code_lang": "javascript"},
 }
 
-tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
-base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16)
+tokenizer = None
+model = None
+_on_gpu = False
 
-model = PeftModel.from_pretrained(
-    base_model, LANG_META["Python"]["repo"], adapter_name="Python", device_map="cpu"
-)
-for lang, meta in LANG_META.items():
-    if lang != "Python":
-        model.load_adapter(meta["repo"], adapter_name=lang, device_map="cpu")
-model.eval()
 
-_on_gpu = False
+def _ensure_loaded():
+    global tokenizer, model
+    if model is not None:
+        return
+    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
+    base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16)
+    m = PeftModel.from_pretrained(base_model, LANG_META["Python"]["repo"], adapter_name="Python")
+    for lang, meta in LANG_META.items():
+        if lang != "Python":
+            m.load_adapter(meta["repo"], adapter_name=lang)
+    m.eval()
+    model = m
 
 
 def _generate(inputs, use_cuda):
@@ -125,12 +130,13 @@ THEME = gr.themes.Base(
 )
 
 
-@spaces.GPU(duration=60)
+@spaces.GPU(duration=120)
 def solve(problem, algorithm_tag, language):
     global _on_gpu
     if not problem.strip():
         return "", "Enter a problem statement first."
 
+    _ensure_loaded()
     model.set_adapter(language)
     lang_name = language
     system_prompt = (