fix(openresponses): bound resume-stream buffer and enforce response ownership (#10569)

The background=true resumable-stream path had two latent issues.

1. Unbounded resume buffer. AppendEvent grew StreamEvents without limit, so
   a long-running or abandoned background generation could consume process
   memory without bound. The store now caps the buffer (event count and total
   bytes, mirroring llama.cpp's byte-capped slot ring), evicting oldest events
   from the front and advancing a droppedThrough watermark. GetEventsAfter
   returns ErrOffsetLost when the requested starting_after is below the
   watermark, and handleStreamResume surfaces that as HTTP 409 before
   committing to the SSE response, so a resuming client gets a clear error
   instead of a silently truncated stream.

2. Missing ownership check (IDOR). GET /responses/:id, its stream resume, and
   /cancel looked up responses purely by ID, letting any caller who knows or
   guesses an ID read or cancel another caller's response. Responses now carry
   the creating caller's identity (auth.GetUser), stamped at creation and
   compared on read/cancel/resume; a mismatch returns 404 (not 403) so
   existence is not leaked. Backward compatible: responses with no owner
   (single-key / no-auth deployments) remain accessible.


Assisted-by: Claude:claude-opus-4-8 [Claude Code]

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>

.github/backend-matrix.yml

@@ -4991,9 +4991,6 @@ includeDarwin:
   - backend: "qwen-tts"
     tag-suffix: "-metal-darwin-arm64-qwen-tts"
     build-type: "mps"
-  - backend: "fish-speech"
-    tag-suffix: "-metal-darwin-arm64-fish-speech"
-    build-type: "mps"
   - backend: "voxcpm"
     tag-suffix: "-metal-darwin-arm64-voxcpm"
     build-type: "mps"

.github/workflows/backend_build_darwin.yml

@@ -99,6 +99,7 @@ jobs:
             /opt/homebrew/Cellar/xxhash
             /opt/homebrew/Cellar/zstd
             /opt/homebrew/Cellar/nlohmann-json
+            /opt/homebrew/Cellar/opus
           key: brew-${{ runner.os }}-${{ runner.arch }}-v1-${{ hashFiles('.github/workflows/backend_build_darwin.yml') }}
 
       - name: Dependencies
@@ -113,7 +114,12 @@ jobs:
           # nlohmann-json is header-only and required by the ds4 backend
           # (dsml_renderer.cpp includes <nlohmann/json.hpp>); on Linux it comes
           # from the apt-installed nlohmann-json3-dev in the build image.
-          brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json
+          # opus + pkg-config are required by the opus go backend: its
+          # Makefile/package.sh call `pkg-config --cflags/--libs opus` to build
+          # libopusshim.dylib and to locate libopus.dylib for bundling. brew's
+          # pkg-config defaults its search path to the Homebrew prefix so the
+          # opus.pc is found.
+          brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json opus pkg-config
           # Force-reinstall ccache so brew re-validates its full runtime-dep
           # closure on every run. This is the durable fix: when the upstream
           # ccache formula gains a new transitive dep (as it has multiple times
@@ -132,7 +138,7 @@ jobs:
           # and decides "already installed" without re-linking, so on a cache-
           # hit run the formulas aren't on PATH. Force-link them; --overwrite
           # tolerates pre-existing symlinks from earlier installs.
-          brew link --overwrite protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json 2>/dev/null || true
+          brew link --overwrite protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm ccache blake3 fmt hiredis xxhash zstd nlohmann-json opus pkg-config 2>/dev/null || true
 
       - name: Save Homebrew cache
         if: github.event_name != 'pull_request' && steps.brew-cache.outputs.cache-hit != 'true'
@@ -153,6 +159,7 @@ jobs:
             /opt/homebrew/Cellar/xxhash
             /opt/homebrew/Cellar/zstd
             /opt/homebrew/Cellar/nlohmann-json
+            /opt/homebrew/Cellar/opus
           key: brew-${{ runner.os }}-${{ runner.arch }}-v1-${{ hashFiles('.github/workflows/backend_build_darwin.yml') }}
 
       # ---- ccache for llama.cpp CMake builds ----

.github/workflows/release.yaml

@@ -24,6 +24,11 @@ jobs:
           args: release --clean
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          MACOS_SIGN_P12: ${{ secrets.MACOS_CERTIFICATE }}
+          MACOS_SIGN_PASSWORD: ${{ secrets.MACOS_CERTIFICATE_PWD }}
+          MACOS_NOTARY_KEY: ${{ secrets.MACOS_NOTARY_KEY }}
+          MACOS_NOTARY_KEY_ID: ${{ secrets.MACOS_NOTARY_KEY_ID }}
+          MACOS_NOTARY_ISSUER_ID: ${{ secrets.MACOS_NOTARY_ISSUER_ID }}
   launcher-build-darwin:
     runs-on: macos-latest
     steps:
@@ -35,9 +40,19 @@ jobs:
         uses: actions/setup-go@v5
         with:
           go-version: 1.23
-      - name: Build launcher for macOS ARM64
-        run: |
-          make build-launcher-darwin
+      - name: Import signing certificate
+        env:
+          MACOS_CERTIFICATE: ${{ secrets.MACOS_CERTIFICATE }}
+          MACOS_CERTIFICATE_PWD: ${{ secrets.MACOS_CERTIFICATE_PWD }}
+          MACOS_CI_KEYCHAIN_PWD: ${{ secrets.MACOS_CI_KEYCHAIN_PWD }}
+        run: bash contrib/macos/sign-and-notarize.sh import-cert
+      - name: Build, sign and notarize the DMG
+        env:
+          MACOS_SIGN_IDENTITY: ${{ secrets.MACOS_SIGN_IDENTITY }}
+          MACOS_NOTARY_KEY: ${{ secrets.MACOS_NOTARY_KEY }}
+          MACOS_NOTARY_KEY_ID: ${{ secrets.MACOS_NOTARY_KEY_ID }}
+          MACOS_NOTARY_ISSUER_ID: ${{ secrets.MACOS_NOTARY_ISSUER_ID }}
+        run: make release-launcher-darwin
       - name: Upload DMG to Release
         uses: softprops/action-gh-release@v3
         with:

.github/workflows/test-extra.yml

@@ -1008,7 +1008,11 @@ jobs:
   # image + working dir.
   tests-vibevoice-cpp-grpc-transcription:
     needs: detect-changes
-    if: needs.detect-changes.outputs.vibevoice-cpp == 'true' || needs.detect-changes.outputs.run-all == 'true'
+    # Skip on release tag pushes: the ASR Q4_K model is ~10 GB and cannot be
+    # pulled from HF within the inner `go test -timeout 30m` budget on a CI
+    # runner, so every tag build hung and timed out. Still runs on PRs/branch
+    # pushes that touch vibevoice-cpp so regressions are caught off the release path.
+    if: (needs.detect-changes.outputs.vibevoice-cpp == 'true' || needs.detect-changes.outputs.run-all == 'true') && !startsWith(github.ref, 'refs/tags/')
     runs-on: bigger-runner
     timeout-minutes: 150
     steps:

.github/workflows/test.yml

@@ -121,3 +121,19 @@ jobs:
           detached: true
           connect-timeout-seconds: 180
           limit-access-to-actor: true
+
+  # Fast standalone unit tests for the backends' pure C++ helpers - currently the
+  # llama-cpp message reconstruction (backend/cpp/llama-cpp/message_content.h),
+  # which guards the OpenAI chat content normalization (mudler/LocalAI#10524,
+  # #7324, #7528). The runner discovers every *_test.cpp under backend/cpp/, so
+  # new pure-C++ unit tests are picked up with no CI changes. These need only the
+  # C++ stdlib + nlohmann/json, so they run on every PR without the full
+  # llama.cpp + gRPC backend build. (The same suite is also wired as an opt-in
+  # CMake/ctest target, -DLLAMA_GRPC_BUILD_TESTS=ON, for in-backend-build runs.)
+  tests-backend-cpp:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Clone
+        uses: actions/checkout@v7
+      - name: Run backend C++ unit tests
+        run: make test-backend-cpp

.gitignore

@@ -94,3 +94,6 @@ core/http/react-ui/test-results/
 
 # SDD / brainstorm scratch (agent-driven development)
 .superpowers/
+
+# Local Apple signing material (never commit)
+.certs/

.goreleaser.yaml

@@ -9,7 +9,8 @@ source:
   enabled: true
   name_template: '{{ .ProjectName }}-{{ .Tag }}-source'
 builds:
-  - main: ./cmd/local-ai
+  - id: local-ai
+    main: ./cmd/local-ai
     env:
       - CGO_ENABLED=0
     ldflags:
@@ -35,3 +36,19 @@ snapshot:
   version_template: "{{ .Tag }}-next"
 changelog:
   use: github-native
+# Sign + notarize the macOS server binary via the quill backend (runs on Linux,
+# no macOS runner needed). Disabled automatically when MACOS_SIGN_P12 is unset
+# (forks / PRs), so those builds stay unsigned and green.
+notarize:
+  macos:
+    - enabled: '{{ isEnvSet "MACOS_SIGN_P12" }}'
+      ids:
+        - local-ai
+      sign:
+        certificate: "{{.Env.MACOS_SIGN_P12}}"
+        password: "{{.Env.MACOS_SIGN_PASSWORD}}"
+      notarize:
+        issuer_id: "{{.Env.MACOS_NOTARY_ISSUER_ID}}"
+        key_id: "{{.Env.MACOS_NOTARY_KEY_ID}}"
+        key: "{{.Env.MACOS_NOTARY_KEY}}"
+        wait: true

Makefile

@@ -103,7 +103,7 @@ COVERAGE_E2E_LABELS?=!real-models
 COVERAGE_EXCLUDE_RE?=grpc/proto/.*[.]pb[.]go
 
 
-.PHONY: all test test-coverage test-coverage-baseline test-coverage-check test-ui test-ui-coverage-baseline test-ui-coverage-check install-hooks build vendor lint lint-all
+.PHONY: all test test-coverage test-coverage-baseline test-coverage-check test-backend-cpp test-ui test-ui-coverage-baseline test-ui-coverage-check install-hooks build vendor lint lint-all
 
 all: help
 
@@ -201,6 +201,13 @@ test: prepare-test
 	OPUS_SHIM_LIBRARY=$(abspath ./pkg/opus/shim/libopusshim.so) \
 	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
 
+## Compiles and runs the standalone C++ unit tests for the backends (pure
+## helpers that depend only on the stdlib + nlohmann/json, no full backend
+## build). Discovers every *_test.cpp under backend/cpp/ - see
+## backend/cpp/run-unit-tests.sh. Set NLOHMANN_INCLUDE to skip the header fetch.
+test-backend-cpp:
+	bash backend/cpp/run-unit-tests.sh
+
 ## Runs the core suite ($(TEST_PATHS)) with statement-coverage instrumentation
 ## and writes a merged profile to $(COVERAGE_PROFILE). Deliberately omits
 ## --fail-fast so a single failure doesn't truncate the coverage number, and
@@ -1453,13 +1460,32 @@ docs: docs/static/gallery.html
 ########################################################
 
 ## fyne cross-platform build
-build-launcher-darwin: build-launcher
-	go run github.com/tiagomelo/macos-dmg-creator/cmd/createdmg@latest \
-	--appName "LocalAI" \
-	--appBinaryPath "$(LAUNCHER_BINARY_NAME)" \
-	--bundleIdentifier "com.localai.launcher" \
-	--iconPath "core/http/static/logo.png" \
-	--outputDir "dist/"
+# Build LocalAI.app from the launcher via fyne (metadata read from cmd/launcher/FyneApp.toml).
+# Signing happens via contrib/macos/sign-and-notarize.sh, which is a no-op when the signing
+# secrets are unset, so unsigned local/fork builds keep working.
+build-launcher-darwin:
+	rm -rf dist/LocalAI.app cmd/launcher/LocalAI.app
+	mkdir -p dist
+	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os darwin -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)
+	mv cmd/launcher/LocalAI.app dist/LocalAI.app
+	bash contrib/macos/sign-and-notarize.sh sign dist/LocalAI.app
+
+# Wrap the (signed) app into a drag-to-Applications DMG via hdiutil, then sign the DMG.
+dmg-launcher-darwin: build-launcher-darwin
+	rm -rf dist/dmg dist/LocalAI.dmg
+	mkdir -p dist/dmg
+	cp -R dist/LocalAI.app dist/dmg/LocalAI.app
+	ln -s /Applications dist/dmg/Applications
+	hdiutil create -volname "LocalAI" -srcfolder dist/dmg -ov -format UDZO dist/LocalAI.dmg
+	bash contrib/macos/sign-and-notarize.sh sign dist/LocalAI.dmg
+
+# Submit the DMG to Apple notarization and staple the ticket (no-op without notary secrets).
+notarize-launcher-darwin: dmg-launcher-darwin
+	bash contrib/macos/sign-and-notarize.sh notarize dist/LocalAI.dmg
+
+# Single entrypoint for CI: build -> sign app -> dmg -> sign dmg -> notarize -> staple.
+release-launcher-darwin: notarize-launcher-darwin
+	@echo "dist/LocalAI.dmg is ready"
 
 build-launcher-linux:
-	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os linux -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)-linux && mv launcher.tar.xz ../../$(LAUNCHER_BINARY_NAME)-linux.tar.xz
+	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os linux -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)-linux && mv LocalAI.tar.xz ../../$(LAUNCHER_BINARY_NAME)-linux.tar.xz

backend/cpp/llama-cpp/CMakeLists.txt

@@ -87,3 +87,18 @@ target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)
 endif()
+
+# Unit test for the message-content normalization helper (message_content.h).
+# Off by default so the normal backend build is untouched; enable with
+# -DLLAMA_GRPC_BUILD_TESTS=ON and run via ctest. It reuses llama.cpp's vendored
+# <nlohmann/json.hpp> (propagated by the common helpers library) so it has no
+# extra dependency beyond what the backend already builds against.
+option(LLAMA_GRPC_BUILD_TESTS "Build grpc-server unit tests" OFF)
+if(LLAMA_GRPC_BUILD_TESTS)
+    enable_testing()
+    add_executable(message_content_test message_content_test.cpp message_content.h)
+    target_include_directories(message_content_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+    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)
+endif()

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

@@ -39,6 +39,7 @@
 #include "common.h"
 #include "arg.h"
 #include "chat-auto-parser.h"
+#include "message_content.h"
 #include <getopt.h>
 #include <grpcpp/ext/proto_server_reflection_plugin.h>
 #include <grpcpp/grpcpp.h>
@@ -1616,242 +1617,20 @@ public:
 
                 for (int i = 0; i < request->messages_size(); i++) {
                     const auto& msg = request->messages(i);
-                    json msg_json;
-                    msg_json["role"] = msg.role();
-
-                    bool is_last_user_msg = (i == last_user_msg_idx);
-                    bool has_images_or_audio = (request->images_size() > 0 || request->audios_size() > 0 || request->videos_size() > 0);
-
-                    // Handle content - can be string, null, or array
-                    // For multimodal content, we'll embed images/audio from separate fields
-                    if (!msg.content().empty()) {
-                        // Try to parse content as JSON to see if it's already an array
-                        json content_val;
-                        try {
-                            content_val = json::parse(msg.content());
-                            // Handle null values - convert to empty string to avoid template errors
-                            if (content_val.is_null()) {
-                                content_val = "";
-                            }
-                        } catch (const json::parse_error&) {
-                            // Not JSON, treat as plain string
-                            content_val = msg.content();
-                        }
-
-                        // If content is an object (e.g., from tool call failures), convert to string
-                        if (content_val.is_object()) {
-                            content_val = content_val.dump();
-                        }
-
-                        // If content is a string and this is the last user message with images/audio, combine them
-                        if (content_val.is_string() && is_last_user_msg && has_images_or_audio) {
-                            json content_array = json::array();
-                            // Add text first
-                            content_array.push_back({{"type", "text"}, {"text", content_val.get<std::string>()}});
-                            // Add images
-                            if (request->images_size() > 0) {
-                                for (int j = 0; j < request->images_size(); j++) {
-                                    json image_chunk;
-                                    image_chunk["type"] = "image_url";
-                                    json image_url;
-                                    image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                    image_chunk["image_url"] = image_url;
-                                    content_array.push_back(image_chunk);
-                                }
-                            }
-                            // Add audios
-                            if (request->audios_size() > 0) {
-                                for (int j = 0; j < request->audios_size(); j++) {
-                                    json audio_chunk;
-                                    audio_chunk["type"] = "input_audio";
-                                    json input_audio;
-                                    input_audio["data"] = request->audios(j);
-                                    input_audio["format"] = "wav"; // default, could be made configurable
-                                    audio_chunk["input_audio"] = input_audio;
-                                    content_array.push_back(audio_chunk);
-                                }
-                            }
-                            if (request->videos_size() > 0) {
-                                for (int j = 0; j < request->videos_size(); j++) {
-                                    json video_chunk;
-                                    video_chunk["type"] = "input_video";
-                                    json input_video;
-                                    input_video["data"] = request->videos(j);
-                                    video_chunk["input_video"] = input_video;
-                                    content_array.push_back(video_chunk);
-                                }
-                            }
-                            msg_json["content"] = content_array;
-                        } else {
-                            // Use content as-is (already array or not last user message)
-                            // Ensure null values are converted to empty string
-                            if (content_val.is_null()) {
-                                msg_json["content"] = "";
-                            } else {
-                                msg_json["content"] = content_val;
-                            }
-                        }
-                    } else if (is_last_user_msg && has_images_or_audio) {
-                        // If no content but this is the last user message with images/audio, create content array
-                        json content_array = json::array();
-                        if (request->images_size() > 0) {
-                            for (int j = 0; j < request->images_size(); j++) {
-                                json image_chunk;
-                                image_chunk["type"] = "image_url";
-                                json image_url;
-                                image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                image_chunk["image_url"] = image_url;
-                                content_array.push_back(image_chunk);
-                            }
-                        }
-                        if (request->audios_size() > 0) {
-                            for (int j = 0; j < request->audios_size(); j++) {
-                                json audio_chunk;
-                                audio_chunk["type"] = "input_audio";
-                                json input_audio;
-                                input_audio["data"] = request->audios(j);
-                                input_audio["format"] = "wav"; // default, could be made configurable
-                                audio_chunk["input_audio"] = input_audio;
-                                content_array.push_back(audio_chunk);
-                            }
-                        }
-                        if (request->videos_size() > 0) {
-                            for (int j = 0; j < request->videos_size(); j++) {
-                                json video_chunk;
-                                video_chunk["type"] = "input_video";
-                                json input_video;
-                                input_video["data"] = request->videos(j);
-                                video_chunk["input_video"] = input_video;
-                                content_array.push_back(video_chunk);
-                            }
-                        }
-                        msg_json["content"] = content_array;
-                    } else if (msg.role() == "tool") {
-                        // Tool role messages must have content field set, even if empty
-                        // Jinja templates expect content to be a string, not null or object
-                        SRV_INF("[CONTENT DEBUG] PredictStream: Message %d is tool role, content_empty=%d\n", i, msg.content().empty() ? 1 : 0);
-                        if (msg.content().empty()) {
-                            msg_json["content"] = "";
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): empty content, set to empty string\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): content exists: %s\n",
-                                    i, msg.content().substr(0, std::min<size_t>(200, msg.content().size())).c_str());
-                            // Content exists, parse and ensure it's a string
-                            json content_val;
-                            try {
-                                content_val = json::parse(msg.content());
-                                SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): parsed JSON, type=%s\n",
-                                        i, content_val.is_null() ? "null" :
-                                           content_val.is_object() ? "object" :
-                                           content_val.is_string() ? "string" :
-                                           content_val.is_array() ? "array" : "other");
-                                // Handle null values - Jinja templates expect content to be a string, not null
-                                if (content_val.is_null()) {
-                                    msg_json["content"] = "";
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): null content, converted to empty string\n", i);
-                                } else if (content_val.is_object()) {
-                                    // If content is an object (e.g., from tool call failures/errors), convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): object content, converted to string: %s\n",
-                                            i, content_val.dump().substr(0, std::min<size_t>(200, content_val.dump().size())).c_str());
-                                } else if (content_val.is_string()) {
-                                    msg_json["content"] = content_val.get<std::string>();
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): string content, using as-is\n", i);
-                                } else {
-                                    // For arrays or other types, convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): %s content, converted to string\n",
-                                            i, content_val.is_array() ? "array" : "other type");
-                                }
-                            } catch (const json::parse_error&) {
-                                // Not JSON, treat as plain string
-                                msg_json["content"] = msg.content();
-                                SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (tool): not JSON, using as string\n", i);
-                            }
-                        }
-                    } else {
-                        // Ensure all messages have content set (fallback for any unhandled cases)
-                        // Jinja templates expect content to be present, default to empty string if not set
-                        if (!msg_json.contains("content")) {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d (role=%s): no content field, adding empty string\n",
-                                    i, msg.role().c_str());
-                            msg_json["content"] = "";
-                        }
+                    llama_grpc::ReconstructedMessageInput rin;
+                    rin.role = msg.role();
+                    rin.content = msg.content();
+                    rin.name = msg.name();
+                    rin.tool_call_id = msg.tool_call_id();
+                    rin.reasoning_content = msg.reasoning_content();
+                    rin.tool_calls = msg.tool_calls();
+                    rin.is_last_user_msg = (i == last_user_msg_idx);
+                    if (rin.is_last_user_msg) {
+                        for (int j = 0; j < request->images_size(); j++) rin.images.push_back(request->images(j));
+                        for (int j = 0; j < request->audios_size(); j++) rin.audios.push_back(request->audios(j));
+                        for (int j = 0; j < request->videos_size(); j++) rin.videos.push_back(request->videos(j));
                     }
-
-                    // Add optional fields for OpenAI-compatible message format
-                    if (!msg.name().empty()) {
-                        msg_json["name"] = msg.name();
-                    }
-                    if (!msg.tool_call_id().empty()) {
-                        msg_json["tool_call_id"] = msg.tool_call_id();
-                    }
-                    if (!msg.reasoning_content().empty()) {
-                        msg_json["reasoning_content"] = msg.reasoning_content();
-                    }
-                    if (!msg.tool_calls().empty()) {
-                        // Parse tool_calls JSON string and add to message
-                        try {
-                            json tool_calls = json::parse(msg.tool_calls());
-                            msg_json["tool_calls"] = tool_calls;
-                            SRV_INF("[TOOL CALLS DEBUG] PredictStream: Message %d has tool_calls: %s\n", i, tool_calls.dump().c_str());
-                            // IMPORTANT: If message has tool_calls but content is empty or not set,
-                            // set content to space " " instead of empty string "", because llama.cpp's
-                            // common_chat_msgs_to_json_oaicompat converts empty strings to null (line 312),
-                            // which causes template errors when accessing message.content[:tool_start_length]
-                            if (!msg_json.contains("content") || (msg_json.contains("content") && msg_json["content"].is_string() && msg_json["content"].get<std::string>().empty())) {
-                                SRV_INF("[CONTENT DEBUG] PredictStream: Message %d has tool_calls but empty content, setting to space\n", i);
-                                msg_json["content"] = " ";
-                            }
-                            // Log each tool call with name and arguments
-                            if (tool_calls.is_array()) {
-                                for (size_t tc_idx = 0; tc_idx < tool_calls.size(); tc_idx++) {
-                                    const auto& tc = tool_calls[tc_idx];
-                                    std::string tool_name = "unknown";
-                                    std::string tool_args = "{}";
-                                    if (tc.contains("function")) {
-                                        const auto& func = tc["function"];
-                                        if (func.contains("name")) {
-                                            tool_name = func["name"].get<std::string>();
-                                        }
-                                        if (func.contains("arguments")) {
-                                            tool_args = func["arguments"].is_string() ?
-                                                func["arguments"].get<std::string>() :
-                                                func["arguments"].dump();
-                                        }
-                                    } else if (tc.contains("name")) {
-                                        tool_name = tc["name"].get<std::string>();
-                                        if (tc.contains("arguments")) {
-                                            tool_args = tc["arguments"].is_string() ?
-                                                tc["arguments"].get<std::string>() :
-                                                tc["arguments"].dump();
-                                        }
-                                    }
-                                    SRV_INF("[TOOL CALLS DEBUG] PredictStream: Message %d, tool_call %zu: name=%s, arguments=%s\n",
-                                            i, tc_idx, tool_name.c_str(), tool_args.c_str());
-                                }
-                            }
-                        } catch (const json::parse_error& e) {
-                            SRV_WRN("Failed to parse tool_calls JSON: %s\n", e.what());
-                        }
-                    }
-
-                    // Debug: Log final content state before adding to array
-                    if (msg_json.contains("content")) {
-                        if (msg_json["content"].is_null()) {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d FINAL STATE: content is NULL - THIS WILL CAUSE ERROR!\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: Message %d FINAL STATE: content type=%s, has_value=%d\n",
-                                    i, msg_json["content"].is_string() ? "string" :
-                                       msg_json["content"].is_array() ? "array" :
-                                       msg_json["content"].is_object() ? "object" : "other",
-                                    msg_json["content"].is_null() ? 0 : 1);
-                        }
-                    } else {
-                        SRV_INF("[CONTENT DEBUG] PredictStream: Message %d FINAL STATE: NO CONTENT FIELD - THIS WILL CAUSE ERROR!\n", i);
-                    }
-
-                    messages_json.push_back(msg_json);
+                    messages_json.push_back(llama_grpc::build_reconstructed_message(rin));
                 }
 
