Upload Dockerfile

Dockerfile

@@ -0,0 +1,85 @@
+# ──────────────────────────────────────────────────────────────────────────
+# Dockerfile – Children's Learning Router Service
+# Target:  Hugging Face Spaces  (CPU-only, Docker SDK)
+# Port:    7860  (required by HF Spaces)
+#
+# Model delivery: via `preload_from_hub` in README.md
+#   HF Spaces downloads Qwen/Qwen2.5-1.5B-Instruct before container start
+#   and places it under /repo-cache (HF_HOME=/repo-cache).
+#   No in-build download is needed or possible (build env has no internet).
+#
+# OOM mitigation: packages are installed in small isolated groups so pip's
+#   dependency resolver never spikes RAM.  --no-cache-dir and --no-compile
+#   keep peak memory low throughout the build.
+# ──────────────────────────────────────────────────────────────────────────
+
+FROM python:3.10-slim
+
+# ── System packages ───────────────────────────────────────────────────────
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        build-essential \
+        git \
+        curl \
+    && rm -rf /var/lib/apt/lists/*
+
+# ── Working directory ─────────────────────────────────────────────────────
+WORKDIR /app
+
+# ── Pip hygiene: upgrade pip/wheel first (small, fast) ───────────────────
+RUN pip install --no-cache-dir --no-compile --upgrade pip wheel
+
+# ── 1 of 5 · CPU-only PyTorch (largest wheel – install alone) ────────────
+RUN pip install --no-cache-dir --no-compile \
+    torch==2.3.1 \
+    --index-url https://download.pytorch.org/whl/cpu
+
+# ── 2 of 5 · HuggingFace stack (transformers pulls in tokenizers etc.) ───
+RUN pip install --no-cache-dir --no-compile \
+    transformers==4.46.3 \
+    accelerate==1.1.1
+
+# ── 3 of 5 · Serialisation libs ──────────────────────────────────────────
+RUN pip install --no-cache-dir --no-compile \
+    sentencepiece==0.2.0 \
+    protobuf==5.28.3
+
+# ── 4 of 5 · Async HTTP client ───────────────────────────────────────────
+RUN pip install --no-cache-dir --no-compile \
+    httpx==0.27.2
+
+# ── 5 of 5 · Web framework + ASGI server ─────────────────────────────────
+RUN pip install --no-cache-dir --no-compile \
+    fastapi==0.115.0 \
+    gunicorn==22.0.0 \
+    uvicorn[standard]==0.30.6
+
+# ── Application code ──────────────────────────────────────────────────────
+COPY app.py .
+
+# ── HuggingFace Spaces: run as non-root user (UID 1000) ──────────────────
+# mkdir -p /repo-cache/hub ensures the cache path exists and is writable
+# by hfuser whether HF Spaces pre-populates it or the model downloads fresh.
+RUN useradd -m -u 1000 hfuser \
+    && mkdir -p /repo-cache/hub \
+    && chown -R hfuser:hfuser /app /repo-cache
+USER hfuser
+
+# ── Runtime config ────────────────────────────────────────────────────────
+# HF Spaces sets HF_HOME=/repo-cache and places preload_from_hub models
+# there before the container starts. HF_HOME alone is sufficient;
+# TRANSFORMERS_CACHE is deprecated since transformers v4 and removed in v5.
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    HF_HOME=/repo-cache
+
+EXPOSE 7860
+
+# ── Start-up command ──────────────────────────────────────────────────────
+# Gunicorn + uvicorn worker serves the FastAPI app.
+# --preload ensures the model is loaded ONCE before workers fork.
+CMD ["gunicorn", "app:app", \
+     "--worker-class", "uvicorn.workers.UvicornWorker", \
+     "--workers",      "1", \
+     "--bind",         "0.0.0.0:7860", \
+     "--timeout",      "300", \
+     "--preload"]