Address Codex review after upstream sync

Author: romgenie

common/common.h

@@ -617,7 +617,11 @@ struct common_params {
     std::map<std::string, std::string> default_template_kwargs;
 
     // UI configs
-    bool ui = true;
+#ifdef LLAMA_UI_DEFAULT_ENABLED
+    bool ui = LLAMA_UI_DEFAULT_ENABLED != 0;
+#else
+    bool ui = false;
+#endif
 
     // Deprecated: use ui, ui_mcp_proxy, ui_config_json instead
     bool webui = ui;

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -713,6 +713,8 @@ struct ggml_backend_opencl_context {
     }
 
     void write_profiling_info() {
+        flush_profiling_batch();
+
         if (profiling_results.empty()) {
             return;
         }

scripts/ui-assets.cmake

@@ -1,9 +1,9 @@
 # Provision UI assets and generate ui.cpp/ui.h.
 #
 # Asset provisioning priority:
-#   1. Pre-built assets in SRC_DIST_DIR (manually built by user)
+#   1. If BUILD_UI=ON: pre-built assets in SRC_DIST_DIR (manually built by user)
 #   2. If BUILD_UI=ON: npm build
-#   3. If above did not produce assets and HF_ENABLED=ON: HF Bucket download
+#   3. If BUILD_UI=ON, above did not produce assets, and HF_ENABLED=ON: HF Bucket download
 
 cmake_minimum_required(VERSION 3.16)
 
@@ -257,7 +257,7 @@ function(emit_files)
 
     if("${LLAMA_UI_EMBED}" STREQUAL "")
         set(h "#pragma once\n\n#include <stddef.h>\n\n")
-        if(present)
+        if(BUILD_UI AND present)
             string(APPEND h "#define LLAMA_UI_HAS_ASSETS 1\n\n")
         endif()
         string(APPEND h
@@ -270,7 +270,7 @@ function(emit_files)
         )
 
         set(cpp "#include \"ui.h\"\n\n#include <string.h>\n\n")
-        if(present)
+        if(BUILD_UI AND present)
             set(asset_index 0)
             foreach(asset ${ASSETS})
                 file(READ "${DIST_DIR}/${asset}" asset_hex HEX)
@@ -318,7 +318,7 @@ function(emit_files)
     endif()
 
     set(args "${UI_CPP}" "${UI_H}")
-    if(present)
+    if(BUILD_UI AND present)
         foreach(asset ${ASSETS})
             list(APPEND args "${asset}" "${DIST_DIR}/${asset}")
         endforeach()
@@ -336,10 +336,12 @@ endfunction()
 # ---------------------------------------------------------------------------
 # 1. Priority 1: pre-built assets supplied in tools/ui/dist
 # ---------------------------------------------------------------------------
-copy_src_dist(SRC_OK)
-if(SRC_OK)
-    emit_files()
-    return()
+if(BUILD_UI)
+    copy_src_dist(SRC_OK)
+    if(SRC_OK)
+        emit_files()
+        return()
+    endif()
 endif()
 
 # ---------------------------------------------------------------------------
@@ -355,9 +357,9 @@ if(BUILD_UI)
 endif()
 
 # ---------------------------------------------------------------------------
-# 3. Priority 3: HF Bucket download (if npm did not produce assets and HF_ENABLED=ON)
+# 3. Priority 3: HF Bucket download (if BUILD_UI=ON, npm did not produce assets, and HF_ENABLED=ON)
 # ---------------------------------------------------------------------------
-if(NOT provisioned AND HF_ENABLED)
+if(BUILD_UI AND NOT provisioned AND HF_ENABLED)
     resolve_version(VERSION)
 
     set(stamp_ok FALSE)
@@ -388,7 +390,7 @@ endif()
 # ---------------------------------------------------------------------------
 # 4. Fallback: warn about stale or missing assets, then emit whatever we have
 # ---------------------------------------------------------------------------
-if(NOT provisioned)
+if(BUILD_UI AND NOT provisioned)
     assets_present(have_assets)
     if(have_assets)
         message(WARNING "UI: provisioning failed; embedding stale assets from ${DIST_DIR}")
@@ -399,6 +401,8 @@ if(NOT provisioned)
                         "https://github.com/ggml-org/llama.cpp/releases and "
                         "extract to tools/ui/dist.")
     endif()
+elseif(NOT BUILD_UI)
+    message(STATUS "UI: LLAMA_BUILD_UI=OFF, building without an embedded UI")
 endif()
 
 emit_files()

tools/server/server-context.cpp

@@ -826,33 +826,32 @@ struct server_context_impl {
                 cparams_dft.n_rs_seq = 0;
 
                 std::vector<ggml_backend_dev_t> devs;
+                std::vector<ggml_backend_dev_t> target_devs;
                 uint32_t hp_ngl = 0;
                 uint32_t hp_nct = 0;
                 uint32_t hp_nex = 0;
                 try {
+                    auto mparams_tgt = common_model_params_to_llama(params_base);
+                    auto cparams_tgt = common_context_params_to_llama(params_base);
+                    (void) common_get_device_memory_data(
+                        params_base.model.path.c_str(), &mparams_tgt, &cparams_tgt,
+                        target_devs, hp_ngl, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
+
                     auto dmd = common_get_device_memory_data(
                         params_dft.model.path.c_str(), &mparams_dft, &cparams_dft,
                         devs, hp_ngl, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
 
                     GGML_ASSERT(!params_base.fit_params_target.empty());
                     size_t total = 0;
 
-                    std::vector<ggml_backend_dev_t> tgt_devices = params.devices;
-
-                    if (tgt_devices.empty()) {
-                        for(size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                           tgt_devices.push_back(ggml_backend_dev_get(i));
-                        }
-                    }
-
                     for (size_t j = 0; j < devs.size(); ++j) {
                         const size_t bytes =
                             (measure_model_bytes ? dmd[j].mb.model : 0) +
                             dmd[j].mb.context +
                             dmd[j].mb.compute;
                         total += bytes;
-                        for (size_t i = 0; i < tgt_devices.size(); i++) {
-                            if (tgt_devices[i] == devs[j]) {
+                        for (size_t i = 0; i < target_devs.size(); ++i) {
+                            if (target_devs[i] == devs[j] && i < params_base.fit_params_target.size()) {
                                 SRV_DBG("[spec] adding %.2f MiB to fit_params_target for device %s\n",
                                         bytes / (1024.0 * 1024.0), ggml_backend_dev_name(devs[j]));
                                 params_base.fit_params_target[i] += bytes;

tools/server/server.cpp

@@ -10,6 +10,7 @@
 #include "fit.h"
 #include "llama.h"
 #include "log.h"
+#include "ui.h"
 
 #include <atomic>
 #include <clocale>
@@ -79,6 +80,10 @@ int llama_server(int argc, char ** argv) {
 
     // own arguments required by this example
     common_params params;
+#if defined(LLAMA_UI_HAS_ASSETS)
+    params.ui = true;
+    params.webui = true;
+#endif
 
     common_init();