                 // Final safety check: Ensure no message has null content (Jinja templates require strings)
@@ -2072,36 +1851,7 @@ public:
                 if (body_json.contains("messages") && body_json["messages"].is_array()) {
                     SRV_INF("[CONTENT DEBUG] PredictStream: Before oaicompat_chat_params_parse - checking %zu messages\n", body_json["messages"].size());
                     for (size_t idx = 0; idx < body_json["messages"].size(); idx++) {
-                        auto& msg = body_json["messages"][idx];
-                        std::string role_str = msg.contains("role") ? msg["role"].get<std::string>() : "unknown";
-                        if (msg.contains("content")) {
-                            if (msg["content"].is_null()) {
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s) has NULL content - FIXING!\n", idx, role_str.c_str());
-                                msg["content"] = ""; // Fix null content
-                            } else if (role_str == "tool" && msg["content"].is_array()) {
-                                // Tool messages must have string content, not array
-                                // oaicompat_chat_params_parse expects tool messages to have string content
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=tool) has array content, converting to string\n", idx);
-                                msg["content"] = msg["content"].dump();
-                            } else if (!msg["content"].is_string() && !msg["content"].is_array()) {
-                                // If content is object or other non-string type, convert to string for templates
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s) content is not string/array, converting\n", idx, role_str.c_str());
-                                if (msg["content"].is_object()) {
-                                    msg["content"] = msg["content"].dump();
-                                } else {
-                                    msg["content"] = "";
-                                }
-                            } else {
-                                SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s): content type=%s\n",
-                                        idx, role_str.c_str(),
-                                        msg["content"].is_string() ? "string" :
-                                        msg["content"].is_array() ? "array" :
-                                        msg["content"].is_object() ? "object" : "other");
-                            }
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] PredictStream: BEFORE TEMPLATE - Message %zu (role=%s) MISSING content field - ADDING!\n", idx, role_str.c_str());
-                            msg["content"] = ""; // Add missing content
-                        }
+                        llama_grpc::normalize_template_message(body_json["messages"][idx]);
                     }
                 }
 
@@ -2433,264 +2183,20 @@ public:
                 SRV_INF("[CONTENT DEBUG] Predict: Processing %d messages\n", request->messages_size());
                 for (int i = 0; i < request->messages_size(); i++) {
                     const auto& msg = request->messages(i);
-                    json msg_json;
-                    msg_json["role"] = msg.role();
-
-                    SRV_INF("[CONTENT DEBUG] Predict: Message %d: role=%s, content_empty=%d, content_length=%zu\n",
-                            i, msg.role().c_str(), msg.content().empty() ? 1 : 0, msg.content().size());
-                    if (!msg.content().empty()) {
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d content (first 200 chars): %s\n",
-                                i, msg.content().substr(0, std::min<size_t>(200, msg.content().size())).c_str());
+                    llama_grpc::ReconstructedMessageInput rin;
+                    rin.role = msg.role();
+                    rin.content = msg.content();
+                    rin.name = msg.name();
+                    rin.tool_call_id = msg.tool_call_id();
+                    rin.reasoning_content = msg.reasoning_content();
+                    rin.tool_calls = msg.tool_calls();
+                    rin.is_last_user_msg = (i == last_user_msg_idx);
+                    if (rin.is_last_user_msg) {
+                        for (int j = 0; j < request->images_size(); j++) rin.images.push_back(request->images(j));
+                        for (int j = 0; j < request->audios_size(); j++) rin.audios.push_back(request->audios(j));
+                        for (int j = 0; j < request->videos_size(); j++) rin.videos.push_back(request->videos(j));
                     }
-
-                    bool is_last_user_msg = (i == last_user_msg_idx);
-                    bool has_images_or_audio = (request->images_size() > 0 || request->audios_size() > 0 || request->videos_size() > 0);
-
-                    // Handle content - can be string, null, or array
-                    // For multimodal content, we'll embed images/audio from separate fields
-                    if (!msg.content().empty()) {
-                        // Try to parse content as JSON to see if it's already an array
-                        json content_val;
-                        try {
-                            content_val = json::parse(msg.content());
-                            // Handle null values - convert to empty string to avoid template errors
-                            if (content_val.is_null()) {
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d parsed JSON is null, converting to empty string\n", i);
-                                content_val = "";
-                            }
-                        } catch (const json::parse_error&) {
-                            // Not JSON, treat as plain string
-                            content_val = msg.content();
-                        }
-
-                        // If content is an object (e.g., from tool call failures), convert to string
-                        if (content_val.is_object()) {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d content is object, converting to string\n", i);
-                            content_val = content_val.dump();
-                        }
-
-                        // If content is a string and this is the last user message with images/audio, combine them
-                        if (content_val.is_string() && is_last_user_msg && has_images_or_audio) {
-                            json content_array = json::array();
-                            // Add text first
-                            content_array.push_back({{"type", "text"}, {"text", content_val.get<std::string>()}});
-                            // Add images
-                            if (request->images_size() > 0) {
-                                for (int j = 0; j < request->images_size(); j++) {
-                                    json image_chunk;
-                                    image_chunk["type"] = "image_url";
-                                    json image_url;
-                                    image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                    image_chunk["image_url"] = image_url;
-                                    content_array.push_back(image_chunk);
-                                }
-                            }
-                            // Add audios
-                            if (request->audios_size() > 0) {
-                                for (int j = 0; j < request->audios_size(); j++) {
-                                    json audio_chunk;
-                                    audio_chunk["type"] = "input_audio";
-                                    json input_audio;
-                                    input_audio["data"] = request->audios(j);
-                                    input_audio["format"] = "wav"; // default, could be made configurable
-                                    audio_chunk["input_audio"] = input_audio;
-                                    content_array.push_back(audio_chunk);
-                                }
-                            }
-                            if (request->videos_size() > 0) {
-                                for (int j = 0; j < request->videos_size(); j++) {
-                                    json video_chunk;
-                                    video_chunk["type"] = "input_video";
-                                    json input_video;
-                                    input_video["data"] = request->videos(j);
-                                    video_chunk["input_video"] = input_video;
-                                    content_array.push_back(video_chunk);
-                                }
-                            }
-                            msg_json["content"] = content_array;
-                        } else {
-                            // Use content as-is (already array or not last user message)
-                            // Ensure null values are converted to empty string
-                            if (content_val.is_null()) {
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d content_val was null, setting to empty string\n", i);
-                                msg_json["content"] = "";
-                            } else {
-                                msg_json["content"] = content_val;
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d content set, type=%s\n",
-                                        i, content_val.is_string() ? "string" :
-                                           content_val.is_array() ? "array" :
-                                           content_val.is_object() ? "object" : "other");
-                            }
-                        }
-                    } else if (is_last_user_msg && has_images_or_audio) {
-                        // If no content but this is the last user message with images/audio, create content array
-                        json content_array = json::array();
-                        if (request->images_size() > 0) {
-                            for (int j = 0; j < request->images_size(); j++) {
-                                json image_chunk;
-                                image_chunk["type"] = "image_url";
-                                json image_url;
-                                image_url["url"] = "data:image/jpeg;base64," + request->images(j);
-                                image_chunk["image_url"] = image_url;
-                                content_array.push_back(image_chunk);
-                            }
-                        }
-                        if (request->audios_size() > 0) {
-                            for (int j = 0; j < request->audios_size(); j++) {
-                                json audio_chunk;
-                                audio_chunk["type"] = "input_audio";
-                                json input_audio;
-                                input_audio["data"] = request->audios(j);
-                                input_audio["format"] = "wav"; // default, could be made configurable
-                                audio_chunk["input_audio"] = input_audio;
-                                content_array.push_back(audio_chunk);
-                            }
-                        }
-                        if (request->videos_size() > 0) {
-                            for (int j = 0; j < request->videos_size(); j++) {
-                                json video_chunk;
-                                video_chunk["type"] = "input_video";
-                                json input_video;
-                                input_video["data"] = request->videos(j);
-                                video_chunk["input_video"] = input_video;
-                                content_array.push_back(video_chunk);
-                            }
-                        }
-                        msg_json["content"] = content_array;
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d created content array with media\n", i);
-                    } else if (!msg.tool_calls().empty()) {
-                        // Tool call messages may have null content, but templates expect string
-                        // IMPORTANT: Set to space " " instead of empty string "", because llama.cpp's
-                        // common_chat_msgs_to_json_oaicompat converts empty strings to null (line 312),
-                        // which causes template errors when accessing message.content[:tool_start_length]
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d has tool_calls, setting content to space (not empty string)\n", i);
-                        msg_json["content"] = " ";
-                    } else if (msg.role() == "tool") {
-                        // Tool role messages must have content field set, even if empty
-                        // Jinja templates expect content to be a string, not null or object
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d is tool role, content_empty=%d\n", i, msg.content().empty() ? 1 : 0);
-                        if (msg.content().empty()) {
-                            msg_json["content"] = "";
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): empty content, set to empty string\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): content exists: %s\n",
-                                    i, msg.content().substr(0, std::min<size_t>(200, msg.content().size())).c_str());
-                            // Content exists, parse and ensure it's a string
-                            json content_val;
-                            try {
-                                content_val = json::parse(msg.content());
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): parsed JSON, type=%s\n",
-                                        i, content_val.is_null() ? "null" :
-                                           content_val.is_object() ? "object" :
-                                           content_val.is_string() ? "string" :
-                                           content_val.is_array() ? "array" : "other");
-                                // Handle null values - Jinja templates expect content to be a string, not null
-                                if (content_val.is_null()) {
-                                    msg_json["content"] = "";
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): null content, converted to empty string\n", i);
-                                } else if (content_val.is_object()) {
-                                    // If content is an object (e.g., from tool call failures/errors), convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): object content, converted to string: %s\n",
-                                            i, content_val.dump().substr(0, std::min<size_t>(200, content_val.dump().size())).c_str());
-                                } else if (content_val.is_string()) {
-                                    msg_json["content"] = content_val.get<std::string>();
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): string content, using as-is\n", i);
-                                } else {
-                                    // For arrays or other types, convert to string
-                                    msg_json["content"] = content_val.dump();
-                                    SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): %s content, converted to string\n",
-                                            i, content_val.is_array() ? "array" : "other type");
-                                }
-                            } catch (const json::parse_error&) {
-                                // Not JSON, treat as plain string
-                                msg_json["content"] = msg.content();
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d (tool): not JSON, using as string\n", i);
-                            }
-                        }
-                    } else {
-                        // Ensure all messages have content set (fallback for any unhandled cases)
-                        // Jinja templates expect content to be present, default to empty string if not set
-                        if (!msg_json.contains("content")) {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d (role=%s): no content field, adding empty string\n",
-                                    i, msg.role().c_str());
-                            msg_json["content"] = "";
-                        }
-                    }
-
-                    // Add optional fields for OpenAI-compatible message format
-                    if (!msg.name().empty()) {
-                        msg_json["name"] = msg.name();
-                    }
-                    if (!msg.tool_call_id().empty()) {
-                        msg_json["tool_call_id"] = msg.tool_call_id();
-                    }
-                    if (!msg.reasoning_content().empty()) {
-                        msg_json["reasoning_content"] = msg.reasoning_content();
-                    }
-                    if (!msg.tool_calls().empty()) {
-                        // Parse tool_calls JSON string and add to message
-                        try {
-                            json tool_calls = json::parse(msg.tool_calls());
-                            msg_json["tool_calls"] = tool_calls;
-                            SRV_INF("[TOOL CALLS DEBUG] Predict: Message %d has tool_calls: %s\n", i, tool_calls.dump().c_str());
-                            // IMPORTANT: If message has tool_calls but content is empty or not set,
-                            // set content to space " " instead of empty string "", because llama.cpp's
-                            // common_chat_msgs_to_json_oaicompat converts empty strings to null (line 312),
-                            // which causes template errors when accessing message.content[:tool_start_length]
-                            if (!msg_json.contains("content") || (msg_json.contains("content") && msg_json["content"].is_string() && msg_json["content"].get<std::string>().empty())) {
-                                SRV_INF("[CONTENT DEBUG] Predict: Message %d has tool_calls but empty content, setting to space\n", i);
-                                msg_json["content"] = " ";
-                            }
-                            // Log each tool call with name and arguments
-                            if (tool_calls.is_array()) {
-                                for (size_t tc_idx = 0; tc_idx < tool_calls.size(); tc_idx++) {
-                                    const auto& tc = tool_calls[tc_idx];
-                                    std::string tool_name = "unknown";
-                                    std::string tool_args = "{}";
-                                    if (tc.contains("function")) {
-                                        const auto& func = tc["function"];
-                                        if (func.contains("name")) {
-                                            tool_name = func["name"].get<std::string>();
-                                        }
-                                        if (func.contains("arguments")) {
-                                            tool_args = func["arguments"].is_string() ?
-                                                func["arguments"].get<std::string>() :
-                                                func["arguments"].dump();
-                                        }
-                                    } else if (tc.contains("name")) {
-                                        tool_name = tc["name"].get<std::string>();
-                                        if (tc.contains("arguments")) {
-                                            tool_args = tc["arguments"].is_string() ?
-                                                tc["arguments"].get<std::string>() :
-                                                tc["arguments"].dump();
-                                        }
-                                    }
-                                    SRV_INF("[TOOL CALLS DEBUG] Predict: Message %d, tool_call %zu: name=%s, arguments=%s\n",
-                                            i, tc_idx, tool_name.c_str(), tool_args.c_str());
-                                }
-                            }
-                        } catch (const json::parse_error& e) {
-                            SRV_WRN("Failed to parse tool_calls JSON: %s\n", e.what());
-                        }
-                    }
-
-                    // Debug: Log final content state before adding to array
-                    if (msg_json.contains("content")) {
-                        if (msg_json["content"].is_null()) {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d FINAL STATE: content is NULL - THIS WILL CAUSE ERROR!\n", i);
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] Predict: Message %d FINAL STATE: content type=%s, has_value=%d\n",
-                                    i, msg_json["content"].is_string() ? "string" :
-                                       msg_json["content"].is_array() ? "array" :
-                                       msg_json["content"].is_object() ? "object" : "other",
-                                    msg_json["content"].is_null() ? 0 : 1);
-                        }
-                    } else {
-                        SRV_INF("[CONTENT DEBUG] Predict: Message %d FINAL STATE: NO CONTENT FIELD - THIS WILL CAUSE ERROR!\n", i);
-                    }
-
-                    messages_json.push_back(msg_json);
+                    messages_json.push_back(llama_grpc::build_reconstructed_message(rin));
                 }
 
