Upload app.py with huggingface_hub

app.py

@@ -13,52 +13,30 @@ MODEL_REPO = "sakamakismile/gemma-4-12B-coder-fable5-composer2.5-GGUF"
 MODEL_FILE = "gemma-4-12B-coder-fable5-composer2.5-Q4_K_M.gguf"
 MODEL_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), MODEL_FILE)
 
-_llm = None
-_downloaded = threading.Event()
-_download_error = None
-
-def download_model():
-    global _download_error
-    if os.path.exists(MODEL_PATH):
-        log.info("Model already downloaded")
-        _downloaded.set()
-        return
-    log.info("Downloading model (7.38 GB)...")
-    try:
+def ensure_model():
+    if not os.path.exists(MODEL_PATH):
+        log.info("Downloading model (7.38 GB)...")
         hf_hub_download(
             repo_id=MODEL_REPO,
             filename=MODEL_FILE,
             local_dir=os.path.dirname(os.path.abspath(__file__)),
         )
         log.info("Download complete")
-    except Exception as e:
-        _download_error = str(e)
-        log.error(f"Download failed: {e}")
-    finally:
-        _downloaded.set()
 
-threading.Thread(target=download_model, daemon=True).start()
+threading.Thread(target=ensure_model, daemon=True).start()
 
-@spaces.GPU
-def cuda_test():
-    import torch
-    return {
-        "cuda_available": torch.cuda.is_available(),
-        "device_count": torch.cuda.device_count(),
-        "device_name": torch.cuda.get_device_name(0) if torch.cuda.is_available() else "N/A",
-    }
+_llm = None
+_llm_lock = threading.Lock()
 
-@spaces.GPU
-def generate(messages, max_tokens=1024, temperature=0.7, top_p=0.95):
+def load_model():
     global _llm
-    _downloaded.wait()
-
-    if _download_error:
-        raise RuntimeError(f"Model download failed: {_download_error}")
-
-    if _llm is None:
-        from llama_cpp import Llama as _Llama
+    if _llm is not None:
+        return _llm
+    with _llm_lock:
+        if _llm is not None:
+            return _llm
         log.info("Loading model into GPU...")
+        from llama_cpp import Llama as _Llama
         _llm = _Llama(
             model_path=MODEL_PATH,
             n_gpu_layers=-1,
@@ -66,9 +44,22 @@ def generate(messages, max_tokens=1024, temperature=0.7, top_p=0.95):
             verbose=False
         )
         log.info("Model loaded")
+        return _llm
 
+@spaces.GPU
+def cuda_test():
+    import torch
+    return {
+        "cuda_available": torch.cuda.is_available(),
+        "device_count": torch.cuda.device_count(),
+        "device_name": torch.cuda.get_device_name(0) if torch.cuda.is_available() else "N/A",
+    }
+
+@spaces.GPU
+def generate(messages, max_tokens=1024, temperature=0.7, top_p=0.95):
+    llm = load_model()
     log.info(f"Generating (max_tokens={max_tokens}, temp={temperature})")
-    output = _llm.create_chat_completion(
+    output = llm.create_chat_completion(
         messages=messages,
         max_tokens=max_tokens,
         temperature=temperature,
@@ -78,6 +69,8 @@ def generate(messages, max_tokens=1024, temperature=0.7, top_p=0.95):
 
 def predict(message, history):
     try:
+        if not os.path.exists(MODEL_PATH):
+            return "Model is still downloading... Please wait ~5 minutes and try again."
         messages = []
         for user_msg, assistant_msg in history:
             messages.append({"role": "user", "content": user_msg})
@@ -85,12 +78,13 @@ def predict(message, history):
         messages.append({"role": "user", "content": message})
         return generate(messages)
     except Exception as e:
-        return f"Error: {e}\n\n{traceback.format_exc()}"
+        return f"Error: {traceback.format_exc()}"
 
 def check_status():
-    dl_status = "downloading" if not _downloaded.is_set() else ("done" if not _download_error else f"failed: {_download_error}")
+    exists = os.path.exists(MODEL_PATH)
+    size = os.path.getsize(MODEL_PATH) if exists else 0
     cuda = cuda_test()
-    return f"Model: {dl_status}\nGPU: {cuda}"
+    return f"Model file: {'exists' if exists else 'missing'} ({size/1e9:.1f} GB)\nGPU: {cuda}"
 
 with gr.Blocks(title="Gemma Coder Zero", theme=gr.themes.Soft()) as demo:
     gr.Markdown("# Gemma 4 12B Coder Zero")