fix(gpu-libs): bundle hipBLASLt TensileLibrary data so ROCm backends stop falling back (#10660)

The ROCm packager copied rocBLAS kernel data (rocblas/library/*.dat) into the
bundled lib/ dir and run.sh pointed ROCBLAS_TENSILE_LIBPATH at it, but the
parallel hipBLASLt data dir (hipblaslt/library/TensileLibrary_lazy_gfx*.dat)
was never packaged and no HIPBLASLT_TENSILE_LIBPATH was set. The bundled
libhipblaslt.so therefore resolved its per-arch kernel data relative to itself,
found nothing, and silently fell back to slow generic kernels, logging:

    rocblaslt error: Cannot read "TensileLibrary_lazy_gfx1201.dat": No such file or directory
    rocblaslt error: Could not load "TensileLibrary_lazy_gfx1201.dat"

Fix, mirroring the existing rocBLAS handling:
- package-gpu-libs.sh: extract the rocblas data-dir copy into a reusable
  copy_rocm_data_dir helper and call it for both rocblas and hipblaslt.
- llama-cpp/turboquant run.sh: export HIPBLASLT_TENSILE_LIBPATH when the
  bundled hipblaslt/library dir exists.

The helper takes an optional ROCM_BASE_DIRS override so the copy is unit
testable without a real ROCm install; add a regression test that runs
package_rocm_libs against a fabricated ROCm tree and asserts both data dirs
are bundled.

Note: this bundles whatever gfx*.dat the build image's ROCm provides. If a
given arch's tensile data is absent from the shipped ROCm, that arch still
needs a ROCm bump; the packaging gap itself is fixed for every supported arch.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-8 [Claude Code]

Dockerfile

@@ -171,6 +171,17 @@ RUN if [ "${BUILD_TYPE}" = "hipblas" ]; then \
     ln -s /opt/rocm-**/lib/llvm/lib/libomp.so /usr/lib/libomp.so \
     ; fi
 
+# ROCm's bundled libdrm_amdgpu is built with a hardcoded fallback lookup path
+# for the ASIC ID table (/opt/amdgpu/share/libdrm/amdgpu.ids), which only exists
+# if AMD's full amdgpu graphics/DKMS stack is installed. This compute-only image
+# doesn't have it, so hipblas/rocBLAS log "No such file or directory" on every
+# model load and can fail to identify the GPU. Point it at the equivalent file
+# Ubuntu's libdrm-common package already ships.
+RUN if [ "${BUILD_TYPE}" = "hipblas" ] && [ -f /usr/share/libdrm/amdgpu.ids ] && [ ! -e /opt/amdgpu/share/libdrm/amdgpu.ids ]; then \
+    mkdir -p /opt/amdgpu/share/libdrm && \
+    ln -s /usr/share/libdrm/amdgpu.ids /opt/amdgpu/share/libdrm/amdgpu.ids \
+    ; fi
+
 RUN expr "${BUILD_TYPE}" = intel && echo "intel" > /run/localai/capability || echo "not intel"
 
 # Cuda

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=f74a6fb87b315b2c3154166e075360e15021a61d
+IK_LLAMA_VERSION?=87fc8701ff4da81a7d2a91ec0695f95eb3066a47
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/CMakeLists.txt

@@ -101,4 +101,13 @@ if(LLAMA_GRPC_BUILD_TESTS)
     target_link_libraries(message_content_test PRIVATE ${_LLAMA_COMMON_TARGET})
     target_compile_features(message_content_test PRIVATE cxx_std_17)
     add_test(NAME message_content_test COMMAND message_content_test)
+
+    # Parent-death watcher test (parent_watch.h) — standard library only, but
+    # needs a threading runtime for std::thread.
+    find_package(Threads REQUIRED)
+    add_executable(parent_watch_test parent_watch_test.cpp parent_watch.h)
+    target_include_directories(parent_watch_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+    target_link_libraries(parent_watch_test PRIVATE Threads::Threads)
+    target_compile_features(parent_watch_test PRIVATE cxx_std_17)
+    add_test(NAME parent_watch_test COMMAND parent_watch_test)
 endif()

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=6f4f53f2b7da54fcdbbecaaa734337c337ad6176
+LLAMA_VERSION?=fdb1db877c526ec90f668eca1b858da5dba85560
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/grpc-server.cpp

@@ -75,6 +75,8 @@
 #include <windows.h>
 #endif
 
+#include "parent_watch.h" // best-effort parent-death backstop (see header)
+
 
 using grpc::Server;
 using grpc::ServerBuilder;
@@ -3442,6 +3444,10 @@ int main(int argc, char** argv) {
     }
   }
 
+    // Best-effort backstop: self-terminate if the LocalAI process that spawned
+    // us dies without cleaning us up (see parent_watch.h).
+    llama_grpc::start_parent_death_watcher();
+
     server_context ctx_server;
     BackendServiceImpl service(ctx_server);
 

backend/cpp/llama-cpp/parent_watch.h

@@ -0,0 +1,179 @@
+// Parent-death watcher (best-effort backstop) for the llama.cpp gRPC backend.
+//
+// LocalAI spawns this backend as a child process and, on a clean shutdown,
+// tears it down itself (SIGTERM -> grace -> SIGKILL). That graceful path only
+// runs when LocalAI receives a catchable signal and lives long enough to run
+// its handlers. If LocalAI is SIGKILLed (e.g. a supervising process's grace
+// period elapses first), that teardown never runs and this backend would be
+// reparented to init and linger, holding VRAM and its listen port.
+//
+// The watcher here is a best-effort backstop for exactly that case: it does
+// NOT replace the graceful teardown, it only covers the "parent vanished
+// without cleaning up" path. It detects reparenting: when the process that
+// spawned this backend dies, the kernel reparents us to the nearest sub-reaper
+// or to init (PID 1), so getppid() stops matching the value captured at
+// startup. This getppid() approach is portable across Linux/macOS (unlike the
+// Linux-only PR_SET_PDEATHSIG), which is why it is used here, mirroring the Go
+// backends' pkg/grpc/parentwatch.go. It is disabled on Windows, which has no
+// equivalent orphan-reparenting semantics.
+//
+// This header is intentionally dependency-free (C++ standard library only) so
+// it can be exercised by a standalone unit test (parent_watch_test.cpp) without
+// building the full llama.cpp + gRPC backend.
+#ifndef LLAMA_GRPC_PARENT_WATCH_H
+#define LLAMA_GRPC_PARENT_WATCH_H
+
+#include <algorithm>
+#include <cctype>
+#include <chrono>
+#include <cstdio>
+#include <cstdlib>
+#include <functional>
+#include <string>
+#include <thread>
+
+#if !defined(_WIN32)
+#include <unistd.h> // getppid(2), _exit(2)
+#endif
+
+namespace llama_grpc {
+
+// Env var names are shared verbatim with the Go and Python backends for
+// consistency across languages.
+inline const char *kEnvParentWatch()         { return "LOCALAI_BACKEND_PARENT_WATCH"; }
+inline const char *kEnvParentWatchInterval() { return "LOCALAI_BACKEND_PARENT_WATCH_INTERVAL"; }
+
+// Default poll interval in milliseconds. Matches the Go side's 2 * time.Second.
+inline long parent_watch_default_interval_ms() { return 2000; }
+
+namespace detail {
+inline std::string trim_lower(const std::string &in, bool lower) {
+    size_t a = in.find_first_not_of(" \t\r\n");
+    size_t b = in.find_last_not_of(" \t\r\n");
+    if (a == std::string::npos) {
+        return "";
+    }
+    std::string s = in.substr(a, b - a + 1);
+    if (lower) {
+        std::transform(s.begin(), s.end(), s.begin(),
+                       [](unsigned char c) { return std::tolower(c); });
+    }
+    return s;
+}
+} // namespace detail
+
+// parent_watch_enabled reports whether the watcher should run. Enabled by
+// default; a falsey value ("false"/"0"/"no"/"off", case-insensitive) disables
+// it, matching the Go implementation's exact semantics.
+inline bool parent_watch_enabled() {
+#if defined(_WIN32)
+    return false;
+#else
+    const char *v = std::getenv(kEnvParentWatch());
+    if (v == nullptr || v[0] == '\0') {
+        return true;
+    }
+    const std::string s = detail::trim_lower(v, true);
+    return !(s == "false" || s == "0" || s == "no" || s == "off");
+#endif
+}
+
+// parent_watch_interval_ms returns the poll interval in milliseconds. Accepts
+// Go-style duration strings ("500ms", "2s", "1m") for cross-language parity, or
+// a bare number interpreted as seconds. Defaults to
+// parent_watch_default_interval_ms().
+inline long parent_watch_interval_ms() {
+    const long def = parent_watch_default_interval_ms();
+    const char *v = std::getenv(kEnvParentWatchInterval());
+    if (v == nullptr || v[0] == '\0') {
+        return def;
+    }
+    const std::string s = detail::trim_lower(v, false);
+    if (s.empty()) {
+        return def;
+    }
+    size_t i = 0;
+    while (i < s.size() && (std::isdigit((unsigned char)s[i]) || s[i] == '.')) {
+        i++;
+    }
+    if (i == 0) {
+        return def;
+    }
+    double num = 0.0;
+    try {
+        num = std::stod(s.substr(0, i));
+    } catch (...) {
+        return def;
+    }
+    const std::string unit = s.substr(i);
+    long ms;
+    if (unit == "ms") {
+        ms = (long)num;
+    } else if (unit == "s" || unit.empty()) {
+        ms = (long)(num * 1000.0);
+    } else if (unit == "m") {
+        ms = (long)(num * 60000.0);
+    } else {
+        return def; // unrecognized unit
+    }
+    return ms > 0 ? ms : def;
+}
+
+#if !defined(_WIN32)
+// parent_died reports whether this process has been reparented away from the
+// parent it had when the watcher started. Reparenting is the standard POSIX
+// signal that the original parent (here, the LocalAI process that spawned this
+// backend) has exited: the orphan is handed to the nearest sub-reaper or to
+// init (PID 1), so getppid() no longer matches the value captured at startup.
+inline bool parent_died(pid_t orig_ppid) {
+    const pid_t ppid = getppid();
+    return ppid != orig_ppid || ppid == 1;
+}
+
+// watch_parent_death polls until parent_died reports the original parent is
+// gone, then invokes on_death. It blocks, so run it on its own thread.
+inline void watch_parent_death(pid_t orig_ppid, long interval_ms,
+                               const std::function<void()> &on_death) {
+    for (;;) {
+        std::this_thread::sleep_for(std::chrono::milliseconds(interval_ms));
+        if (parent_died(orig_ppid)) {
+            on_death();
+            return;
+        }
+    }
+}
+#endif
+
+// start_parent_death_watcher installs the best-effort safety net described in
+// the file header on the calling backend process. It is a no-op when disabled,
+// on Windows, or when the process is already orphaned at startup
+// (getppid() <= 1). This is a backstop alongside — never a replacement for —
+// LocalAI's graceful teardown.
+inline void start_parent_death_watcher() {
+#if !defined(_WIN32)
+    if (!parent_watch_enabled()) {
+        return;
+    }
+    const pid_t orig_ppid = getppid();
+    // A parent of 1 (or less) at startup means we were already orphaned (or
+    // launched directly under init) — there is no original parent to watch for.
+    if (orig_ppid <= 1) {
+        return;
+    }
+    const long interval_ms = parent_watch_interval_ms();
+    std::thread([orig_ppid, interval_ms]() {
+        watch_parent_death(orig_ppid, interval_ms, [orig_ppid]() {
+            fprintf(stderr,
+                    "backend parent process (pid %d) exited without stopping "
+                    "this backend; self-terminating to avoid orphaning\n",
+                    (int)orig_ppid);
+            fflush(stderr);
+            _exit(1);
+        });
+    }).detach();
+#endif
+}
+
+} // namespace llama_grpc
+
+#endif // LLAMA_GRPC_PARENT_WATCH_H

backend/cpp/llama-cpp/parent_watch_test.cpp

@@ -0,0 +1,197 @@
+// Unit tests for the parent-death watcher (parent_watch.h).
+//
+// Build & run standalone (C++ standard library only, no nlohmann/json needed):
+//   g++ -std=c++17 -pthread parent_watch_test.cpp -o t && ./t
+//
+// The core test (TestDetectsReparent) builds a genuine two-level process tree
+// (test -> middle -> grandchild), lets the middle process die, and asserts the
+// grandchild's watch_parent_death detects the reparenting and self-terminates —
+// mirroring the Go test in pkg/grpc/parentwatch_test.go, but with fork(2).
+//
+// On Windows this file compiles to a no-op success (the watcher is unsupported
+// there), matching parent_watch.h's platform gating.
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+
+#include "parent_watch.h"
+
+static int failures = 0;
+
+static void check(bool ok, const std::string &name) {
+    if (!ok) {
+        failures++;
+        fprintf(stderr, "FAIL: %s\n", name.c_str());
+    } else {
+        fprintf(stderr, "ok:   %s\n", name.c_str());
+    }
+}
+
+// Env-parsing tests are platform-independent and always run.
+static void test_env_parsing() {
+    using namespace llama_grpc;
+
+    // Interval: default when unset.
+    unsetenv("LOCALAI_BACKEND_PARENT_WATCH_INTERVAL");
+    check(parent_watch_interval_ms() == 2000, "interval default 2000ms");
+
+    setenv("LOCALAI_BACKEND_PARENT_WATCH_INTERVAL", "500ms", 1);
+    check(parent_watch_interval_ms() == 500, "interval 500ms");
+
+    setenv("LOCALAI_BACKEND_PARENT_WATCH_INTERVAL", "2s", 1);
+    check(parent_watch_interval_ms() == 2000, "interval 2s");
+
+    setenv("LOCALAI_BACKEND_PARENT_WATCH_INTERVAL", "1m", 1);
+    check(parent_watch_interval_ms() == 60000, "interval 1m");
+
+    setenv("LOCALAI_BACKEND_PARENT_WATCH_INTERVAL", "3", 1); // bare number -> seconds
+    check(parent_watch_interval_ms() == 3000, "interval bare 3 -> 3000ms");
+
+    setenv("LOCALAI_BACKEND_PARENT_WATCH_INTERVAL", "garbage", 1);
+    check(parent_watch_interval_ms() == 2000, "interval garbage -> default");
+    unsetenv("LOCALAI_BACKEND_PARENT_WATCH_INTERVAL");
+
+#if !defined(_WIN32)
+    // Enabled semantics (POSIX only; always false on Windows).
+    unsetenv("LOCALAI_BACKEND_PARENT_WATCH");
+    check(parent_watch_enabled(), "enabled by default");
+
+    for (const char *falsey : {"false", "0", "no", "off", "OFF", " False "}) {
+        setenv("LOCALAI_BACKEND_PARENT_WATCH", falsey, 1);
+        check(!parent_watch_enabled(), std::string("disabled by '") + falsey + "'");
+    }
+    setenv("LOCALAI_BACKEND_PARENT_WATCH", "true", 1);
+    check(parent_watch_enabled(), "enabled by 'true'");
+    setenv("LOCALAI_BACKEND_PARENT_WATCH", "1", 1);
+    check(parent_watch_enabled(), "enabled by '1'");
+    unsetenv("LOCALAI_BACKEND_PARENT_WATCH");
+#endif
+}
+
+#if !defined(_WIN32)
+
+#include <atomic>
+#include <ctime>
+#include <sys/stat.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+static bool file_exists(const std::string &p) {
+    struct stat st;
+    return ::stat(p.c_str(), &st) == 0;
+}
+
+static bool wait_for_file(const std::string &p, int timeout_ms) {
+    int waited = 0;
+    while (waited < timeout_ms) {
+        if (file_exists(p)) {
+            return true;
+        }
+        usleep(20 * 1000);
+        waited += 20;
+    }
+    return false;
+}
+
+static void write_file(const std::string &p, const std::string &content) {
+    FILE *f = fopen(p.c_str(), "w");
+    if (f) {
+        fwrite(content.data(), 1, content.size(), f);
+        fclose(f);
+    }
+}
+
+// Builds test -> middle -> grandchild via fork(2). The grandchild arms the REAL
+// watch_parent_death against middle; middle exits, orphaning the grandchild;
+// the watcher must detect the reparenting and self-terminate.
+static void test_detects_reparent() {
+    char tmpl[] = "/tmp/parentwatch_test_XXXXXX";
+    char *dir = mkdtemp(tmpl);
+    if (dir == nullptr) {
+        check(false, "mkdtemp");
+        return;
+    }
+    const std::string ready_file = std::string(dir) + "/ready";
+    const std::string exited_file = std::string(dir) + "/exited";
+
+    pid_t middle = fork();
+    if (middle < 0) {
+        check(false, "fork middle");
+        return;
+    }
+
+    if (middle == 0) {
+        // ---- middle process ----
+        pid_t grandchild = fork();
+        if (grandchild < 0) {
+            _exit(4);
+        }
+        if (grandchild == 0) {
+            // ---- grandchild process ----
+            pid_t orig_ppid = getppid(); // == middle
+            std::thread([&]() {
+                llama_grpc::watch_parent_death(orig_ppid, 50 /*ms*/, [&]() {
+                    write_file(exited_file, "1");
+                    _exit(7);
+                });
+            }).detach();
+
+            // Safety valve: never linger if something goes wrong.
+            std::thread([]() {
+                usleep(30 * 1000 * 1000);
+                _exit(2);
+            }).detach();
+
+            // Signal readiness only after the watcher captured orig_ppid.
+            write_file(ready_file, std::to_string(getpid()));
+            for (;;) {
+                pause();
+            }
+        }
+        // middle: wait until grandchild is ready, then exit to orphan it.
+        if (!wait_for_file(ready_file, 10000)) {
+            _exit(5);
+        }
+        _exit(0);
+    }
+
+    // ---- test (top) process ----
+    int status = 0;
+    waitpid(middle, &status, 0); // reap middle only; grandchild is orphaned
+
+    check(file_exists(ready_file), "grandchild signaled readiness");
+
+    bool detected = wait_for_file(exited_file, 10000);
+    check(detected, "watcher detected parent death and self-terminated");
+
+    // Best-effort cleanup: kill the grandchild if it somehow survived.
+    if (file_exists(ready_file)) {
+        FILE *f = fopen(ready_file.c_str(), "r");
+        if (f) {
+            int pid = 0;
+            if (fscanf(f, "%d", &pid) == 1 && pid > 1) {
+                kill(pid, SIGKILL);
+            }
+            fclose(f);
+        }
+    }
+    unlink(ready_file.c_str());
+    unlink(exited_file.c_str());
+    rmdir(dir);
+}
+
+#endif // !_WIN32
+
+int main() {
+    test_env_parsing();
+#if !defined(_WIN32)
+    test_detects_reparent();
+#endif
+    if (failures == 0) {
+        fprintf(stderr, "\nAll parent_watch tests passed.\n");
+        return 0;
+    }
+    fprintf(stderr, "\n%d parent_watch test(s) failed.\n", failures);
+    return 1;
+}

backend/cpp/llama-cpp/prepare.sh

@@ -22,6 +22,10 @@ cp -r grpc-server.cpp llama.cpp/tools/grpc-server/
 # unit test (compiled only when -DLLAMA_GRPC_BUILD_TESTS=ON).
 cp -r message_content.h llama.cpp/tools/grpc-server/
 cp -r message_content_test.cpp llama.cpp/tools/grpc-server/
+# Parent-death watcher (included by grpc-server.cpp) and its standalone unit
+# test (run via backend/cpp/run-unit-tests.sh; also buildable under ctest).
+cp -r parent_watch.h llama.cpp/tools/grpc-server/
+cp -r parent_watch_test.cpp llama.cpp/tools/grpc-server/
 cp -rfv llama.cpp/vendor/nlohmann/json.hpp llama.cpp/tools/grpc-server/
 cp -rfv llama.cpp/vendor/cpp-httplib/httplib.h llama.cpp/tools/grpc-server/
 

backend/cpp/llama-cpp/run.sh

@@ -36,6 +36,12 @@ else
 	if [ -d "$CURDIR/lib/rocblas/library" ]; then
 		export ROCBLAS_TENSILE_LIBPATH="$CURDIR"/lib/rocblas/library
 	fi
+	# Same for hipBLASLt (rocblaslt): the bundled libhipblaslt.so resolves its
+	# TensileLibrary_lazy_gfx*.dat kernel data relative to itself, so point it at
+	# the bundled data or it falls back to slow generic kernels (issue #10660).
+	if [ -d "$CURDIR/lib/hipblaslt/library" ]; then
+		export HIPBLASLT_TENSILE_LIBPATH="$CURDIR"/lib/hipblaslt/library
+	fi
 fi
 
 # If there is a lib/ld.so, use it

backend/cpp/privacy-filter/Makefile

@@ -8,7 +8,7 @@
 # Local development: point at a working checkout instead of cloning, e.g.
 #   make PRIVACY_FILTER_SRC=$HOME/c/privacy-filter.cpp grpc-server
 
-PRIVACY_FILTER_VERSION?=595f59630c69d361b5196f2aba2c71c873d0c13c
+PRIVACY_FILTER_VERSION?=735a6c28607ee82afc3a670383f41b55266a3b9a
 PRIVACY_FILTER_REPO?=https://github.com/localai-org/privacy-filter.cpp
 PRIVACY_FILTER_SRC?=
 

backend/cpp/run-unit-tests.sh

@@ -54,7 +54,7 @@ for test_src in "${tests[@]}"; do
     name="$(basename "$test_src" .cpp)"
     bin="$(mktemp -d)/$name"
     echo "==> $test_src"
-    if ! "$CXX" -std=c++17 -Wall -Wextra \
+    if ! "$CXX" -std=c++17 -Wall -Wextra -pthread \
         -I"$JSON_INC" -I"$(dirname "$test_src")" \
         "$test_src" -o "$bin"; then
         echo "COMPILE FAILED: $test_src" >&2

backend/cpp/turboquant/run.sh

@@ -34,6 +34,12 @@ else
 	if [ -d "$CURDIR/lib/rocblas/library" ]; then
 		export ROCBLAS_TENSILE_LIBPATH="$CURDIR"/lib/rocblas/library
 	fi
+	# Same for hipBLASLt (rocblaslt): the bundled libhipblaslt.so resolves its
+	# TensileLibrary_lazy_gfx*.dat kernel data relative to itself, so point it at
+	# the bundled data or it falls back to slow generic kernels (issue #10660).
+	if [ -d "$CURDIR/lib/hipblaslt/library" ]; then
+		export HIPBLASLT_TENSILE_LIBPATH="$CURDIR"/lib/hipblaslt/library
+	fi
 fi
 
 # If there is a lib/ld.so, use it

backend/go/cloud-proxy/provider_anthropic.go

@@ -142,19 +142,12 @@ func buildAnthropicRequest(opts *pb.PredictOptions, cfg *proxyConfig, stream boo
 	if req.MaxTokens <= 0 {
 		req.MaxTokens = anthropicDefaultMaxTokens
 	}
-	// Newer Anthropic models 400 when both temperature and top_p are
-	// set ("`temperature` and `top_p` cannot both be specified for
-	// this model. Please use only one.") even though their docs only
-	// "recommend" picking one. The OpenAI-compatible chat UI almost
-	// always sends both with default values, so prefer temperature
-	// and drop top_p when both are present.
-	if t := opts.GetTemperature(); t != 0 {
-		v := float64(t)
-		req.Temperature = &v
-	} else if t := opts.GetTopP(); t != 0 {
-		v := float64(t)
-		req.TopP = &v
-	}
+	// Do not forward temperature/top_p. Newer Anthropic reasoning models reject
+	// requests that carry temperature ("`temperature` is deprecated for this
+	// model"), and the OpenAI-compatible clients typically send only the
+	// server-side DEFAULT sampling values rather than user intent — dropping
+	// them loses nothing and lets the upstream apply its own defaults.
+	_ = opts
 
 	req.Tools = convertOpenAITools(opts.GetTools())
 	req.ToolChoice = convertOpenAIToolChoice(opts.GetToolChoice())

backend/go/cloud-proxy/provider_anthropic_test.go

@@ -3,7 +3,6 @@ package main
 import (
 	"encoding/json"
 	"io"
-	"math"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -75,15 +74,16 @@ func TestPredict_Anthropic_BasicMessages(t *testing.T) {
 	g.Expect(captured.Messages).To(HaveLen(1))
 	g.Expect(captured.Messages[0].Role).To(Equal("user"))
 	g.Expect(captured.MaxTokens).To(Equal(int32(32)))
-	g.Expect(captured.Temperature).NotTo(BeNil())
-	g.Expect(*captured.Temperature).To(Equal(0.5))
-	// Anthropic 400s when both temperature and top_p are set; the
-	// translator must prefer temperature and drop top_p.
+	// Newer Anthropic reasoning models reject requests carrying temperature
+	// ("`temperature` is deprecated for this model"); clients typically send
+	// only default sampling values, so the translator forwards neither.
+	g.Expect(captured.Temperature).To(BeNil())
 	g.Expect(captured.TopP).To(BeNil())
 	g.Expect(captured.Stream).To(BeFalse())
 }
 
-// When only top_p is set, it should be forwarded.
+// Sampling parameters are not forwarded at all — the upstream applies its
+// own defaults (newest models reject explicit temperature/top_p).
 func TestPredict_Anthropic_TopPOnly(t *testing.T) {
 	g := NewWithT(t)
 	srv, captured := fakeAnthropicUpstream(t, func(_ anthropicRequest) (int, string, string) {
@@ -99,11 +99,7 @@ func TestPredict_Anthropic_TopPOnly(t *testing.T) {
 	})
 	g.Expect(err).NotTo(HaveOccurred())
 	g.Expect(captured.Temperature).To(BeNil())
-	// PredictOptions.TopP is float32 on the wire; the translator widens
-	// to float64 so 0.9 round-trips as 0.8999999761581421… — compare
-	// with a small tolerance rather than exact equality.
-	g.Expect(captured.TopP).NotTo(BeNil())
-	g.Expect(math.Abs(*captured.TopP - 0.9)).To(BeNumerically("<=", 1e-6))
+	g.Expect(captured.TopP).To(BeNil())
 }
 
 func TestPredict_Anthropic_DefaultsMaxTokens(t *testing.T) {

backend/go/cloud-proxy/provider_openai.go

@@ -30,7 +30,7 @@ type openAIRequest struct {
 	Stream           bool            `json:"stream,omitempty"`
 	Temperature      *float64        `json:"temperature,omitempty"`
 	TopP             *float64        `json:"top_p,omitempty"`
-	MaxTokens        *int32          `json:"max_tokens,omitempty"`
+	MaxTokens        *int32          `json:"max_completion_tokens,omitempty"` // newer OpenAI models reject max_tokens ("use max_completion_tokens instead")
 	Stop             []string        `json:"stop,omitempty"`
 	FrequencyPenalty *float64        `json:"frequency_penalty,omitempty"`
 	PresencePenalty  *float64        `json:"presence_penalty,omitempty"`
@@ -107,14 +107,10 @@ func buildOpenAIRequest(opts *pb.PredictOptions, cfg *proxyConfig, stream bool)
 		Tools:      parseRawJSON(opts.GetTools()),
 		ToolChoice: parseRawJSON(opts.GetToolChoice()),
 	}
-	if t := opts.GetTemperature(); t != 0 {
-		v := float64(t)
-		req.Temperature = &v
-	}
-	if t := opts.GetTopP(); t != 0 {
-		v := float64(t)
-		req.TopP = &v
-	}
+	// Do not forward temperature/top_p. Newer OpenAI reasoning models reject
+	// temperature as deprecated, and clients typically send only default
+	// sampling values rather than user intent — let the upstream apply its
+	// own defaults.
 	if n := opts.GetTokens(); n > 0 {
 		req.MaxTokens = &n
 	}

backend/go/cloud-proxy/provider_openai_test.go

@@ -74,8 +74,9 @@ func TestPredict_OpenAI_BasicChat(t *testing.T) {
 	g.Expect(captured.Messages).To(HaveLen(2))
 	g.Expect(captured.Messages[0].Role).To(Equal("system"))
 	g.Expect(captured.Messages[1].Role).To(Equal("user"))
-	g.Expect(captured.Temperature).NotTo(BeNil())
-	g.Expect(*captured.Temperature).To(Equal(0.5))
+	// Sampling parameters are not forwarded (newest models reject explicit
+	// temperature); token limit is serialized as max_completion_tokens.
+	g.Expect(captured.Temperature).To(BeNil())
 	g.Expect(captured.MaxTokens).NotTo(BeNil())
 	g.Expect(*captured.MaxTokens).To(Equal(int32(32)))
 	g.Expect(captured.Stream).To(BeFalse())

backend/go/crispasr/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=3b93758f9725d400eca82976f895e4cec3f31260
+CRISPASR_VERSION?=9a26976a8c8cf5af0afcdd04463cf8ba91e96a54
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
+# Upstream pin lives below as PARAKEET_VERSION?=e8acc6172a94e20a952cf1843decace5d771a94b
 # (.github/bump_deps.sh) can find and update it - matches the
 # whisper.cpp / ds4 / vibevoice-cpp convention.
 #
@@ -15,7 +15,7 @@
 # That's what the L0 smoke test uses. The default target below does the
 # proper clone-at-pin + cmake build so CI doesn't need a side-checkout.
 
-PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
+PARAKEET_VERSION?=e8acc6172a94e20a952cf1843decace5d771a94b
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go

backend/go/stablediffusion-ggml/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=3b6c9ca97cfcda8e68e719e6670d06379fcbe943
+STABLEDIFFUSION_GGML_VERSION?=2574f5936571645f784b77623e1f09bad97d948a
 
 CMAKE_ARGS+=-DGGML_MAX_NAME=128
 

backend/go/stablediffusion-ggml/cpp/gosd.cpp

@@ -798,6 +798,7 @@ void sd_img_gen_params_set_seed(sd_img_gen_params_t *params, int64_t seed) {
 int gen_image(sd_img_gen_params_t *p, int steps, char *dst, float cfg_scale, char *src_image, float strength, char *mask_image, char* ref_images[], int ref_images_count) {
 
     sd_image_t* results;
+    int num_results_out = 0;
 
     std::vector<int> skip_layers = {7, 8, 9};
 
@@ -994,10 +995,14 @@ int gen_image(sd_img_gen_params_t *p, int steps, char *dst, float cfg_scale, cha
             sd_ctx_params_to_str(&ctx_params),
             sd_img_gen_params_to_str(p));
 
-    results = generate_image(sd_c, p);
+    bool gen_ok = generate_image(sd_c, p, &results, &num_results_out);
 
     std::free(p);
 
+    if (!gen_ok || num_results_out == 0) {
+        results = NULL;
+    }
+
     if (results == NULL) {
         fprintf (stderr, "NO results\n");
         if (input_image_buffer) free(input_image_buffer);

backend/go/whisper/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=0ae02cdb2c7317b50991367c165736ce42ed96ac
+WHISPER_CPP_VERSION?=6fc7c33b4c3a2cec83e4b65abd5e96a890480375
 SO_TARGET?=libgowhisper.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/python/common/grpc_auth.py

@@ -11,6 +11,8 @@ import os
 
 import grpc
 
+from parent_watch import start_parent_death_watcher
+
 
 class _AbortHandler(grpc.RpcMethodHandler):
     """A method handler that immediately aborts with UNAUTHENTICATED."""
@@ -70,6 +72,13 @@ def get_auth_interceptors(*, aio: bool = False):
 
     Returns an empty list when LOCALAI_GRPC_AUTH_TOKEN is not set.
     """
+    # Arm the best-effort parent-death backstop here: this is the single helper
+    # every LocalAI Python backend invokes exactly once while building its gRPC
+    # server (mirroring how the Go watcher arms in pkg/grpc's shared serve path).
+    # start_parent_death_watcher() is idempotent and a no-op when disabled or on
+    # unsupported platforms — see parent_watch.py.
+    start_parent_death_watcher()
+
     token = os.environ.get("LOCALAI_GRPC_AUTH_TOKEN", "")
     if not token:
         return []

backend/python/common/mlx_utils.py

@@ -20,7 +20,15 @@ def split_reasoning(text, think_start, think_end):
     Returns ``(reasoning_content, remaining_text)``. When ``think_start`` is
     empty or not found, returns ``("", text)`` unchanged.
     """
-    if not think_start or not text or think_start not in text:
+    if not think_start or not text:
+        return "", text
+    if think_start not in text:
+        # Models like Qwen3.5 open assistant turns already INSIDE thinking, so
+        # the generated text carries only the closing tag. Everything before it
+        # is reasoning that would otherwise leak into the content.
+        if think_end and think_end in text:
+            head, _, tail = text.partition(think_end)
+            return head.strip(), tail.strip()
         return "", text
     pattern = re.compile(
         re.escape(think_start) + r"(.*?)" + re.escape(think_end or ""),

backend/python/common/mlx_utils_test.py

@@ -0,0 +1,75 @@
+"""Unit tests for the mlx/mlx-vlm shared helpers (mlx_utils.py).
+
+Run standalone (Python standard library only, no backend venv needed):
+    python3 -m unittest mlx_utils_test
+
+These mirror the server-less helper tests in backend/python/mlx/test.py
+(TestSharedHelpers), but live here so they run on any platform: the mlx
+test module imports grpc/backend_pb2 at import time and needs the MLX venv,
+whereas mlx_utils only needs the standard library.
+"""
+
+import types
+import unittest
+
+from mlx_utils import parse_tool_calls, split_reasoning
+
+
+class TestSplitReasoning(unittest.TestCase):
+    def test_both_tags(self):
+        r, c = split_reasoning(
+            "<think>step 1\nstep 2</think>The answer is 42.", "<think>", "</think>"
+        )
+        self.assertEqual(r, "step 1\nstep 2")
+        self.assertEqual(c, "The answer is 42.")
+
+    def test_implicit_opener_only_closing_tag(self):
+        # Qwen3.5 opens the assistant turn already inside thinking, so the
+        # output carries only the closing tag; everything before it is reasoning.
+        r, c = split_reasoning(
+            "The user is asking about the weather.\n</think>\n\nThe weather in Rome is sunny.",
+            "<think>",
+            "</think>",
+        )
+        self.assertEqual(r, "The user is asking about the weather.")
+        self.assertEqual(c, "The weather in Rome is sunny.")
+
+    def test_no_tags_at_all(self):
+        r, c = split_reasoning("just text", "<think>", "</think>")
+        self.assertEqual(r, "")
+        self.assertEqual(c, "just text")
+
+    def test_empty_think_end_and_no_opener_match(self):
+        # No think_end to anchor on, and the opener is absent → return unchanged.
+        r, c = split_reasoning("no opener here", "<think>", "")
+        self.assertEqual(r, "")
+        self.assertEqual(c, "no opener here")
+
+    def test_empty_text(self):
+        r, c = split_reasoning("", "<think>", "</think>")
+        self.assertEqual(r, "")
+        self.assertEqual(c, "")
+
+
+class TestParseToolCalls(unittest.TestCase):
+    def test_with_shim(self):
+        tm = types.SimpleNamespace(
+            tool_call_start="<tool_call>",
+            tool_call_end="</tool_call>",
+            parse_tool_call=lambda body, tools: {
+                "name": "get_weather",
+                "arguments": {"location": body.strip()},
+            },
+        )
+        calls, remaining = parse_tool_calls(
+            "Sure: <tool_call>Paris</tool_call>", tm, tools=None
+        )
+        self.assertEqual(len(calls), 1)
+        self.assertEqual(calls[0]["name"], "get_weather")
+        self.assertEqual(calls[0]["arguments"], '{"location": "Paris"}')
+        self.assertEqual(calls[0]["index"], 0)
+        self.assertNotIn("<tool_call>", remaining)
+
+
+if __name__ == "__main__":
+    unittest.main()

backend/python/common/parent_watch.py

@@ -0,0 +1,149 @@
+"""Parent-death watcher (best-effort backstop) for LocalAI Python backends.
+
+LocalAI spawns each backend as a child process and, on a clean shutdown, tears
+it down itself (SIGTERM -> grace -> SIGKILL). That graceful path only runs when
+LocalAI receives a catchable signal and lives long enough to run its handlers.
+If LocalAI is SIGKILLed (e.g. a supervising process's grace period elapses
+first), that teardown never runs and this backend would be reparented to init
+and linger, holding GPU/VRAM and its listen port.
+
+The watcher here is a best-effort backstop for exactly that case: it does NOT
+replace the graceful teardown, it only covers the "parent vanished without
+cleaning up" path. It detects reparenting: when the process that spawned this
+backend dies, the kernel reparents us to the nearest sub-reaper or to init
+(PID 1), so os.getppid() stops matching the value captured at startup. This
+getppid() approach is portable across Linux/macOS (unlike the Linux-only
+PR_SET_PDEATHSIG), which is why it is used here, mirroring the Go backends'
+pkg/grpc/parentwatch.go and the C++ backends' parent_watch.h. It is disabled on
+Windows, which has no equivalent orphan-reparenting semantics.
+
+Env vars (shared verbatim across the Go, C++ and Python backends):
+  LOCALAI_BACKEND_PARENT_WATCH           enabled by default; a falsey value
+                                         ("false"/"0"/"no"/"off", case-insensitive)
+                                         disables it.
+  LOCALAI_BACKEND_PARENT_WATCH_INTERVAL  poll interval as a Go-style duration
+                                         string ("500ms", "2s", "1m") or a bare
+                                         number of seconds. Defaults to 2s.
+"""
+
+import os
+import sys
+import threading
+
+ENV_PARENT_WATCH = "LOCALAI_BACKEND_PARENT_WATCH"
+ENV_PARENT_WATCH_INTERVAL = "LOCALAI_BACKEND_PARENT_WATCH_INTERVAL"
+
+_DEFAULT_INTERVAL_SECONDS = 2.0
+
+# Guard so repeated calls (e.g. get_auth_interceptors invoked more than once)
+# only ever arm a single watcher thread per process.
+_started = False
+_started_lock = threading.Lock()
+
+
+def _enabled():
+    """Report whether the watcher should run in this process."""
+    # Windows does not reparent orphans to a well-known init PID, so the
+    # getppid() heuristic used here doesn't apply there.
+    if os.name == "nt" or sys.platform.startswith("win"):
+        return False
+    val = os.environ.get(ENV_PARENT_WATCH, "").strip().lower()
+    if val in ("false", "0", "no", "off"):
+        return False
+    return True
+
+
+def _interval_seconds():
+    """Return the configured poll interval in seconds, or the default.
+
+    Accepts Go-style duration strings ("500ms", "2s", "1m") for cross-language
+    parity, or a bare number interpreted as seconds.
+    """
+    raw = os.environ.get(ENV_PARENT_WATCH_INTERVAL, "").strip()
+    if not raw:
+        return _DEFAULT_INTERVAL_SECONDS
+    # Split numeric prefix from unit suffix.
+    i = 0
+    while i < len(raw) and (raw[i].isdigit() or raw[i] == "." or (i == 0 and raw[i] in "+-")):
+        i += 1
+    if i == 0:
+        return _DEFAULT_INTERVAL_SECONDS
+    try:
+        num = float(raw[:i])
+    except ValueError:
+        return _DEFAULT_INTERVAL_SECONDS
+    unit = raw[i:].lower()
+    if unit == "ms":
+        seconds = num / 1000.0
+    elif unit in ("s", ""):
+        seconds = num
+    elif unit == "m":
+        seconds = num * 60.0
+    else:
+        return _DEFAULT_INTERVAL_SECONDS
+    return seconds if seconds > 0 else _DEFAULT_INTERVAL_SECONDS
+
+
+def _parent_died(orig_ppid):
+    """Report whether this process has been reparented away from orig_ppid.
+
+    Reparenting is the standard POSIX signal that the original parent (here, the
+    LocalAI process that spawned this backend) has exited: the orphan is handed
+    to the nearest sub-reaper or to init (PID 1), so os.getppid() no longer
+    matches the value captured at startup.
+    """
+    ppid = os.getppid()
+    return ppid != orig_ppid or ppid == 1
+
+
+def _watch(orig_ppid, interval, on_death):
+    """Poll until _parent_died reports the original parent is gone, then call
+    on_death. Blocks, so run it on its own (daemon) thread."""
+    import time
+
+    while True:
+        time.sleep(interval)
+        if _parent_died(orig_ppid):
+            on_death()
+            return
+
+
+def start_parent_death_watcher():
+    """Install the best-effort safety net described in this module's docstring.
+
+    No-op when disabled, on Windows, when already orphaned at startup
+    (os.getppid() <= 1), or if already started. This is a backstop alongside —
+    never a replacement for — LocalAI's graceful teardown.
+    """
+    global _started
+    if not _enabled():
+        return
+    with _started_lock:
+        if _started:
+            return
+        orig_ppid = os.getppid()
+        # A parent of 1 (or less) at startup means we were already orphaned (or
+        # launched directly under init) — there is no original parent to watch.
+        if orig_ppid <= 1:
+            return
+        interval = _interval_seconds()
+
+        def on_death():
+            print(
+                "backend parent process (pid {}) exited without stopping this "
+                "backend; self-terminating to avoid orphaning".format(orig_ppid),
+                file=sys.stderr,
+                flush=True,
+            )
+            # Immediate, non-cleanup exit: this is a shutdown safety net and the
+            # normal graceful path is already gone.
+            os._exit(1)
+
+        thread = threading.Thread(
+            target=_watch,
+            args=(orig_ppid, interval, on_death),
+            name="parent-death-watcher",
+            daemon=True,
+        )
+        thread.start()
+        _started = True

backend/python/common/parent_watch_test.py

@@ -0,0 +1,150 @@
+"""Unit tests for the parent-death watcher (parent_watch.py).
+
+Run standalone (Python standard library only, no backend venv needed):
+    python3 -m unittest parent_watch_test
+
+The core test (test_detects_reparent) builds a genuine two-level process tree
+(test -> middle -> grandchild) with os.fork, lets the middle process die, and
+asserts the grandchild's parent_watch._watch detects the reparenting and
+self-terminates — mirroring the Go test in pkg/grpc/parentwatch_test.go and the
+C++ test in backend/cpp/llama-cpp/parent_watch_test.cpp.
+"""
+
+import os
+import sys
+import tempfile
+import threading
+import time
+import unittest
+
+import parent_watch
+
+
+class TestParentWatchEnvParsing(unittest.TestCase):
+    def setUp(self):
+        self._saved = {
+            k: os.environ.get(k)
+            for k in (parent_watch.ENV_PARENT_WATCH, parent_watch.ENV_PARENT_WATCH_INTERVAL)
+        }
+        for k in self._saved:
+            os.environ.pop(k, None)
+
+    def tearDown(self):
+        for k, v in self._saved.items():
+            if v is None:
+                os.environ.pop(k, None)
+            else:
+                os.environ[k] = v
+
+    def test_interval_default(self):
+        self.assertEqual(parent_watch._interval_seconds(), 2.0)
+
+    def test_interval_units(self):
+        cases = {"500ms": 0.5, "2s": 2.0, "1m": 60.0, "3": 3.0, "0.5s": 0.5}
+        for raw, expected in cases.items():
+            os.environ[parent_watch.ENV_PARENT_WATCH_INTERVAL] = raw
+            self.assertAlmostEqual(parent_watch._interval_seconds(), expected, msg=raw)
+
+    def test_interval_garbage_falls_back(self):
+        os.environ[parent_watch.ENV_PARENT_WATCH_INTERVAL] = "garbage"
+        self.assertEqual(parent_watch._interval_seconds(), 2.0)
+
+    @unittest.skipIf(os.name == "nt" or sys.platform.startswith("win"), "POSIX only")
+    def test_enabled_default(self):
+        self.assertTrue(parent_watch._enabled())
+
+    @unittest.skipIf(os.name == "nt" or sys.platform.startswith("win"), "POSIX only")
+    def test_disabled_by_falsey(self):
+        for val in ("false", "0", "no", "off", "OFF", " False "):
+            os.environ[parent_watch.ENV_PARENT_WATCH] = val
+            self.assertFalse(parent_watch._enabled(), msg=val)
+
+    @unittest.skipIf(os.name == "nt" or sys.platform.startswith("win"), "POSIX only")
+    def test_enabled_by_truthy(self):
+        for val in ("true", "1", "yes", "on"):
+            os.environ[parent_watch.ENV_PARENT_WATCH] = val
+            self.assertTrue(parent_watch._enabled(), msg=val)
+
+
+@unittest.skipIf(os.name == "nt" or sys.platform.startswith("win"), "fork/reparent is POSIX only")
+class TestParentWatchReparent(unittest.TestCase):
+    def _wait_for_file(self, path, timeout=10.0):
+        deadline = time.time() + timeout
+        while time.time() < deadline:
+            if os.path.exists(path):
+                return True
+            time.sleep(0.02)
+        return False
+
+    def test_detects_reparent(self):
+        tmpdir = tempfile.mkdtemp(prefix="parentwatch_test_")
+        ready_file = os.path.join(tmpdir, "ready")
+        exited_file = os.path.join(tmpdir, "exited")
+
+        middle = os.fork()
+        if middle == 0:
+            # ---- middle process ----
+            grandchild = os.fork()
+            if grandchild == 0:
+                # ---- grandchild process: arm the REAL watcher against middle ----
+                orig_ppid = os.getppid()
+
+                def on_death():
+                    with open(exited_file, "w") as f:
+                        f.write("1")
+                    os._exit(7)
+
+                threading.Thread(
+                    target=parent_watch._watch,
+                    args=(orig_ppid, 0.05, on_death),
+                    daemon=True,
+                ).start()
+
+                # Safety valve: never linger if something goes wrong.
+                def bail():
+                    time.sleep(30)
+                    os._exit(2)
+
+                threading.Thread(target=bail, daemon=True).start()
+
+                # Signal readiness only after the watcher captured orig_ppid.
+                with open(ready_file, "w") as f:
+                    f.write(str(os.getpid()))
+                while True:
+                    time.sleep(1)
+            else:
+                # middle: wait until grandchild is ready, then exit to orphan it.
+                if not self._wait_for_file(ready_file):
+                    os._exit(5)
+                os._exit(0)
+
+        # ---- test (top) process ----
+        os.waitpid(middle, 0)  # reap middle only; grandchild is orphaned
+
+        self.assertTrue(os.path.exists(ready_file), "grandchild never signaled readiness")
+        self.assertTrue(
+            self._wait_for_file(exited_file),
+            "watcher did not detect parent death within timeout",
+        )
+
+        # Best-effort cleanup: kill the grandchild if it somehow survived.
+        try:
+            with open(ready_file) as f:
+                pid = int(f.read().strip())
+            if pid > 1:
+                os.kill(pid, 9)
+        except (OSError, ValueError):
+            pass
+        for p in (ready_file, exited_file):
+            try:
+                os.remove(p)
+            except OSError:
+                pass
+        try:
+            os.rmdir(tmpdir)
+        except OSError:
+            pass
+
+
+if __name__ == "__main__":
+    unittest.main()

backend/python/common/python_utils.py

@@ -58,7 +58,18 @@ def messages_to_dicts(proto_messages):
             d["reasoning_content"] = msg.reasoning_content
         if msg.tool_calls:
             try:
-                d["tool_calls"] = json.loads(msg.tool_calls)
+                tool_calls = json.loads(msg.tool_calls)
+                # Chat templates (e.g. Qwen) iterate function.arguments as a
+                # mapping, but the OpenAI wire format carries it as a JSON
+                # string — decode it back so the template's .items() works.
+                for tc in tool_calls:
+                    fn = tc.get("function") if isinstance(tc, dict) else None
+                    if isinstance(fn, dict) and isinstance(fn.get("arguments"), str):
+                        try:
+                            fn["arguments"] = json.loads(fn["arguments"])
+                        except json.JSONDecodeError:
+                            pass
+                d["tool_calls"] = tool_calls
             except json.JSONDecodeError:
                 pass
         result.append(d)

backend/python/common/python_utils_test.py

@@ -0,0 +1,122 @@
+"""Unit tests for the shared python backend helpers (python_utils.py).
+
+Run standalone (Python standard library only, no backend venv needed):
+    python3 -m unittest python_utils_test
+
+These mirror the server-less helper tests in backend/python/mlx/test.py
+(TestSharedHelpers), but live here so they run on any platform: the mlx
+test module imports grpc/backend_pb2 at import time and needs the MLX venv,
+whereas python_utils has no third-party dependency. Proto Message objects
+are faked with types.SimpleNamespace (real proto fields default to "").
+"""
+
+import json
+import types
+import unittest
+
+from python_utils import messages_to_dicts, parse_options
+
+
+def _msg(**fields):
+    """Fake a proto Message: every unset field is the empty string, as protobuf."""
+    defaults = {
+        "role": "",
+        "content": "",
+        "name": "",
+        "tool_call_id": "",
+        "reasoning_content": "",
+        "tool_calls": "",
+    }
+    defaults.update(fields)
+    return types.SimpleNamespace(**defaults)
+
+
+class TestParseOptions(unittest.TestCase):
+    def test_type_inference(self):
+        opts = parse_options(
+            ["temperature:0.7", "max_tokens:128", "trust:true", "name:hello", "no_colon_skipped"]
+        )
+        self.assertEqual(opts["temperature"], 0.7)
+        self.assertEqual(opts["max_tokens"], 128)
+        self.assertIs(opts["trust"], True)
+        self.assertEqual(opts["name"], "hello")
+        self.assertNotIn("no_colon_skipped", opts)
+
+
+class TestMessagesToDicts(unittest.TestCase):
+    def test_basic_fields(self):
+        out = messages_to_dicts(
+            [
+                _msg(role="user", content="hi"),
+                _msg(role="tool", content="42", tool_call_id="call_1", name="f"),
+            ]
+        )
+        self.assertEqual(out[0], {"role": "user", "content": "hi"})
+        self.assertEqual(out[1]["tool_call_id"], "call_1")
+        self.assertEqual(out[1]["name"], "f")
+
+    def test_tool_call_arguments_string_decoded_to_mapping(self):
+        # OpenAI wire format ships function.arguments as a JSON *string*; chat
+        # templates iterate it as a mapping, so it must come back as a dict.
+        out = messages_to_dicts(
+            [
+                _msg(
+                    role="assistant",
+                    tool_calls=json.dumps(
+                        [
+                            {
+                                "id": "call_1",
+                                "type": "function",
+                                "function": {
+                                    "name": "get_weather",
+                                    "arguments": '{"location": "Rome"}',
+                                },
+                            }
+                        ]
+                    ),
+                )
+            ]
+        )
+        args = out[0]["tool_calls"][0]["function"]["arguments"]
+        self.assertEqual(args, {"location": "Rome"})
+        self.assertEqual(dict(args.items()), {"location": "Rome"})
+
+    def test_tool_call_arguments_already_mapping_is_idempotent(self):
+        out = messages_to_dicts(
+            [
+                _msg(
+                    role="assistant",
+                    tool_calls=json.dumps(
+                        [{"function": {"name": "f", "arguments": {"a": 1}}}]
+                    ),
+                )
+            ]
+        )
+        self.assertEqual(out[0]["tool_calls"][0]["function"]["arguments"], {"a": 1})
+
+    def test_tool_call_arguments_invalid_json_left_as_string(self):
+        out = messages_to_dicts(
+            [
+                _msg(
+                    role="assistant",
+                    tool_calls=json.dumps(
+                        [{"function": {"name": "f", "arguments": "not-json"}}]
+                    ),
+                )
+            ]
+        )
+        self.assertEqual(out[0]["tool_calls"][0]["function"]["arguments"], "not-json")
+
+    def test_tool_call_without_function_key(self):
+        out = messages_to_dicts(
+            [_msg(role="assistant", tool_calls=json.dumps([{"id": "call_1"}]))]
+        )
+        self.assertEqual(out[0]["tool_calls"], [{"id": "call_1"}])
+
+    def test_tool_calls_invalid_json_dropped(self):
+        out = messages_to_dicts([_msg(role="assistant", tool_calls="{not json")])
+        self.assertNotIn("tool_calls", out[0])
+
+
+if __name__ == "__main__":
+    unittest.main()

backend/python/vllm/backend.py

@@ -748,7 +748,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         # When (A) native streaming ran cleanly, per-delta yields above already
         # delivered everything — do NOT extract again on the full text or we'd
         # duplicate content/tool_calls into the final chunk.
-        if has_tool_parser and not (native_streaming and not native_streaming_error):
+        # NOTE: `native_streaming` is a capability flag ("streaming parser is
+        # available"), not a state flag ("streaming actually ran"). For
+        # non-streaming requests it is still True but the per-delta loop was
+        # never entered, so we MUST still run extract_tool_calls here. Hence
+        # the explicit `streaming and …` guard on both branches.
+        if has_tool_parser and not (streaming and native_streaming and not native_streaming_error):
             try:
                 tp = tp_instance
                 if tp is None:
@@ -770,7 +775,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                         ))
             except Exception as e:
                 print(f"Tool parser error: {e}", file=sys.stderr)
-        elif native_streaming and not native_streaming_error:
+        elif streaming and native_streaming and not native_streaming_error:
             # Per-delta path already emitted content + tool_calls; the final
             # chat_delta should carry only metadata (token counts, logprobs).
             content = ""

backend/python/vllm/install.sh

@@ -35,6 +35,21 @@ if [ "x${BUILD_PROFILE}" == "xcpu" ]; then
     EXTRA_PIP_INSTALL_FLAGS+=" --index-strategy=unsafe-best-match"
 fi
 
+# AMD ROCm: vLLM ships prebuilt ROCm wheels, but on a DEDICATED index
+# (https://wheels.vllm.ai/rocm/), NOT PyPI, and ONLY for CPython 3.12. On any
+# other Python the installer silently falls back to the CUDA-only PyPI wheel,
+# which is unusable on an AMD GPU (import fails, so the backend never finds the
+# vllm module). Force Python 3.12 before the venv is created (matches the
+# intel/l4t13 cp312 bump); the hipblas branch below pulls vllm from the ROCm
+# wheel index. unsafe-best-match lets uv consult that index and PyPI together.
+# https://docs.vllm.ai/en/latest/getting_started/installation/gpu.html?device=rocm
+if [ "x${BUILD_TYPE}" == "xhipblas" ]; then
+    PYTHON_VERSION="3.12"
+    PYTHON_PATCH="12"
+    PY_STANDALONE_TAG="20251120"
+    EXTRA_PIP_INSTALL_FLAGS+=" --index-strategy=unsafe-best-match"
+fi
+
 # cublas13 pulls the vLLM wheel from a per-tag cu130 index (PyPI's vllm wheel
 # is built against CUDA 12 and won't load on cu130). uv's default per-package
 # first-match strategy would still pick the PyPI wheel, so allow it to consult
@@ -104,7 +119,7 @@ if [ "$(uname -s)" = "Darwin" ]; then
     # can rewrite it. Darwin therefore follows vllm-metal and can lag the Linux
     # vllm pin (requirements-cublas13-after.txt, bumped independently against
     # vllm/vllm) until vllm-metal supports a newer vLLM.
-    VLLM_METAL_VERSION="v0.3.0.dev20260628073537"
+    VLLM_METAL_VERSION="v0.3.0.dev20260701212152"
 
     # The coupled vLLM source version is whatever this vllm-metal release builds
     # against -- it declares it in its own installer as `vllm_v=`. Derive it from
@@ -194,6 +209,22 @@ elif [ "x${BUILD_TYPE}" == "xintel" ]; then
         export CMAKE_PREFIX_PATH="$(python -c 'import site; print(site.getsitepackages()[0])'):${CMAKE_PREFIX_PATH:-}"
         VLLM_TARGET_DEVICE=xpu uv pip install ${EXTRA_PIP_INSTALL_FLAGS:-} --no-deps .
     popd
+# AMD ROCm: install vllm from its dedicated ROCm wheel index instead of the
+# CUDA-only PyPI wheel. installRequirements brings the base ROCm
+# torch/transformers (requirements-hipblas.txt), then we pull vllm (plus the
+# matching ROCm torch, via --upgrade) from wheels.vllm.ai/rocm. This is the
+# method upstream prescribes for AMD; the Python-3.12 pin is set above.
+# There is intentionally no requirements-hipblas-after.txt: a bare `vllm`
+# there would resolve to the CUDA wheel, and installRequirements never loads
+# a ${BUILD_TYPE}-after file for hipblas anyway (BUILD_TYPE == BUILD_PROFILE).
+# https://docs.vllm.ai/en/latest/getting_started/installation/gpu.html?device=rocm
+elif [ "x${BUILD_TYPE}" == "xhipblas" ]; then
+    installRequirements
+
+    # --upgrade reconciles the base ROCm torch to whatever the vllm ROCm wheel
+    # pins; --extra-index-url adds the ROCm wheel repository on top of PyPI.
+    uv pip install ${EXTRA_PIP_INSTALL_FLAGS:-} \
+        --extra-index-url https://wheels.vllm.ai/rocm/ --upgrade vllm
 # FROM_SOURCE=true on a CPU build skips the prebuilt vllm wheel in
 # requirements-cpu-after.txt and compiles vllm locally against the host's
 # actual CPU. Not used by default because it takes ~30-40 minutes, but

backend/python/vllm/requirements-cublas13-after.txt

@@ -3,8 +3,8 @@
 # on a cu130 host. Pull the cu130-flavoured wheel from vLLM's per-tag index
 # instead — the cublas13 case in install.sh adds --index-strategy=unsafe-best-match
 # so uv consults this index alongside PyPI.
---extra-index-url https://wheels.vllm.ai/0.23.0/cu130
+--extra-index-url https://wheels.vllm.ai/0.24.0/cu130
 # VERSION COUPLING: darwin/Apple-Silicon builds use vllm-metal (see install.sh),
 # which pins this exact vLLM version. Bumping vllm here means coordinating with a
 # vllm-metal release that supports the new version, or macOS/Metal builds break.
-vllm==0.23.0
+vllm==0.24.0

backend/python/vllm/requirements-hipblas-after.txt

@@ -1 +0,0 @@
-vllm

backend/rust/kokoros/src/service.rs

@@ -351,6 +351,16 @@ impl Backend for KokorosService {
         Err(Status::unimplemented("Not supported"))
     }
 
+    type AudioTranscriptionLiveStream =
+        ReceiverStream<Result<backend::TranscriptLiveResponse, Status>>;
+
+    async fn audio_transcription_live(
+        &self,
+        _: Request<tonic::Streaming<backend::TranscriptLiveRequest>>,
+    ) -> Result<Response<Self::AudioTranscriptionLiveStream>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
     async fn diarize(
         &self,
         _: Request<backend::DiarizeRequest>,

cmd/launcher/internal/launcher.go

@@ -207,12 +207,20 @@ func (l *Launcher) StartLocalAI() error {
 	}
 
 	// Build command arguments
+	dataPath := l.GetDataPath()
 	args := []string{
 		"run",
 		"--models-path", l.config.ModelsPath,
 		"--backends-path", l.config.BackendsPath,
 		"--address", l.config.Address,
 		"--log-level", l.config.LogLevel,
+		// Keep persistent data and dynamic config under the launcher's data
+		// directory (~/.localai) rather than letting the server resolve them
+		// to ${basepath}/{data,configuration}. ${basepath} expands to the
+		// launcher process's CWD (often the user's home root), which puts
+		// ~/data and ~/configuration outside ~/.localai. See #10610.
+		"--data-path", filepath.Join(dataPath, "data"),
+		"--localai-config-dir", filepath.Join(dataPath, "configuration"),
 	}
 
 	l.localaiCmd = exec.CommandContext(l.ctx, binaryPath, args...)

core/application/distributed.go

@@ -356,6 +356,12 @@ func initDistributed(cfg *config.ApplicationConfig, authDB *gorm.DB, configLoade
 		PrefixConfig:     prefixCfg,
 		Pressure:         pressure,
 		SharedModels:     cfg.Distributed.SharedModels,
+		// Cap how long a cold load may hold the per-model advisory lock: the
+		// configured backend.install deadline plus a margin for file staging and
+		// the remote LoadModel. Derived from the install timeout so raising it
+		// (for slow links pulling multi-GB images) widens the ceiling too,
+		// instead of letting the static default cut a legitimately slow load.
+		ModelLoadCeiling: cfg.Distributed.BackendInstallTimeoutOrDefault() + 10*time.Minute,
 	})
 
 	// Wire staging-progress broadcasting so file-staging shows up on every

core/application/startup.go

@@ -369,7 +369,7 @@ func New(opts ...config.AppOption) (*Application, error) {
 	}
 
 	for _, backend := range options.ExternalBackends {
-		if err := galleryop.InstallExternalBackend(options.Context, options.BackendGalleries, options.SystemState, application.ModelLoader(), nil, backend, "", "", options.RequireBackendIntegrity); err != nil {
+		if err := galleryop.InstallExternalBackend(options.Context, options.BackendGalleries, options.SystemState, application.ModelLoader(), nil, backend, "", "", false, options.RequireBackendIntegrity); err != nil {
 			xlog.Error("error installing external backend", "error", err)
 		}
 	}
@@ -473,20 +473,13 @@ func New(opts ...config.AppOption) (*Application, error) {
 
 	if options.LoadToMemory != nil && !options.SingleBackend {
 		for _, m := range options.LoadToMemory {
-			cfg, err := application.ModelConfigLoader().LoadModelConfigFileByNameDefaultOptions(m, options)
-			if err != nil {
+			xlog.Debug("Auto loading model into memory from file", "model", m)
+			// Same path as POST /backend/load: a realtime pipeline model expands
+			// to its sub-models, and load failures are recorded as model_load
+			// traces.
+			if _, err := backend.PreloadModelByName(options.Context, application.ModelConfigLoader(), application.ModelLoader(), options, m); err != nil {
 				return nil, err
 			}
-
-			xlog.Debug("Auto loading model into memory from file", "model", m, "file", cfg.Model)
-
-			o := backend.ModelOptions(*cfg, options)
-
-			var backendErr error
-			_, backendErr = application.ModelLoader().Load(o...)
-			if backendErr != nil {
-				return nil, backendErr
-			}
 		}
 	}
 

core/backend/options.go

@@ -52,6 +52,22 @@ func ModelLoadTraceObserver(appConfig *config.ApplicationConfig) func(model.Back
 	}
 }
 
+// PreloadModel warms a model into memory without running any inference, so the
+// first real request doesn't pay the backend's cold-start load cost. It uses
+// the same ModelOptions + ml.Load path the modality functions use, so a
+// subsequent inference call hits the loader cache instead of reloading. Load
+// failures are recorded and returned; callers that warm models opportunistically
+// (e.g. realtime session warm-up) typically log and continue, since the lazy
+// path will retry on first use.
+func PreloadModel(ctx context.Context, ml *model.ModelLoader, modelConfig config.ModelConfig, appConfig *config.ApplicationConfig) error {
+	opts := ModelOptions(modelConfig, appConfig, model.WithContext(ctx))
+	if _, err := ml.Load(opts...); err != nil {
+		recordModelLoadFailure(appConfig, modelConfig.Name, modelConfig.Backend, err, nil)
+		return err
+	}
+	return nil
+}
+
 // recordModelLoadFailure records a backend trace when model loading fails.
 func recordModelLoadFailure(appConfig *config.ApplicationConfig, modelName, backend string, err error, data map[string]any) {
 	if !appConfig.EnableTracing {

core/backend/preload.go

@@ -0,0 +1,122 @@
+package backend
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"sync"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/xlog"
+)
+
+// PreloadModelByName loads the named model into memory so the first request
+// that uses it pays no cold-start load cost — the inverse of shutting a model
+// down. If the model is a realtime pipeline (its config declares a `pipeline:`
+// block), each configured sub-model (VAD, transcription, LLM, TTS,
+// sound_detection, voice_recognition) is loaded concurrently instead of the
+// pipeline stub, which has no backend of its own. It returns the model names
+// actually loaded and a joined error naming each sub-model that failed (nil on
+// full success); a partial pipeline load reports both the loaded names and the
+// failures so the caller can surface exactly what is and isn't resident.
+// Compaction's summary_model is deliberately left cold: it is only invoked off
+// the response path, so it can stay lazy.
+func PreloadModelByName(ctx context.Context, cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig, name string) ([]string, error) {
+	cfg, err := cl.LoadModelConfigFileByNameDefaultOptions(name, appConfig)
+	if err != nil {
+		return nil, err
+	}
+
+	stages, err := pipelineStages(cl, &cfg.Pipeline, ml.ModelPath)
+	if err != nil {
+		return nil, err
+	}
+	if len(stages) == 0 {
+		// Not a pipeline: load the model's own backend directly.
+		if err := PreloadModel(ctx, ml, *cfg, appConfig); err != nil {
+			return nil, err
+		}
+		return []string{cfg.Name}, nil
+	}
+	return PreloadStages(ctx, ml, appConfig, stages)
+}
+
+// PreloadStage names one pipeline sub-model to preload and the resolved config
+// to load it from (nil = stage absent, skipped). Role labels the pipeline slot
+// in errors and logs.
+type PreloadStage struct {
+	Role string
+	Cfg  *config.ModelConfig
+}
+
+// loadStage is PreloadModel behind a seam so PreloadStages can be unit-tested
+// without spawning real backends.
+var loadStage = PreloadModel
+
+// pipelineStages resolves each populated pipeline stage to its concrete model
+// config, following a single alias hop — the same resolution the realtime
+// pipeline itself uses. A stage that fails to resolve is a misconfiguration,
+// so it fails fast rather than being deferred to load. A pipeline with no
+// stages set returns nil, which callers treat as "not a pipeline".
+func pipelineStages(cl *config.ModelConfigLoader, p *config.Pipeline, modelPath string) ([]PreloadStage, error) {
+	voiceRec := ""
+	if p.VoiceRecognition != nil {
+		voiceRec = p.VoiceRecognition.Model
+	}
+	var stages []PreloadStage
+	for _, s := range []struct{ role, name string }{
+		{"vad", p.VAD},
+		{"transcription", p.Transcription},
+		{"llm", p.LLM},
+		{"tts", p.TTS},
+		{"sound_detection", p.SoundDetection},
+		{"voice_recognition", voiceRec},
+	} {
+		if s.name == "" {
+			continue
+		}
+		cfg, err := cl.LoadResolvedModelConfig(s.name, modelPath)
+		if err != nil {
+			return nil, fmt.Errorf("%s (%s): %w", s.role, s.name, err)
+		}
+		stages = append(stages, PreloadStage{Role: s.role, Cfg: cfg})
+	}
+	return stages, nil
+}
+
+// PreloadStages loads every present stage at once and waits for all of them, so
+// a pipeline warms in the time of its slowest stage rather than the sum. Absent
+// (nil-config) stages are skipped. A failed stage does not cancel the others —
+// they all run to completion so the joined error names every broken stage at
+// once, alongside the names that did load.
+func PreloadStages(ctx context.Context, ml *model.ModelLoader, appConfig *config.ApplicationConfig, stages []PreloadStage) ([]string, error) {
+	var (
+		wg     sync.WaitGroup
+		mu     sync.Mutex
+		loaded []string
+		errs   []error
+	)
+	for _, s := range stages {
+		if s.Cfg == nil {
+			continue
+		}
+		wg.Add(1)
+		go func(s PreloadStage) {
+			defer wg.Done()
+			if err := loadStage(ctx, ml, *s.Cfg, appConfig); err != nil {
+				xlog.Warn("preload: failed to load pipeline sub-model", "stage", s.Role, "model", s.Cfg.Name, "error", err)
+				mu.Lock()
+				errs = append(errs, fmt.Errorf("%s (%s): %w", s.Role, s.Cfg.Name, err))
+				mu.Unlock()
+				return
+			}
+			xlog.Debug("preload: loaded pipeline sub-model", "stage", s.Role, "model", s.Cfg.Name)
+			mu.Lock()
+			loaded = append(loaded, s.Cfg.Name)
+			mu.Unlock()
+		}(s)
+	}
+	wg.Wait()
+	return loaded, errors.Join(errs...)
+}

core/backend/preload_internal_test.go

@@ -0,0 +1,146 @@
+package backend
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"sync"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/pkg/model"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("pipelineStages", func() {
+	seed := func(dir string, names ...string) *config.ModelConfigLoader {
+		for _, n := range names {
+			yaml := "name: " + n + "\nbackend: fake-backend\n"
+			Expect(os.WriteFile(filepath.Join(dir, n+".yaml"), []byte(yaml), 0o644)).To(Succeed())
+		}
+		cl := config.NewModelConfigLoader(dir)
+		Expect(cl.LoadModelConfigsFromPath(dir)).To(Succeed())
+		return cl
+	}
+
+	It("resolves only the populated stages, in load order", func() {
+		dir := GinkgoT().TempDir()
+		cl := seed(dir, "vad-m", "stt-m", "llm-m", "tts-m")
+
+		stages, err := pipelineStages(cl, &config.Pipeline{
+			VAD:           "vad-m",
+			Transcription: "stt-m",
+			LLM:           "llm-m",
+			TTS:           "tts-m",
+		}, dir)
+		Expect(err).ToNot(HaveOccurred())
+
+		roles := make([]string, len(stages))
+		names := make([]string, len(stages))
+		for i, s := range stages {
+			roles[i] = s.Role
+			names[i] = s.Cfg.Name
+		}
+		Expect(roles).To(Equal([]string{"vad", "transcription", "llm", "tts"}))
+		Expect(names).To(Equal([]string{"vad-m", "stt-m", "llm-m", "tts-m"}))
+	})
+
+	It("skips unset stages and includes sound_detection and voice_recognition when set", func() {
+		dir := GinkgoT().TempDir()
+		cl := seed(dir, "stt-m", "ced", "spk")
+
+		stages, err := pipelineStages(cl, &config.Pipeline{
+			Transcription:    "stt-m",
+			SoundDetection:   "ced",
+			VoiceRecognition: &config.PipelineVoiceRecognition{Model: "spk"},
+		}, dir)
+		Expect(err).ToNot(HaveOccurred())
+
+		roles := make([]string, len(stages))
+		for i, s := range stages {
+			roles[i] = s.Role
+		}
+		Expect(roles).To(ConsistOf("transcription", "sound_detection", "voice_recognition"))
+	})
+
+	It("returns nil for a pipeline with no stages (not a pipeline)", func() {
+		dir := GinkgoT().TempDir()
+		cl := seed(dir)
+
+		stages, err := pipelineStages(cl, &config.Pipeline{}, dir)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(stages).To(BeNil())
+	})
+})
+
+var _ = Describe("PreloadStages", func() {
+	var (
+		mu   sync.Mutex
+		seen []string
+	)
+
+	// stubLoader swaps the loadStage seam for a recorder so no real backends
+	// are spawned; errFor injects per-model failures.
+	stubLoader := func(errFor map[string]error) {
+		loadStage = func(_ context.Context, _ *model.ModelLoader, cfg config.ModelConfig, _ *config.ApplicationConfig) error {
+			mu.Lock()
+			seen = append(seen, cfg.Name)
+			mu.Unlock()
+			return errFor[cfg.Name]
+		}
+	}
+
+	BeforeEach(func() {
+		seen = nil
+	})
+	AfterEach(func() {
+		loadStage = PreloadModel
+	})
+
+	mkStage := func(role, name string) PreloadStage {
+		return PreloadStage{Role: role, Cfg: &config.ModelConfig{Name: name}}
+	}
+
+	It("loads every present stage, skips absent (nil-config) ones, and returns the loaded names", func() {
+		stubLoader(nil)
+
+		loaded, err := PreloadStages(context.Background(), nil, nil, []PreloadStage{
+			mkStage("vad", "vad-m"),
+			{Role: "transcription"}, // absent stage
+			mkStage("llm", "llm-m"),
+		})
+
+		Expect(err).ToNot(HaveOccurred())
+		Expect(loaded).To(ConsistOf("vad-m", "llm-m"))
+		// Barrier: every stage has run by the time PreloadStages returns, so
+		// reading seen without the lock here is safe.
+		Expect(seen).To(ConsistOf("vad-m", "llm-m"))
+	})
+
+	It("reports a joined error naming each failed stage while still loading the rest", func() {
+		stubLoader(map[string]error{
+			"vad-m": errors.New("vad boom"),
+			"tts-m": errors.New("tts boom"),
+		})
+
+		loaded, err := PreloadStages(context.Background(), nil, nil, []PreloadStage{
+			mkStage("vad", "vad-m"),
+			mkStage("llm", "llm-m"),
+			mkStage("tts", "tts-m"),
+		})
+
+		// Every stage ran (a failure does not cancel the others)...
+		Expect(seen).To(ConsistOf("vad-m", "llm-m", "tts-m"))
+		// ...the stage that loaded fine is reported as loaded...
+		Expect(loaded).To(ConsistOf("llm-m"))
+		// ...and the joined error names every broken stage and its cause.
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("vad (vad-m)"))
+		Expect(err.Error()).To(ContainSubstring("vad boom"))
+		Expect(err.Error()).To(ContainSubstring("tts (tts-m)"))
+		Expect(err.Error()).To(ContainSubstring("tts boom"))
+		Expect(err.Error()).ToNot(ContainSubstring("llm"))
+	})
+})

core/cli/backends.go

@@ -127,7 +127,7 @@ func (bi *BackendsInstall) Run(ctx *cliContext.Context) error {
 	}
 
 	modelLoader := model.NewModelLoader(systemState)
-	err = galleryop.InstallExternalBackend(context.Background(), galleries, systemState, modelLoader, progressCallback, bi.BackendArgs, bi.Name, bi.Alias, bi.RequireBackendIntegrity)
+	err = galleryop.InstallExternalBackend(context.Background(), galleries, systemState, modelLoader, progressCallback, bi.BackendArgs, bi.Name, bi.Alias, false, bi.RequireBackendIntegrity)
 	if err != nil {
 		return err
 	}

core/config/meta/registry.go

@@ -599,6 +599,13 @@ func DefaultRegistry() map[string]FieldMetaOverride {
 			Component:   "toggle",
 			Order:       89,
 		},
+		"pipeline.disable_warmup": {
+			Section:     "pipeline",
+			Label:       "Disable Warmup",
+			Description: "Turn off eager pre-loading of the pipeline's sub-models at realtime session start. By default LocalAI loads every configured sub-model backend (VAD, transcription, LLM, TTS, sound detection, voice recognition) before the session starts and blocks until they are ready, so the first turn pays no cold-start cost and a model that fails to load is reported at session start instead of mid-call. Enable this to restore the lazy 'load on first use' behavior — session start no longer waits on loading and load errors surface on the first turn instead. Useful to keep idle sessions from holding model memory they may never use.",
+			Component:   "toggle",
+			Order:       90,
+		},
 
 		// --- Functions ---
 		"function.grammar.parallel_calls": {

core/config/model_config.go

@@ -656,6 +656,18 @@ type Pipeline struct {
 	// to benefit. A client session.update still overrides type and eagerness
 	// per session; retranscribe is server-side only. Unset keeps server_vad.
 	TurnDetection PipelineTurnDetection `yaml:"turn_detection,omitempty" json:"turn_detection,omitempty"`
+
+	// DisableWarmup turns off eager pre-loading of the pipeline's sub-models at
+	// realtime session start. By default (false) LocalAI loads every configured
+	// sub-model backend (VAD, transcription, LLM, TTS, sound detection, voice
+	// recognition) into memory (concurrently) before the
+	// session is announced and blocks until they are ready, so the first turn
+	// pays no cold-start cost and a model that fails to load surfaces as an error
+	// at session start rather than mid-call. Set true to restore the lazy "load
+	// on first use" behavior — session start no longer blocks on loading and
+	// load errors surface on first use instead (e.g. to keep idle sessions from
+	// holding model memory they may never use).
+	DisableWarmup bool `yaml:"disable_warmup,omitempty" json:"disable_warmup,omitempty"`
 }
 
 // PipelineCompaction configures summarize-then-drop for a realtime pipeline.

core/config/model_config_loader.go

@@ -155,6 +155,25 @@ func (bcl *ModelConfigLoader) LoadModelConfigFileByNameDefaultOptions(modelName
 		ModelPath(appConfig.SystemState.Model.ModelsPath))
 }
 
+// LoadResolvedModelConfig loads a model config by name and follows a single
+// alias hop, so a caller that references an alias (e.g. a pipeline with
+// `llm: default`) gets the alias target's full config (Backend, Model, ...)
+// rather than the alias stub with an empty Backend. Without this the alias
+// survives unresolved into model loading and fails downstream — notably in
+// distributed mode with "backend name is empty". Mirrors the top-level alias
+// resolution in core/http/middleware/request.go.
+func (bcl *ModelConfigLoader) LoadResolvedModelConfig(modelName, modelPath string) (*ModelConfig, error) {
+	cfg, err := bcl.LoadModelConfigFileByName(modelName, modelPath)
+	if err != nil {
+		return nil, err
+	}
+	resolved, _, err := bcl.ResolveAlias(cfg)
+	if err != nil {
+		return nil, err
+	}
+	return resolved, nil
+}
+
 // This format is currently only used when reading a single file at startup, passed in via ApplicationConfig.ConfigFile
 func (bcl *ModelConfigLoader) LoadMultipleModelConfigsSingleFile(file string, opts ...ConfigLoaderOption) error {
 	bcl.Lock()

core/config/model_config_loader_resolve_test.go

@@ -1,4 +1,4 @@
-package openai
+package config_test
 
 import (
 	"os"
@@ -10,14 +10,14 @@ import (
 	"github.com/mudler/LocalAI/core/config"
 )
 
-// loadPipelineSubModel must resolve a pipeline sub-model that references an
-// alias (e.g. `llm: default`) one hop to the alias target's full config — so
-// the effective backend is the target's backend, not the empty backend of the
-// alias stub. This mirrors the top-level alias resolution done in
-// core/http/middleware/request.go, which the realtime pipeline previously
+// LoadResolvedModelConfig must resolve a model that references an alias
+// (e.g. a pipeline with `llm: default`) one hop to the alias target's full
+// config — so the effective backend is the target's backend, not the empty
+// backend of the alias stub. This mirrors the top-level alias resolution done
+// in core/http/middleware/request.go, which the realtime pipeline previously
 // skipped (failing in distributed mode with "backend name is empty").
-var _ = Describe("loadPipelineSubModel", func() {
-	It("resolves a sub-model alias one hop to the target's config", func() {
+var _ = Describe("LoadResolvedModelConfig", func() {
+	It("resolves an alias one hop to the target's config", func() {
 		tmpDir := GinkgoT().TempDir()
 