                 // Final safety check: Ensure no message has null content (Jinja templates require strings)
@@ -2911,36 +2417,7 @@ public:
                 if (body_json.contains("messages") && body_json["messages"].is_array()) {
                     SRV_INF("[CONTENT DEBUG] Predict: Before oaicompat_chat_params_parse - checking %zu messages\n", body_json["messages"].size());
                     for (size_t idx = 0; idx < body_json["messages"].size(); idx++) {
-                        auto& msg = body_json["messages"][idx];
-                        std::string role_str = msg.contains("role") ? msg["role"].get<std::string>() : "unknown";
-                        if (msg.contains("content")) {
-                            if (msg["content"].is_null()) {
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s) has NULL content - FIXING!\n", idx, role_str.c_str());
-                                msg["content"] = ""; // Fix null content
-                            } else if (role_str == "tool" && msg["content"].is_array()) {
-                                // Tool messages must have string content, not array
-                                // oaicompat_chat_params_parse expects tool messages to have string content
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=tool) has array content, converting to string\n", idx);
-                                msg["content"] = msg["content"].dump();
-                            } else if (!msg["content"].is_string() && !msg["content"].is_array()) {
-                                // If content is object or other non-string type, convert to string for templates
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s) content is not string/array, converting\n", idx, role_str.c_str());
-                                if (msg["content"].is_object()) {
-                                    msg["content"] = msg["content"].dump();
-                                } else {
-                                    msg["content"] = "";
-                                }
-                            } else {
-                                SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s): content type=%s\n",
-                                        idx, role_str.c_str(),
-                                        msg["content"].is_string() ? "string" :
-                                        msg["content"].is_array() ? "array" :
-                                        msg["content"].is_object() ? "object" : "other");
-                            }
-                        } else {
-                            SRV_INF("[CONTENT DEBUG] Predict: BEFORE TEMPLATE - Message %zu (role=%s) MISSING content field - ADDING!\n", idx, role_str.c_str());
-                            msg["content"] = ""; // Add missing content
-                        }
+                        llama_grpc::normalize_template_message(body_json["messages"][idx]);
                     }
                 }
 

backend/cpp/llama-cpp/message_content.h

@@ -0,0 +1,192 @@
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include <nlohmann/json.hpp>
+
+namespace llama_grpc {
+
+// Normalizes a proto message's content string into the JSON value used when
+// reconstructing OpenAI-format messages for the tokenizer (jinja) template.
+//
+// Shared by the streaming (PredictStream) and non-streaming (Predict) message
+// reconstruction paths so the two cannot drift.
+//
+// LocalAI's Go layer (schema.Messages.ToProto) always sends content as a plain
+// text string; multimodal media travels in separate proto fields, never inside
+// content. So user/system/developer content is *only ever* opaque text and must
+// NOT be JSON-sniffed: a prompt that merely looks like JSON (e.g. an ingredient
+// list ["1/4 cup sugar", ...]) would otherwise be reinterpreted as structured
+// content parts and rejected by oaicompat_chat_params_parse with
+// "unsupported content[].type" (https://github.com/mudler/LocalAI/issues/10524).
+// (developer is OpenAI's modern system alias - same "human-authored text" nature.)
+//
+// For assistant/tool messages we still collapse a literal JSON null/object
+// (tool-call bookkeeping) to a string, but we never turn a plain string into an
+// array/scalar. The array defense is therefore role-independent (arrays/scalars
+// fall through for every role); the role gate only governs the null/object case.
+inline nlohmann::ordered_json normalize_message_content(const std::string& role,
+                                                        const std::string& content) {
+    nlohmann::ordered_json content_val = content;
+    if (role != "user" && role != "system" && role != "developer") {
+        try {
+            nlohmann::ordered_json parsed = nlohmann::ordered_json::parse(content);
+            if (parsed.is_null()) {
+                content_val = "";
+            } else if (parsed.is_object()) {
+                content_val = parsed.dump();
+            }
+            // arrays / scalars: keep the original plain-text string as-is
+        } catch (const nlohmann::ordered_json::parse_error&) {
+            // Not JSON, already the plain string
+        }
+    }
+    return content_val;
+}
+
+// Final safety pass applied to each reconstructed OpenAI message right before it
+// is handed to oaicompat_chat_params_parse (jinja templating). Jinja templates
+// assume content is a string: a literal null breaks slicing such as
+// message.content[:N] (#7324), and a tool message with array content is rejected
+// (#7528). A multimodal user message legitimately carries a typed-part array
+// ({type:text}, {type:image_url}, ...), which must be left intact. Shared by the
+// streaming and non-streaming paths so this invariant cannot drift between them.
+inline void normalize_template_message(nlohmann::ordered_json& msg) {
+    if (!msg.contains("content")) {
+        msg["content"] = ""; // templates expect the field to exist
+        return;
+    }
+    nlohmann::ordered_json& content = msg["content"];
+    const std::string role = (msg.contains("role") && msg["role"].is_string())
+                                 ? msg["role"].get<std::string>()
+                                 : std::string();
+    if (content.is_null()) {
+        content = ""; // #7324: null would crash content[:N] slicing
+    } else if (role == "tool" && content.is_array()) {
+        content = content.dump(); // #7528: tool messages must have string content
+    } else if (!content.is_string() && !content.is_array()) {
+        if (content.is_object()) {
+            content = content.dump(); // tool-call bookkeeping object -> string
+        } else {
+            content = ""; // other scalar (number/bool) -> empty
+        }
+    }
+    // string, or a non-tool (multimodal) typed-part array: leave untouched
+}
+
+// One proto message's data, flattened to plain types so the reconstruction logic
+// can be shared and unit-tested without protobuf. The streaming and non-streaming
+// predict paths both populate this from proto::Message + the request's media.
+struct ReconstructedMessageInput {
+    std::string role;
+    std::string content;            // proto.Message.content (always a plain string)
+    std::string name;
+    std::string tool_call_id;
+    std::string reasoning_content;
+    std::string tool_calls;         // tool_calls as a JSON string, or empty
+    bool is_last_user_msg = false;  // attach request media to this message
+    std::vector<std::string> images; // base64 (jpeg)
+    std::vector<std::string> audios; // base64 (wav)
+    std::vector<std::string> videos; // base64
+};
+
+// Appends the request's media as OpenAI typed content parts. Imperative (not
+// brace-init) to avoid nlohmann's object-vs-array initializer-list ambiguity.
+inline void append_media_parts(nlohmann::ordered_json& content_array,
+                               const std::vector<std::string>& images,
+                               const std::vector<std::string>& audios,
+                               const std::vector<std::string>& videos) {
+    for (const auto& img : images) {
+        nlohmann::ordered_json image_chunk;
+        image_chunk["type"] = "image_url";
+        nlohmann::ordered_json image_url;
+        image_url["url"] = "data:image/jpeg;base64," + img;
+        image_chunk["image_url"] = image_url;
+        content_array.push_back(image_chunk);
+    }
+    for (const auto& aud : audios) {
+        nlohmann::ordered_json audio_chunk;
+        audio_chunk["type"] = "input_audio";
+        nlohmann::ordered_json input_audio;
+        input_audio["data"] = aud;
+        input_audio["format"] = "wav"; // default; could be made configurable
+        audio_chunk["input_audio"] = input_audio;
+        content_array.push_back(audio_chunk);
+    }
+    for (const auto& vid : videos) {
+        nlohmann::ordered_json video_chunk;
+        video_chunk["type"] = "input_video";
+        nlohmann::ordered_json input_video;
+        input_video["data"] = vid;
+        video_chunk["input_video"] = input_video;
+        content_array.push_back(video_chunk);
+    }
+}
+
+// Reconstructs a single OpenAI-format message (the object fed to
+// oaicompat_chat_params_parse) from a proto message. Shared by PredictStream and
+// Predict so the content/multimodal/tool_calls handling cannot drift between the
+// two stream modes (it previously lived as two ~150-line copies with a redundant
+// Predict-only tool_calls->" " branch). Guarantees content is always a string or
+// a typed-part array, never null/missing.
+inline nlohmann::ordered_json build_reconstructed_message(const ReconstructedMessageInput& in) {
+    nlohmann::ordered_json msg_json;
+    msg_json["role"] = in.role;
+    const bool has_media = !in.images.empty() || !in.audios.empty() || !in.videos.empty();
+
+    if (!in.content.empty()) {
+        nlohmann::ordered_json content_val = normalize_message_content(in.role, in.content);
+        if (content_val.is_string() && in.is_last_user_msg && has_media) {
+            // Last user message + media: build a typed-part array (text first).
+            nlohmann::ordered_json content_array = nlohmann::ordered_json::array();
+            nlohmann::ordered_json text_part;
+            text_part["type"] = "text";
+            text_part["text"] = content_val.get<std::string>();
+            content_array.push_back(text_part);
+            append_media_parts(content_array, in.images, in.audios, in.videos);
+            msg_json["content"] = content_array;
+        } else if (content_val.is_null()) {
+            msg_json["content"] = "";
+        } else {
+            msg_json["content"] = content_val;
+        }
+    } else if (in.is_last_user_msg && has_media) {
+        // No text but media on the last user message: media-only typed array.
+        nlohmann::ordered_json content_array = nlohmann::ordered_json::array();
+        append_media_parts(content_array, in.images, in.audios, in.videos);
+        msg_json["content"] = content_array;
+    } else {
+        // Empty content (any role, incl. tool/assistant): templates need a string.
+        msg_json["content"] = "";
+    }
+
+    if (!in.name.empty()) {
+        msg_json["name"] = in.name;
+    }
+    if (!in.tool_call_id.empty()) {
+        msg_json["tool_call_id"] = in.tool_call_id;
+    }
+    if (!in.reasoning_content.empty()) {
+        msg_json["reasoning_content"] = in.reasoning_content;
+    }
+    if (!in.tool_calls.empty()) {
+        try {
+            nlohmann::ordered_json tool_calls = nlohmann::ordered_json::parse(in.tool_calls);
+            msg_json["tool_calls"] = tool_calls;
+            // tool_calls + empty/blank content: use " " not "", because llama.cpp's
+            // common_chat_msgs_to_json_oaicompat turns "" into null, which breaks
+            // templates that slice message.content[:tool_start_length] (#7324).
+            if (!msg_json.contains("content") ||
+                (msg_json["content"].is_string() && msg_json["content"].get<std::string>().empty())) {
+                msg_json["content"] = " ";
+            }
+        } catch (const nlohmann::ordered_json::parse_error&) {
+            // Malformed tool_calls JSON: leave content as-is (prior behavior).
+        }
+    }
+
+    return msg_json;
+}
+
+}  // namespace llama_grpc

backend/cpp/llama-cpp/message_content_test.cpp

@@ -0,0 +1,234 @@
+// Unit tests for the shared message-reconstruction helpers (message_content.h).
+//
+// Build & run standalone (nlohmann/json single header on the include path):
+//   g++ -std=c++17 -I<dir-with-nlohmann> message_content_test.cpp -o t && ./t
+// or via CMake: -DLLAMA_GRPC_BUILD_TESTS=ON then ctest.
+//
+// Regression coverage for:
+//   #10524 - a user/system prompt that is itself a JSON-array string must stay
+//            plain text, never be reinterpreted as OpenAI structured parts.
+//   #7324  - assistant/tool null content -> "" (templates slice content[:N]);
+//            assistant+tool_calls+empty content -> " " (not "", which becomes null).
+//   #7528  - tool message array content must reach the template as a string.
+//   multimodal - last user message text + media -> typed-part array, media kept.
+
+#include <cassert>
+#include <iostream>
+#include <string>
+
+#include "message_content.h"
+
+using nlohmann::ordered_json;
+using llama_grpc::normalize_message_content;
+using llama_grpc::normalize_template_message;
+using llama_grpc::build_reconstructed_message;
+using llama_grpc::ReconstructedMessageInput;
+
+static int failures = 0;
+
+static void check(bool ok, const std::string& name, const std::string& detail = "") {
+    if (!ok) {
+        std::cerr << "FAIL " << name << (detail.empty() ? "" : ": " + detail) << "\n";
+        failures++;
+    }
+}
+
+// ---- normalize_message_content -------------------------------------------
+
+static void expect_norm_string(const char* name, const std::string& role,
+                               const std::string& content, const std::string& want) {
+    auto got = normalize_message_content(role, content);
+    if (!got.is_string()) {
+        check(false, name, "expected a JSON string, got " +
+                               std::string(got.is_array() ? "array" : got.is_object() ? "object" : "other") +
+                               " (" + got.dump() + ")");
+        return;
+    }
+    check(got.get<std::string>() == want, name, "expected \"" + want + "\", got \"" + got.get<std::string>() + "\"");
+}
+
+static void test_normalize() {
+    const std::string ingredients = R"(["1/4 cup brown sugar, packed","1 pound ground beef"])";
+
+    // #10524 - JSON-array text must stay a string. Role-INDEPENDENT array defense.
+    for (const char* role : {"user", "system", "developer", "function", "assistant", "tool"}) {
+        expect_norm_string((std::string("json_array_stays_text:") + role).c_str(), role, ingredients, ingredients);
+    }
+
+    // #10524 - user/system/developer JSON-object text stays verbatim (NOT re-dumped).
+    expect_norm_string("user_json_object_verbatim", "user", R"({"a":1})", R"({"a":1})");
+    expect_norm_string("system_json_object_verbatim", "system", R"({"a":1})", R"({"a":1})");
+    expect_norm_string("developer_json_object_verbatim", "developer", R"({"a":1})", R"({"a":1})");
+
+    // Plain text unchanged for all roles.
+    expect_norm_string("user_plain_text", "user", "hello world", "hello world");
+    expect_norm_string("assistant_non_json_text_kept", "assistant", "hi [unclosed", "hi [unclosed");
+
+    // #7324 boundary - user/system/developer literal "null" preserved (never parsed).
+    expect_norm_string("user_literal_null_stays", "user", "null", "null");
+    expect_norm_string("system_literal_null_stays", "system", "null", "null");
+    expect_norm_string("developer_literal_null_stays", "developer", "null", "null");
+
+    // #7324 - assistant/tool literal null collapses to empty string.
+    expect_norm_string("assistant_null_to_empty", "assistant", "null", "");
+    expect_norm_string("tool_null_to_empty", "tool", "null", "");
+
+    // #7324/#7528 - assistant/tool object bookkeeping stringified (stays a string).
+    check(normalize_message_content("assistant", R"({"tool":"x"})").is_string(), "assistant_object_stringified");
+    check(normalize_message_content("tool", R"({"error":"boom"})").is_string(), "tool_object_stringified");
+
+    // #10524-family - a bare scalar that parses as a JSON number stays the string.
+    expect_norm_string("assistant_scalar_number_stays_string", "assistant", "42", "42");
+
+    // baseline - empty content stays empty.
+    expect_norm_string("user_empty_stays_empty", "user", "", "");
+}
+
+// ---- normalize_template_message (BEFORE TEMPLATE sanitizer) ---------------
+
+static void test_template_sanitizer() {
+    // #7528 - a tool message with an ACTUAL array becomes a string.
+    {
+        ordered_json msg = {{"role", "tool"}, {"content", ordered_json::array({{{"type", "text"}, {"text", "r"}}})}};
+        normalize_template_message(msg);
+        check(msg["content"].is_string(), "before_template_tool_array_to_string", "got " + msg["content"].dump());
+    }
+    // #7324 - null content -> "" for any role.
+    {
+        ordered_json msg = {{"role", "assistant"}, {"content", nullptr}};
+        normalize_template_message(msg);
+        check(msg["content"].is_string() && msg["content"] == "", "before_template_null_to_empty");
+    }
+    // object content -> dumped string (would otherwise throw at the template).
+    {
+        ordered_json msg = {{"role", "assistant"}, {"content", {{"x", 1}}}};
+        normalize_template_message(msg);
+        check(msg["content"].is_string(), "before_template_object_to_string", "got " + msg["content"].dump());
+    }
+    // missing content field -> "".
+    {
+        ordered_json msg = {{"role", "user"}};
+        normalize_template_message(msg);
+        check(msg.contains("content") && msg["content"] == "", "before_template_missing_to_empty");
+    }
+    // multimodal: a well-typed user array must be left UNTOUCHED (role!=tool).
+    {
+        ordered_json parts = ordered_json::array();
+        parts.push_back({{"type", "text"}, {"text", "x"}});
+        ordered_json img; img["type"] = "image_url"; img["image_url"] = {{"url", "data:..."}};
+        parts.push_back(img);
+        ordered_json msg = {{"role", "user"}, {"content", parts}};
+        normalize_template_message(msg);
+        check(msg["content"].is_array() && msg["content"].size() == 2, "before_template_user_typed_array_preserved",
+              "got " + msg["content"].dump());
+    }
+    // a plain string is left untouched.
+    {
+        ordered_json msg = {{"role", "user"}, {"content", "hello"}};
+        normalize_template_message(msg);
+        check(msg["content"] == "hello", "before_template_string_untouched");
+    }
+}
+
+// ---- build_reconstructed_message ----------------------------------------
+
+static void test_reconstruction() {
+    const std::string ingredients = R"(["1/4 cup brown sugar","1 pound ground beef"])";
+
+    // #10524 end-state - user JSON-array text, no media -> string content.
+    {
+        ReconstructedMessageInput in;
+        in.role = "user"; in.content = ingredients;
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == ingredients, "recon_user_json_array_string",
+              "got " + m["content"].dump());
+    }
+    // multimodal - user text + one image on last user msg -> typed array, image kept.
+    {
+        ReconstructedMessageInput in;
+        in.role = "user"; in.content = ingredients; in.is_last_user_msg = true;
+        in.images.push_back("BASE64IMG");
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_array() && m["content"].size() == 2, "recon_multimodal_text_plus_image",
+              "got " + m["content"].dump());
+        check(m["content"][0]["type"] == "text" && m["content"][0]["text"] == ingredients, "recon_multimodal_text_first");
+        check(m["content"][1]["type"] == "image_url", "recon_multimodal_image_kept");
+    }
+    // multimodal media-only - empty text + image on last user msg.
+    {
+        ReconstructedMessageInput in;
+        in.role = "user"; in.content = ""; in.is_last_user_msg = true;
+        in.images.push_back("BASE64IMG");
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_array() && m["content"].size() == 1 && m["content"][0]["type"] == "image_url",
+              "recon_media_only", "got " + m["content"].dump());
+    }
+    // #7528 - tool array-string content stays a string.
+    {
+        ReconstructedMessageInput in;
+        in.role = "tool"; in.content = R"(["a","b"])"; in.tool_call_id = "call_1";
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == R"(["a","b"])", "recon_tool_array_string",
+              "got " + m["content"].dump());
+        check(m["tool_call_id"] == "call_1", "recon_tool_call_id_set");
+    }
+    // tool empty content -> "".
+    {
+        ReconstructedMessageInput in;
+        in.role = "tool"; in.content = "";
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == "", "recon_tool_empty_to_string");
+    }
+    // #7324 - assistant + tool_calls + empty content -> " " (single space, not "").
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "";
+        in.tool_calls = R"([{"id":"c1","type":"function","function":{"name":"f","arguments":"{}"}}])";
+        auto m = build_reconstructed_message(in);
+        check(m["content"].is_string() && m["content"] == " ", "recon_toolcalls_empty_content_space",
+              "got " + m["content"].dump());
+        check(m["tool_calls"].is_array() && m["tool_calls"].size() == 1, "recon_toolcalls_parsed");
+    }
+    // assistant + tool_calls + real content keeps the content.
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "I'll call f";
+        in.tool_calls = R"([{"id":"c1","type":"function","function":{"name":"f","arguments":"{}"}}])";
+        auto m = build_reconstructed_message(in);
+        check(m["content"] == "I'll call f", "recon_toolcalls_with_content_kept");
+    }
+    // assistant null content -> "".
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "null";
+        auto m = build_reconstructed_message(in);
+        check(m["content"] == "", "recon_assistant_null_to_empty");
+    }
+    // malformed tool_calls JSON must not throw; content preserved.
+    {
+        ReconstructedMessageInput in;
+        in.role = "assistant"; in.content = "hi"; in.tool_calls = "{not json";
+        auto m = build_reconstructed_message(in);
+        check(m["content"] == "hi" && !m.contains("tool_calls"), "recon_malformed_toolcalls_safe");
+    }
+    // optional fields: name + reasoning carried through.
+    {
+        ReconstructedMessageInput in;
+        in.role = "tool"; in.content = "result"; in.name = "get_weather"; in.reasoning_content = "thinking";
+        auto m = build_reconstructed_message(in);
+        check(m["name"] == "get_weather" && m["reasoning_content"] == "thinking", "recon_optional_fields");
+    }
+}
+
+int main() {
+    test_normalize();
+    test_template_sanitizer();
+    test_reconstruction();
+
+    if (failures == 0) {
+        std::cout << "OK: all message_content tests passed\n";
+        return 0;
+    }
+    std::cerr << failures << " test(s) failed\n";
+    return 1;
+}

backend/cpp/llama-cpp/prepare.sh

@@ -18,6 +18,10 @@ done
 
 cp -r CMakeLists.txt llama.cpp/tools/grpc-server/
 cp -r grpc-server.cpp llama.cpp/tools/grpc-server/
+# Shared message-reconstruction helpers (included by grpc-server.cpp) and their
+# 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/
 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/run-unit-tests.sh

