Upload 2 files

Dockerfile

@@ -21,11 +21,9 @@ RUN pip install --no-cache-dir \
         pydantic==2.7.1
 
 # Install llama-cpp-python CPU-only (compiled from source)
-# 0.3.4: new enough to read current GGUF format, stable enough to avoid
-# the malloc/GGML_ASSERT corruption seen in 0.3.9 on CPU-only builds
-RUN CMAKE_ARGS="-DGGML_CUDA=OFF -DGGML_METAL=OFF -DGGML_OPENCL=OFF" \
+RUN CMAKE_ARGS="-DLLAMA_CUBLAS=OFF -DLLAMA_METAL=OFF -DLLAMA_OPENCL=OFF" \
     FORCE_CMAKE=1 \
-    pip install --no-cache-dir llama-cpp-python==0.3.4
+    pip install --no-cache-dir llama-cpp-python==0.2.77
 
 # ── App code ─────────────────────────────────────────────────
 COPY app.py .
@@ -39,7 +37,7 @@ COPY app.py .
 RUN mkdir -p /models
 
 # ── Runtime env defaults (override with -e or docker-compose) ─
-ENV MODEL_PATH=/models/qwen3-14b-q4_k_m.gguf \
+ENV MODEL_PATH=/data/qwen3-14b-q4_k_m.gguf \
     MODEL_URL=https://huggingface.co/bartowski/Qwen_Qwen3-14B-GGUF/resolve/main/Qwen_Qwen3-14B-Q4_K_M.gguf \
     MODEL_ID=qwen3-14b \
     N_CTX=4096 \

docker-compose.yml

@@ -8,12 +8,10 @@ services:
     ports:
       - "8000:8000"
     volumes:
-      # Named volume — model is downloaded here on first boot and reused after
-      - qwen3_models:/models
-      # Alternative: mount your own pre-downloaded folder instead:
-      # - ./models:/models
+      # /data is the HF Spaces persistent storage bucket — model lives here
+      - /data:/data
     environment:
-      - MODEL_PATH=/models/qwen3-14b-q4_k_m.gguf
+      - MODEL_PATH=/data/qwen3-14b-q4_k_m.gguf
       - MODEL_URL=https://huggingface.co/bartowski/Qwen_Qwen3-14B-GGUF/resolve/main/Qwen_Qwen3-14B-Q4_K_M.gguf
       - MODEL_ID=qwen3-14b
       - N_CTX=4096
@@ -22,6 +20,3 @@ services:
       - VERBOSE=false
       - HF_TOKEN
     restart: unless-stopped
-
-volumes:
-  qwen3_models: