Upload 2 files

Dockerfile

@@ -21,9 +21,9 @@ 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" \
+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==0.3.9
 
 # ── App code ─────────────────────────────────────────────────
 COPY app.py .

requirements.txt

@@ -3,4 +3,4 @@ uvicorn[standard]==0.29.0
 pydantic==2.7.1
 # llama-cpp-python is installed via pip in the Dockerfile
 # with CPU-only build flags (no CUDA / Metal)
-llama-cpp-python==0.2.77
+llama-cpp-python==0.3.9