out of memory fix

Dockerfile

@@ -33,17 +33,15 @@ 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
-RUN python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='facebook/mms-1b-all')" \
- && python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='facebook/mms-tts-hau')" \
+# Pre-download models at build time (split into smaller chunks to avoid memory issues)
+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)
-RUN python -c "from transformers import Wav2Vec2Processor; Wav2Vec2Processor.from_pretrained('facebook/mms-1b-all')" \
- && python -c "from transformers import WhisperProcessor; WhisperProcessor.from_pretrained('NCAIR1/Igbo-ASR')" \
+# 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')" \
  && 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')"