 		// A real model config with a concrete backend.
@@ -38,13 +38,13 @@ alias: real-llm
 		Expect(cl.LoadModelConfigsFromPath(tmpDir)).To(Succeed())
 
 		// Resolving the alias must follow the hop to the target's full config.
-		resolved, err := loadPipelineSubModel(cl, "default", tmpDir)
+		resolved, err := cl.LoadResolvedModelConfig("default", tmpDir)
 		Expect(err).NotTo(HaveOccurred())
 		Expect(resolved.IsAlias()).To(BeFalse())
 		Expect(resolved.Backend).To(Equal("llama-cpp"))
 
 		// A non-alias name must load unchanged.
-		direct, err := loadPipelineSubModel(cl, "real-llm", tmpDir)
+		direct, err := cl.LoadResolvedModelConfig("real-llm", tmpDir)
 		Expect(err).NotTo(HaveOccurred())
 		Expect(direct.Backend).To(Equal("llama-cpp"))
 		Expect(direct.Name).To(Equal("real-llm"))

core/http/endpoints/localai/backend.go

@@ -65,6 +65,10 @@ type BackendEndpointService struct {
 
 type GalleryBackend struct {
 	ID string `json:"id"`
+	// Force reinstalls the backend even when it is already installed and
+	// runnable. Off by default so apply stays idempotent for supervising
+	// apps that ensure their backend on every boot.
+	Force bool `json:"force"`
 }
 
 func CreateBackendEndpointService(galleries []config.Gallery, systemState *system.SystemState, backendApplier *galleryop.GalleryService, upgradeChecker UpgradeInfoProvider) BackendEndpointService {
@@ -103,7 +107,9 @@ func (mgs *BackendEndpointService) GetAllStatusEndpoint() echo.HandlerFunc {
 	}
 }
 
-// ApplyBackendEndpoint installs a new backend to a LocalAI instance
+// ApplyBackendEndpoint installs a new backend to a LocalAI instance. The op is
+// idempotent: an already-installed, runnable backend is left alone unless the
+// request sets "force": true (explicit reinstall).
 // @Summary Install backends to LocalAI.
 // @Tags backends
 // @Param request body GalleryBackend true "query params"
@@ -137,6 +143,7 @@ func (mgs *BackendEndpointService) ApplyBackendEndpoint(systemState *system.Syst
 			ID:                 uuid.String(),
 			GalleryElementName: input.ID,
 			Galleries:          mgs.galleries,
+			Force:              input.Force,
 		}
 
 		return c.JSON(200, schema.BackendResponse{ID: uuid.String(), StatusURL: fmt.Sprintf("%sbackends/jobs/%s", middleware.BaseURL(c), uuid.String())})

core/http/endpoints/localai/backend_apply_test.go

@@ -0,0 +1,87 @@
+package localai_test
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"strings"
+
+	"github.com/labstack/echo/v4"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/gallery"
+	. "github.com/mudler/LocalAI/core/http/endpoints/localai"
+	"github.com/mudler/LocalAI/core/services/galleryop"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/system"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// POST /backends/apply must be idempotent by default: supervising apps call it
+// on every boot to ensure a backend exists, and forcing a reinstall there
+// re-downloads the whole artifact each time. Reinstall stays available behind
+// the explicit force flag.
+var _ = Describe("POST /backends/apply force plumbing", func() {
+	var (
+		app     *echo.Echo
+		gs      *galleryop.GalleryService
+		tmpDir  string
+		received chan galleryop.ManagementOp[gallery.GalleryBackend, any]
+	)
+
+	BeforeEach(func() {
+		app = echo.New()
+
+		var err error
+		tmpDir, err = os.MkdirTemp("", "backends-apply-test-*")
+		Expect(err).NotTo(HaveOccurred())
+
+		systemState, err := system.GetSystemState(system.WithBackendPath(tmpDir))
+		Expect(err).NotTo(HaveOccurred())
+		appConfig := &config.ApplicationConfig{SystemState: systemState}
+
+		// The service is deliberately not started: the test reads the op off
+		// the (unbuffered) channel itself.
+		gs = galleryop.NewGalleryService(appConfig, model.NewModelLoader(systemState))
+		svc := CreateBackendEndpointService(nil, systemState, gs, nil)
+		app.POST("/backends/apply", svc.ApplyBackendEndpoint(systemState))
+
+		received = make(chan galleryop.ManagementOp[gallery.GalleryBackend, any], 1)
+		go func() {
+			op := <-gs.BackendGalleryChannel
+			received <- op
+		}()
+	})
+
+	AfterEach(func() {
+		Expect(os.RemoveAll(tmpDir)).To(Succeed())
+	})
+
+	apply := func(body string) *httptest.ResponseRecorder {
+		req := httptest.NewRequest(http.MethodPost, "/backends/apply", strings.NewReader(body))
+		req.Header.Set(echo.HeaderContentType, echo.MIMEApplicationJSON)
+		rec := httptest.NewRecorder()
+		app.ServeHTTP(rec, req)
+		return rec
+	}
+
+	It("enqueues a non-forced op by default", func() {
+		rec := apply(`{"id":"llama-cpp"}`)
+		Expect(rec.Code).To(Equal(http.StatusOK))
+
+		var op galleryop.ManagementOp[gallery.GalleryBackend, any]
+		Eventually(received).Should(Receive(&op))
+		Expect(op.GalleryElementName).To(Equal("llama-cpp"))
+		Expect(op.Force).To(BeFalse())
+	})
+
+	It("enqueues a forced op when the request sets force", func() {
+		rec := apply(`{"id":"llama-cpp","force":true}`)
+		Expect(rec.Code).To(Equal(http.StatusOK))
+
+		var op galleryop.ManagementOp[gallery.GalleryBackend, any]
+		Eventually(received).Should(Receive(&op))
+		Expect(op.GalleryElementName).To(Equal("llama-cpp"))
+		Expect(op.Force).To(BeTrue())
+	})
+})

core/http/endpoints/localai/backend_load.go

@@ -0,0 +1,54 @@
+package localai
+
+import (
+	"net/http"
+
+	"github.com/labstack/echo/v4"
+	"github.com/mudler/LocalAI/core/backend"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/xlog"
+)
+
+// LoadModelEndpoint pre-loads a model into memory by name — the inverse of
+// /backend/shutdown. For a realtime pipeline model every configured sub-model
+// (VAD, transcription, LLM, TTS, sound_detection, voice_recognition) is loaded; for a regular
+// model its own backend is loaded. The call blocks until loading finishes so
+// clients can drive warm-up explicitly and learn up front whether a model
+// fails to load.
+// @Summary Pre-load a model into memory
+// @Description Loads the named model (or, for a realtime pipeline, all of its sub-models) into memory so subsequent requests pay no cold-start cost. The inverse of /backend/shutdown.
+// @Tags monitoring
+// @Accept json
+// @Produce json
+// @Param request body schema.ModelLoadRequest true "Model to load"
+// @Success 200 {object} schema.ModelLoadResponse "Model loaded"
+// @Failure 400 {object} schema.ModelLoadResponse "Missing model name"
+// @Failure 500 {object} schema.ModelLoadResponse "Load failed (Loaded lists any sub-models that did load)"
+// @Router /backend/load [post]
+func LoadModelEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) echo.HandlerFunc {
+	return func(c echo.Context) error {
+		input := new(schema.ModelLoadRequest)
+		if err := c.Bind(input); err != nil {
+			return err
+		}
+		if input.Model == "" {
+			return c.JSON(http.StatusBadRequest, schema.ModelLoadResponse{Message: "model is required"})
+		}
+
+		loaded, err := backend.PreloadModelByName(c.Request().Context(), cl, ml, appConfig, input.Model)
+		if err != nil {
+			xlog.Error("failed to pre-load model", "model", input.Model, "loaded", loaded, "error", err)
+			return c.JSON(http.StatusInternalServerError, schema.ModelLoadResponse{
+				Loaded:  loaded,
+				Message: "failed to load model: " + err.Error(),
+			})
+		}
+
+		return c.JSON(http.StatusOK, schema.ModelLoadResponse{
+			Loaded:  loaded,
+			Message: "model loaded",
+		})
+	}
+}

core/http/endpoints/localai/backend_load_test.go

@@ -0,0 +1,102 @@
+package localai_test
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+
+	"github.com/labstack/echo/v4"
+	"github.com/mudler/LocalAI/core/config"
+	. "github.com/mudler/LocalAI/core/http/endpoints/localai"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/system"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("LoadModelEndpoint (/backend/load)", func() {
+	var (
+		app          *echo.Echo
+		tempDir      string
+		configLoader *config.ModelConfigLoader
+		modelLoader  *model.ModelLoader
+		appConfig    *config.ApplicationConfig
+	)
+
+	post := func(body string) *httptest.ResponseRecorder {
+		req := httptest.NewRequest(http.MethodPost, "/backend/load", bytes.NewBufferString(body))
+		req.Header.Set(echo.HeaderContentType, echo.MIMEApplicationJSON)
+		rec := httptest.NewRecorder()
+		app.ServeHTTP(rec, req)
+		return rec
+	}
+
+	decode := func(rec *httptest.ResponseRecorder) schema.ModelLoadResponse {
+		var resp schema.ModelLoadResponse
+		Expect(json.Unmarshal(rec.Body.Bytes(), &resp)).To(Succeed())
+		return resp
+	}
+
+	writeConfig := func(name, contents string) {
+		Expect(os.WriteFile(filepath.Join(tempDir, name+".yaml"), []byte(contents), 0o600)).To(Succeed())
+	}
+
+	BeforeEach(func() {
+		var err error
+		tempDir, err = os.MkdirTemp("", "backend-load-test-*")
+		Expect(err).NotTo(HaveOccurred())
+
+		systemState, err := system.GetSystemState(system.WithModelPath(tempDir))
+		Expect(err).NotTo(HaveOccurred())
+
+		appConfig = config.NewApplicationConfig(config.WithSystemState(systemState))
+		configLoader = config.NewModelConfigLoader(tempDir)
+		modelLoader = model.NewModelLoader(systemState) // no backends installed
+
+		app = echo.New()
+		app.POST("/backend/load", LoadModelEndpoint(configLoader, modelLoader, appConfig))
+	})
+
+	AfterEach(func() {
+		_ = os.RemoveAll(tempDir)
+	})
+
+	It("rejects a request with no model name", func() {
+		rec := post(`{}`)
+		Expect(rec.Code).To(Equal(http.StatusBadRequest))
+		Expect(decode(rec).Message).To(ContainSubstring("model is required"))
+	})
+
+	It("reports a load failure for a regular model with nothing loaded", func() {
+		writeConfig("solo", "name: solo\n")
+
+		rec := post(`{"model":"solo"}`)
+		Expect(rec.Code).To(Equal(http.StatusInternalServerError))
+
+		resp := decode(rec)
+		Expect(resp.Loaded).To(BeEmpty())
+		Expect(resp.Message).To(ContainSubstring("failed to load model"))
+	})
+
+	It("expands a pipeline model and reports each sub-model that failed to load", func() {
+		writeConfig("voicebot", "name: voicebot\npipeline:\n  vad: vad-m\n  transcription: stt-m\n  llm: llm-m\n  tts: tts-m\n")
+		writeConfig("vad-m", "name: vad-m\n")
+		writeConfig("stt-m", "name: stt-m\n")
+		writeConfig("llm-m", "name: llm-m\n")
+		writeConfig("tts-m", "name: tts-m\n")
+
+		rec := post(`{"model":"voicebot"}`)
+		Expect(rec.Code).To(Equal(http.StatusInternalServerError))
+
+		resp := decode(rec)
+		Expect(resp.Message).To(ContainSubstring("failed to load model"))
+		// The pipeline stub itself is never loaded; its sub-models are what the
+		// endpoint tries, so the error names them rather than "voicebot".
+		Expect(resp.Message).To(ContainSubstring("vad-m"))
+		Expect(resp.Message).ToNot(ContainSubstring("voicebot"))
+	})
+})

core/http/endpoints/mcp/localai_assistant_test.go

@@ -51,6 +51,9 @@ func (stubClient) EditModelConfig(_ context.Context, _ string, _ map[string]any)
 	return nil
 }
 func (stubClient) ReloadModels(_ context.Context) error { return nil }
+func (stubClient) LoadModel(_ context.Context, model string) ([]string, error) {
+	return []string{model}, nil
+}
 func (stubClient) SetAlias(_ context.Context, _, _ string) error {
 	return nil
 }

core/http/endpoints/openai/realtime.go

@@ -7,6 +7,7 @@ import (
 	"encoding/binary"
 	"encoding/hex"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"math"
 	"os"
@@ -266,6 +267,12 @@ type Model interface {
 	// grpcerrors.IsLiveTranscriptionUnsupported.
 	TranscribeLive(ctx context.Context, language string, onEvent func(backend.LiveTranscriptionEvent)) (backend.LiveTranscriptionSession, error)
 	PredictConfig() *config.ModelConfig
+	// Warmup eagerly loads the pipeline's sub-model backends into memory so the
+	// first realtime turn doesn't pay each backend's cold-start load cost. Loads
+	// run concurrently; Warmup blocks until they all finish and returns a joined
+	// error naming every stage that failed to load (nil if all succeeded), so a
+	// caller can surface model-load failures at session start instead of mid-call.
+	Warmup(ctx context.Context) error
 }
 
 var upgrader = websocket.Upgrader{
@@ -583,18 +590,8 @@ func runRealtimeSession(application *application.Application, t Transport, model
 	}
 	session.ModelInterface = m
 
-	if session.SummaryModel != "" {
-		summaryModelName := session.SummaryModel
-		sid := sessionID
-		session.summarizerFactory = func() (Model, error) {
-			summaryCfg, lerr := application.ModelConfigLoader().LoadModelConfigFileByNameDefaultOptions(summaryModelName, application.ApplicationConfig())
-			if lerr != nil {
-				return nil, fmt.Errorf("load summary model config %q: %w", summaryModelName, lerr)
-			}
-			return newModel(&summaryCfg.Pipeline, application.ModelConfigLoader(), application.ModelLoader(), application.ApplicationConfig(), evaluator, buildRealtimeRoutingContext(application, sid))
-		}
-	}
-
+	// The voice gate is built before the warm-up below so its
+	// speaker-recognition model can warm alongside the pipeline stages.
 	if cfg.Pipeline.VoiceGateEnabled() {
 		gate, gerr := newVoiceGate(
 			*cfg.Pipeline.VoiceRecognition,
@@ -612,6 +609,47 @@ func runRealtimeSession(application *application.Application, t Transport, model
 		xlog.Info("realtime voice recognition gate enabled", "mode", gate.cfg.Mode, "when", gate.cfg.When)
 	}
 
+	// Warm the pipeline's sub-model backends before announcing the session.
+	// Loads run concurrently but we block here until they all finish, so a model
+	// that fails to load (missing weights, bad backend, OOM) surfaces as an error
+	// at session start rather than stalling — or failing — mid-call on the first
+	// turn (VAD on the first audio chunk, STT at end-of-speech, LLM on the first
+	// reply, TTS on the first spoken output). On success the backends are already
+	// resident, so the first turn pays no cold-start cost. Opt out per pipeline
+	// with `pipeline.disable_warmup: true` to restore lazy load-on-first-use
+	// (errors then surface on first use instead of at session start).
+	if !cfg.Pipeline.DisableWarmup {
+		warmErr := make(chan error, 1)
+		go func() { warmErr <- m.Warmup(context.Background()) }()
+		// The voice-gate model warms concurrently with the pipeline stages: an
+		// enforced gate blocks each utterance on speaker resolution, so its
+		// cold-start would otherwise land on the first turn too. (Compaction's
+		// summary_model stays lazy — it only runs off the response path.)
+		var gateErr error
+		if session.voiceGate != nil {
+			_, gateErr = backend.PreloadStages(context.Background(), application.ModelLoader(), application.ApplicationConfig(), []backend.PreloadStage{
+				{Role: "voice_recognition", Cfg: session.voiceGate.recCfg},
+			})
+		}
+		if err := errors.Join(<-warmErr, gateErr); err != nil {
+			xlog.Error("realtime warmup failed", "model", model, "error", err)
+			sendError(t, "model_load_error", "Failed to load pipeline models: "+err.Error(), "", "")
+			return
+		}
+	}
+
+	if session.SummaryModel != "" {
+		summaryModelName := session.SummaryModel
+		sid := sessionID
+		session.summarizerFactory = func() (Model, error) {
+			summaryCfg, lerr := application.ModelConfigLoader().LoadModelConfigFileByNameDefaultOptions(summaryModelName, application.ApplicationConfig())
+			if lerr != nil {
+				return nil, fmt.Errorf("load summary model config %q: %w", summaryModelName, lerr)
+			}
+			return newModel(&summaryCfg.Pipeline, application.ModelConfigLoader(), application.ModelLoader(), application.ApplicationConfig(), evaluator, buildRealtimeRoutingContext(application, sid))
+		}
+	}
+
 	// Store the session and notify the transport (for WebRTC audio track handling)
 	sessionLock.Lock()
 	sessions[sessionID] = session
@@ -1125,6 +1163,21 @@ func updateSession(session *Session, update *types.SessionUnion, cl *config.Mode
 			return err
 		}
 		session.ModelInterface = m
+		// A session.update that swaps the model/voice rebuilds the pipeline, so
+		// warm the new backends too (unless opted out) — otherwise the next turn
+		// pays the cold-start load the original session warm-up already avoided.
+		// Unlike session start this stays non-blocking: updateSession runs under
+		// the global sessionLock, so blocking on a multi-second load here would
+		// stall every other session. Load errors are logged (and still surface on
+		// first use); per-stage failures are already warned inside
+		// backend.PreloadStages.
+		if !session.ModelConfig.Pipeline.DisableWarmup {
+			go func() {
+				if err := m.Warmup(context.Background()); err != nil {
+					xlog.Error("realtime warmup failed after session.update", "error", err)
+				}
+			}()
+		}
 	}
 
 	if rt.Audio != nil && rt.Audio.Input != nil && rt.Audio.Input.TurnDetectionSet {

core/http/endpoints/openai/realtime_doubles_test.go

@@ -174,6 +174,8 @@ func (m *fakeModel) TranscribeLive(_ context.Context, _ string, onEvent func(bac
 
 func (m *fakeModel) PredictConfig() *config.ModelConfig { return m.cfg }
 
+func (m *fakeModel) Warmup(ctx context.Context) error { return nil }
+
 // fakeLiveSession records what semantic_vad fed and closed; closeEvents are
 // replayed through onEvent during Close, mimicking the backend's finalize
 // flush (trailing delta + Final) landing before Close returns.

core/http/endpoints/openai/realtime_model.go

@@ -110,6 +110,15 @@ func (m *transcriptOnlyModel) PredictConfig() *config.ModelConfig {
 	return nil
 }
 
+func (m *transcriptOnlyModel) Warmup(ctx context.Context) error {
+	_, err := backend.PreloadStages(ctx, m.modelLoader, m.appConfig, []backend.PreloadStage{
+		{Role: "vad", Cfg: m.VADConfig},
+		{Role: "transcription", Cfg: m.TranscriptionConfig},
+		{Role: "sound_detection", Cfg: m.SoundDetectionConfig},
+	})
+	return err
+}
+
 func (m *wrappedModel) VAD(ctx context.Context, request *schema.VADRequest) (*schema.VADResponse, error) {
 	return backend.VAD(request, ctx, m.modelLoader, m.appConfig, *m.VADConfig)
 }
@@ -360,6 +369,17 @@ func (m *wrappedModel) PredictConfig() *config.ModelConfig {
 	return m.LLMConfig
 }
 
+func (m *wrappedModel) Warmup(ctx context.Context) error {
+	_, err := backend.PreloadStages(ctx, m.modelLoader, m.appConfig, []backend.PreloadStage{
+		{Role: "vad", Cfg: m.VADConfig},
+		{Role: "transcription", Cfg: m.TranscriptionConfig},
+		{Role: "llm", Cfg: m.LLMConfig},
+		{Role: "tts", Cfg: m.TTSConfig},
+		{Role: "sound_detection", Cfg: m.SoundDetectionConfig},
+	})
+	return err
+}
+
 // wavStreamHeaderBytes is the size of the WAV header that backend.ModelTTSStream
 // emits as its first audio callback; the sample rate lives at byte offset 24.
 const wavStreamHeaderBytes = 44
@@ -440,7 +460,7 @@ func loadSoundDetectionConfig(pipeline *config.Pipeline, cl *config.ModelConfigL
 	if pipeline.SoundDetection == "" {
 		return nil, nil
 	}
-	cfg, err := loadPipelineSubModel(cl, pipeline.SoundDetection, ml.ModelPath)
+	cfg, err := cl.LoadResolvedModelConfig(pipeline.SoundDetection, ml.ModelPath)
 	if err != nil {
 		return nil, fmt.Errorf("failed to load sound detection config: %w", err)
 	}
@@ -451,7 +471,7 @@ func loadSoundDetectionConfig(pipeline *config.Pipeline, cl *config.ModelConfigL
 }
 
 func newTranscriptionOnlyModel(pipeline *config.Pipeline, cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) (Model, *config.ModelConfig, error) {
-	cfgVAD, err := loadPipelineSubModel(cl, pipeline.VAD, ml.ModelPath)
+	cfgVAD, err := cl.LoadResolvedModelConfig(pipeline.VAD, ml.ModelPath)
 	if err != nil {
 
 		return nil, nil, fmt.Errorf("failed to load backend config: %w", err)
@@ -461,7 +481,7 @@ func newTranscriptionOnlyModel(pipeline *config.Pipeline, cl *config.ModelConfig
 		return nil, nil, fmt.Errorf("failed to validate config: %w", err)
 	}
 
-	cfgSST, err := loadPipelineSubModel(cl, pipeline.Transcription, ml.ModelPath)
+	cfgSST, err := cl.LoadResolvedModelConfig(pipeline.Transcription, ml.ModelPath)
 	if err != nil {
 
 		return nil, nil, fmt.Errorf("failed to load backend config: %w", err)
@@ -550,30 +570,11 @@ func buildRealtimeRoutingContext(a *application.Application, sessionID string) *
 	}
 }
 
-// loadPipelineSubModel loads a pipeline sub-model config by name and follows a
-// single alias hop, so a pipeline that references an alias (e.g. `llm: default`)
-// gets the alias target's full config (Backend, Model, ...) rather than the
-// alias stub with an empty Backend. Without this the alias survives unresolved
-// into model loading and fails downstream — notably in distributed mode with
-// "backend name is empty". Mirrors the top-level alias resolution in
-// core/http/middleware/request.go.
-func loadPipelineSubModel(cl *config.ModelConfigLoader, name, modelPath string) (*config.ModelConfig, error) {
-	cfg, err := cl.LoadModelConfigFileByName(name, modelPath)
-	if err != nil {
-		return nil, err
-	}
-	resolved, _, err := cl.ResolveAlias(cfg)
-	if err != nil {
-		return nil, err
-	}
-	return resolved, nil
-}
-
 // returns and loads either a wrapped model or a model that support audio-to-audio
 func newModel(pipeline *config.Pipeline, cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig, evaluator *templates.Evaluator, routing *RealtimeRoutingContext) (Model, error) {
 	xlog.Debug("Creating new model pipeline model", "pipeline", pipeline)
 
-	cfgVAD, err := loadPipelineSubModel(cl, pipeline.VAD, ml.ModelPath)
+	cfgVAD, err := cl.LoadResolvedModelConfig(pipeline.VAD, ml.ModelPath)
 	if err != nil {
 
 		return nil, fmt.Errorf("failed to load backend config: %w", err)
@@ -584,7 +585,7 @@ func newModel(pipeline *config.Pipeline, cl *config.ModelConfigLoader, ml *model
 	}
 
 	// TODO: Do we always need a transcription model? It can be disabled. Note that any-to-any instruction following models don't transcribe as such, so if transcription is required it is a separate process
-	cfgSST, err := loadPipelineSubModel(cl, pipeline.Transcription, ml.ModelPath)
+	cfgSST, err := cl.LoadResolvedModelConfig(pipeline.Transcription, ml.ModelPath)
 	if err != nil {
 
 		return nil, fmt.Errorf("failed to load backend config: %w", err)
@@ -616,7 +617,7 @@ func newModel(pipeline *config.Pipeline, cl *config.ModelConfigLoader, ml *model
 	xlog.Debug("Loading a wrapped model")
 
 	// Otherwise we want to return a wrapped model, which is a "virtual" model that re-uses other models to perform operations
-	cfgLLM, err := loadPipelineSubModel(cl, pipeline.LLM, ml.ModelPath)
+	cfgLLM, err := cl.LoadResolvedModelConfig(pipeline.LLM, ml.ModelPath)
 	if err != nil {
 
 		return nil, fmt.Errorf("failed to load backend config: %w", err)
@@ -631,7 +632,7 @@ func newModel(pipeline *config.Pipeline, cl *config.ModelConfigLoader, ml *model
 	applyPipelineReasoning(cfgLLM, *pipeline)
 	applyPipelineThinking(cfgLLM, *pipeline)
 
-	cfgTTS, err := loadPipelineSubModel(cl, pipeline.TTS, ml.ModelPath)
+	cfgTTS, err := cl.LoadResolvedModelConfig(pipeline.TTS, ml.ModelPath)
 	if err != nil {
 
 		return nil, fmt.Errorf("failed to load backend config: %w", err)

core/http/endpoints/openai/realtime_voicegate.go

@@ -21,6 +21,7 @@ type namedEmbedding struct {
 // drive the realtime pipeline.
 type voiceGate struct {
 	cfg       config.PipelineVoiceRecognition // normalized
+	recCfg    *config.ModelConfig             // resolved speaker-recognition model, for warm-up
 	registry  voicerecognition.Registry       // identify mode (nil otherwise)
 	refEmbeds []namedEmbedding                // verify mode, pre-embedded refs
 	refAudios []config.VoiceReference         // verify + anti-spoofing: ref paths
@@ -72,7 +73,9 @@ func newVoiceGate(
 		return nil, err
 	}
 
