Fetch unsloth bnb-4bit Gemma + Dramabox checkpoints from HF on startup

Replaces hardcoded local paths (which only existed on the dev machine)
with model_downloader.get_all_paths(). Gemma now resolves to
unsloth/gemma-3-12b-it-bnb-4bit, matching the bnb_4bit=True path.

app.py

@@ -16,11 +16,17 @@ import gradio as gr
 # Local src import.
 sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "src"))
 from inference_server import TTSServer  # noqa: E402
+from model_downloader import get_all_paths  # noqa: E402
 
 
 logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+logging.info("Fetching DramaBox checkpoints from HuggingFace (cached after first run)...")
+paths = get_all_paths()
 logging.info("Loading DramaBox warm server (Gemma + DiT + VAE + Decoder)...")
 tts = TTSServer(
+    checkpoint=paths["transformer"],
+    full_checkpoint=paths["audio_components"],
+    gemma_root=paths["gemma_root"],
     device="cuda",
     dtype=os.environ.get("LTX_DTYPE", "bf16"),
     compile_model=os.environ.get("LTX_COMPILE", "0") == "1",

src/inference_server.py

@@ -65,8 +65,10 @@ class TTSServer:
         self.checkpoint = checkpoint or str(MODELS / "ltx-2.3-22b-dev-audio-only-v13-merged.safetensors")
         self.full_checkpoint = full_checkpoint or os.environ.get(
             "LTX_FULL_CHECKPOINT", "/mnt/persistent0/manmay/models/ltx23/ltx-2.3-22b-dev.safetensors")
-        self.gemma_root = gemma_root or os.environ.get(
-            "GEMMA_DIR", "/mnt/persistent0/manmay/models/gemma-3-12b-it-qat-q4_0-unquantized")
+        if gemma_root is None and not os.environ.get("GEMMA_DIR"):
+            from model_downloader import get_gemma_path
+            gemma_root = get_gemma_path()
+        self.gemma_root = gemma_root or os.environ["GEMMA_DIR"]
         self.device = torch.device(device)
         self.dtype = torch.float16 if dtype == "fp16" else torch.bfloat16
         self.compile_model = compile_model