push

Dockerfile

@@ -46,16 +46,7 @@ RUN python -c "from transformers import pipeline; pipeline('text-to-speech', mod
  && 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')"
 
-# Pre-load N-ATLaS model during build
-RUN python -c "import os; token = os.getenv('HF_TOKEN'); print(f'HF_TOKEN available: {bool(token)}'); print(f'Token length: {len(token) if token else 0}'); from huggingface_hub import snapshot_download; snapshot_download(repo_id='NCAIR1/N-ATLaS', token=token)" \
- && python -c "import os; from transformers import AutoTokenizer, AutoModelForCausalLM; import torch; token = os.getenv('HF_TOKEN'); tokenizer = AutoTokenizer.from_pretrained('NCAIR1/N-ATLaS', token=token); model = AutoModelForCausalLM.from_pretrained('NCAIR1/N-ATLaS', torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, device_map='auto' if torch.cuda.is_available() else None, trust_remote_code=True, low_cpu_mem_usage=True, use_cache=True, token=token); print('N-ATLaS model loaded successfully')"
-
-# Pre-download ASR models (will be lazy-loaded at runtime)
-RUN python -c "import os; token = os.getenv('HF_TOKEN'); print(f'ASR download - HF_TOKEN available: {bool(token)}'); from huggingface_hub import snapshot_download; snapshot_download(repo_id='NCAIR1/Hausa-ASR', token=token)" \
- && python -c "import os; token = os.getenv('HF_TOKEN'); from huggingface_hub import snapshot_download; snapshot_download(repo_id='NCAIR1/Yoruba-ASR', token=token)" \
- && python -c "import os; token = os.getenv('HF_TOKEN'); from huggingface_hub import snapshot_download; snapshot_download(repo_id='NCAIR1/Igbo-ASR', token=token)" \
- && python -c "import os; token = os.getenv('HF_TOKEN'); from huggingface_hub import snapshot_download; snapshot_download(repo_id='NCAIR1/NigerianAccentedEnglish', token=token)" \
- && find /models/huggingface -name '*.lock' -delete
+# NCAIR1 models will be downloaded at runtime when HF_TOKEN is available
 
 # Copy project files
 COPY . .

app.py

@@ -214,7 +214,7 @@ def _load_natlas():
     
     try:
         logger.info("Loading N-ATLaS language identification model...")
-        logger.info("Model files are pre-cached from Docker build, loading should be faster...")
+        logger.info("Downloading model files from Hugging Face...")
         
         natlas_tokenizer = AutoTokenizer.from_pretrained("NCAIR1/N-ATLaS", token=hf_token)
         natlas_model = AutoModelForCausalLM.from_pretrained(