@@ -0,0 +1,71 @@
+#!/bin/bash
+#
+# Discovers and runs every standalone C++ unit test under backend/cpp/.
+#
+# A "standalone" unit test is a *_test.cpp that depends only on the C++ standard
+# library and nlohmann/json (single header) - i.e. it exercises pure helpers and
+# does not need the full llama.cpp + gRPC backend build. Tests that DO need the
+# backend build use the CMake/ctest path (e.g. -DLLAMA_GRPC_BUILD_TESTS=ON)
+# instead and are skipped here.
+#
+# This keeps CI generic: adding a new pure-C++ unit test file named *_test.cpp in
+# an active backend source dir is picked up automatically, with no CI edits.
+#
+# Env:
+#   NLOHMANN_INCLUDE  include dir that contains nlohmann/json.hpp. If unset, the
+#                     nlohmann/json single header is fetched to a temp dir.
+#   CXX               compiler (default: g++).
+#   JSON_VERSION      nlohmann/json tag to fetch when NLOHMANN_INCLUDE is unset
+#                     (default: v3.11.3).
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "$0")" && pwd)"
+CXX="${CXX:-g++}"
+JSON_VERSION="${JSON_VERSION:-v3.11.3}"
+
+JSON_INC="${NLOHMANN_INCLUDE:-}"
+if [ -z "$JSON_INC" ]; then
+    JSON_INC="$(mktemp -d)"
+    mkdir -p "$JSON_INC/nlohmann"
+    echo "Fetching nlohmann/json ${JSON_VERSION} single header..."
+    if ! curl -L -sf \
+        "https://raw.githubusercontent.com/nlohmann/json/${JSON_VERSION}/single_include/nlohmann/json.hpp" \
+        -o "$JSON_INC/nlohmann/json.hpp"; then
+        echo "ERROR: failed to fetch nlohmann/json header" >&2
+        exit 1
+    fi
+fi
+
+# Active source dirs only - exclude per-variant build copies, dev snapshots and
+# the vendored upstream llama.cpp tree.
+mapfile -t tests < <(find "$ROOT" -name '*_test.cpp' \
+    -not -path '*/llama.cpp/*' \
+    -not -path '*-build/*' \
+    -not -path '*-dev/*' \
+    -not -path '*fallback*' | sort)
+
+if [ "${#tests[@]}" -eq 0 ]; then
+    echo "No standalone C++ unit tests found under $ROOT"
+    exit 0
+fi
+
+fail=0
+for test_src in "${tests[@]}"; do
+    name="$(basename "$test_src" .cpp)"
+    bin="$(mktemp -d)/$name"
+    echo "==> $test_src"
+    if ! "$CXX" -std=c++17 -Wall -Wextra \
+        -I"$JSON_INC" -I"$(dirname "$test_src")" \
+        "$test_src" -o "$bin"; then
+        echo "COMPILE FAILED: $test_src" >&2
+        fail=1
+        continue
+    fi
+    if ! "$bin"; then
+        echo "TEST FAILED: $test_src" >&2
+        fail=1
+    fi
+done
+
+echo "Ran ${#tests[@]} standalone C++ unit test file(s)"
+exit "$fail"

backend/go/opus/Makefile

@@ -1,13 +1,30 @@
 GOCMD?=go
 GO_TAGS?=
 
+# The opus shim is a small C wrapper around libopus' variadic
+# opus_encoder_ctl (see csrc/opus_shim.c). It is built as a shared library
+# and dlopen'd at runtime by the Go backend (codec.go). The extension is
+# OS-specific: Linux uses .so, macOS uses .dylib. OS is exported by the root
+# Makefile (`export OS := $(shell uname -s)`).
+SHIM_EXT=so
+
 OPUS_CFLAGS := $(shell pkg-config --cflags opus)
 OPUS_LIBS := $(shell pkg-config --libs opus)
+SHIM_LDFLAGS := $(OPUS_LIBS)
 
-libopusshim.so: csrc/opus_shim.c
-	$(CC) -shared -fPIC -o $@ $< $(OPUS_CFLAGS) $(OPUS_LIBS)
+ifeq ($(OS),Darwin)
+	SHIM_EXT=dylib
+	# Resolve libopus symbols lazily from the already globally-loaded
+	# libopus (codec.go dlopens it RTLD_GLOBAL before the shim) rather than
+	# recording an absolute Homebrew path in the dylib. This keeps the
+	# packaged shim relocatable on machines that have no Homebrew.
+	SHIM_LDFLAGS := -undefined dynamic_lookup
+endif
 
-opus: libopusshim.so
+libopusshim.$(SHIM_EXT): csrc/opus_shim.c
+	$(CC) -shared -fPIC -o $@ $< $(OPUS_CFLAGS) $(SHIM_LDFLAGS)
+
+opus: libopusshim.$(SHIM_EXT)
 	$(GOCMD) build -tags "$(GO_TAGS)" -o opus ./
 
 package: opus
@@ -16,4 +33,7 @@ package: opus
 build: package
 
 clean:
-	rm -f opus libopusshim.so
+	rm -f opus libopusshim.$(SHIM_EXT)
+	rm -rf package
+
+.PHONY: build package clean

backend/go/opus/package.sh

@@ -8,13 +8,23 @@ mkdir -p $CURDIR/package/lib
 cp -avf $CURDIR/opus $CURDIR/package/
 cp -avf $CURDIR/run.sh $CURDIR/package/
 
-# Copy the opus shim library
-cp -avf $CURDIR/libopusshim.so $CURDIR/package/lib/
+# The shim extension is OS-specific (.so on Linux, .dylib on macOS).
+SHIM_EXT=so
+if [ "$(uname)" = "Darwin" ]; then
+    SHIM_EXT=dylib
+fi
 
-# Copy system libopus
+# Copy the opus shim library
+cp -avf $CURDIR/libopusshim.$SHIM_EXT $CURDIR/package/lib/
+
+# Copy system libopus so the backend is self-contained: the runtime base
+# image has neither libopus-dev (Linux) nor Homebrew (macOS), so codec.go's
+# dlopen would otherwise fail. Both name patterns are attempted; only the
+# host's matching one exists.
 if command -v pkg-config >/dev/null 2>&1 && pkg-config --exists opus; then
     LIBOPUS_DIR=$(pkg-config --variable=libdir opus)
-    cp -avfL $LIBOPUS_DIR/libopus.so* $CURDIR/package/lib/ 2>/dev/null || true
+    cp -avf $LIBOPUS_DIR/libopus.so* $CURDIR/package/lib/ 2>/dev/null || true
+    cp -avf $LIBOPUS_DIR/libopus*.dylib $CURDIR/package/lib/ 2>/dev/null || true
 fi
 
 # Detect architecture and copy appropriate libraries
@@ -38,6 +48,8 @@ elif [ -f "/lib/ld-linux-aarch64.so.1" ]; then
     cp -arfLv /lib/aarch64-linux-gnu/libdl.so.2 $CURDIR/package/lib/libdl.so.2
     cp -arfLv /lib/aarch64-linux-gnu/librt.so.1 $CURDIR/package/lib/librt.so.1
     cp -arfLv /lib/aarch64-linux-gnu/libpthread.so.0 $CURDIR/package/lib/libpthread.so.0
+elif [ "$(uname -s)" = "Darwin" ]; then
+    echo "Detected Darwin — system libraries linked dynamically, no bundled loader needed"
 else
     echo "Warning: Could not detect architecture for system library bundling"
 fi

backend/go/opus/run.sh

@@ -3,8 +3,13 @@ set -ex
 
 CURDIR=$(dirname "$(realpath "$0")")
 
-export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
-export OPUS_SHIM_LIBRARY="$CURDIR"/lib/libopusshim.so
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+	export OPUS_SHIM_LIBRARY="$CURDIR"/lib/libopusshim.dylib
+else
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
+	export OPUS_SHIM_LIBRARY="$CURDIR"/lib/libopusshim.so
+fi
 
 # If there is a lib/ld.so, use it
 if [ -f "$CURDIR"/lib/ld.so ]; then

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=89f5e2977b4d8bccd45e7bcc6f2ef7c4ed49e89a
+# Upstream pin lives below as PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
 # (.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?=89f5e2977b4d8bccd45e7bcc6f2ef7c4ed49e89a
+PARAKEET_VERSION?=f469a57270a1cc4554acb15febf60e56619673b9
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go

backend/go/piper/package.sh

@@ -16,7 +16,15 @@ cp -rfv $CURDIR/run.sh $CURDIR/package/
 cp -rfLv $CURDIR/sources/go-piper/piper-phonemize/pi/lib/* $CURDIR/package/lib/
 
 # Detect architecture and copy appropriate libraries
-if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
+if [ "$(uname)" = "Darwin" ]; then
+    # macOS has no glibc loader to bundle. The piper binary links its bundled
+    # libs (libucd, libespeak-ng, libpiper_phonemize, libonnxruntime) via
+    # @rpath but ships with no LC_RPATH, so dyld aborts at launch with
+    # "Library not loaded: @rpath/libucd.dylib ... no LC_RPATH's found".
+    # Add an @loader_path/lib rpath so @rpath resolves to package/lib/.
+    echo "Detected macOS; adding @loader_path/lib rpath so bundled libs resolve via @rpath..."
+    install_name_tool -add_rpath @loader_path/lib "$CURDIR/package/piper"
+elif [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
     # x86_64 architecture
     echo "Detected x86_64 architecture, copying x86_64 libraries..."
     cp -arfLv /lib64/ld-linux-x86-64.so.2 $CURDIR/package/lib/ld.so

backend/go/piper/run.sh

@@ -4,7 +4,12 @@ set -ex
 CURDIR=$(dirname "$(realpath "$0")")
 
 export ESPEAK_NG_DATA="$CURDIR"/espeak-ng-data
-export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
+
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
+fi
 
 # If there is a lib/ld.so, use it
 if [ -f "$CURDIR"/lib/ld.so ]; then

backend/go/silero-vad/package.sh

@@ -15,7 +15,14 @@ cp -avf $CURDIR/run.sh $CURDIR/package/
 cp -rfLv $CURDIR/backend-assets/lib/* $CURDIR/package/lib/
 
 # Detect architecture and copy appropriate libraries
-if [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
+if [ "$(uname)" = "Darwin" ]; then
+    # macOS has no glibc loader to bundle. silero-vad links its bundled
+    # libonnxruntime via @rpath but ships with no LC_RPATH, so dyld can't find
+    # it at runtime. Add an @loader_path/lib rpath so @rpath resolves to
+    # package/lib/ (matching the piper darwin fix, #10525).
+    echo "Detected macOS; adding @loader_path/lib rpath so bundled libs resolve via @rpath..."
+    install_name_tool -add_rpath @loader_path/lib "$CURDIR/package/silero-vad"
+elif [ -f "/lib64/ld-linux-x86-64.so.2" ]; then
     # x86_64 architecture
     echo "Detected x86_64 architecture, copying x86_64 libraries..."
     cp -arfLv /lib64/ld-linux-x86-64.so.2 $CURDIR/package/lib/ld.so

backend/go/silero-vad/run.sh

@@ -3,7 +3,11 @@ set -ex
 
 CURDIR=$(dirname "$(realpath "$0")")
 
-export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
+if [ "$(uname)" = "Darwin" ]; then
+	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
+else
+	export LD_LIBRARY_PATH="$CURDIR"/lib:$LD_LIBRARY_PATH
+fi
 
 # If there is a lib/ld.so, use it
 if [ -f "$CURDIR"/lib/ld.so ]; then

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?=8caa3f908ae6d4a4bef531e73b9a969f266a3d1f
+STABLEDIFFUSION_GGML_VERSION?=9956436c925a367daeab097598b1ea1f32d3503f
 
 CMAKE_ARGS+=-DGGML_MAX_NAME=128
 

backend/go/whisper/run.sh

@@ -13,8 +13,14 @@ if [ "$(uname)" != "Darwin" ]; then
 fi
 
 if [ "$(uname)" = "Darwin" ]; then
-	# macOS: single dylib variant (Metal or Accelerate)
-	LIBRARY="$CURDIR/libgowhisper-fallback.dylib"
+	# macOS: single fallback variant (Metal/Accelerate). The cmake build emits a
+	# Mach-O named .so, but tolerate .dylib too — pick whichever exists so the Go
+	# loader doesn't panic on a hardcoded name that isn't on disk.
+	if [ -e "$CURDIR/libgowhisper-fallback.dylib" ]; then
+		LIBRARY="$CURDIR/libgowhisper-fallback.dylib"
+	else
+		LIBRARY="$CURDIR/libgowhisper-fallback.so"
+	fi
 	export DYLD_LIBRARY_PATH="$CURDIR"/lib:$DYLD_LIBRARY_PATH
 else
 	LIBRARY="$CURDIR/libgowhisper-fallback.so"

backend/index.yaml

@@ -1356,7 +1356,6 @@
     intel: "intel-fish-speech"
     amd: "rocm-fish-speech"
     nvidia-l4t: "nvidia-l4t-fish-speech"
-    metal: "metal-fish-speech"
     default: "cpu-fish-speech"
     nvidia-cuda-13: "cuda13-fish-speech"
     nvidia-cuda-12: "cuda12-fish-speech"
@@ -4870,7 +4869,6 @@
     intel: "intel-fish-speech-development"
     amd: "rocm-fish-speech-development"
     nvidia-l4t: "nvidia-l4t-fish-speech-development"
-    metal: "metal-fish-speech-development"
     default: "cpu-fish-speech-development"
     nvidia-cuda-13: "cuda13-fish-speech-development"
     nvidia-cuda-12: "cuda12-fish-speech-development"
@@ -4946,16 +4944,6 @@
   uri: "quay.io/go-skynet/local-ai-backends:master-nvidia-l4t-cuda-13-arm64-fish-speech"
   mirrors:
     - localai/localai-backends:master-nvidia-l4t-cuda-13-arm64-fish-speech
-- !!merge <<: *fish-speech
-  name: "metal-fish-speech"
-  uri: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-fish-speech"
-  mirrors:
-    - localai/localai-backends:latest-metal-darwin-arm64-fish-speech
-- !!merge <<: *fish-speech
-  name: "metal-fish-speech-development"
-  uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-fish-speech"
-  mirrors:
-    - localai/localai-backends:master-metal-darwin-arm64-fish-speech
 ## faster-qwen3-tts
 - !!merge <<: *faster-qwen3-tts
   name: "faster-qwen3-tts-development"

backend/python/fish-speech/requirements-mps.txt

@@ -1,2 +0,0 @@
-torch
-torchaudio

backend/python/fish-speech/requirements.txt

@@ -7,3 +7,7 @@ setuptools
 six
 scipy
 numpy
+# fish-speech is installed editable with --no-build-isolation, so the build
+# backends of its transitive deps must already be in the venv. One of them
+# builds a Rust extension and needs setuptools-rust present at metadata time.
+setuptools-rust

backend/python/llama-cpp-quantization/install.sh

@@ -11,14 +11,31 @@ fi
 EXTRA_PIP_INSTALL_FLAGS+=" --upgrade "
 installRequirements
 
-# Fetch convert_hf_to_gguf.py from llama.cpp
+# Fetch convert_hf_to_gguf.py from llama.cpp.
+# Upstream split the model-specific logic out of the single file into a
+# sibling `conversion/` package (convert_hf_to_gguf.py now does
+# `from conversion import ...`), so a single-file download no longer runs —
+# it fails with `ModuleNotFoundError: No module named 'conversion'`. We clone
+# the repo and copy both the script and the package; Python puts the script's
+# own directory on sys.path[0], so the package resolves when placed beside it.
 LLAMA_CPP_CONVERT_VERSION="${LLAMA_CPP_CONVERT_VERSION:-master}"
+LLAMA_CPP_SRC="${EDIR}/llama.cpp"
 CONVERT_SCRIPT="${EDIR}/convert_hf_to_gguf.py"
-if [ ! -f "${CONVERT_SCRIPT}" ]; then
-    echo "Downloading convert_hf_to_gguf.py from llama.cpp (${LLAMA_CPP_CONVERT_VERSION})..."
-    curl -L --fail --retry 3 \
-        "https://raw.githubusercontent.com/ggml-org/llama.cpp/${LLAMA_CPP_CONVERT_VERSION}/convert_hf_to_gguf.py" \
-        -o "${CONVERT_SCRIPT}" || echo "Warning: Failed to download convert_hf_to_gguf.py."
+
+cloneLlamaCpp() {
+    if [ ! -d "${LLAMA_CPP_SRC}/.git" ]; then
+        git clone --depth 1 --branch "${LLAMA_CPP_CONVERT_VERSION}" \
+            https://github.com/ggml-org/llama.cpp.git "${LLAMA_CPP_SRC}" 2>/dev/null || \
+        git clone --depth 1 https://github.com/ggml-org/llama.cpp.git "${LLAMA_CPP_SRC}"
+    fi
+}
+
+if [ ! -f "${CONVERT_SCRIPT}" ] || [ ! -d "${EDIR}/conversion" ]; then
+    echo "Fetching convert_hf_to_gguf.py + conversion/ from llama.cpp (${LLAMA_CPP_CONVERT_VERSION})..."
+    cloneLlamaCpp
+    cp "${LLAMA_CPP_SRC}/convert_hf_to_gguf.py" "${CONVERT_SCRIPT}"
+    rm -rf "${EDIR}/conversion"
+    cp -r "${LLAMA_CPP_SRC}/conversion" "${EDIR}/conversion"
 fi
 
 # Install gguf package from the same llama.cpp commit to keep them in sync
@@ -41,12 +58,7 @@ QUANTIZE_BIN="${EDIR}/llama-quantize"
 if [ ! -x "${QUANTIZE_BIN}" ] && ! command -v llama-quantize &>/dev/null; then
     if command -v cmake &>/dev/null; then
         echo "Building llama-quantize from llama.cpp (${LLAMA_CPP_CONVERT_VERSION})..."
-        LLAMA_CPP_SRC="${EDIR}/llama.cpp"
-        if [ ! -d "${LLAMA_CPP_SRC}" ]; then
-            git clone --depth 1 --branch "${LLAMA_CPP_CONVERT_VERSION}" \
-                https://github.com/ggml-org/llama.cpp.git "${LLAMA_CPP_SRC}" 2>/dev/null || \
-            git clone --depth 1 https://github.com/ggml-org/llama.cpp.git "${LLAMA_CPP_SRC}"
-        fi
+        cloneLlamaCpp  # reuses the clone fetched for convert_hf_to_gguf.py
         cmake -B "${LLAMA_CPP_SRC}/build" -S "${LLAMA_CPP_SRC}" -DGGML_NATIVE=OFF -DBUILD_SHARED_LIBS=OFF
         cmake --build "${LLAMA_CPP_SRC}/build" --target llama-quantize -j"$(nproc 2>/dev/null || echo 2)"
         cp "${LLAMA_CPP_SRC}/build/bin/llama-quantize" "${QUANTIZE_BIN}"

backend/python/sglang/install.sh

@@ -85,9 +85,15 @@ if [ "x${BUILD_TYPE}" == "x" ] || [ "x${FROM_SOURCE:-}" == "xtrue" ]; then
     # The resulting binary still requires an AVX-512 capable CPU at runtime,
     # same constraint sglang upstream documents in docker/xeon.Dockerfile.
 
+    # Pin the source build to the same release the GPU path floors on
+    # (0.5.11, see requirements-cublas12-after.txt). An unpinned master clone
+    # pulls in newer CPU kernels (e.g. mamba/fla.cpp) that fail to compile
+    # (constexpr non-constant + kineto_LIBRARY-NOTFOUND). Bump deliberately.
+    SGLANG_VERSION="${SGLANG_VERSION:-v0.5.11}"
     _sgl_src=$(mktemp -d)
     trap 'rm -rf "${_sgl_src}"' EXIT
-    git clone --depth 1 https://github.com/sgl-project/sglang "${_sgl_src}/sglang"
+    git clone --depth 1 --branch "${SGLANG_VERSION}" \
+        https://github.com/sgl-project/sglang "${_sgl_src}/sglang"
 
     # Patch -march=native → -march=sapphirerapids in the CPU kernel CMakeLists
     sed -i 's/-march=native/-march=sapphirerapids/g' \

backend/rust/kokoros/src/service.rs

@@ -570,6 +570,43 @@ impl Backend for KokorosService {
     ) -> Result<Response<backend::Result>, Status> {
         Err(Status::unimplemented("Not supported"))
     }
+
+    async fn sound_detection(
+        &self,
+        _: Request<backend::SoundDetectionRequest>,
+    ) -> Result<Response<backend::SoundDetectionResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn depth(
+        &self,
+        _: Request<backend::DepthRequest>,
+    ) -> Result<Response<backend::DepthResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn token_classify(
+        &self,
+        _: Request<backend::TokenClassifyRequest>,
+    ) -> Result<Response<backend::TokenClassifyResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    async fn score(
+        &self,
+        _: Request<backend::ScoreRequest>,
+    ) -> Result<Response<backend::ScoreResponse>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
+
+    type ForwardStream = ReceiverStream<Result<backend::ForwardReply, Status>>;
+
+    async fn forward(
+        &self,
+        _: Request<tonic::Streaming<backend::ForwardRequest>>,
+    ) -> Result<Response<Self::ForwardStream>, Status> {
+        Err(Status::unimplemented("Not supported"))
+    }
 }
 
 #[cfg(test)]

cmd/launcher/FyneApp.toml

@@ -0,0 +1,8 @@
+Website = "https://localai.io"
+
+[Details]
+Icon = "../../core/http/static/logo.png"
+Name = "LocalAI"
+ID = "com.localai.launcher"
+Version = "0.0.0"
+Build = 1

contrib/macos/Launcher.entitlements

@@ -0,0 +1,14 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>com.apple.security.network.client</key>
+    <true/>
+    <key>com.apple.security.network.server</key>
+    <true/>
+    <key>com.apple.security.cs.allow-jit</key>
+    <true/>
+    <key>com.apple.security.cs.allow-unsigned-executable-memory</key>
+    <true/>
+</dict>
+</plist>

contrib/macos/sign-and-notarize.sh

@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+# Code-sign and notarize macOS artifacts for LocalAI.
+# Every sub-command is a no-op (exit 0) when its required secret is unset,
+# so unsigned builds (forks, local dev, PRs) keep working.
+set -euo pipefail
+
+ENTITLEMENTS="contrib/macos/Launcher.entitlements"
+KEYCHAIN="localai-ci.keychain-db"
+
+cmd_import_cert() {
+  if [ -z "${MACOS_CERTIFICATE:-}" ]; then
+    echo "[sign] MACOS_CERTIFICATE unset: skipping cert import (unsigned build)"
+    return 0
+  fi
+  local certfile keychain_pwd default_keychain
+  certfile="$(mktemp).p12"
+  keychain_pwd="${MACOS_CI_KEYCHAIN_PWD:?MACOS_CI_KEYCHAIN_PWD required when signing}"
+  echo "$MACOS_CERTIFICATE" | base64 --decode > "$certfile"
+  security create-keychain -p "$keychain_pwd" "$KEYCHAIN"
+  security set-keychain-settings -lut 21600 "$KEYCHAIN"
+  security unlock-keychain -p "$keychain_pwd" "$KEYCHAIN"
+  security import "$certfile" -k "$KEYCHAIN" -P "${MACOS_CERTIFICATE_PWD:?}" \
+    -T /usr/bin/codesign -T /usr/bin/security
+  security set-key-partition-list -S apple-tool:,apple:,codesign: \
+    -s -k "$keychain_pwd" "$KEYCHAIN" >/dev/null
+  default_keychain="$(security default-keychain | tr -d ' "')"
+  security list-keychains -d user -s "$KEYCHAIN" "$default_keychain"
+  rm -f "$certfile"
+  echo "[sign] certificate imported into $KEYCHAIN"
+}
+
+cmd_sign() {
+  local target="$1"
+  if [ -z "${MACOS_SIGN_IDENTITY:-}" ]; then
+    echo "[sign] MACOS_SIGN_IDENTITY unset: skipping codesign of $target"
+    return 0
+  fi
+  case "$target" in
+    *.app)
+      # Hardened runtime + entitlements are required for notarizing the app bundle.
+      codesign --deep --force --options runtime --timestamp \
+        --entitlements "$ENTITLEMENTS" \
+        --sign "$MACOS_SIGN_IDENTITY" "$target"
+      ;;
+    *)
+      # A disk image carries no entitlements/runtime; just sign the container.
+      codesign --force --timestamp --sign "$MACOS_SIGN_IDENTITY" "$target"
+      ;;
+  esac
+  codesign --verify --strict --verbose=2 "$target"
+  echo "[sign] signed $target"
+}
+
+cmd_notarize() {
+  local dmg="$1"
+  if [ -z "${MACOS_NOTARY_KEY:-}" ]; then
+    echo "[notarize] MACOS_NOTARY_KEY unset: skipping notarization of $dmg"
+    return 0
+  fi
+  local keyfile
+  keyfile="$(mktemp).p8"
+  echo "$MACOS_NOTARY_KEY" | base64 --decode > "$keyfile"
+  xcrun notarytool submit "$dmg" \
+    --key "$keyfile" \
+    --key-id "${MACOS_NOTARY_KEY_ID:?}" \
+    --issuer "${MACOS_NOTARY_ISSUER_ID:?}" \
+    --wait
+  rm -f "$keyfile"
+  xcrun stapler staple "$dmg"
+  xcrun stapler validate "$dmg"
+  echo "[notarize] notarized and stapled $dmg"
+}
+
+main() {
+  local sub="${1:-}"; shift || true
+  case "$sub" in
+    import-cert) cmd_import_cert ;;
+    sign)        cmd_sign "$@" ;;
+    notarize)    cmd_notarize "$@" ;;
+    *) echo "usage: $0 {import-cert|sign <path>|notarize <dmg>}" >&2; exit 2 ;;
+  esac
+}
+
+main "$@"

core/application/agent_jobs.go

@@ -37,6 +37,8 @@ func (a *Application) RestartAgentJobService() error {
 		if d.JobStore != nil {
 			agentJobService.SetDistributedJobStore(d.JobStore)
 		}
+		// Keep agent tasks consistent across replicas (same client the dispatcher uses).
+		agentJobService.SetTaskSyncNATS(d.Nats)
 	}
 