-	recCfg, err := cl.LoadModelConfigFileByName(cfg.Model, ml.ModelPath)
+	// Resolved like every other pipeline sub-model (one alias hop), so an
+	// aliased voice_recognition model gets its target's backend.
+	recCfg, err := cl.LoadResolvedModelConfig(cfg.Model, ml.ModelPath)
 	if err != nil {
 		return nil, fmt.Errorf("voice_recognition: failed to load model %q: %w", cfg.Model, err)
 	}
@@ -82,6 +85,7 @@ func newVoiceGate(
 
 	g := &voiceGate{
 		cfg:      cfg,
+		recCfg:   recCfg,
 		registry: registry,
 		embedFn: func(ctx context.Context, wavPath string) ([]float32, error) {
 			res, err := backend.VoiceEmbed(ctx, wavPath, ml, appConfig, *recCfg)

core/http/endpoints/openai/realtime_warmup_test.go

@@ -0,0 +1,64 @@
+package openai
+
+import (
+	"context"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/system"
+)
+
+// Warmup delegates to backend.PreloadStages (its concurrency, nil-skipping and
+// error-joining semantics are pinned in core/backend). These specs pin the
+// wiring instead: each realtime model type must warm exactly its configured
+// stages under the right pipeline-role labels. No backends are installed, so
+// every attempted stage fails to load — the joined error is the proof of which
+// stages were attempted and how they were labeled.
+var _ = Describe("realtime model Warmup wiring", func() {
+	newLoader := func() (*model.ModelLoader, *config.ApplicationConfig) {
+		systemState, err := system.GetSystemState(system.WithModelPath(GinkgoT().TempDir()))
+		Expect(err).ToNot(HaveOccurred())
+		appConfig := config.NewApplicationConfig(config.WithSystemState(systemState))
+		return model.NewModelLoader(systemState), appConfig
+	}
+
+	It("wrappedModel warms every configured stage under its pipeline role", func() {
+		ml, appConfig := newLoader()
+		m := &wrappedModel{
+			VADConfig:            &config.ModelConfig{Name: "vad-m"},
+			TranscriptionConfig:  &config.ModelConfig{Name: "stt-m"},
+			LLMConfig:            &config.ModelConfig{Name: "llm-m"},
+			TTSConfig:            &config.ModelConfig{Name: "tts-m"},
+			SoundDetectionConfig: &config.ModelConfig{Name: "ced-m"},
+			modelLoader:          ml,
+			appConfig:            appConfig,
+		}
+
+		err := m.Warmup(context.Background())
+		Expect(err).To(HaveOccurred())
+		for _, stage := range []string{"vad (vad-m)", "transcription (stt-m)", "llm (llm-m)", "tts (tts-m)", "sound_detection (ced-m)"} {
+			Expect(err.Error()).To(ContainSubstring(stage))
+		}
+	})
+
+	It("transcriptOnlyModel warms its stages and skips absent ones", func() {
+		ml, appConfig := newLoader()
+		m := &transcriptOnlyModel{
+			VADConfig:           &config.ModelConfig{Name: "vad-m"},
+			TranscriptionConfig: &config.ModelConfig{Name: "stt-m"},
+			// SoundDetectionConfig nil: an absent stage must be skipped, not
+			// fail the warm-up.
+			modelLoader: ml,
+			appConfig:   appConfig,
+		}
+
+		err := m.Warmup(context.Background())
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("vad (vad-m)"))
+		Expect(err.Error()).To(ContainSubstring("transcription (stt-m)"))
+		Expect(err.Error()).ToNot(ContainSubstring("sound_detection"))
+	})
+})

core/http/openresponses_test.go

@@ -7,6 +7,7 @@ import (
 	"io"
 	"net/http"
 	"os"
+	"path/filepath"
 	"strings"
 	"time"
 
@@ -29,6 +30,8 @@ const testModel = "Qwen3-VL-2B-Instruct-Q4_K_M"
 
 var _ = Describe("Open Responses API", func() {
 	var app *echo.Echo
+	var localApp *application.Application
+	var localModelDir string
 	var c context.Context
 	var cancel context.CancelFunc
 
@@ -38,28 +41,47 @@ var _ = Describe("Open Responses API", func() {
 
 	Context("API with ephemeral models", func() {
 		BeforeEach(func(sc SpecContext) {
-			var err error
+			// This suite exercises the /v1/responses HTTP/protocol contract
+			// (Content-Type, SSE framing, response envelope, error shapes),
+			// not real inference — so it runs against the same prebuilt
+			// mock-backend the rest of the http suite uses instead of
+			// downloading a real model. Skip cleanly when it isn't built.
+			if mockBackendPath == "" {
+				Skip("mock-backend binary not built; run 'make build-mock-backend'")
+			}
 
-			backendPath := os.Getenv("BACKENDS_PATH")
+			var err error
 
 			c, cancel = context.WithCancel(context.Background())
 
+			// Isolated model dir carrying a single config named after testModel
+			// but served by the mock backend, so the responses endpoint can
+			// resolve and load the model without any real backend build.
+			localModelDir, err = os.MkdirTemp("", "openresponses-models-")
+			Expect(err).ToNot(HaveOccurred())
+
+			mockModelYAML := "name: " + testModel + "\n" +
+				"backend: mock-backend\n" +
+				"parameters:\n" +
+				"  model: mock-model.bin\n"
+			Expect(os.WriteFile(filepath.Join(localModelDir, testModel+".yaml"), []byte(mockModelYAML), 0644)).To(Succeed())
+
 			systemState, err := system.GetSystemState(
-				system.WithBackendPath(backendPath),
-				system.WithModelPath(modelDir),
+				system.WithBackendPath(backendDir),
+				system.WithModelPath(localModelDir),
 			)
 			Expect(err).ToNot(HaveOccurred())
 
-			application, err := application.New(
+			localApp, err = application.New(
 				append(commonOpts,
 					config.WithContext(c),
 					config.WithSystemState(systemState),
 					config.WithApiKeys([]string{apiKey}),
-					config.WithModelsURL("https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF"),
 				)...)
 			Expect(err).ToNot(HaveOccurred())
+			localApp.ModelLoader().SetExternalBackend("mock-backend", mockBackendPath)
 
-			app, err = API(application)
+			app, err = API(localApp)
 			Expect(err).ToNot(HaveOccurred())
 
 			go func() {
@@ -80,14 +102,24 @@ var _ = Describe("Open Responses API", func() {
 		})
 
 		AfterEach(func(sc SpecContext) {
+			// Synchronous app shutdown first — context-cancel cleanup is async
+			// and races test-binary exit, orphaning mock-backend children.
+			if localApp != nil {
+				_ = localApp.Shutdown()
+				localApp = nil
+			}
 			cancel()
 			if app != nil {
 				ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 				defer cancel()
 				err := app.Shutdown(ctx)
 				Expect(err).ToNot(HaveOccurred())
+				app = nil
+			}
+			if localModelDir != "" {
+				_ = os.RemoveAll(localModelDir)
+				localModelDir = ""
 			}
-
 		})
 
 		Context("HTTP Protocol Compliance", func() {
@@ -969,13 +1001,16 @@ var _ = Describe("Open Responses API", func() {
 				Expect(ok).To(BeTrue())
 				Expect(itemID).ToNot(BeEmpty())
 
-				// Now create a new response with item_reference
+				// Now create a new response with item_reference. Per the OpenAI
+				// Responses spec (and this server's parser in
+				// endpoints/openresponses/responses.go) an item_reference carries
+				// the referenced item in the "id" field, not "item_id".
 				reqBody2 := map[string]any{
 					"model": testModel,
 					"input": []any{
 						map[string]any{
-							"type":    "item_reference",
-							"item_id": itemID,
+							"type": "item_reference",
+							"id":   itemID,
 						},
 						map[string]any{
 							"type":    "message",
@@ -1005,8 +1040,8 @@ var _ = Describe("Open Responses API", func() {
 					"model": testModel,
 					"input": []any{
 						map[string]any{
-							"type":    "item_reference",
-							"item_id": "nonexistent_item_id",
+							"type": "item_reference",
+							"id":   "nonexistent_item_id",
 						},
 					},
 				}

core/http/react-ui/e2e/forking-chat.spec.js

@@ -0,0 +1,133 @@
+import { test, expect } from './coverage-fixtures.js'
+
+// Seeds two-message chat into localStorage so we don't need a live model.
+async function seedChat(page, history) {
+  await page.addInitScript((h) => {
+    const chat = {
+      id: 'seed1', name: 'Seeded Chat', model: 'test-model',
+      history: h, systemPrompt: '', mcpMode: false, mcpServers: [],
+      clientMCPServers: [], temperature: null, topP: null, topK: null,
+      tokenUsage: { prompt: 0, completion: 0, total: 0 },
+      contextSize: null, createdAt: Date.now(), updatedAt: Date.now(),
+    }
+    localStorage.setItem('localai_chats_data', JSON.stringify({
+      chats: [chat], activeChatId: 'seed1', lastSaved: Date.now(),
+    }))
+  }, history)
+}
+
+async function mockModels(page) {
+  await page.route('**/api/models/capabilities', (route) => route.fulfill({
+    contentType: 'application/json',
+    body: JSON.stringify({ data: [{ id: 'test-model', capabilities: ['FLAG_CHAT'] }] }),
+  }))
+  await page.route('**/api/operations', (route) => route.fulfill({
+    contentType: 'application/json', body: JSON.stringify({ operations: [] }),
+  }))
+}
+
+const TWO_TURNS = [
+  { role: 'user', content: 'first question' },
+  { role: 'assistant', content: 'first answer' },
+  { role: 'user', content: 'second question' },
+  { role: 'assistant', content: 'second answer' },
+]
+
+test('duplicate creates an independent copy and switches to it', async ({ page }) => {
+  await mockModels(page)
+  await seedChat(page, TWO_TURNS)
+  await page.goto('/app/chat')
+
+  // Open the chats menu (Ctrl/Cmd+K) and duplicate the seeded chat.
+  // Wait for the menu trigger to mount so its global keydown listener is armed
+  // before we dispatch the shortcut.
+  await page.getByTitle('Conversations (Ctrl/Cmd+K)').waitFor()
+  await page.keyboard.press('Control+k')
+  await page.getByTitle('Duplicate chat').first().click()
+
+  // A new active chat named "Seeded Chat (fork)" with the same 4 messages.
+  await expect(page.locator('.chat-header-title')).toHaveText('Seeded Chat (fork)')
+  await expect(page.locator('.chat-message-user')).toHaveCount(2)
+  await expect(page.locator('.chat-message-assistant')).toHaveCount(2)
+})
+
+async function mockCompletion(page, replyText) {
+  await page.route('**/v1/chat/completions', (route) => {
+    const sse =
+      `data: ${JSON.stringify({ choices: [{ delta: { content: replyText } }] })}\n\n` +
+      `data: ${JSON.stringify({ choices: [{ delta: {}, finish_reason: 'stop' }], usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 } })}\n\n` +
+      `data: [DONE]\n\n`
+    route.fulfill({ status: 200, contentType: 'text/event-stream', body: sse })
+  })
+}
+
+test('retry regenerates the first answer and drops the later turn', async ({ page }) => {
+  await mockModels(page)
+  // Capture the outbound request body so we can assert the model receives the
+  // truncated history (not the stale downstream turns).
+  let sentMessages = null
+  await page.route('**/v1/chat/completions', (route) => {
+    sentMessages = route.request().postDataJSON()?.messages || []
+    const sse =
+      `data: ${JSON.stringify({ choices: [{ delta: { content: 'REGENERATED first answer' } }] })}\n\n` +
+      `data: ${JSON.stringify({ choices: [{ delta: {}, finish_reason: 'stop' }], usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 } })}\n\n` +
+      `data: [DONE]\n\n`
+    route.fulfill({ status: 200, contentType: 'text/event-stream', body: sse })
+  })
+  await seedChat(page, TWO_TURNS)
+  await page.goto('/app/chat')
+
+  // Hover the FIRST assistant message and click its retry button.
+  const firstAssistant = page.locator('.chat-message-assistant').first()
+  await firstAssistant.hover()
+  await firstAssistant.getByTitle('Regenerate').click()
+
+  // History is truncated to the first user turn, then the new answer streams in;
+  // the second Q/A turn is gone.
+  await expect(page.locator('.chat-message-assistant')).toContainText(['REGENERATED first answer'])
+  await expect(page.locator('.chat-message-user')).toHaveCount(1)
+  await expect(page.locator('.chat-message-assistant')).toHaveCount(1)
+
+  // The OUTBOUND payload must also be truncated: the resent user turn is present,
+  // but the downstream turn and the stale first answer must be gone.
+  const contents = (sentMessages || []).map(m =>
+    typeof m.content === 'string' ? m.content : JSON.stringify(m.content)
+  )
+  expect(contents.join('\n')).toContain('first question')
+  expect(contents.join('\n')).not.toContain('second question')
+  expect(contents.join('\n')).not.toContain('first answer')
+})
+
+test('copy chat puts the whole conversation on the clipboard', async ({ page, context }) => {
+  await context.grantPermissions(['clipboard-read', 'clipboard-write'])
+  await mockModels(page)
+  await seedChat(page, TWO_TURNS)
+  await page.goto('/app/chat')
+
+  // Wait for the menu trigger to mount so its global keydown listener is armed
+  // before we dispatch the shortcut (same mount-race guard as the duplicate test).
+  await page.getByTitle('Conversations (Ctrl/Cmd+K)').waitFor()
+  await page.keyboard.press('Control+k')
+  await page.getByTitle('Copy chat').first().click()
+
+  const clip = await page.evaluate(() => navigator.clipboard.readText())
+  expect(clip).toContain('# Seeded Chat')
+  expect(clip).toContain('first answer')
+  expect(clip).toContain('second answer')
+})
+
+test('branch from the first answer forks history up to that point', async ({ page }) => {
+  await mockModels(page)
+  await seedChat(page, TWO_TURNS)
+  await page.goto('/app/chat')
+
+  const firstAssistant = page.locator('.chat-message-assistant').first()
+  await firstAssistant.hover()
+  await firstAssistant.getByTitle('Branch from here').click()
+
+  // New active chat "Seeded Chat (fork)" contains only the first Q/A turn.
+  await expect(page.locator('.chat-header-title')).toHaveText('Seeded Chat (fork)')
+  await expect(page.locator('.chat-message-user')).toHaveCount(1)
+  await expect(page.locator('.chat-message-assistant')).toHaveCount(1)
+  await expect(page.locator('.chat-message-assistant')).toContainText(['first answer'])
+})

core/http/react-ui/public/locales/en/chat.json

@@ -72,6 +72,7 @@
   "actions": {
     "copy": "Copy",
     "regenerate": "Regenerate",
+    "branch": "Branch from here",
     "jumpToLatest": "Jump to latest"
   },
   "streaming": {
@@ -100,7 +101,9 @@
   "toasts": {
     "selectModel": "Please select a model",
     "copied": "Copied to clipboard",
-    "copyFailed": "Could not copy to clipboard"
+    "copyFailed": "Could not copy to clipboard",
+    "chatCopied": "Chat copied to clipboard",
+    "forked": "Created a new chat"
   },
   "menu": {
     "trigger": "Chats",
@@ -110,6 +113,8 @@
     "noMatch": "No conversations match your search",
     "noConversations": "No conversations yet",
     "rename": "Rename",
+    "duplicate": "Duplicate chat",
+    "copyChat": "Copy chat",
     "exportMarkdown": "Export as Markdown",
     "deleteChat": "Delete chat",
     "newChat": "New chat",

core/http/react-ui/src/components/ChatsMenu.jsx

@@ -24,6 +24,8 @@ const ChatsMenu = forwardRef(function ChatsMenu({
   onDeleteAll,
   onRename,
   onExport,
+  onCopyChat,
+  onDuplicate,
 }, ref) {
   const { t } = useTranslation('chat')
   const [open, setOpen] = useState(false)
@@ -230,6 +232,24 @@ const ChatsMenu = forwardRef(function ChatsMenu({
                   >
                     <i className="fas fa-pen" />
                   </button>
+                  {onDuplicate && (
+                    <button
+                      type="button"
+                      onClick={(e) => { e.stopPropagation(); onDuplicate(chat); setOpen(false) }}
+                      title={t('menu.duplicate')}
+                    >
+                      <i className="fas fa-clone" />
+                    </button>
+                  )}
+                  {(chat.history?.length || 0) > 0 && onCopyChat && (
+                    <button
+                      type="button"
+                      onClick={(e) => { e.stopPropagation(); onCopyChat(chat) }}
+                      title={t('menu.copyChat')}
+                    >
+                      <i className="fas fa-clipboard" />
+                    </button>
+                  )}
                   {(chat.history?.length || 0) > 0 && onExport && (
                     <button
                       type="button"

core/http/react-ui/src/hooks/useChat.js

@@ -141,6 +141,24 @@ export function useChat(initialModel = '') {
     return chat
   }, [])
 
+  const forkChat = useCallback((chatId, uptoIndex) => {
+    const src = chats.find(c => c.id === chatId)
+    if (!src) return null
+    const end = typeof uptoIndex === 'number' ? uptoIndex : src.history.length
+    const forked = {
+      ...src,
+      id: generateId(),
+      name: `${src.name} (fork)`,
+      history: structuredClone(src.history.slice(0, end)),
+      tokenUsage: { prompt: 0, completion: 0, total: 0 },
+      createdAt: Date.now(),
+      updatedAt: Date.now(),
+    }
+    setChats(prev => [forked, ...prev])
+    setActiveChatId(forked.id)
+    return forked
+  }, [chats])
+
   const switchChat = useCallback((chatId) => {
     setActiveChatId(chatId)
     setStreamingContent('')
@@ -260,8 +278,12 @@ export function useChat(initialModel = '') {
     if (chat?.systemPrompt) {
       messages.push({ role: 'system', content: chat.systemPrompt })
     }
-    // Filter out thinking/reasoning/tool_call/tool_result messages
-    const historyForApi = (chat?.history || []).filter(m =>
+    // Filter out thinking/reasoning/tool_call/tool_result messages.
+    // options.baseHistory lets callers (e.g. mid-conversation retry) pass the
+    // intended truncated history synchronously; the closure `chat` still holds
+    // the stale pre-truncation state because setChats only schedules an update.
+    const baseHistory = options.baseHistory || chat?.history || []
+    const historyForApi = baseHistory.filter(m =>
       m.role !== 'thinking' && m.role !== 'reasoning' && m.role !== 'tool_call' && m.role !== 'tool_result'
     )
     messages.push(...historyForApi, { role: 'user', content: messageContent })
@@ -793,6 +815,7 @@ export function useChat(initialModel = '') {
     tokensPerSecond,
     maxTokensPerSecond,
     addChat,
+    forkChat,
     switchChat,
     deleteChat,
     deleteAllChats,

core/http/react-ui/src/pages/Chat.jsx

@@ -33,7 +33,7 @@ function getLastMessagePreview(chat) {
   return ''
 }
 
-function exportChatAsMarkdown(chat) {
+function serializeChatAsMarkdown(chat) {
   let md = `# ${chat.name}\n\n`
   md += `Model: ${chat.model || 'Unknown'}\n`
   md += `Date: ${new Date(chat.createdAt).toLocaleString()}\n\n---\n\n`
@@ -47,7 +47,11 @@ function exportChatAsMarkdown(chat) {
       md += `<details><summary>Thinking</summary>\n\n${msg.content}\n\n</details>\n\n`
     }
   }
-  const blob = new Blob([md], { type: 'text/markdown' })
+  return md
+}
+
+function downloadChatAsMarkdown(chat) {
+  const blob = new Blob([serializeChatAsMarkdown(chat)], { type: 'text/markdown' })
   const url = URL.createObjectURL(blob)
   const a = document.createElement('a')
   a.href = url
@@ -294,7 +298,7 @@ export default function Chat() {
   const {
     chats, activeChat, activeChatId, isStreaming, streamingChatId, streamingContent,
     streamingReasoning, streamingToolCalls, tokensPerSecond, maxTokensPerSecond,
-    addChat, switchChat, deleteChat, deleteAllChats, renameChat, updateChatSettings,
+    addChat, forkChat, switchChat, deleteChat, deleteAllChats, renameChat, updateChatSettings,
     sendMessage, stopGeneration, clearHistory, getContextUsagePercent, addMessage,
   } = useChat(urlModel || '')
 
@@ -795,34 +799,27 @@ export default function Chat() {
     await sendMessage(msg, files, mcpOptions)
   }, [input, files, activeChat, sendMessage, addToast, getToolsForLLM, isClientTool, executeTool, hasAppUI, getAppResource, getToolDefinition])
 
-  const handleRegenerate = useCallback(async () => {
+  const handleRegenerate = useCallback(async (targetIndex) => {
     if (!activeChat || isStreaming) return
     const history = activeChat.history
-    let lastUserMsg = null
-    let lastUserFiles = null
-    for (let i = history.length - 1; i >= 0; i--) {
-      if (history[i].role === 'user') {
-        lastUserMsg = typeof history[i].content === 'string' ? history[i].content : history[i].content?.[0]?.text || ''
-        lastUserFiles = history[i].files || []
-        break
-      }
+    const end = typeof targetIndex === 'number' ? targetIndex : history.length
+    // Nearest user message at or before the target answer.
+    let userIdx = -1
+    for (let i = Math.min(end, history.length) - 1; i >= 0; i--) {
+      if (history[i].role === 'user') { userIdx = i; break }
     }
-    if (!lastUserMsg) return
-
-    // Remove everything after and including the last user message
-    const newHistory = []
-    let foundLastUser = false
-    for (let i = history.length - 1; i >= 0; i--) {
-      if (!foundLastUser && history[i].role === 'user') {
-        foundLastUser = true
-        continue
-      }
-      if (foundLastUser) {
-        newHistory.unshift(history[i])
-      }
-    }
-    updateChatSettings(activeChat.id, { history: newHistory })
-    await sendMessage(lastUserMsg, lastUserFiles)
+    if (userIdx === -1) return
+    const userMsg = typeof history[userIdx].content === 'string'
+      ? history[userIdx].content
+      : history[userIdx].content?.[0]?.text || ''
+    const userFiles = history[userIdx].files || []
+    // Drop the user turn and everything after it; sendMessage re-appends it.
+    // Thread the truncated history through explicitly: updateChatSettings only
+    // schedules a state update, so sendMessage's closure would otherwise read
+    // the stale pre-truncation history for the outbound API payload.
+    const baseHistory = history.slice(0, userIdx)
+    updateChatSettings(activeChat.id, { history: baseHistory })
+    await sendMessage(userMsg, userFiles, { baseHistory })
   }, [activeChat, isStreaming, sendMessage, updateChatSettings])
 
   const handleKeyDown = (e) => {
@@ -852,6 +849,11 @@ export default function Chat() {
     }
   }
 
+  const copyChatAsMarkdown = async (chat) => {
+    const ok = await copyToClipboard(serializeChatAsMarkdown(chat))
+    addToast(ok ? t('toasts.chatCopied') : t('toasts.copyFailed'), ok ? 'success' : 'error', ok ? 2000 : 3000)
+  }
+
   const contextPercent = getContextUsagePercent()
 
   // Recent chats for the empty state — exclude the current chat and any
@@ -892,7 +894,9 @@ export default function Chat() {
             onDelete={deleteChat}
             onDeleteAll={promptDeleteAll}
             onRename={renameChat}
-            onExport={(chat) => exportChatAsMarkdown(chat)}
+            onExport={(chat) => downloadChatAsMarkdown(chat)}
+            onCopyChat={(chat) => copyChatAsMarkdown(chat)}
+            onDuplicate={(chat) => { if (forkChat(chat.id)) addToast(t('toasts.forked'), 'success', 2000) }}
           />
           {activeChat.localaiAssistant && (
             <span
@@ -1184,11 +1188,19 @@ export default function Chat() {
                       <button onClick={() => copyMessage(msg.content)} title={t('actions.copy')}>
                         <i className="fas fa-copy" />
                       </button>
-                      {msg.role === 'assistant' && i === activeChat.history.length - 1 && !isStreaming && (
-                        <button onClick={handleRegenerate} title={t('actions.regenerate')}>
+                      {msg.role === 'assistant' && !isStreaming && (
+                        <button onClick={() => handleRegenerate(i)} title={t('actions.regenerate')}>
                           <i className="fas fa-rotate" />
                         </button>
                       )}
+                      {msg.role === 'assistant' && !isStreaming && (
+                        <button
+                          onClick={() => { forkChat(activeChat.id, i + 1); addToast(t('toasts.forked'), 'success', 2000) }}
+                          title={t('actions.branch')}
+                        >
+                          <i className="fas fa-code-branch" />
+                        </button>
+                      )}
                     </div>
                   </div>
                 </div>

core/http/react-ui/src/pages/Manage.jsx

@@ -146,6 +146,7 @@ export default function Manage() {
   const [distributedMode, setDistributedMode] = useState(false)
   const [togglingModels, setTogglingModels] = useState(new Set())
   const [pinningModels, setPinningModels] = useState(new Set())
+  const [loadingModels, setLoadingModels] = useState(new Set())
   // Expanded row state — keyed by `${tab}:${id}` so switching tabs doesn't
   // collide and a single row is open at a time per tab.
   const [expandedKey, setExpandedKey] = useState(null)
@@ -313,6 +314,26 @@ export default function Manage() {
     })
   }
 
