feat: downloading only needed models

app.py

@@ -83,6 +83,34 @@ from pathlib import Path
 
 logger = logging.getLogger(__name__)
 
+# ─── Auto-download models on first run ────────────────────────
+# Ensures the app works out-of-the-box (e.g. HF Spaces) without
+# baking 26 GB of model weights into git.
+def _auto_download_models():
+    """Download all models if none are present yet."""
+    models_dir = Path(
+        os.environ.get("EUROPALEX_MODELS_DIR", ".local/models")
+    )
+    # Check for GGUF files (llama-cpp-python models). If missing, download
+    # everything — this covers the fresh-install case. Flux safetensors are
+    # not checked separately because if GGUF is missing but safetensors exist,
+    # the user still needs MiniCPM/tiny-aya and a full download is correct.
+    if not any(models_dir.rglob("*.gguf")):
+        print(f"No models found in {models_dir}. Downloading...")
+        try:
+            from models.download_models import download_all
+            download_all(output_dir=str(models_dir))
+        except Exception as e:
+            logger.error("Auto-download failed: %s", e)
+            raise RuntimeError(
+                f"Model download failed. Please run:\n"
+                f"  python -m models.download_models\n"
+                f"Or set EUROPALEX_MODELS_DIR to a directory with model files."
+            ) from e
+
+
+_auto_download_models()
+
 # ─── Phase State ────────────────────────────────────────────────────
 
 _phase1_texts: list[str] = []    # English texts from Phase 1, passed to Phase 2

models/download_models.py

@@ -2,13 +2,16 @@
 
 Usage:
     python -m models.download_models                  # Download all models
-    python -m models.download_models minicpm tiny_aya  # Download specific models
+    python -m models.download_models minicpm tiny_aya flux  # Download specific models
 
 Models:
     minicpm         — MiniCPM5-1B Q8_0 (llama-cpp-python)
     tiny_aya        — tiny-aya-water Q4_K_M (llama-cpp-python)
-    omnivoice       — OmniVoice Q8_0 TTS (omnivoice.cpp, requires base + tokenizer)
     flux            — FLUX.2-klein 4B image gen (diffusers)
+
+Note: OmniVoice TTS is loaded at runtime via
+    omnivoice.OmniVoice.from_pretrained("k2-fsa/OmniVoice")
+and cached in ~/.cache/huggingface/ — no manual download needed.
 """
 
 import argparse
@@ -28,11 +31,6 @@ MODELS = {
         "files": ["tiny-aya-water-q4_k_m.gguf"],
         "description": "tiny-aya-water q4_k_m translation (llama-cpp-python)",
     },
-    "omnivoice": {
-        "repo": "Serveurperso/OmniVoice-GGUF",
-        "files": ["omnivoice-base-Q8_0.gguf", "omnivoice-tokenizer-Q8_0.gguf"],
-        "description": "OmniVoice Q8_0 TTS (base + tokenizer, omnivoice.cpp)",
-    },
     "flux": {
         "repo": "black-forest-labs/FLUX.2-klein-4B",
         "files": None,  # Download all files — safetensors weights + configs (~10–12 GB)
@@ -66,6 +64,23 @@ def download_model(name: str, target_dir: Path) -> None:
     print(f"  ✓ Done — {output_dir}\n")
 
 
+def download_all(output_dir: Path | str = ".local/models") -> None:
+    """Download all models from HF Hub. Convenience wrapper around download_model."""
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    errors = []
+    for name in MODELS:
+        try:
+            download_model(name, output_dir)
+        except Exception as e:
+            errors.append((name, str(e)))
+            print(f"  ✗ {name} failed: {e}\n")
+
+    if errors:
+        raise RuntimeError(f"{len(errors)} model(s) failed to download: {[n for n, _ in errors]}")
+
+
 def main():
     parser = argparse.ArgumentParser(description="Download models from HF Hub")
     parser.add_argument(