Upload 2 files

Dockerfile

@@ -17,12 +17,17 @@ WORKDIR /app
 COPY requirements.txt .
 
 # ── Python deps (single source of truth: requirements.txt) ───
-# CMAKE_ARGS/FORCE_CMAKE apply to the whole install so llama-cpp-python
-# gets built from source as CPU-only; the other packages are unaffected.
-# See requirements.txt for why the llama-cpp-python version matters.
+# Install llama-cpp-python from abetlen's prebuilt CPU wheel index
+# instead of compiling it from source: building llama.cpp's C++ tree
+# from scratch spawns several parallel compiler processes and was
+# OOMing the build (exit 137) on the platform's build container.
+# CMAKE_ARGS / CMAKE_BUILD_PARALLEL_LEVEL only matter if pip ever has
+# to fall back to a source build (e.g. no matching wheel for this
+# platform yet) -- they keep that fallback CPU-only and memory-bounded.
 RUN CMAKE_ARGS="-DGGML_CUDA=OFF -DGGML_METAL=OFF -DGGML_OPENCL=OFF" \
-    FORCE_CMAKE=1 \
-    pip install --no-cache-dir -r requirements.txt
+    CMAKE_BUILD_PARALLEL_LEVEL=1 \
+    pip install --no-cache-dir -r requirements.txt \
+        --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
 
 # ── App code ─────────────────────────────────────────────────
 COPY app.py .

requirements.txt

@@ -4,5 +4,6 @@ pydantic==2.7.1
 # Must be >=0.3.9 -- that's the release that added Qwen3 architecture
 # support to llama.cpp. Anything older will fail to load a Qwen3 GGUF
 # with "Failed to load model from file".
-# Built from source by the Dockerfile with CPU-only CMAKE_ARGS (no CUDA / Metal).
+# Installed by the Dockerfile from abetlen's prebuilt CPU wheel index
+# (not compiled from source -- compiling timed out/OOMed the build).
 llama-cpp-python==0.3.30