+  // Pre-load a model (or all of a realtime pipeline's sub-models) into memory.
+  // The /backend/load call blocks until loading finishes, so the menu item shows
+  // a loading state while in flight and reports the outcome on completion.
+  const handleLoadModel = async (modelName) => {
+    setLoadingModels(prev => new Set(prev).add(modelName))
+    try {
+      await backendControlApi.load({ model: modelName })
+      addToast(`Loaded ${modelName}`, 'success')
+      setTimeout(fetchLoadedModels, 500)
+    } catch (err) {
+      addToast(`Failed to load: ${err.message}`, 'error')
+    } finally {
+      setLoadingModels(prev => {
+        const next = new Set(prev)
+        next.delete(modelName)
+        return next
+      })
+    }
+  }
+
   const handleDeleteModel = (modelName) => {
     setConfirmDialog({
       title: 'Delete Model',
@@ -687,6 +708,11 @@ export default function Manage() {
                               label: model.disabled ? 'Enable model' : 'Disable model',
                               onClick: () => handleToggleModel(model.id, model.disabled),
                               disabled: togglingModels.has(model.id) },
+                            { key: 'load', icon: 'fa-bolt',
+                              label: loadingModels.has(model.id) ? 'Loading…' : 'Load into memory',
+                              onClick: () => handleLoadModel(model.id),
+                              hidden: isRunning || !!model.disabled,
+                              disabled: loadingModels.has(model.id) },
                             { key: 'stop', icon: 'fa-stop', label: 'Stop model',
                               onClick: () => handleStopModel(model.id), hidden: !isRunning },
                             { key: 'pin', icon: 'fa-thumbtack',

core/http/react-ui/src/utils/api.js

@@ -352,6 +352,9 @@ export const realtimeApi = {
 // Backend control
 export const backendControlApi = {
   shutdown: (body) => postJSON(API_CONFIG.endpoints.backendShutdown, body),
+  // Pre-load a model (or all of a realtime pipeline's sub-models) into memory.
+  // body: { model: "<name>" }. Inverse of shutdown.
+  load: (body) => postJSON(API_CONFIG.endpoints.backendLoad, body),
 }
 
 // System info

core/http/react-ui/src/utils/config.js

@@ -106,6 +106,7 @@ export const API_CONFIG = {
     video: '/video',
     backendMonitor: '/backend/monitor',
     backendShutdown: '/backend/shutdown',
+    backendLoad: '/backend/load',
     modelsApply: '/models/apply',
     modelsDelete: (name) => `/models/delete/${name}`,
     modelsAvailable: '/models/available',

core/http/routes/localai.go

@@ -207,9 +207,14 @@ func RegisterLocalAIRoutes(router *echo.Echo,
 	backendMonitorService := monitoring.NewBackendMonitorService(ml, cl, appConfig) // Split out for now
 	router.GET("/backend/monitor", localai.BackendMonitorEndpoint(backendMonitorService), adminMiddleware)
 	router.POST("/backend/shutdown", localai.BackendShutdownEndpoint(backendMonitorService), adminMiddleware)
+	// /backend/load is the inverse of /backend/shutdown: pre-load a model (or all
+	// of a realtime pipeline's sub-models) into memory so clients can drive
+	// warm-up explicitly instead of paying the cold-start cost on first use.
+	router.POST("/backend/load", localai.LoadModelEndpoint(cl, ml, appConfig), adminMiddleware)
 	// The v1/* urls are exactly the same as above - makes local e2e testing easier if they are registered.
 	router.GET("/v1/backend/monitor", localai.BackendMonitorEndpoint(backendMonitorService), adminMiddleware)
 	router.POST("/v1/backend/shutdown", localai.BackendShutdownEndpoint(backendMonitorService), adminMiddleware)
+	router.POST("/v1/backend/load", localai.LoadModelEndpoint(cl, ml, appConfig), adminMiddleware)
 
 	// Traces and backend logs (monitoring)
 	router.GET("/api/traces", localai.GetAPITracesEndpoint(), adminMiddleware)
@@ -245,6 +250,7 @@ func RegisterLocalAIRoutes(router *echo.Echo,
 			"metrics":              "/metrics",
 			"backend_monitor":      "/backend/monitor",
 			"backend_shutdown":     "/backend/shutdown",
+			"backend_load":         "/backend/load",
 			"system":               "/system",
 			"version":              "/version",
 			"traces":               "/api/traces",

core/http/routes/ui_api.go

@@ -1243,6 +1243,9 @@ func RegisterUIAPIRoutes(app *echo.Echo, cl *config.ModelConfigLoader, ml *model
 			Galleries:          appConfig.BackendGalleries,
 			Context:            ctx,
 			CancelFunc:         cancelFunc,
+			// The React UI's "Reinstall backend" action reuses this route, so
+			// the op must force even when the backend is already installed.
+			Force: true,
 		}
 		// Store cancellation function immediately so queued operations can be cancelled
 		galleryService.StoreCancellation(uid, cancelFunc)

core/schema/localai.go

@@ -11,6 +11,24 @@ type BackendMonitorRequest struct {
 	BasicModelRequest
 }
 
+// ModelLoadRequest asks LocalAI to pre-load a model into memory by name, so the
+// first request that uses it pays no cold-start load cost. For a realtime
+// pipeline model, every configured sub-model (VAD, transcription, LLM, TTS,
+// sound_detection, voice_recognition) is loaded instead of the pipeline stub.
+// It is the inverse of the /backend/shutdown request.
+type ModelLoadRequest struct {
+	BasicModelRequest
+}
+
+// ModelLoadResponse reports the outcome of a /backend/load call.
+type ModelLoadResponse struct {
+	// Loaded lists the model names actually resident in memory after the call.
+	// For a pipeline model these are its sub-models, not the pipeline name.
+	Loaded []string `json:"loaded"`
+	// Message is a short human-readable status ("model loaded", or an error).
+	Message string `json:"message"`
+}
+
 type TokenMetricsRequest struct {
 	BasicModelRequest
 }

core/services/advisorylock/advisorylock.go

@@ -6,10 +6,39 @@ import (
 	"hash/fnv"
 	"strings"
 	"sync"
+	"time"
 
 	"gorm.io/gorm"
 )
 
+// advisoryLockWaitBackstop bounds, server-side, how long we will wait to
+// acquire a blocking advisory lock when the caller's context carries no
+// deadline (e.g. a startup schema migration using context.Background()). It
+// only exists so such a caller cannot hang forever behind a holder whose
+// session never releases the lock; it is far longer than any legitimate
+// guarded section. A var (not const) so tests can shrink it.
+var advisoryLockWaitBackstop = 30 * time.Minute
+
+// advisoryLockTimeoutMargin is added to a context's remaining budget when
+// deriving the server-side lock_timeout, so the Go context's own (cleaner)
+// cancellation fires first and the server bound is only ever a backstop.
+const advisoryLockTimeoutMargin = 30 * time.Second
+
+// advisoryLockWaitBudget returns the server-side lock_timeout to use for a
+// blocking acquire: the caller context's remaining time plus a margin (so the
+// Go context still governs), or the backstop when the context has no deadline.
+// Never returns zero - "wait forever" must not be possible.
+func advisoryLockWaitBudget(ctx context.Context) time.Duration {
+	if dl, ok := ctx.Deadline(); ok {
+		budget := time.Until(dl) + advisoryLockTimeoutMargin
+		if budget < time.Second {
+			budget = time.Second
+		}
+		return budget
+	}
+	return advisoryLockWaitBackstop
+}
+
 // localLocks holds one buffered channel (capacity 1) per lock key, used as an
 // in-process mutex for non-PostgreSQL dialects (SQLite). A SQLite auth DB is
 // effectively single-process, so serializing guarded sections within this
@@ -130,6 +159,27 @@ func WithLockCtx(ctx context.Context, db *gorm.DB, key int64, fn func() error) e
 	}
 	defer conn.Close()
 
+	// Override any deployment-wide lock_timeout on this dedicated connection.
+	// Operators commonly set a short global lock_timeout (on the role or
+	// database) to bound ordinary row-lock waits. Applied to the blocking
+	// pg_advisory_lock below, it aborts the wait with SQLSTATE 55P03 and turns
+	// LocalAI's intentional cross-replica "wait your turn, then re-check"
+	// coordination into a hard error for the caller (e.g. a chat request that
+	// just wanted to reuse a model another replica is loading).
+	//
+	// We do NOT disable it outright (lock_timeout = 0 would wait forever, which
+	// is unsafe for the schema-migration callers that pass context.Background()).
+	// Instead we set a bound derived from the caller's context: its remaining
+	// budget plus a margin so the Go context's cancellation wins with a clean
+	// error, or a finite backstop when the context has no deadline.
+	waitBudget := advisoryLockWaitBudget(ctx)
+	if _, err := conn.ExecContext(ctx,
+		fmt.Sprintf("SET lock_timeout = %d", waitBudget.Milliseconds())); err != nil {
+		return fmt.Errorf("advisorylock: setting lock_timeout: %w", err)
+	}
+	// Restore the session default before this pooled connection is reused.
+	defer func() { _, _ = conn.ExecContext(context.Background(), "RESET lock_timeout") }()
+
 	if _, err := conn.ExecContext(ctx, "SELECT pg_advisory_lock($1)", key); err != nil {
 		return fmt.Errorf("advisorylock: acquiring lock %d: %w", key, err)
 	}

core/services/advisorylock/advisorylock_test.go

@@ -158,6 +158,87 @@ var _ = Describe("AdvisoryLock", func() {
 			Expect(err).To(HaveOccurred())
 		})
 
+		It("waits out a short server-side lock_timeout instead of failing with 55P03", func() {
+			const lockKey int64 = 703
+
+			// Reproduce the production deployment that triggered this: a short
+			// global lock_timeout set on the database. Without the fix, a waiter
+			// blocked on pg_advisory_lock() is aborted by the server after this
+			// window and surfaces SQLSTATE 55P03 ("canceling statement due to
+			// lock timeout") to the caller instead of waiting for its turn.
+			Expect(db.Exec("ALTER DATABASE testdb SET lock_timeout = '300ms'").Error).ToNot(HaveOccurred())
+			sqlDB, err := db.DB()
+			Expect(err).ToNot(HaveOccurred())
+			// Drop pooled connections so subsequent ones reconnect and inherit
+			// the new database-level lock_timeout default.
+			sqlDB.SetMaxIdleConns(0)
+
+			holding := make(chan struct{})
+			released := make(chan struct{})
+			go func() {
+				defer GinkgoRecover()
+				herr := WithLockCtx(context.Background(), db, lockKey, func() error {
+					close(holding)
+					// Hold well past the 300ms server lock_timeout.
+					time.Sleep(1 * time.Second)
+					return nil
+				})
+				Expect(herr).ToNot(HaveOccurred())
+				close(released)
+			}()
+
+			<-holding // ensure the holder owns the lock before we contend
+
+			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+			defer cancel()
+			executed := false
+			start := time.Now()
+			werr := WithLockCtx(ctx, db, lockKey, func() error {
+				executed = true
+				return nil
+			})
+			Expect(werr).ToNot(HaveOccurred(),
+				"waiter should wait out the in-progress hold, not fail with lock_timeout (55P03)")
+			Expect(executed).To(BeTrue())
+			Expect(time.Since(start)).To(BeNumerically(">=", 400*time.Millisecond),
+				"waiter should have actually waited for the holder to release")
+			<-released
+		})
+
+		It("bounds a deadline-less waiter with the backstop instead of waiting forever", func() {
+			const lockKey int64 = 704
+
+			// A caller with no context deadline (e.g. startup schema migration
+			// passing context.Background()) must not hang forever if the holder
+			// never releases. Shrink the backstop so the test is fast.
+			origBackstop := advisoryLockWaitBackstop
+			advisoryLockWaitBackstop = 500 * time.Millisecond
+			DeferCleanup(func() { advisoryLockWaitBackstop = origBackstop })
+
+			holding := make(chan struct{})
+			release := make(chan struct{})
+			go func() {
+				defer GinkgoRecover()
+				_ = WithLockCtx(context.Background(), db, lockKey, func() error {
+					close(holding)
+					<-release // hold until the test releases us
+					return nil
+				})
+			}()
+			defer close(release)
+
+			<-holding
+
+			start := time.Now()
+			err := WithLockCtx(context.Background(), db, lockKey, func() error {
+				Fail("waiter should not have acquired the still-held lock")
+				return nil
+			})
+			Expect(err).To(HaveOccurred(), "deadline-less waiter should give up at the backstop, not hang")
+			Expect(time.Since(start)).To(BeNumerically("<", 5*time.Second),
+				"backstop must cap the wait well under the test timeout")
+		})
+
 		It("serializes concurrent WithLockCtx on same key", func() {
 			const lockKey int64 = 702
 

core/services/galleryop/backend_force_test.go

@@ -0,0 +1,114 @@
+package galleryop_test
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/gallery"
+	"github.com/mudler/LocalAI/core/services/galleryop"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/system"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+	"gopkg.in/yaml.v3"
+)
+
+// The install op must be idempotent unless Force is set: API clients call
+// POST /backends/apply on every boot to make sure the backend exists, and an
+// unconditional force here re-downloads the whole backend artifact each time.
+// Reinstall is an explicit, opted-in action.
+var _ = Describe("LocalBackendManager force semantics", func() {
+	var (
+		backendsDir string
+		srcDir      string
+		mgr         *galleryop.LocalBackendManager
+		systemState *system.SystemState
+		ml          *model.ModelLoader
+	)
+
+	const installedRunSh = "#!/bin/sh\necho installed\n"
+	const galleryRunSh = "#!/bin/sh\necho from-gallery\n"
+
+	installedRunShPath := func() string {
+		return filepath.Join(backendsDir, "test-backend", "run.sh")
+	}
+
+	BeforeEach(func() {
+		var err error
+		backendsDir, err = os.MkdirTemp("", "force-backends-*")
+		Expect(err).NotTo(HaveOccurred())
+		srcDir, err = os.MkdirTemp("", "force-src-*")
+		Expect(err).NotTo(HaveOccurred())
+
+		// The gallery serves test-backend from a plain directory (offline).
+		// The gallery yaml itself must live under the backends path: file://
+		// galleries outside the trusted root are rejected by the downloader.
+		Expect(os.WriteFile(filepath.Join(srcDir, "run.sh"), []byte(galleryRunSh), 0o755)).To(Succeed())
+		entries := []map[string]any{{"name": "test-backend", "uri": srcDir}}
+		data, err := yaml.Marshal(entries)
+		Expect(err).NotTo(HaveOccurred())
+		galleryYAML := filepath.Join(backendsDir, "gallery.yaml")
+		Expect(os.WriteFile(galleryYAML, data, 0o644)).To(Succeed())
+
+		// test-backend is already installed, with content that differs from
+		// the gallery's so a reinstall is observable.
+		Expect(os.MkdirAll(filepath.Join(backendsDir, "test-backend"), 0o755)).To(Succeed())
+		Expect(os.WriteFile(installedRunShPath(), []byte(installedRunSh), 0o755)).To(Succeed())
+
+		systemState, err = system.GetSystemState(system.WithBackendPath(backendsDir))
+		Expect(err).NotTo(HaveOccurred())
+		appConfig := &config.ApplicationConfig{
+			SystemState:      systemState,
+			BackendGalleries: []config.Gallery{{Name: "test", URL: "file://" + galleryYAML}},
+		}
+		ml = model.NewModelLoader(systemState)
+		mgr = galleryop.NewLocalBackendManager(appConfig, ml)
+	})
+
+	AfterEach(func() {
+		Expect(os.RemoveAll(backendsDir)).To(Succeed())
+		Expect(os.RemoveAll(srcDir)).To(Succeed())
+	})
+
+	It("skips an already-installed backend when Force is not set", func() {
+		op := &galleryop.ManagementOp[gallery.GalleryBackend, any]{
+			ID:                 "op-1",
+			GalleryElementName: "test-backend",
+		}
+		Expect(mgr.InstallBackend(context.Background(), op, nil)).To(Succeed())
+
+		content, err := os.ReadFile(installedRunShPath())
+		Expect(err).NotTo(HaveOccurred())
+		Expect(string(content)).To(Equal(installedRunSh), "install without Force must not overwrite an installed backend")
+	})
+
+	It("reinstalls an already-installed backend when Force is set", func() {
+		op := &galleryop.ManagementOp[gallery.GalleryBackend, any]{
+			ID:                 "op-2",
+			GalleryElementName: "test-backend",
+			Force:              true,
+		}
+		Expect(mgr.InstallBackend(context.Background(), op, nil)).To(Succeed())
+
+		content, err := os.ReadFile(installedRunShPath())
+		Expect(err).NotTo(HaveOccurred())
+		Expect(string(content)).To(Equal(galleryRunSh), "install with Force must overwrite the installed backend")
+	})
+
+	// The LOCALAI_EXTERNAL_BACKENDS boot loop goes through
+	// InstallExternalBackend's gallery-name path on EVERY startup; it must not
+	// force, or each boot re-downloads every listed backend.
+	It("skips an already-installed backend on the non-forced external gallery-name path", func() {
+		err := galleryop.InstallExternalBackend(context.Background(),
+			[]config.Gallery{{Name: "test", URL: "file://" + filepath.Join(backendsDir, "gallery.yaml")}},
+			systemState, ml, nil, "test-backend", "", "", false, false)
+		Expect(err).NotTo(HaveOccurred())
+
+		content, err := os.ReadFile(installedRunShPath())
+		Expect(err).NotTo(HaveOccurred())
+		Expect(string(content)).To(Equal(installedRunSh), "non-forced external install must not overwrite an installed backend")
+	})
+})

core/services/galleryop/backends.go

@@ -144,7 +144,12 @@ func (g *GalleryService) backendHandler(op *ManagementOp[gallery.GalleryBackend,
 // InstallExternalBackend installs a backend from an external source (OCI image, URL, or path).
 // This method contains the logic to detect the input type and call the appropriate installation function.
 // It can be used by both CLI and Web UI for installing backends from external sources.
-func InstallExternalBackend(ctx context.Context, galleries []config.Gallery, systemState *system.SystemState, modelLoader *model.ModelLoader, downloadStatus func(string, string, string, float64), backend, name, alias string, requireIntegrity bool) error {
+//
+// force applies only to the gallery-name fallback: a URI install (dir/OCI/file)
+// always writes, but a bare gallery name is an "ensure installed" — the
+// LOCALAI_EXTERNAL_BACKENDS boot loop runs it on every start and must not
+// re-download an installed, runnable backend.
+func InstallExternalBackend(ctx context.Context, galleries []config.Gallery, systemState *system.SystemState, modelLoader *model.ModelLoader, downloadStatus func(string, string, string, float64), backend, name, alias string, force, requireIntegrity bool) error {
 	uri := downloader.URI(backend)
 	switch {
 	case uri.LooksLikeDir():
@@ -202,7 +207,7 @@ func InstallExternalBackend(ctx context.Context, galleries []config.Gallery, sys
 		if name != "" || alias != "" {
 			return fmt.Errorf("specifying a name or alias is not supported for gallery backends")
 		}
-		err := gallery.InstallBackendFromGallery(ctx, galleries, systemState, modelLoader, backend, downloadStatus, true, requireIntegrity)
+		err := gallery.InstallBackendFromGallery(ctx, galleries, systemState, modelLoader, backend, downloadStatus, force, requireIntegrity)
 		if err != nil {
 			return fmt.Errorf("error installing backend %s: %w", backend, err)
 		}

core/services/galleryop/backends_test.go

@@ -70,6 +70,7 @@ var _ = Describe("InstallExternalBackend", func() {
 				"test-backend", // gallery name
 				"custom-name",  // name should not be allowed
 				"",
+				false, // force
 				false,
 			)
 			Expect(err).To(HaveOccurred())
@@ -86,6 +87,7 @@ var _ = Describe("InstallExternalBackend", func() {
 				"non-existent-backend",
 				"",
 				"",
+				false, // force
 				false,
 			)
 			Expect(err).To(HaveOccurred())
@@ -103,6 +105,7 @@ var _ = Describe("InstallExternalBackend", func() {
 				"oci://quay.io/mudler/tests:localai-backend-test",
 				"", // name is required for OCI images
 				"",
+				false, // force
 				false,
 			)
 			Expect(err).To(HaveOccurred())
@@ -136,6 +139,7 @@ var _ = Describe("InstallExternalBackend", func() {
 				testBackendPath,
 				"", // name should be inferred as "source-backend"
 				"",
+				false, // force
 				false,
 			)
 			// The function should at least attempt to install with the inferred name
@@ -155,6 +159,7 @@ var _ = Describe("InstallExternalBackend", func() {
 				testBackendPath,
 				"custom-backend-name",
 				"",
+				false, // force
 				false,
 			)
 			// The function should use the provided name
@@ -173,6 +178,7 @@ var _ = Describe("InstallExternalBackend", func() {
 				testBackendPath,
 				"custom-backend-name",
 				"custom-alias",
+				false, // force
 				false,
 			)
 			// The function should accept alias for directory paths

core/services/galleryop/managers_local.go

@@ -110,10 +110,13 @@ func (b *LocalBackendManager) CheckUpgrades(ctx context.Context) (map[string]gal
 func (b *LocalBackendManager) InstallBackend(ctx context.Context, op *ManagementOp[gallery.GalleryBackend, any], progressCb ProgressCallback) error {
 	if op.ExternalURI != "" {
 		return InstallExternalBackend(ctx, b.backendGalleries, b.systemState, b.modelLoader,
-			progressCb, op.ExternalURI, op.ExternalName, op.ExternalAlias, b.requireBackendIntegrity)
+			progressCb, op.ExternalURI, op.ExternalName, op.ExternalAlias, op.Force, b.requireBackendIntegrity)
 	}
+	// op.Force distinguishes an explicit reinstall from an idempotent
+	// "make sure it's installed" op; the latter must not re-download an
+	// already-runnable backend (supervisors apply on every boot).
 	return gallery.InstallBackendFromGallery(ctx, b.backendGalleries, b.systemState,
-		b.modelLoader, op.GalleryElementName, progressCb, true, b.requireBackendIntegrity)
+		b.modelLoader, op.GalleryElementName, progressCb, op.Force, b.requireBackendIntegrity)
 }
 
 func (b *LocalBackendManager) IsDistributed() bool { return false }

core/services/galleryop/operation.go

@@ -45,6 +45,13 @@ type ManagementOp[T any, E any] struct {
 
 	// Upgrade is true if this is an upgrade operation (not a fresh install)
 	Upgrade bool
+
+	// Force reinstalls a backend even when it is already installed and
+	// runnable. Without it a backend install op is idempotent — API clients
+	// that ensure a backend exists on every boot must not trigger a full
+	// artifact re-download each time. The UI's explicit "Reinstall backend"
+	// action sets it.
+	Force bool
 }
 
 type OpStatus struct {

core/services/nodes/router.go

@@ -68,6 +68,13 @@ type SmartRouterOptions struct {
 	// the absolute model paths untouched so the worker loads them directly from
 	// the shared volume (#10556). See config.DistributedConfig.SharedModels.
 	SharedModels bool
+	// ModelLoadCeiling is the hard upper bound on how long a single cold-load
+	// attempt (node selection -> backend install -> file staging -> LoadModel)
+	// may run while holding the per-model advisory lock. It backstops every
+	// sub-step's own timeout so a wedged worker can never pin the lock - and
+	// every other replica's request for that model - indefinitely. Zero selects
+	// defaultModelLoadCeiling.
+	ModelLoadCeiling time.Duration
 }
 
 // SmartRouter routes inference requests to the best available backend node.
@@ -101,8 +108,18 @@ type SmartRouter struct {
 	// sharedModels skips file staging when all nodes mount the same models
 	// directory at the same path (see SmartRouterOptions.SharedModels).
 	sharedModels bool
+	// modelLoadCeiling bounds how long a cold load may hold the per-model
+	// advisory lock (see SmartRouterOptions.ModelLoadCeiling).
+	modelLoadCeiling time.Duration
 }
 
+// defaultModelLoadCeiling is the fallback hold ceiling for a cold model load.
+// It must comfortably exceed the slowest legitimate load - a multi-GB backend
+// install (DefaultBackendInstallTimeout, 15m) plus staging and the remote
+// LoadModel (5m) - so it never cuts a real load short; it only ever fires when
+// a step is genuinely wedged (e.g. a worker that died mid-install).
+const defaultModelLoadCeiling = 25 * time.Minute
+
 // probeCacheTTL is how long a successful gRPC HealthCheck on a backend is
 // trusted before the next request re-probes. Matches healthCheckTTL in
 // pkg/model/model.go so the single-process and distributed paths share a
@@ -117,6 +134,10 @@ func NewSmartRouter(registry ModelRouter, opts SmartRouterOptions) *SmartRouter
 	if factory == nil {
 		factory = &tokenClientFactory{token: opts.AuthToken}
 	}
+	ceiling := opts.ModelLoadCeiling
+	if ceiling <= 0 {
+		ceiling = defaultModelLoadCeiling
+	}
 	return &SmartRouter{
 		registry:         registry,
 		unloader:         opts.Unloader,
@@ -131,6 +152,7 @@ func NewSmartRouter(registry ModelRouter, opts SmartRouterOptions) *SmartRouter
 		prefixConfig:     opts.PrefixConfig,
 		pressure:         opts.Pressure,
 		sharedModels:     opts.SharedModels,
+		modelLoadCeiling: ceiling,
 	}
 }
 
