gated model

Dockerfile

@@ -33,16 +33,14 @@ ENV HF_HOME=/models/huggingface \
 # Created cache dir and set permissions
 RUN mkdir -p /models/huggingface && chmod -R 777 /models/huggingface
 
-# Pre-download models at build time (split into smaller chunks to avoid memory issues)
+# Pre-download models at build time (only non-gated models)
 RUN python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='facebook/mms-tts-hau')" \
  && python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='facebook/mms-tts-eng')" \
  && python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='facebook/mms-tts-yor')" \
- && python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='NCAIR1/Igbo-ASR')" \
  && find /models/huggingface -name '*.lock' -delete
 
-# Preload tokenizers (avoid runtime delays) - MMS will be lazy-loaded
-RUN python -c "from transformers import WhisperProcessor; WhisperProcessor.from_pretrained('NCAIR1/Igbo-ASR')" \
- && python -c "from transformers import pipeline; pipeline('text-to-speech', model='facebook/mms-tts-hau')" \
+# Preload tokenizers (avoid runtime delays) - MMS and Igbo ASR will be lazy-loaded
+RUN python -c "from transformers import pipeline; pipeline('text-to-speech', model='facebook/mms-tts-hau')" \
  && python -c "from transformers import pipeline; pipeline('text-to-speech', model='facebook/mms-tts-eng')" \
  && python -c "from transformers import pipeline; pipeline('text-to-speech', model='facebook/mms-tts-yor')"
 

app.py

@@ -102,17 +102,21 @@ def _get_igbo_asr():
         return igbo_model, igbo_processor
     
     hf_token = os.getenv("HF_TOKEN")
+    if not hf_token:
+        logger.warning("HF_TOKEN not set - Igbo ASR model requires authentication")
+        return None, None
+    
     try:
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        logger.info("Lazy-loading Igbo ASR model...")
+        logger.info("Lazy-loading Igbo ASR model (gated model)...")
         igbo_processor = WhisperProcessor.from_pretrained("NCAIR1/Igbo-ASR", token=hf_token)
         igbo_model = WhisperForConditionalGeneration.from_pretrained("NCAIR1/Igbo-ASR", token=hf_token)
         igbo_model.to(device)
         igbo_model.eval()
         logger.info("Loaded Igbo ASR model")
         return igbo_model, igbo_processor
-    except Exception:
-        logger.exception("Failed to load Igbo ASR model")
+    except Exception as e:
+        logger.exception(f"Failed to load Igbo ASR model: {e}")
         igbo_model, igbo_processor = None, None
 def _run_whisper(model: WhisperForConditionalGeneration, proc: WhisperProcessor, audio_array: np.ndarray) -> str:
     try: