Fix Dockerfile: use standalone download script

Dockerfile

@@ -9,27 +9,10 @@ COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
 COPY app.py .
+COPY download_models.py .
 
 # Pre-download commonly used TTS models so first requests are fast.
-# Individual failures are non-fatal.
-RUN python -c "\
-from transformers import VitsModel, AutoTokenizer; \
-models = [ \
-    'facebook/mms-tts-yor', 'facebook/mms-tts-swh', 'facebook/mms-tts-hau', \
-    'facebook/mms-tts-pcm', 'facebook/mms-tts-aka', 'facebook/mms-tts-lug', \
-    'facebook/mms-tts-amh', 'facebook/mms-tts-som', 'facebook/mms-tts-sna', \
-    'khof312/mms-tts-lin', 'facebook/mms-tts-ara', \
-]; \
-ok = 0; \
-for m in models: \
-    try: \
-        VitsModel.from_pretrained(m); AutoTokenizer.from_pretrained(m); \
-        print(f'  OK {m}'); ok += 1 \
-    except Exception as e: \
-        print(f'  SKIP {m}: {e}') \
-; \
-print(f'{ok}/{len(models)} models cached') \
-"
+RUN python download_models.py
 
 EXPOSE 7860
 

download_models.py

@@ -0,0 +1,37 @@
+"""Pre-download MMS-TTS VITS models at Docker build time.
+
+This runs during `docker build` so that models are baked into the image.
+At runtime, from_pretrained() loads from the local cache (fast) instead
+of downloading from HuggingFace (slow cold-start).
+
+Individual failures are non-fatal — the build continues with whatever
+models succeed.
+"""
+
+from transformers import VitsModel, AutoTokenizer
+
+MODELS = [
+    "facebook/mms-tts-yor",   # Yorùbá
+    "facebook/mms-tts-swh",   # Swahili
+    "facebook/mms-tts-hau",   # Hausa
+    "facebook/mms-tts-pcm",   # Pidgin
+    "facebook/mms-tts-aka",   # Twi (Akan)
+    "facebook/mms-tts-lug",   # Luganda
+    "facebook/mms-tts-amh",   # Amharic
+    "facebook/mms-tts-som",   # Somali
+    "facebook/mms-tts-sna",   # Shona
+    "khof312/mms-tts-lin",    # Lingala
+    "facebook/mms-tts-ara",   # Arabic
+]
+
+ok = 0
+for m in MODELS:
+    try:
+        VitsModel.from_pretrained(m)
+        AutoTokenizer.from_pretrained(m)
+        print(f"  OK {m}")
+        ok += 1
+    except Exception as e:
+        print(f"  SKIP {m}: {e}")
+
+print(f"{ok}/{len(MODELS)} models cached")