Fix NLTK punkt_tab by setting NLTK_DATA and pre-downloading

Dockerfile

@@ -4,40 +4,54 @@ ENV DEBIAN_FRONTEND=noninteractive PIP_NO_CACHE_DIR=1
 WORKDIR /app
 ENV PYTHONPATH=/app/backend:$PYTHONPATH
 
-# OpenMP runtime
+# OpenMP + caches
 RUN apt-get update && apt-get install -y libgomp1 && rm -rf /var/lib/apt/lists/*
 
-# Writable caches + thread caps + NLTK data path
+# Writable caches + NLTK path + thread caps
 ENV HF_HOME=/app/.cache \
-    TRANSFORMERS_CACHE=/app/.cache/transformers \
     HUGGINGFACE_HUB_CACHE=/app/.cache/huggingface \
     SENTENCE_TRANSFORMERS_HOME=/app/.cache/sentence-transformers \
     XDG_CACHE_HOME=/app/.cache \
     NLTK_DATA=/usr/local/nltk_data \
-    OMP_NUM_THREADS=1 \
-    OPENBLAS_NUM_THREADS=1 \
-    MKL_NUM_THREADS=1 \
-    NUMEXPR_NUM_THREADS=1
-
+    OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 NUMEXPR_NUM_THREADS=1
 RUN mkdir -p /app/.cache "$NLTK_DATA" && chmod -R 777 /app/.cache "$NLTK_DATA"
 
 COPY . .
 
-# Install deps (CUDA 12.1 torch per requirements.txt)
+# Install deps (CUDA 12.1 wheel for Torch is in requirements.txt)
 RUN pip install -r requirements.txt
 
-# Preload spaCy + NLTK data into the path NLTK will search
+# Preload spaCy + NLTK
 RUN python -m spacy download en_core_web_sm
 RUN python - <<'PY'
 import os, nltk
 path = os.environ.get("NLTK_DATA", "/usr/local/nltk_data")
-for pkg in ["punkt", "punkt_tab", "wordnet", "omw-1.4"]:
+for pkg in ("punkt","punkt_tab","wordnet","omw-1.4"):
     nltk.download(pkg, download_dir=path)
-print("Downloaded NLTK data to:", path)
+print("Downloaded NLTK to", path)
+PY
+
+# ---------- PRE-DOWNLOAD MODELS TO AVOID BOOT TIMEOUT ----------
+RUN python - <<'PY'
+import os
+from transformers import AutoTokenizer, AutoModelForTokenClassification
+from sentence_transformers import SentenceTransformer
+
+cache = os.environ.get("HF_HOME", "/app/.cache")
+
+# Oliver Guhr punctuation model
+m_punc = "oliverguhr/fullstop-punctuation-multilang-large"
+AutoTokenizer.from_pretrained(m_punc, cache_dir=cache)
+AutoModelForTokenClassification.from_pretrained(m_punc, cache_dir=cache)
+
+# SBERT encoder
+m_sbert = "sentence-transformers/all-MiniLM-L6-v2"
+SentenceTransformer(m_sbert, cache_folder=os.path.join(cache, "sentence-transformers"))
+print("Pre-downloaded:", m_punc, "and", m_sbert)
 PY
+# ---------------------------------------------------------------
 
-# HF Spaces port
 EXPOSE 7860
 
-# Single worker + a few threads
-CMD ["gunicorn","-w","1","-k","gthread","--threads","4","-b","0.0.0.0:7860","backend.app:app"]
+# Give startup more time (models/GPU init) and keep single worker
+CMD ["gunicorn","-w","1","-k","gthread","--threads","4","--timeout","300","-b","0.0.0.0:7860","backend.app:app"]