@@ -383,11 +405,19 @@ func (r *SmartRouter) Route(ctx context.Context, modelID, modelName, backendType
 	// the request context. If staging were bound to it, the multi-GB upload
 	// aborts with "context canceled" mid-transfer and large models can never
 	// finish staging (the model-load outage). WithoutCancel keeps the request's
-	// values (prefix chain, etc.) but drops its cancellation/deadline. Each
-	// long step still has its own bound (the file stager's resume budget,
-	// LoadModel's 5m timeout), and the per-model advisory lock below de-dupes
-	// concurrent loaders across replicas.
-	loadCtx := context.WithoutCancel(ctx)
+	// values (prefix chain, etc.) but drops its cancellation/deadline.
+	//
+	// Detaching from the caller is necessary, but it must not be unbounded: the
+	// load runs while holding the per-model advisory lock, and a worker that
+	// dies mid-install (its backend.install never replies) would otherwise pin
+	// that lock (and every other replica's request for the same model) until
+	// the NATS install deadline alone expires. Re-impose a single hard ceiling
+	// over the whole sequence so the lock is always released in bounded time,
+	// even if a sub-step wedges. Each long step still has its own (tighter)
+	// bound; this only backstops them. The per-model advisory lock below
+	// de-dupes concurrent loaders across replicas.
+	loadCtx, cancelLoad := context.WithTimeout(context.WithoutCancel(ctx), r.modelLoadCeiling)
+	defer cancelLoad()
 	loadModel := func(ctx context.Context) (*RouteResult, error) {
 		// Re-check after acquiring lock — another request may have loaded it
 		node, nm, err := r.registry.FindAndLockNodeWithModel(ctx, trackingKey, candidateNodeIDs, pref)
@@ -916,7 +946,14 @@ func (r *SmartRouter) installBackendOnNode(ctx context.Context, node *BackendNod
 	}
 
 	key := fmt.Sprintf("%s|%s|%s|%d", node.ID, backendType, modelID, replicaIndex)
-	v, err, _ := r.installFlight.Do(key, func() (any, error) {
+	// DoChan rather than Do so this wait honors ctx cancellation. InstallBackend
+	// blocks for its full NATS deadline (15m by default) when a worker accepts
+	// the request but never replies (e.g. it died mid-install). Without ctx
+	// awareness the caller (holding the per-model advisory lock) would sit there
+	// the whole time; here a cancelled ctx (typically the model-load ceiling)
+	// frees the caller promptly. The shared install keeps running in the
+	// background and still coalesces other callers via singleflight.
+	resCh := r.installFlight.DoChan(key, func() (any, error) {
 		reply, err := r.unloader.InstallBackend(node.ID, backendType, modelID, r.galleriesJSON, "", "", "", replicaIndex, "", nil)
 		if err != nil {
 			return "", err
@@ -931,10 +968,15 @@ func (r *SmartRouter) installBackendOnNode(ctx context.Context, node *BackendNod
 		}
 		return addr, nil
 	})
-	if err != nil {
-		return "", err
+	select {
+	case <-ctx.Done():
+		return "", ctx.Err()
+	case res := <-resCh:
+		if res.Err != nil {
+			return "", res.Err
+		}
+		return res.Val.(string), nil
 	}
-	return v.(string), nil
 }
 
 func (r *SmartRouter) buildClientForAddr(node *BackendNode, addr string, parallel bool) grpc.Backend {

core/services/nodes/router_test.go

@@ -493,6 +493,44 @@ var _ = Describe("SmartRouter", func() {
 				Expect(result.Node.ID).To(Equal("n3"))
 			})
 		})
+
+		Context("worker wedges mid-install (dead node holding the lock)", func() {
+			It("aborts the load at the ModelLoadCeiling instead of blocking forever", func() {
+				// Simulate the production incident: the chosen worker accepts the
+				// backend.install but never replies (it died), so InstallBackend
+				// would otherwise block for its full NATS deadline (15m by
+				// default) while pinning the per-model advisory lock. Route must
+				// give up at the ceiling so the lock is released promptly.
+				reg.findAndLockErr = errors.New("not found")
+				reg.findIdleNode = &BackendNode{ID: "n4", Name: "dead-node", Address: "10.0.0.4:50051"}
+
+				block := make(chan struct{})
+				defer close(block) // let the background install goroutine drain at test end
+				unloader.installHook = func() { <-block }
+
+				router := NewSmartRouter(reg, SmartRouterOptions{
+					Unloader:         unloader,
+					ClientFactory:    factory,
+					ModelLoadCeiling: 200 * time.Millisecond,
+				})
+
+				done := make(chan error, 1)
+				start := time.Now()
+				go func() {
+					defer GinkgoRecover()
+					_, err := router.Route(context.Background(), "wedged-model",
+						"models/wedged.gguf", "llama-cpp",
+						&pb.ModelOptions{Model: "models/wedged.gguf"}, false)
+					done <- err
+				}()
+
+				var routeErr error
+				Eventually(done, 5*time.Second).Should(Receive(&routeErr),
+					"Route must not block on a wedged install past the ceiling")
+				Expect(routeErr).To(HaveOccurred())
+				Expect(time.Since(start)).To(BeNumerically("<", 5*time.Second))
+			})
+		})
 	})
 
 	Describe("scheduleNewModel (mock-based, via Route)", func() {

core/services/worker/install.go

@@ -134,7 +134,7 @@ func (s *backendSupervisor) installBackend(req messaging.BackendInstallRequest,
 		if req.URI != "" {
 			xlog.Info("Installing backend from external URI", "backend", req.Backend, "uri", req.URI, "force", force)
 			if err := galleryop.InstallExternalBackend(
-				context.Background(), galleries, s.systemState, s.ml, downloadCb, req.URI, req.Name, req.Alias, s.cfg.RequireBackendIntegrity,
+				context.Background(), galleries, s.systemState, s.ml, downloadCb, req.URI, req.Name, req.Alias, force, s.cfg.RequireBackendIntegrity,
 			); err != nil {
 				return "", fmt.Errorf("installing backend from gallery: %w", err)
 			}
@@ -201,7 +201,7 @@ func (s *backendSupervisor) upgradeBackend(req messaging.BackendUpgradeRequest)
 	if req.URI != "" {
 		xlog.Info("Upgrading backend from external URI", "backend", req.Backend, "uri", req.URI)
 		if err := galleryop.InstallExternalBackend(
-			context.Background(), galleries, s.systemState, s.ml, downloadCb, req.URI, req.Name, req.Alias, s.cfg.RequireBackendIntegrity,
+			context.Background(), galleries, s.systemState, s.ml, downloadCb, req.URI, req.Name, req.Alias, true, s.cfg.RequireBackendIntegrity,
 		); err != nil {
 			return fmt.Errorf("upgrading backend from external URI: %w", err)
 		}

core/trace/audio_snippet.go

@@ -14,6 +14,16 @@ import (
 // MaxSnippetSeconds is the maximum number of seconds of audio captured per trace.
 const MaxSnippetSeconds = 30
 
+// silenceFloorDBFS is the dBFS value reported for digital silence (RMS or peak
+// of zero). The true level is -∞ dBFS; reporting a finite floor keeps the
+// metric present and meaningful in the Traces UI (a scrubbed nil would read as
+// "missing" rather than "silent"). -120 dBFS sits well below 16-bit PCM's
+// ~-90 dBFS least-significant-bit floor, so it reads unambiguously as
+// "effectively silent". JSON-marshal safety for any non-finite float that does
+// reach a trace is owned centrally by RecordBackendTrace's sanitizer — this
+// floor is about presentation, not transport.
+const silenceFloorDBFS = -120.0
+
 // AudioSnippet captures the first MaxSnippetSeconds of a WAV file and computes
 // quality metrics. The result is a map suitable for merging into a BackendTrace
 // Data field. maxBytes caps the embedded base64 waveform so a single TTS or
@@ -63,7 +73,7 @@ func AudioSnippetFromPCM(pcm []byte, sampleRate, totalPCMBytes, maxBytes int) ma
 	snippetDuration := float64(len(samples)) / float64(sampleRate)
 
 	rms := sound.CalculateRMS16(samples)
-	rmsDBFS := -math.Inf(1)
+	rmsDBFS := silenceFloorDBFS
 	if rms > 0 {
 		rmsDBFS = 20 * math.Log10(rms/32768.0)
 	}
@@ -78,7 +88,7 @@ func AudioSnippetFromPCM(pcm []byte, sampleRate, totalPCMBytes, maxBytes int) ma
 		}
 		dcSum += int64(s)
 	}
-	peakDBFS := -math.Inf(1)
+	peakDBFS := silenceFloorDBFS
 	if peak > 0 {
 		peakDBFS = 20 * math.Log10(float64(peak)/32768.0)
 	}

core/trace/audio_snippet_test.go

@@ -1,6 +1,9 @@
 package trace_test
 
 import (
+	"encoding/json"
+	"math"
+
 	. "github.com/onsi/ginkgo/v2"
 	. "github.com/onsi/gomega"
 
@@ -47,3 +50,32 @@ var _ = Describe("AudioSnippetFromPCM byte cap", func() {
 		Expect(out).To(HaveKey("audio_wav_base64"))
 	})
 })
+
+// Silent audio (RMS/peak of zero) has a true level of -∞ dBFS, but emitting
+// -Inf made the whole /api/backend-traces response fail to JSON-marshal and
+// blanked the Traces UI. The metrics must instead be finite and serializable.
+var _ = Describe("AudioSnippetFromPCM silent audio dBFS", func() {
+	pcm := makePCM(snippetSeconds, snippetSampleRate) // all zeros == digital silence
+	totalPCM := len(pcm)
+
+	It("reports finite dBFS for silence instead of -Inf", func() {
+		out := trace.AudioSnippetFromPCM(pcm, snippetSampleRate, totalPCM, 0)
+
+		rms, ok := out["audio_rms_dbfs"].(float64)
+		Expect(ok).To(BeTrue())
+		Expect(math.IsInf(rms, 0)).To(BeFalse(), "silent RMS must not be ±Inf")
+		Expect(math.IsNaN(rms)).To(BeFalse())
+
+		peak, ok := out["audio_peak_dbfs"].(float64)
+		Expect(ok).To(BeTrue())
+		Expect(math.IsInf(peak, 0)).To(BeFalse(), "silent peak must not be ±Inf")
+		Expect(math.IsNaN(peak)).To(BeFalse())
+	})
+
+	It("produces a snippet that round-trips through encoding/json", func() {
+		out := trace.AudioSnippetFromPCM(pcm, snippetSampleRate, totalPCM, 0)
+
+		_, err := json.Marshal(out)
+		Expect(err).ToNot(HaveOccurred(), "silent-audio metrics must be JSON-marshalable")
+	})
+})

core/trace/backend_trace.go

@@ -3,6 +3,8 @@ package trace
 import (
 	"encoding/json"
 	"fmt"
+	"maps"
+	"math"
 	"slices"
 	"sync"
 	"time"
@@ -116,8 +118,13 @@ func RecordBackendTrace(t BackendTrace) {
 	backendMu.Lock()
 	maxBody := backendMaxBodyBytes
 	backendMu.Unlock()
-	if t.Data != nil && maxBody > 0 {
-		t.Data = capDataStrings(t.Data, maxBody)
+	// Always walk Data, even with no body cap configured: besides capping
+	// oversized strings (maxBody > 0), the walk replaces non-finite floats
+	// (Inf/NaN) that encoding/json cannot marshal. A single such value — e.g. a
+	// -Inf dBFS audio metric from a silent clip — would otherwise fail the whole
+	// /api/backend-traces response and blank the Traces UI.
+	if t.Data != nil {
+		t.Data = sanitizeData(t.Data, maxBody)
 	}
 	select {
 	case backendLogChan <- &t:
@@ -126,32 +133,90 @@ func RecordBackendTrace(t BackendTrace) {
 	}
 }
 
-// capDataStrings walks a trace Data map and replaces any string value (at any
-// depth) that exceeds maxBytes with a fixed-size marker that names the
-// original byte count. The replacement is intentionally short and not valid
-// base64/JSON: the goal is to flag "this was dropped" cheaply, not to keep a
-// partial value that the UI might try to render. Non-string scalars and
-// non-map containers pass through untouched so structural fields like
-// total_deltas or audio_sample_rate remain useful.
-func capDataStrings(data map[string]any, maxBytes int) map[string]any {
-	out := make(map[string]any, len(data))
-	for k, v := range data {
-		out[k] = capValue(v, maxBytes)
-	}
+// sanitizeData walks a trace Data map (recursing into nested maps and slices)
+// and makes every value safe for the /api/backend-traces JSON response:
+//
+//   - When maxBytes > 0, any string longer than maxBytes is replaced with a
+//     fixed-size marker that names the original byte count. The replacement is
+//     intentionally short and not valid base64/JSON: it flags "this was dropped"
+//     cheaply rather than keeping a partial value the UI might try to render.
+//   - Non-finite floats (Inf/NaN) are replaced with nil regardless of maxBytes,
+//     because encoding/json refuses to marshal them and one bad value would fail
+//     the entire response.
+//
+// Other scalars (ints, bools, finite floats) pass through untouched so
+// structural fields like total_deltas or audio_sample_rate remain useful.
+//
+// The walk is copy-on-write: it runs on every RecordBackendTrace call, and in
+// the common case nothing needs rewriting, so containers are only re-allocated
+// on the paths that actually changed and untouched values keep their original
+// interface boxes instead of paying a per-value re-boxing allocation.
+func sanitizeData(data map[string]any, maxBytes int) map[string]any {
+	out, _ := sanitizeMap(data, maxBytes)
 	return out
 }
 
-func capValue(v any, maxBytes int) any {
+func sanitizeMap(m map[string]any, maxBytes int) (map[string]any, bool) {
+	var out map[string]any
+	for k, v := range m {
+		nv, changed := sanitizeValue(v, maxBytes)
+		if changed && out == nil {
+			// First change: fork the map. Entries already visited were
+			// unchanged, so a full copy then overwriting as we go is exact.
+			out = make(map[string]any, len(m))
+			maps.Copy(out, m)
+		}
+		if out != nil {
+			out[k] = nv
+		}
+	}
+	if out == nil {
+		return m, false
+	}
+	return out, true
+}
+
+func sanitizeSlice(s []any, maxBytes int) ([]any, bool) {
+	var out []any
+	for i, v := range s {
+		nv, changed := sanitizeValue(v, maxBytes)
+		if changed && out == nil {
+			out = make([]any, len(s))
+			copy(out, s)
+		}
+		if out != nil {
+			out[i] = nv
+		}
+	}
+	if out == nil {
+		return s, false
+	}
+	return out, true
+}
+
+func sanitizeValue(v any, maxBytes int) (any, bool) {
 	switch val := v.(type) {
 	case string:
-		if len(val) > maxBytes {
-			return fmt.Sprintf("<truncated: %d bytes>", len(val))
+		if maxBytes > 0 && len(val) > maxBytes {
+			return fmt.Sprintf("<truncated: %d bytes>", len(val)), true
 		}
-		return val
+		return v, false
+	case float64:
+		if math.IsInf(val, 0) || math.IsNaN(val) {
+			return nil, true
+		}
+		return v, false
+	case float32:
+		if f := float64(val); math.IsInf(f, 0) || math.IsNaN(f) {
+			return nil, true
+		}
+		return v, false
 	case map[string]any:
-		return capDataStrings(val, maxBytes)
+		return sanitizeMap(val, maxBytes)
+	case []any:
+		return sanitizeSlice(val, maxBytes)
 	default:
-		return v
+		return v, false
 	}
 }
 

core/trace/backend_trace_sanitize_test.go

@@ -0,0 +1,80 @@
+package trace_test
+
+import (
+	"encoding/json"
+	"math"
+	"time"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/trace"
+)
+
+// encoding/json cannot marshal ±Inf or NaN. The /api/backend-traces endpoint
+// serializes the whole buffer with one json call, so a single non-finite float
+// in any trace's Data map (e.g. a -Inf dBFS audio metric from a silent clip)
+// would fail the entire response and blank the Traces UI. RecordBackendTrace
+// must scrub those values regardless of whether a body cap is configured.
+var _ = Describe("RecordBackendTrace non-finite float sanitization", func() {
+	BeforeEach(func() {
+		// maxBodyBytes 0 == no body cap: float sanitization must still run.
+		trace.InitBackendTracingIfEnabled(64, 0)
+		trace.ClearBackendTraces()
+	})
+
+	It("replaces ±Inf and NaN with nil so the response stays JSON-marshalable", func() {
+		trace.RecordBackendTrace(trace.BackendTrace{
+			Timestamp: time.Now(),
+			Type:      trace.BackendTraceTranscription,
+			ModelName: "m",
+			Data: map[string]any{
+				"audio_rms_dbfs":   math.Inf(-1),
+				"audio_peak_dbfs":  math.Inf(1),
+				"weird":            math.NaN(),
+				"audio_duration_s": 1.5, // finite siblings must survive
+			},
+		})
+
+		Eventually(trace.GetBackendTraces).Should(HaveLen(1))
+		got := trace.GetBackendTraces()[0]
+
+		Expect(got.Data["audio_rms_dbfs"]).To(BeNil())
+		Expect(got.Data["audio_peak_dbfs"]).To(BeNil())
+		Expect(got.Data["weird"]).To(BeNil())
+		Expect(got.Data["audio_duration_s"]).To(Equal(1.5), "finite floats must pass through untouched")
+
+		_, err := json.Marshal(trace.GetBackendTraces())
+		Expect(err).ToNot(HaveOccurred(), "the whole trace buffer must marshal even with non-finite inputs")
+	})
+
+	It("scrubs non-finite floats nested in maps and slices", func() {
+		trace.RecordBackendTrace(trace.BackendTrace{
+			Timestamp: time.Now(),
+			Type:      trace.BackendTraceLLM,
+			ModelName: "m",
+			Data: map[string]any{
+				"nested": map[string]any{
+					"logprob": math.Inf(-1),
+					"ok":      0.25,
+				},
+				"scores": []any{1.0, math.Inf(1), math.NaN()},
+			},
+		})
+
+		Eventually(trace.GetBackendTraces).Should(HaveLen(1))
+		got := trace.GetBackendTraces()[0]
+
+		nested := got.Data["nested"].(map[string]any)
+		Expect(nested["logprob"]).To(BeNil())
+		Expect(nested["ok"]).To(Equal(0.25))
+
+		scores := got.Data["scores"].([]any)
+		Expect(scores[0]).To(Equal(1.0))
+		Expect(scores[1]).To(BeNil())
+		Expect(scores[2]).To(BeNil())
+
+		_, err := json.Marshal(trace.GetBackendTraces())
+		Expect(err).ToNot(HaveOccurred())
+	})
+})

docs/content/advanced/vram-management.md

@@ -381,6 +381,8 @@ curl -X POST http://localhost:8080/backend/shutdown \
 
 To stop all models, you'll need to call the endpoint for each loaded model individually, or use the web UI to stop all models at once.
 
+Conversely, you can pre-load a model into memory ahead of its first request with `POST /backend/load` (the inverse of shutdown) — see [Backend Monitor]({{%relref "features/backend-monitor" %}}).
+
 ### Best Practices
 
 1. **Monitor VRAM usage**: Use `nvidia-smi` (for NVIDIA GPUs) or similar tools to monitor actual VRAM usage

docs/content/features/authentication.md

@@ -166,7 +166,7 @@ When authentication is enabled, the following endpoints require admin role:
 - `GET /api/backend-traces`, `POST /api/backend-traces/clear`
 - `GET /api/backend-logs/*`, `POST /api/backend-logs/*/clear`
 - `GET /api/resources`, `GET /api/settings`, `POST /api/settings`
-- `GET /system`, `GET /backend/monitor`, `POST /backend/shutdown`
+- `GET /system`, `GET /backend/monitor`, `POST /backend/shutdown`, `POST /backend/load`
 
 **P2P:**
 - `GET /api/p2p/*`

docs/content/features/backend-monitor.md

@@ -5,7 +5,9 @@ weight = 20
 url = "/features/backend-monitor/"
 +++
 
-LocalAI provides endpoints to monitor and manage running backends. The `/backend/monitor` endpoint reports the status and resource usage of loaded models, and `/backend/shutdown` allows stopping a model's backend process.
+LocalAI provides endpoints to monitor and manage running backends. The `/backend/monitor` endpoint reports the status and resource usage of loaded models, `/backend/load` pre-loads a model into memory, and `/backend/shutdown` allows stopping a model's backend process.
+
+All three are admin-only.
 
