Upload Dockerfile

Dockerfile

@@ -21,20 +21,19 @@ RUN pip install --no-cache-dir \
         pydantic==2.7.1
 
 # Install llama-cpp-python CPU-only (compiled from source)
-RUN CMAKE_ARGS="-DLLAMA_CUBLAS=OFF -DLLAMA_METAL=OFF -DLLAMA_OPENCL=OFF" \
+# 0.3.4: new enough to read current GGUF format, avoids the
+# malloc/GGML_ASSERT corruption crash introduced in 0.3.9
+RUN CMAKE_ARGS="-DGGML_CUDA=OFF -DGGML_METAL=OFF -DGGML_OPENCL=OFF" \
     FORCE_CMAKE=1 \
-    pip install --no-cache-dir llama-cpp-python==0.2.77
+    pip install --no-cache-dir llama-cpp-python==0.3.4
 
 # ── App code ─────────────────────────────────────────────────
 COPY app.py .
 
-# ── Model volume ─────────────────────────────────────────────
-# The app auto-downloads the model on first boot into /models.
-# Mount a named volume here so the download persists across restarts:
-#   docker run -v qwen3_models:/models ...
-# Or pre-populate with your own GGUF:
-#   docker run -v /path/to/models:/models ...
-RUN mkdir -p /models
+# ── Storage ──────────────────────────────────────────────────
+# /data is the HF Spaces persistent storage bucket.
+# Model is downloaded here on first boot and reused on restarts.
+RUN mkdir -p /data
 
 # ── Runtime env defaults (override with -e or docker-compose) ─
 ENV MODEL_PATH=/data/qwen3-14b-q4_k_m.gguf \
@@ -45,10 +44,10 @@ ENV MODEL_PATH=/data/qwen3-14b-q4_k_m.gguf \
     N_BATCH=512 \
     VERBOSE=false
 
-EXPOSE 8000
+EXPOSE 7860
 
 # Health check — /health returns {"ready": true} once the model is loaded
 HEALTHCHECK --interval=30s --timeout=10s --start-period=600s --retries=20 \
-    CMD wget -qO- http://localhost:8000/health | grep -q '"ready": true' || exit 1
+    CMD wget -qO- http://localhost:7860/health | grep -q '"ready": true' || exit 1
 
 CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]