Deploy PixelLock GPU Space (llama.cpp + GGUF + GBNF + custom UI)

start.sh

@@ -7,14 +7,16 @@ set -uo pipefail
 
 MODEL_PATH="${MODEL_PATH:-/home/user/app/model.gguf}"
 
-# Locate the llama-server binary (path differs across base images).
-LLAMA_BIN="$(command -v llama-server || true)"
-if [ -z "$LLAMA_BIN" ]; then
-  for p in /app/llama-server /llama-server /usr/local/bin/llama-server /usr/bin/llama-server; do
-    [ -x "$p" ] && LLAMA_BIN="$p" && break
-  done
-fi
-echo "[start] using llama-server at: ${LLAMA_BIN:-NOT FOUND}"
+# The base image ships llama-server in /app alongside its shared libs
+# (libllama-server-impl.so, libggml*.so). Its RPATH is relative to the binary, so
+# we MUST run the real /app path (not a symlink) and point the loader at /app.
+export LD_LIBRARY_PATH="/app:${LD_LIBRARY_PATH:-}"
+LLAMA_BIN=""
+for p in /app/llama-server /llama-server /usr/local/bin/llama-server /usr/bin/llama-server; do
+  [ -x "$p" ] && LLAMA_BIN="$p" && break
+done
+[ -z "$LLAMA_BIN" ] && LLAMA_BIN="$(command -v llama-server || true)"
+echo "[start] using llama-server at: ${LLAMA_BIN:-NOT FOUND}  (LD_LIBRARY_PATH=$LD_LIBRARY_PATH)"
 
 serve() {  # $1 = context window
   echo "[start] llama-server -c $1 loading ${MODEL_PATH}"