Fix: Disable LLAMA_CURL in Hugging Face space environment

This PR addresses an issue with the recent change in llama.cpp that sets LLAMA_CURL=ON by default.

The default setting breaks the Hugging Face space environment for gguf-my-repo because libcurl dependencies are not available in the HF space runtime.

This fix explicitly sets LLAMA_CURL=OFF when running in the Hugging Face space environment while maintaining the ability to use CURL when running locally.

Changes:
- Added LLAMA_CURL environment variable control similar to the existing GGML_CUDA control
- Set LLAMA_CURL=OFF by default for safety
- Only enable LLAMA_CURL when NOT running locally

Tested by deploying a duplicate space with these changes applied, which successfully resolves the build failure.

start.sh

@@ -6,13 +6,16 @@ if [ ! -d "llama.cpp" ]; then
 fi
 
 export GGML_CUDA=OFF
+export LLAMA_CURL=OFF
 if [[ -z "${RUN_LOCALLY}" ]]; then
   # enable CUDA if NOT running locally
   export GGML_CUDA=ON
+  # enable CURL if NOT running locally
+  export LLAMA_CURL=ON
 fi
 
 cd llama.cpp
-cmake -B build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=${GGML_CUDA}
+cmake -B build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=${GGML_CUDA} -DLLAMA_CURL=${LLAMA_CURL}
 cmake --build build --config Release -j --target llama-quantize llama-gguf-split llama-imatrix
 cp ./build/bin/llama-* .
 rm -rf build