Package portable Linux runtime and verify Google CPU search

crates/searchi-semantic/native/CMakeLists.txt

@@ -4,11 +4,11 @@ if(NOT DEFINED LLAMA_CPP_DIR OR NOT EXISTS "${LLAMA_CPP_DIR}/include/llama.h")
     message(FATAL_ERROR "Supply -DLLAMA_CPP_DIR=<pinned patched llama.cpp b11461 source>")
 endif()
 set(CMAKE_CXX_STANDARD 17)
-# CPU-only builds remain small internal VM tools. Normal Windows x64 builds
+# CPU-only builds remain small internal VM tools. Normal Windows/Linux x64 builds
 # ship baseline and optimized CPU modules in the same GPU-capable package.
 set(SEARCHI_ACCELERATION "auto" CACHE STRING "auto; explicit developer overrides: cpu, metal, vulkan, cuda")
 set(SEARCHI_CPU_DISPATCH OFF)
-if(WIN32 AND CMAKE_SIZEOF_VOID_P EQUAL 8 AND
+if((WIN32 OR (CMAKE_SYSTEM_NAME STREQUAL "Linux" AND SEARCHI_ACCELERATION MATCHES "^(auto|vulkan)$")) AND CMAKE_SIZEOF_VOID_P EQUAL 8 AND
    (CMAKE_SYSTEM_PROCESSOR MATCHES "^(AMD64|amd64|x86_64|X86_64)$" OR CMAKE_GENERATOR_PLATFORM STREQUAL "x64") AND
    NOT SEARCHI_ACCELERATION STREQUAL "cpu")
     set(SEARCHI_CPU_DISPATCH ON)
@@ -22,12 +22,18 @@ if(MSVC)
         set(CMAKE_MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>" CACHE STRING "Internal static worker CRT" FORCE)
     endif()
 endif()
-set(BUILD_SHARED_LIBS ${SEARCHI_CPU_DISPATCH} CACHE BOOL "Shared runtime for Windows CPU dispatch" FORCE)
+set(BUILD_SHARED_LIBS ${SEARCHI_CPU_DISPATCH} CACHE BOOL "Shared runtime for portable CPU dispatch" FORCE)
 set(GGML_BACKEND_DL ${SEARCHI_CPU_DISPATCH} CACHE BOOL "Load compatible packaged CPU backend" FORCE)
 set(GGML_CPU_ALL_VARIANTS ${SEARCHI_CPU_DISPATCH} CACHE BOOL "Baseline and optimized CPU variants" FORCE)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}")
+if(CMAKE_SYSTEM_NAME STREQUAL "Linux" AND SEARCHI_CPU_DISPATCH)
+    # Adjacent packaged libraries, independent of the build tree or cwd.
+    set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
+    set(CMAKE_INSTALL_RPATH "$ORIGIN")
+    set(CMAKE_INSTALL_RPATH_USE_LINK_PATH OFF)
+endif()
 set(GGML_OPENMP OFF CACHE BOOL "Use bounded native llama thread pools without external OpenMP runtime" FORCE)
-set(GGML_NATIVE OFF CACHE BOOL "Portable CPU kernels")
+set(GGML_NATIVE OFF CACHE BOOL "Portable CPU kernels" FORCE)
 # Normal distribution builds include GPU support and CPU fallback. CPU-only
 # configuration is explicitly reserved for development/testing in the VM.
 set(GGML_CUDA OFF CACHE BOOL "Production CUDA backend")
@@ -62,7 +68,7 @@ set(LLAMA_OPENSSL OFF CACHE BOOL "No network runtime dependencies" FORCE)
 add_subdirectory("${LLAMA_CPP_DIR}" llama-build)
 add_executable(searchi-gemma-worker worker.cpp)
 target_include_directories(searchi-gemma-worker PRIVATE "${LLAMA_CPP_DIR}/vendor")
-target_link_libraries(searchi-gemma-worker PRIVATE llama mtmd)
+target_link_libraries(searchi-gemma-worker PRIVATE llama mtmd ${CMAKE_DL_LIBS})
 if(SEARCHI_CPU_DISPATCH)
     target_compile_definitions(searchi-gemma-worker PRIVATE SEARCHI_CPU_DISPATCH=1)
 endif()

crates/searchi-semantic/native/worker.cpp

@@ -16,10 +16,17 @@
 #include <string>
 #include <vector>
 #ifdef SEARCHI_CPU_DISPATCH
+#include <filesystem>
+#ifdef _WIN32
 #define WIN32_LEAN_AND_MEAN
 #define NOMINMAX
 #include <windows.h>
-#include <filesystem>
+#elif defined(__linux__)
+#include <dlfcn.h>
+#include <unistd.h>
+#else
+#error Unsupported platform for packaged CPU dispatch
+#endif
 #endif
 
 using json = nlohmann::json;
@@ -91,10 +98,19 @@ public:
 #ifdef SEARCHI_CPU_DISPATCH
         // Only load our adjacent modules: no current-directory or environment
         // plugin discovery. CPU-only startup never loads the Vulkan module.
+#ifdef _WIN32
         std::vector<wchar_t> executable(32768);
         DWORD length = GetModuleFileNameW(nullptr, executable.data(), DWORD(executable.size()));
         require(length > 0 && length < executable.size(), "Cannot locate packaged worker backends");
         const auto directory = std::filesystem::path(std::wstring(executable.data(), length)).parent_path().u8string();
+#else
+        // /proc identifies this executable even when launched through PATH or
+        // from another working directory. Reject truncation before discovery.
+        std::vector<char> executable(32768);
+        const auto length = readlink("/proc/self/exe", executable.data(), executable.size());
+        require(length > 0 && size_t(length) < executable.size(), "Cannot locate packaged worker backends");
+        const auto directory = std::filesystem::path(std::string(executable.data(), size_t(length))).parent_path().string();
+#endif
         if (want_gpu) {
             ggml_backend_load_all_from_path(directory.c_str());
         } else {
@@ -103,6 +119,7 @@ public:
         require(ggml_backend_reg_by_name("CPU") != nullptr, "Cannot load compatible packaged CPU backend");
         const auto cpu_reg = ggml_backend_reg_by_name("CPU");
         const auto feature_proc = ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_get_features");
+#ifdef _WIN32
         HMODULE cpu_module = nullptr;
         if (feature_proc && GetModuleHandleExW(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS | GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT,
                 reinterpret_cast<LPCWSTR>(feature_proc), &cpu_module)) {
@@ -112,6 +129,19 @@ public:
             }
         }
         vulkan_module_loaded = GetModuleHandleW(L"ggml-vulkan.dll") != nullptr;
+#else
+        Dl_info cpu_module{};
+        if (feature_proc && dladdr(feature_proc, &cpu_module) && cpu_module.dli_fname) {
+            cpu_backend_module = std::filesystem::path(cpu_module.dli_fname).filename().string();
+        }
+        // RTLD_NOLOAD observes module presence without initiating provider
+        // loading; forced CPU startup must remain independent of Vulkan.
+        const auto vulkan_path = std::filesystem::path(directory) / "libggml-vulkan.so";
+        if (void * module = dlopen(vulkan_path.c_str(), RTLD_NOW | RTLD_NOLOAD)) {
+            vulkan_module_loaded = true;
+            dlclose(module);
+        }
+#endif
 #endif
         llama_backend_init();
         bool use_gpu = want_gpu && llama_supports_gpu_offload();