Non-root + writable caches, NLTK punkt_tab, cached models, timeout=300

Dockerfile

@@ -1,17 +1,20 @@
 FROM python:3.12-slim
 
-ENV DEBIAN_FRONTEND=noninteractive PIP_NO_CACHE_DIR=1
+ENV DEBIAN_FRONTEND=noninteractive \
+    PIP_NO_CACHE_DIR=1 \
+    PYTHONUNBUFFERED=1
 
-# Create non-root user
+# Non-root user so caches are writable at runtime
 RUN useradd -m -u 1000 appuser
 ENV PATH="/home/appuser/.local/bin:$PATH"
 
 WORKDIR /app
 
-# Minimal OS deps
-RUN apt-get update && apt-get install -y --no-install-recommends libgomp1 && rm -rf /var/lib/apt/lists/*
+# Minimal OS deps (OpenMP runtime)
+RUN apt-get update && apt-get install -y --no-install-recommends libgomp1 \
+ && rm -rf /var/lib/apt/lists/*
 
-# Put ALL caches in the user's home; also set Python path
+# Put ALL caches in the user's home; plus perf/env knobs
 ENV HF_HOME=/home/appuser/.cache \
     TRANSFORMERS_CACHE=/home/appuser/.cache/transformers \
     HUGGINGFACE_HUB_CACHE=/home/appuser/.cache/huggingface \
@@ -21,18 +24,19 @@ ENV HF_HOME=/home/appuser/.cache \
     OMP_NUM_THREADS=1 \
     OPENBLAS_NUM_THREADS=1 \
     MKL_NUM_THREADS=1 \
+    HF_HUB_ENABLE_HF_TRANSFER=1 \
     PYTHONPATH=/app/backend:$PYTHONPATH
 
 # Copy code with correct ownership
 COPY --chown=appuser:appuser . .
 
-# Switch to non-root BEFORE installs so any caches/dirs are writable
+# Switch to non-root BEFORE installs so caches land in /home/appuser
 USER appuser
 
-# Python deps (install to user site)
+# Python deps
 RUN pip install --no-cache-dir --user -r requirements.txt
 
-# Preload models/data so first boot doesn’t timeout
+# Preload NLP data
 RUN python -m spacy download en_core_web_sm
 RUN python - <<'PY'
 import nltk
@@ -41,15 +45,19 @@ for pkg in ["punkt","punkt_tab","wordnet","omw-1.4"]:
 print("NLTK OK")
 PY
 
-# (Optional) prefetch light-weight bits; safe to skip
-# RUN python - <<'PY'
-# from sentence_transformers import SentenceTransformer
-# SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
-# from transformers import AutoTokenizer
-# AutoTokenizer.from_pretrained("oliverguhr/fullstop-punctuation-multilang-large")
-# PY
+# Pre-fetch models to avoid cold-start timeouts
+RUN python - <<'PY'
+from sentence_transformers import SentenceTransformer
+from transformers import AutoTokenizer, AutoModelForTokenClassification
+# semantic encoder
+SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+# punctuation model
+AutoTokenizer.from_pretrained("oliverguhr/fullstop-punctuation-multilang-large")
+AutoModelForTokenClassification.from_pretrained("oliverguhr/fullstop-punctuation-multilang-large")
+print("HF models cached")
+PY
 
 EXPOSE 7860
 
-# Extra time for cold-start model downloads
-CMD ["gunicorn","-w","1","-k","gthread","--threads","4","--timeout","180","-b","0.0.0.0:7860","backend.app:app"]
+# Give the worker time for initial GPU warmup
+CMD ["gunicorn","-w","1","-k","gthread","--threads","4","--timeout","300","-b","0.0.0.0:7860","backend.app:app"]