 ## Monitor API
 
@@ -62,6 +64,42 @@ curl "http://localhost:8080/backend/monitor?model=my-model"
 }
 ```
 
+## Load API
+
+Pre-loads a model into memory ahead of its first request, so that request pays no cold-start load cost. It is the inverse of the Shutdown API and works for any model, not just realtime pipelines.
+
+- **Method:** `POST`
+- **Endpoints:** `/backend/load`, `/v1/backend/load`
+
+### Request
+
+| Parameter | Type     | Required | Description                  |
+|-----------|----------|----------|------------------------------|
+| `model`   | `string` | Yes      | Name of the model to load    |
+
+### Behavior
+
+- For a regular model, its own backend is loaded.
+- For a **realtime pipeline** model (a config with a `pipeline:` block), every configured sub-model (VAD, transcription, LLM, TTS, sound_detection, voice_recognition) is loaded concurrently instead of the pipeline stub, which has no backend of its own.
+
+The call blocks until loading finishes and reports which model names became resident, so partial failures are visible.
+
+### Usage
+
+```bash
+curl -X POST http://localhost:8080/backend/load \
+  -H "Content-Type: application/json" \
+  -d '{"model": "my-model"}'
+```
+
+### Example response
+
+```json
+{ "loaded": ["my-model"], "message": "model loaded" }
+```
+
+On failure the call returns `500` with `loaded` listing whichever sub-models did load and `message` naming the failures.
+
 ## Shutdown API
 
 - **Method:** `POST`

docs/content/features/openai-realtime.md

@@ -56,6 +56,39 @@ pipeline:
 
 All streaming flags are off by default, so existing pipelines are unaffected.
 
+### Model warm-up (cold start)
+
+Without warm-up the pipeline's models are loaded into memory only on first use *within* a session: the VAD on the first audio chunk, transcription at the first end-of-speech, the LLM on the first reply, and TTS on the first spoken output. On a cold session this staggers a load delay across those first few interactions — and a model that fails to load (missing weights, wrong backend, out of memory) only fails part-way through the first turn.
+
+To avoid that, LocalAI **warms the pipeline by default**: it loads the VAD, transcription, LLM and TTS backends into memory *before* the session is announced, and the session start **blocks until they are all ready**. The loads run concurrently, so the wait is the slowest single model, not the sum. This means:
+
+- The first turn pays no cold-start cost — every backend is already resident.
+- **Model-load errors surface at session start.** If any stage fails to load, the session is not started and the client receives a `model_load_error` instead of `session.created`, so a broken pipeline fails fast and visibly rather than mid-call.
+
+Set `disable_warmup: true` to restore the lazy "load on first use" behavior — session start no longer waits on loading and load errors surface on the first turn instead. Useful if you want idle sessions to avoid holding model memory they may never use:
+
+```yaml
+name: gpt-realtime
+pipeline:
+  vad: silero-vad-ggml
+  transcription: whisper-large-turbo
+  llm: qwen3-4b
+  tts: tts-1
+  disable_warmup: true   # lazily load each model on first use instead of at session start
+```
+
+#### Pre-loading a pipeline on demand
+
+Warm-up only fires when a realtime session opens. To load a pipeline into memory ahead of time — e.g. to warm it right after boot, or when running with `disable_warmup: true` — POST the model name to the admin-only `/backend/load` endpoint. For a pipeline model it loads every configured sub-model (VAD, transcription, LLM, TTS, sound_detection, voice_recognition) concurrently:
+
+```bash
+curl -X POST http://localhost:8080/backend/load \
+  -H "Content-Type: application/json" \
+  -d '{"model": "gpt-realtime"}'
+```
+
+The endpoint is not realtime-specific — it pre-loads any model. See [Backend Monitor]({{%relref "features/backend-monitor" %}}) for the full request/response reference (it is the inverse of `/backend/shutdown`).
+
 ### Turn detection
 
 Turn detection decides when the user has finished speaking and the pipeline should respond. Two modes are supported, matching the OpenAI session schema:

docs/data/version.json

@@ -1,3 +1,3 @@
 {
-  "version": "v4.5.5"
+  "version": "v4.5.6"
 }

gallery/index.yaml

@@ -1,4 +1,113 @@
 ---
+- name: "qwopus3.6-35b-a3b-coder-mtp"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF
+  description: |
+    # 🌟 Qwopus3.6-35B-A3B-v1
+
+    ## 💡 Base Model Overview
+
+    **Qwen3.6-35B-A3B** is an advanced hybrid sparse MoE (Mixture-of-Experts) model developed by Alibaba Cloud. It features 35B total parameters with only 3B active parameters per token, ensuring high inference efficiency. Architecturally, it combines Gated DeltaNet linear attention with standard gated attention layers, routing tokens across **256 experts**. It natively supports a massive **262k context window** and is specifically designed for high-performance agentic coding, deep reasoning, and multimodal tasks.
+
+    ## 🚀 Model Refinement & Logic Tuning （Qwopus3.6-35B-A3B-v1）
+
+    🪐**Qwopus3.6-35B-A3B-v1** is a reasoning-enhanced MoE (Mixture of Experts) model fine-tuned on top of **Qwen3.6-35B-A3B**.
+
+    ### 🛠 Training Strategy
+
+    The fine-tuning process for this model is structured into **three distinct stages of distributed SFT (Supervised Fine-Tuning)**, progressively scaling reasoning complexity and data diversity. This systematic approach ensures the model inherits the base MoE capabilities while sharpening its logic-handling depth.
+
+    ...
+  license: "apache-2.0"
+  tags:
+    - llm
+    - gguf
+    - vision
+    - multimodal
+  icon: https://cdn-uploads.huggingface.co/production/uploads/66309bd090589b7c65950665/ztbyGV_zGhzcLuTCSVyq3.png
+  overrides:
+    backend: llama-cpp
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    mmproj: llama-cpp/mmproj/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M/mmproj-F32.gguf
+    options:
+      - use_jinja:true
+      - spec_type:draft-mtp
+      - spec_n_max:6
+      - spec_p_min:0.75
+    parameters:
+      model: llama-cpp/models/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf
+      sha256: c283cd2321a3cb4c6e7faf9481ac7d946913e4f02e20172eb2872112f567d8d4
+      uri: https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF/resolve/main/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M.gguf
+    - filename: llama-cpp/mmproj/Qwopus3.6-35B-A3B-Coder-MTP-Q4_K_M/mmproj-F32.gguf
+      sha256: 5c82c8095717b39f29c88ebfec3607a10307785b1e14a87744603d6c582cd497
+      uri: https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF/resolve/main/mmproj-F32.gguf
+- name: "ornith-1.0-9b-mtp"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/protoLabsAI/Ornith-1.0-9B-MTP-GGUF
+  description: |
+    [](https://deep-reinforce.com/ornith.html)
+
+    # Ornith-1.0-9B
+
+    Aloha! 🌺 Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding.
+
+    Highlights:
+
+      - **State-of-the-Art Coding Agents**: Available in 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE (post-trained on top of Gemma 4 and Qwen 3.5), achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo and OpenClaw.
+      - **Self-Improving Training Framework**:  Ornith-1.0 employs RL to learn to generate not only solution rollouts, but also the scallfold that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model  discovers better search trajectories and generates higher-quality solutions.
+      - **Licence**: MIT licensed, globally accessible, and free from regional limitations.
+
+    ## Ornith 1.0 9B
+
+    This model card documents **Ornith-1.0-9B**, the most lightweight member of the Ornith family, designed for efficient single-GPU deployment.
+
+    ### Benchmarks
+
+    Ornith-1.0-9B
+    Qwen3.5-9B
+    Qwen3.5-35B
+    Gemma4-12B
+    Gemma4-31B
+
+    Agentic Coding
+
+    ...
+  license: "mit"
+  tags:
+    - llm
+    - gguf
+  overrides:
+    backend: llama-cpp
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    options:
+      - use_jinja:true
+      - spec_type:draft-mtp
+      - spec_n_max:6
+      - spec_p_min:0.75
+    parameters:
+      model: llama-cpp/models/ornith-9b-mtp-kl-Q4_K_M/ornith-9b-mtp-kl-Q4_K_M.gguf
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/ornith-9b-mtp-kl-Q4_K_M/ornith-9b-mtp-kl-Q4_K_M.gguf
+      sha256: 03de14e9849a02258b9ac9ec2830af39dd7ac48f1ca5492b35539d1838707bc8
+      uri: https://huggingface.co/protoLabsAI/Ornith-1.0-9B-MTP-GGUF/resolve/main/ornith-9b-mtp-kl-Q4_K_M.gguf
 - name: "qwen-agentworld-35b-a3b"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -1716,7 +1825,7 @@
       - use_jinja:true
     parameters:
       min_p: 0.15
-      model: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q4_K_M.gguf
+      model: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0.gguf
       repeat_penalty: 1.05
       temperature: 0.1
       top_k: 50
@@ -1724,9 +1833,9 @@
     template:
       use_tokenizer_template: true
   files:
-    - filename: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q4_K_M.gguf
-      uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF/resolve/main/LFM2.5-8B-A1B-Q4_K_M.gguf
-      sha256: 4923ec14f06b968b74d663e5949867d2d9c3bf13a20b8be1a9f9af39989b2bb0
+    - filename: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0.gguf
+      uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF/resolve/main/LFM2.5-8B-A1B-Q8_0.gguf
+      sha256: 33ab3b8ce6a964fb8ebac89360c9b3cf72c4fa418d5e4c0a94d46883124d5c02
 - name: "qwopus3.5-9b-coder-mtp"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -1758,8 +1867,8 @@
       use_tokenizer_template: true
   files:
     - filename: llama-cpp/models/Qwopus3.5-9B-Coder-MTP-GGUF/Qwopus3.5-9B-Coder-MTP-Q4_K_M.gguf
-      sha256: f6fc5d193045796d9e1870cbc40f827fe55f53f70593c3f5c1968b82b9331991
       uri: https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/resolve/main/Qwopus3.5-9B-Coder-MTP-Q4_K_M.gguf
+      sha256: 9ea3ecd122a5165b8b81655f29eaf09d71daf841503e4c4212bdfadb36ab3712
     - filename: llama-cpp/mmproj/Qwopus3.5-9B-Coder-MTP-GGUF/Qwopus3.5-9B-Coder-MTP-mmproj.gguf
       sha256: f48daca405a1c768a9514e392c3955dcc4a9d66a5cf64cf45e064092b5f20ee4
       uri: https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF/resolve/main/Qwopus3.5-9B-Coder-MTP-mmproj.gguf

pkg/functions/peg/parser.go

@@ -461,10 +461,7 @@ func (p *RuleParser) parse(arena *Arena, ctx *ParseContext, start int) ParseResu
 	if result.Type != Fail {
 		text := ""
 		if result.Start < len(ctx.Input) {
-			end := result.End
-			if end > len(ctx.Input) {
-				end = len(ctx.Input)
-			}
+			end := min(result.End, len(ctx.Input))
 			text = ctx.Input[result.Start:end]
 		}
 
@@ -514,10 +511,7 @@ func (p *TagParser) parse(arena *Arena, ctx *ParseContext, start int) ParseResul
 	if result.Type != Fail {
 		text := ""
 		if result.Start < len(ctx.Input) {
-			end := result.End
-			if end > len(ctx.Input) {
-				end = len(ctx.Input)
-			}
+			end := min(result.End, len(ctx.Input))
 			text = ctx.Input[result.Start:end]
 		}
 

pkg/grpc/grpcerrors/errors.go

@@ -58,6 +58,23 @@ func IsLiveTranscriptionUnsupported(err error) bool {
 	return strings.Contains(strings.ToLower(err.Error()), "unimplemented")
 }
 
+// IsUnimplemented reports whether err is a gRPC Unimplemented status — the
+// signal a backend gives for an RPC it does not implement. The generated
+// UnimplementedBackendServer stub returns exactly this for any RPC a backend
+// (e.g. a Python or external backend) has not overridden, so callers can treat
+// an optional RPC as a no-op rather than a failure. Prefers the typed status
+// code and falls back to the message for paths that lose the status (e.g. errors
+// wrapped across non-gRPC boundaries).
+func IsUnimplemented(err error) bool {
+	if err == nil {
+		return false
+	}
+	if status.Code(err) == codes.Unimplemented {
+		return true
+	}
+	return strings.Contains(strings.ToLower(err.Error()), "unimplemented")
+}
+
 // StreamTranscriptionUnsupported returns the canonical error a backend returns
 // when it (or the loaded model) cannot serve the server-streaming
 // AudioTranscriptionStream RPC. It carries codes.Unimplemented like the live

pkg/grpc/grpcerrors/errors_test.go

@@ -55,6 +55,18 @@ var _ = Describe("grpcerrors", func() {
 		Expect(grpcerrors.IsModelNotLoaded(err)).To(BeFalse())
 	})
 
+	DescribeTable("IsUnimplemented",
+		func(err error, want bool) {
+			Expect(grpcerrors.IsUnimplemented(err)).To(Equal(want))
+		},
+		Entry("nil", nil, false),
+		Entry("typed code", status.Error(codes.Unimplemented, "method Free not implemented"), true),
+		Entry("stale stub message (Unknown code)", errors.New("rpc error: code = Unimplemented desc = "), true),
+		Entry("unrelated error", errors.New("context deadline exceeded"), false),
+		Entry("unrelated grpc code", status.Error(codes.Unavailable, "connection refused"), false),
+		Entry("model not loaded is NOT unimplemented", grpcerrors.ModelNotLoaded("parakeet-cpp"), false),
+	)
+
 	It("StreamTranscriptionUnsupported carries Unimplemented and is not ModelNotLoaded", func() {
 		err := grpcerrors.StreamTranscriptionUnsupported("parakeet-cpp", "not a streaming model")
 		Expect(status.Code(err)).To(Equal(codes.Unimplemented))

pkg/grpc/parentwatch.go

@@ -0,0 +1,105 @@
+package grpc
+
+import (
+	"log"
+	"os"
+	"runtime"
+	"strings"
+	"time"
+)
+
+// Backend worker processes (the per-model gRPC servers LocalAI spawns) are
+// deliberately placed in their own process group by the process manager so
+// LocalAI's graceful shutdown can signal the whole group. That graceful path
+// (SIGTERM -> grace -> SIGKILL, driven by pkg/signals + pkg/model) only runs
+// when LocalAI itself receives a catchable signal and lives long enough to run
+// its handlers. If LocalAI is SIGKILLed (e.g. a supervising process's
+// graceful-shutdown grace period elapses first), that teardown never runs and
+// this backend would be reparented to init and linger, holding VRAM and its
+// listen port.
+//
+// The watcher below is a best-effort backstop for exactly that case: it does
+// NOT replace the graceful teardown, it only covers the "parent vanished
+// without cleaning up" path. It works by detecting reparenting: when the
+// process that spawned this backend dies, the kernel reparents us to the
+// nearest sub-reaper or to init (PID 1), so getppid() stops matching the value
+// we captured at startup. This getppid() approach is portable across
+// Linux/macOS (unlike Linux-only PR_SET_PDEATHSIG), which is why it's used
+// here rather than a kernel parent-death signal.
+const (
+	// EnvBackendParentWatch toggles the parent-death watcher. It is enabled by
+	// default; set it to a falsey value ("false", "0", "no", "off") to disable
+	// (e.g. when running a backend standalone for debugging under a shell whose
+	// lifetime shouldn't govern the backend).
+	EnvBackendParentWatch = "LOCALAI_BACKEND_PARENT_WATCH"
+	// EnvBackendParentWatchInterval overrides the poll interval as a Go
+	// duration string (e.g. "500ms"). Defaults to defaultParentWatchInterval.
+	EnvBackendParentWatchInterval = "LOCALAI_BACKEND_PARENT_WATCH_INTERVAL"
+
+	defaultParentWatchInterval = 2 * time.Second
+)
+
+// parentWatchEnabled reports whether the watcher should run in this process.
+func parentWatchEnabled() bool {
+	switch strings.ToLower(strings.TrimSpace(os.Getenv(EnvBackendParentWatch))) {
+	case "false", "0", "no", "off":
+		return false
+	}
+	// Windows does not reparent orphans to a well-known init PID, so the
+	// getppid() heuristic used here doesn't apply there.
+	return runtime.GOOS != "windows"
+}
+
+// parentWatchInterval returns the configured poll interval, or the default.
+func parentWatchInterval() time.Duration {
+	if v := os.Getenv(EnvBackendParentWatchInterval); v != "" {
+		if d, err := time.ParseDuration(v); err == nil && d > 0 {
+			return d
+		}
+	}
+	return defaultParentWatchInterval
+}
+
+// parentDied reports whether this process has been reparented away from the
+// parent it had when the watcher started. Reparenting is the standard POSIX
+// signal that the original parent (here, the LocalAI process that spawned this
+// backend) has exited: the orphan is handed to the nearest sub-reaper or to
+// init (PID 1), so getppid() no longer matches the value captured at startup.
+func parentDied(origPPID int) bool {
+	ppid := os.Getppid()
+	return ppid != origPPID || ppid == 1
+}
+
+// watchParentDeath polls until parentDied reports the original parent is gone,
+// then invokes onDeath. It blocks, so run it in its own goroutine.
+func watchParentDeath(origPPID int, interval time.Duration, onDeath func()) {
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+	for range ticker.C {
+		if parentDied(origPPID) {
+			onDeath()
+			return
+		}
+	}
+}
+
+// startParentDeathWatcher installs the best-effort safety net described above
+// on the calling backend process. It is a no-op when disabled or on platforms
+// where the mechanism doesn't apply. This is a backstop alongside — never a
+// replacement for — LocalAI's graceful SIGTERM->grace->SIGKILL teardown.
+func startParentDeathWatcher() {
+	if !parentWatchEnabled() {
+		return
+	}
+	origPPID := os.Getppid()
+	// A parent of 1 at startup means we were already orphaned (or launched
+	// directly under init) — there's no original parent to watch for.
+	if origPPID <= 1 {
+		return
+	}
+	interval := parentWatchInterval()
+	go watchParentDeath(origPPID, interval, func() {
+		log.Printf("backend parent process (pid %d) exited without stopping this backend; self-terminating to avoid orphaning", origPPID)
+		os.Exit(1)
+	})
+}

pkg/grpc/parentwatch_test.go

@@ -0,0 +1,163 @@
+//go:build !windows
+
+package grpc
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"syscall"
+	"time"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// These env vars drive the helper roles this test binary re-executes itself as
+// (see the init() dispatcher). They are only set for the spawned child/
+// grandchild processes, never for the normal `go test` invocation.
+const (
+	envRole   = "LOCALAI_PARENTWATCH_TEST_ROLE"
+	envReady  = "LOCALAI_PARENTWATCH_TEST_READY"  // grandchild writes its PID here once the watcher is armed
+	envExited = "LOCALAI_PARENTWATCH_TEST_EXITED" // grandchild writes here when it detects reparenting
+)
+
+// init dispatches the helper roles when this test binary is re-executed with a
+// role set. It runs before the testing/Ginkgo machinery, and is a no-op during
+// a normal test run (role unset).
+func init() {
+	switch os.Getenv(envRole) {
+	case "middle":
+		runMiddleRole()
+	case "grandchild":
+		runGrandchildRole()
+	}
+}
+
+// childEnv returns the current environment with the parentwatch test role set
+// to the given value (replacing any inherited role), leaving the ready/exited
+// file paths inherited.
+func childEnv(role string) []string {
+	out := make([]string, 0, len(os.Environ())+1)
+	for _, kv := range os.Environ() {
+		if len(kv) > len(envRole) && kv[:len(envRole)+1] == envRole+"=" {
+			continue
+		}
+		out = append(out, kv)
+	}
+	return append(out, envRole+"="+role)
+}
+
+// runGrandchildRole arms the REAL watchParentDeath against its current parent
+// (the "middle" process), signals readiness, then blocks. When middle exits and
+// we are reparented, the watcher fires and we record it before exiting.
+func runGrandchildRole() {
+	exitedFile := os.Getenv(envExited)
+	readyFile := os.Getenv(envReady)
+
+	origPPID := os.Getppid()
+	go watchParentDeath(origPPID, 50*time.Millisecond, func() {
+		_ = os.WriteFile(exitedFile, []byte("1"), 0o644)
+		os.Exit(7)
+	})
+
+	// Safety valve: never linger if something goes wrong with the test.
+	go func() {
+		time.Sleep(30 * time.Second)
+		os.Exit(2)
+	}()
+
+	// Signal readiness only after the watcher captured origPPID, so middle
+	// won't exit before we've recorded it as our original parent.
+	_ = os.WriteFile(readyFile, []byte(strconv.Itoa(os.Getpid())), 0o644)
+
+	select {} // block until the watcher terminates us
+}
+
+// runMiddleRole spawns the grandchild (which arms the watcher against us),
+// waits until it is ready, then exits — orphaning the grandchild so it gets
+// reparented, which is what the watcher must detect.
+func runMiddleRole() {
+	readyFile := os.Getenv(envReady)
+
+	self, err := os.Executable()
+	if err != nil {
+		os.Exit(3)
+	}
+	cmd := exec.Command(self)
+	cmd.Env = childEnv("grandchild")
+	// Own process group, mirroring how real backends are spawned, and discard
+	// std streams so the grandchild doesn't keep any parent pipe open.
+	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	if err := cmd.Start(); err != nil {
+		os.Exit(4)
+	}
+
+	if !waitForFile(readyFile, 10*time.Second) {
+		os.Exit(5)
+	}
+	os.Exit(0) // orphan the grandchild
+}
+
+func waitForFile(path string, timeout time.Duration) bool {
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
+		if _, err := os.Stat(path); err == nil {
+			return true
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	return false
+}
+
+// This spec builds a genuine two-level process tree (test -> middle ->
+// grandchild), lets the middle process die, and asserts the grandchild's
+// watchParentDeath detects the reparenting and self-terminates.
+var _ = Describe("watchParentDeath", func() {
+	It("detects reparenting and self-terminates the orphaned process", func() {
+		if runtime.GOOS == "windows" {
+			Skip("parent-death watcher is not supported on windows")
+		}
+
+		dir := GinkgoT().TempDir()
+		readyFile := filepath.Join(dir, "ready")
+		exitedFile := filepath.Join(dir, "exited")
+
+		self, err := os.Executable()
+		Expect(err).NotTo(HaveOccurred(), "cannot resolve test executable")
+
+		middle := exec.Command(self)
+		middle.Env = append(childEnv("middle"),
+			envReady+"="+readyFile,
+			envExited+"="+exitedFile,
+		)
+		// Discard the helpers' output; keep the test log clean.
+		middle.Stdout = nil
+		middle.Stderr = nil
+
+		Expect(middle.Start()).To(Succeed(), "failed to start middle helper")
+		// Wait only for the middle process; the grandchild is intentionally left
+		// orphaned. No pipes are shared, so this returns as soon as middle exits.
+		Expect(middle.Wait()).To(Succeed(), "middle helper exited with error")
+
+		// The grandchild must have armed the watcher (and thus captured middle as
+		// its parent) before middle exited.
+		_, err = os.Stat(readyFile)
+		Expect(err).NotTo(HaveOccurred(), "grandchild never signaled readiness")
+
+		// Best-effort cleanup in case the watcher somehow doesn't fire.
+		DeferCleanup(func() {
+			if b, err := os.ReadFile(readyFile); err == nil {
+				if pid, err := strconv.Atoi(string(b)); err == nil {
+					_ = syscall.Kill(pid, syscall.SIGKILL)
+				}
+			}
+		})
+
+		// Now that middle is gone, the grandchild has been reparented; the watcher
+		// must notice and write the exited marker.
+		Expect(waitForFile(exitedFile, 10*time.Second)).To(BeTrue(), "watcher did not detect parent death within timeout")
+	})
+})

pkg/grpc/server.go

@@ -939,6 +939,9 @@ func StartServer(address string, model AIModel) error {
 	s := grpc.NewServer(serverOpts()...)
 	pb.RegisterBackendServer(s, &server{llm: model})
 	log.Printf("gRPC Server listening at %v", lis.Addr())
+	// Safety net: self-terminate if the LocalAI process that spawned this
+	// backend dies without running its graceful teardown (see parentwatch.go).
+	startParentDeathWatcher()
 	if err := s.Serve(lis); err != nil {
 		return err
 	}
@@ -954,6 +957,9 @@ func RunServer(address string, model AIModel) (func() error, error) {
 	s := grpc.NewServer(serverOpts()...)
 	pb.RegisterBackendServer(s, &server{llm: model})
 	log.Printf("gRPC Server listening at %v", lis.Addr())
+	// Safety net: self-terminate if the LocalAI process that spawned this
+	// backend dies without running its graceful teardown (see parentwatch.go).
+	startParentDeathWatcher()
 	if err = s.Serve(lis); err != nil {
 		return func() error {
 			return lis.Close()

pkg/mcp/localaitools/client.go

@@ -36,6 +36,10 @@ type LocalAIClient interface {
 	DeleteModel(ctx context.Context, name string) error
 	EditModelConfig(ctx context.Context, name string, patch map[string]any) error
 	ReloadModels(ctx context.Context) error
+	// LoadModel pre-loads a model into memory by name (the inverse of shutting
+	// it down). For a realtime pipeline model every configured sub-model is
+	// loaded; it returns the model names that became resident.
+	LoadModel(ctx context.Context, model string) ([]string, error)
 	ImportModelURI(ctx context.Context, req ImportModelURIRequest) (*ImportModelURIResponse, error)
 
 	// ---- Model aliases ----

pkg/mcp/localaitools/coverage_test.go

@@ -49,6 +49,7 @@ var toolToHTTPRoute = map[string]string{
 	ToolDeleteModel:       "POST /models/delete/:name",
 	ToolEditModelConfig:   "PATCH /api/models/config-json/:name",
 	ToolReloadModels:      "POST /models/reload",
+	ToolLoadModel:         "POST /backend/load",
 	ToolInstallBackend:    "POST /backends/apply",
 	ToolUpgradeBackend:    "POST /backends/upgrade/:name",
 	ToolToggleModelState:  "PUT /models/toggle-state/:name/:action",

pkg/mcp/localaitools/fakes_test.go

@@ -35,6 +35,7 @@ type fakeClient struct {
 	setAlias            func(string, string) error
 	listAliases         func() ([]AliasInfo, error)
 	reloadModels        func() error
+	loadModel           func(string) ([]string, error)
 	listBackends        func() ([]Backend, error)
 	listKnownBackends   func() ([]schema.KnownBackend, error)
 	installBackend      func(InstallBackendRequest) (string, error)
@@ -169,6 +170,14 @@ func (f *fakeClient) ReloadModels(_ context.Context) error {
 	return nil
 }
 
+func (f *fakeClient) LoadModel(_ context.Context, model string) ([]string, error) {
+	f.record("LoadModel", model)
+	if f.loadModel != nil {
+		return f.loadModel(model)
+	}
+	return []string{model}, nil
+}
+
 func (f *fakeClient) ListBackends(_ context.Context) ([]Backend, error) {
 	f.record("ListBackends", nil)
 	if f.listBackends != nil {

pkg/mcp/localaitools/httpapi/client.go

@@ -338,6 +338,16 @@ func (c *Client) ReloadModels(ctx context.Context) error {
 	return c.do(ctx, http.MethodPost, routeModelsReload, nil, nil)
 }
 
+func (c *Client) LoadModel(ctx context.Context, model string) ([]string, error) {
+	// On a load failure the endpoint returns a non-2xx whose body (carrying the
+	// per-sub-model failure detail) is folded into the HTTPError by c.do.
+	var resp schema.ModelLoadResponse
+	if err := c.do(ctx, http.MethodPost, routeBackendLoad, map[string]string{"model": model}, &resp); err != nil {
+		return nil, err
+	}
+	return resp.Loaded, nil
+}
+
 // ---- Model aliases ----
 
 // SetAlias is swap-first: it PATCHes the alias config (a deep-merge that

pkg/mcp/localaitools/httpapi/routes.go

@@ -19,6 +19,7 @@ const (
 	routeModelImport     = "/models/import"
 	routeAliases         = "/api/aliases"
 	routeModelsReload    = "/models/reload"
+	routeBackendLoad     = "/backend/load"
 	routeBackends        = "/backends"
 	routeBackendsKnown   = "/backends/known"
 	routeBackendsApply   = "/backends/apply"

pkg/mcp/localaitools/inproc/client.go

@@ -13,6 +13,7 @@ import (
 	"path/filepath"
 
 	"github.com/google/uuid"
+	"github.com/mudler/LocalAI/core/backend"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/gallery"
 	"github.com/mudler/LocalAI/core/gallery/importers"
@@ -302,6 +303,16 @@ func (c *Client) ReloadModels(_ context.Context) error {
 	return c.ConfigLoader.LoadModelConfigsFromPath(c.SystemState.Model.ModelsPath)
 }
 
+func (c *Client) LoadModel(ctx context.Context, model string) ([]string, error) {
+	if c.ConfigLoader == nil || c.ModelLoader == nil {
+		return nil, errors.New("model loader not available")
+	}
+	// Reuse the same preload path the REST /backend/load endpoint uses, so a
+	// pipeline model loads all its sub-models and the behaviour stays identical
+	// across the in-process and HTTP clients.
+	return backend.PreloadModelByName(ctx, c.ConfigLoader, c.ModelLoader, c.AppConfig, model)
+}
+
 // ---- Model aliases ----
 
 // SetAlias is swap-first to match the httpapi client: PatchConfig swaps an

pkg/mcp/localaitools/inproc/load_model_test.go

@@ -0,0 +1,71 @@
+package inproc
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/system"
+)
+
+var _ = Describe("inproc.Client LoadModel", func() {
+	var (
+		ctx       context.Context
+		tempDir   string
+		cl        *config.ModelConfigLoader
+		ml        *model.ModelLoader
+		c         *Client
+		seedModel func(name, body string)
+	)
+
+	BeforeEach(func() {
+		ctx = context.Background()
+		tempDir = GinkgoT().TempDir()
+		systemState, err := system.GetSystemState(system.WithModelPath(tempDir))
+		Expect(err).ToNot(HaveOccurred())
+		appConfig := config.NewApplicationConfig(config.WithSystemState(systemState))
+		cl = config.NewModelConfigLoader(tempDir)
+		ml = model.NewModelLoader(systemState) // no backends installed
+		c = New(appConfig, systemState, cl, ml, nil)
+
+		seedModel = func(name, body string) {
+			Expect(os.WriteFile(filepath.Join(tempDir, name+".yaml"), []byte(body), 0o644)).To(Succeed())
+			Expect(cl.LoadModelConfigsFromPath(tempDir)).To(Succeed())
+		}
+	})
+
+	It("errors when the model loader is unavailable", func() {
+		noLoader := New(c.AppConfig, c.SystemState, cl, nil, nil)
+		_, err := noLoader.LoadModel(ctx, "anything")
+		Expect(err).To(MatchError(ContainSubstring("model loader not available")))
+	})
+
+	It("loads a regular model through the model loader", func() {
+		seedModel("solo", "name: solo\n")
+		// No backend is installed in the test env, so the load itself fails — but
+		// the call must exercise the single-model path and surface that error
+		// rather than panicking or silently succeeding.
+		loaded, err := c.LoadModel(ctx, "solo")
+		Expect(err).To(HaveOccurred())
+		Expect(loaded).To(BeEmpty())
+	})
+
+	It("expands a pipeline model into its sub-models", func() {
+		seedModel("voicebot", "name: voicebot\npipeline:\n  vad: vad-m\n  llm: llm-m\n")
+		seedModel("vad-m", "name: vad-m\n")
+		seedModel("llm-m", "name: llm-m\n")
+
+		loaded, err := c.LoadModel(ctx, "voicebot")
+		// Sub-models can't load without backends, so the joined error names them
+		// — proving the pipeline stub was expanded rather than loaded directly.
+		Expect(err).To(HaveOccurred())
+		Expect(err.Error()).To(ContainSubstring("vad-m"))
+		Expect(err.Error()).ToNot(ContainSubstring("voicebot"))
+		Expect(loaded).To(BeEmpty())
+	})
+})

pkg/mcp/localaitools/prompts/10_safety.md

@@ -2,7 +2,7 @@
 
 These rules are non-negotiable. The user trusts you to operate their server without unintended changes.
 
-1. **Confirm before mutating.** Before calling any of these tools — `install_model`, `import_model_uri`, `delete_model`, `install_backend`, `upgrade_backend`, `edit_model_config`, `reload_models`, `toggle_model_state`, `toggle_model_pinned` — first state in plain language what you are about to do (which tool, which target, which arguments) and wait for the user's explicit confirmation in the next turn. "Yes", "do it", "go ahead", "proceed" all count as confirmation. Anything else does not.
+1. **Confirm before mutating.** Before calling any of these tools — `install_model`, `import_model_uri`, `delete_model`, `install_backend`, `upgrade_backend`, `edit_model_config`, `reload_models`, `load_model`, `toggle_model_state`, `toggle_model_pinned` — first state in plain language what you are about to do (which tool, which target, which arguments) and wait for the user's explicit confirmation in the next turn. "Yes", "do it", "go ahead", "proceed" all count as confirmation. Anything else does not.
 
 2. **Disambiguate before mutating.** If the user's request is ambiguous (several gallery candidates match, the model name has multiple installed versions, the backend has variants), present the candidates as a numbered list and ask the user to pick before calling any mutating tool.