 	// Start the service

core/application/application.go

@@ -604,6 +604,10 @@ func (a *Application) StartAgentPool() {
 			usm.SetJobDBStore(s)
 		}
 	}
+	// Keep per-user agent tasks consistent across replicas (nil in standalone).
+	if d := a.Distributed(); d != nil {
+		usm.SetJobSyncNATS(d.Nats)
+	}
 	aps.SetUserServicesManager(usm)
 
 	a.agentPoolService.Store(aps)

core/application/distributed.go

@@ -355,6 +355,7 @@ func initDistributed(cfg *config.ApplicationConfig, authDB *gorm.DB, configLoade
 		PrefixProvider:   prefixProvider,
 		PrefixConfig:     prefixCfg,
 		Pressure:         pressure,
+		SharedModels:     cfg.Distributed.SharedModels,
 	})
 
 	// Wire staging-progress broadcasting so file-staging shows up on every

core/application/startup.go

@@ -16,6 +16,7 @@ import (
 	"github.com/mudler/LocalAI/core/services/galleryop"
 	"github.com/mudler/LocalAI/core/services/jobs"
 	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/core/services/modeladmin"
 	"github.com/mudler/LocalAI/core/services/monitoring"
 	"github.com/mudler/LocalAI/core/services/nodes"
 	"github.com/mudler/LocalAI/core/services/routing/admission"
@@ -279,6 +280,9 @@ func New(opts ...config.AppOption) (*Application, error) {
 		if application.agentJobService != nil {
 			application.agentJobService.SetDistributedBackends(distSvc.Dispatcher)
 			application.agentJobService.SetDistributedJobStore(distSvc.JobStore)
+			// Keep agent tasks consistent across replicas (jobs already sync via the
+			// dispatcher + DB read-through). Same NATS client the dispatcher uses.
+			application.agentJobService.SetTaskSyncNATS(distSvc.Nats)
 		}
 		// Wire skill store into AgentPoolService (wired at pool start time via closure)
 		// The actual wiring happens in StartAgentPool since the pool doesn't exist yet.
@@ -330,9 +334,14 @@ func New(opts ...config.AppOption) (*Application, error) {
 			gs := application.galleryService
 			sys := options.SystemState
 			cfgLoaderOpts := options.ToConfigLoaderOptions()
-			gs.OnModelsChanged = func(_ messaging.CacheInvalidateEvent) {
-				if err := application.ModelConfigLoader().LoadModelConfigsFromPath(sys.Model.ModelsPath, cfgLoaderOpts...); err != nil {
-					xlog.Warn("Failed to reload model configs after peer invalidation", "error", err)
+			gs.OnModelsChanged = func(evt messaging.CacheInvalidateEvent) {
+				// ApplyRemoteChange honors the op: a "delete" prunes the element
+				// (a reload-from-path is additive and cannot drop it), anything
+				// else reloads from disk; a named element's running instance is
+				// shut down so the new config takes effect. The originating
+				// replica reloads inline and never depends on this path.
+				if err := modeladmin.ApplyRemoteChange(application.ModelConfigLoader(), application.modelLoader, sys.Model.ModelsPath, evt, cfgLoaderOpts...); err != nil {
+					xlog.Warn("Failed to apply peer model config change", "error", err)
 				}
 			}
 			if err := application.galleryService.SubscribeBroadcasts(); err != nil {

core/cli/run.go

@@ -160,6 +160,7 @@ type RunCMD struct {
 	RegistrationRequireAuth   bool   `env:"LOCALAI_REGISTRATION_REQUIRE_AUTH" default:"false" help:"Fail startup when distributed mode is enabled but LOCALAI_REGISTRATION_TOKEN is empty (node endpoints and worker file-transfer server would otherwise be unauthenticated)" group:"distributed"`
 	DistributedRequireAuth    bool   `env:"LOCALAI_DISTRIBUTED_REQUIRE_AUTH" default:"false" help:"Umbrella switch: require BOTH NATS JWT credentials and a registration token when distributed mode is enabled (implies --nats-require-auth and --registration-require-auth)" group:"distributed"`
 	AutoApproveNodes          bool   `env:"LOCALAI_AUTO_APPROVE_NODES" default:"false" help:"Auto-approve new worker nodes (skip admin approval)" group:"distributed"`
+	DistributedSharedModels   bool   `env:"LOCALAI_DISTRIBUTED_SHARED_MODELS" default:"false" help:"Assert that every node mounts the SAME models directory at the SAME path (shared volume). When true, the router skips staging model files to workers and loads them directly from the shared path, avoiding re-downloads." group:"distributed"`
 	DistributedPrefixCache    bool   `env:"LOCALAI_DISTRIBUTED_PREFIX_CACHE" default:"true" help:"Enable prefix-cache-aware routing in distributed mode (default true). When false, routing falls back to round-robin." group:"distributed"`
 	DistributedPrefixCacheTTL string `env:"LOCALAI_DISTRIBUTED_PREFIX_CACHE_TTL" help:"Idle-timeout for prefix-cache index entries; also drives the background eviction cadence (every TTL/2). Default 5m." group:"distributed"`
 	BackendInstallTimeout     string `env:"LOCALAI_NATS_BACKEND_INSTALL_TIMEOUT" help:"NATS round-trip timeout for backend.install requests sent to worker nodes (default 15m). Increase for slow links pulling multi-GB images." group:"distributed"`
@@ -310,6 +311,9 @@ func (r *RunCMD) Run(ctx *cliContext.Context) error {
 	if r.DistributedRequireAuth {
 		opts = append(opts, config.EnableDistributedRequireAuth)
 	}
+	if r.DistributedSharedModels {
+		opts = append(opts, config.EnableDistributedSharedModels)
+	}
 	if r.NatsAccountSeed != "" {
 		opts = append(opts, config.WithNatsAccountSeed(r.NatsAccountSeed))
 	}

core/config/defaults.go

@@ -12,14 +12,12 @@ package config
 // these; config never imports backend.
 const (
 	// DefaultContextSize is the fallback context window when none is configured
-	// or estimable from the model.
+	// or estimable from the model. It is also the fallback for a GGUF whose
+	// metadata yields no usable estimate or that the parser cannot read at all
+	// (e.g. a quant type it does not know, such as NVFP4): a model-agnostic
+	// safe default beats a tiny, surprising window that truncates real prompts.
 	DefaultContextSize = 4096
 
-	// GGUFFallbackContextSize is the context window for a GGUF model whose
-	// metadata yields no usable estimate (see guessGGUFFromFile). Deliberately
-	// smaller than DefaultContextSize to stay conservative on memory there.
-	GGUFFallbackContextSize = 1024
-
 	// DefaultNGPULayers means "offload all layers"; the backend (fit_params)
 	// clamps to what actually fits in device memory.
 	DefaultNGPULayers = 99999999

core/config/distributed_config.go

@@ -31,6 +31,14 @@ type DistributedConfig struct {
 	// available to enforce just one layer.
 	RequireAuth      bool // LOCALAI_DISTRIBUTED_REQUIRE_AUTH
 	AutoApproveNodes bool // --auto-approve-nodes / LOCALAI_AUTO_APPROVE_NODES (skip admin approval for new workers)
+	// SharedModels asserts that every node (frontend and workers) mounts the
+	// SAME models directory at the SAME path (e.g. a shared volume, as in
+	// docker-compose.distributed.yaml). When true, the router skips staging
+	// model files to workers entirely: the frontend's absolute model paths are
+	// already valid on the worker, so re-uploading them into a per-model
+	// subdirectory only re-downloads what is already present (#10556). Default
+	// false preserves the historical per-node staging behavior.
+	SharedModels bool // --distributed-shared-models / LOCALAI_DISTRIBUTED_SHARED_MODELS
 
 	// NATS JWT auth (optional; see pkg/natsauth and docs/features/distributed-mode.md)
 	NatsAccountSeed  string        // LOCALAI_NATS_ACCOUNT_SEED — account signing seed to mint per-node worker JWTs
@@ -282,6 +290,13 @@ var EnableAutoApproveNodes = func(o *ApplicationConfig) {
 	o.Distributed.AutoApproveNodes = true
 }
 
+// EnableDistributedSharedModels marks the cluster as sharing one models
+// directory across all nodes, so the router skips staging model files to
+// workers (see DistributedConfig.SharedModels).
+var EnableDistributedSharedModels = func(o *ApplicationConfig) {
+	o.Distributed.SharedModels = true
+}
+
 // DisablePrefixCache turns off prefix-cache-aware routing (falls back to
 // round-robin). Prefix-cache routing is enabled by default in distributed mode.
 var DisablePrefixCache = func(o *ApplicationConfig) {

core/config/gguf.go

@@ -33,7 +33,7 @@ func guessGGUFFromFile(cfg *ModelConfig, f *gguf.GGUFFile, defaultCtx int) {
 			cSize := int(ctxSize)
 			cfg.ContextSize = &cSize
 		} else {
-			defaultCtx = GGUFFallbackContextSize
+			defaultCtx = DefaultContextSize
 			cfg.ContextSize = &defaultCtx
 		}
 	}

core/config/hooks_llamacpp.go

@@ -34,7 +34,7 @@ func llamaCppDefaults(cfg *ModelConfig, modelPath string) {
 	// Default context size if not set, regardless of whether GGUF parsing succeeds
 	defer func() {
 		if cfg.ContextSize == nil {
-			ctx := GGUFFallbackContextSize
+			ctx := DefaultContextSize
 			cfg.ContextSize = &ctx
 		}
 	}()

core/config/hooks_test.go

@@ -248,7 +248,11 @@ var _ = Describe("Backend hooks and parser defaults", func() {
 			}
 			cfg.SetDefaults(ModelPath(dir))
 
+			// An unreadable/unparseable GGUF (e.g. a quant type the parser does
+			// not know, such as NVFP4) yields no estimate, so the hook must fall
+			// back to DefaultContextSize rather than a tiny, surprising value.
 			Expect(cfg.ContextSize).NotTo(BeNil())
+			Expect(*cfg.ContextSize).To(Equal(DefaultContextSize))
 		})
 	})
 

core/gallery/importers/llama-cpp.go

@@ -25,8 +25,8 @@ var (
 
 type LlamaCPPImporter struct{}
 
-func (i *LlamaCPPImporter) Name() string     { return "llama-cpp" }
-func (i *LlamaCPPImporter) Modality() string { return "text" }
+func (i *LlamaCPPImporter) Name() string      { return "llama-cpp" }
+func (i *LlamaCPPImporter) Modality() string  { return "text" }
 func (i *LlamaCPPImporter) AutoDetects() bool { return true }
 
 // AdditionalBackends advertises drop-in replacements that share the
@@ -293,7 +293,7 @@ func pickPreferredGroup(groups []hfapi.ShardGroup, prefs []string) *hfapi.ShardG
 	for _, pref := range prefs {
 		lower := strings.ToLower(pref)
 		for i := range groups {
-			if strings.Contains(strings.ToLower(groups[i].Base), lower) {
+			if quantTokenMatches(strings.ToLower(groups[i].Base), lower) {
 				return &groups[i]
 			}
 		}
@@ -301,6 +301,39 @@ func pickPreferredGroup(groups []hfapi.ShardGroup, prefs []string) *hfapi.ShardG
 	return &groups[len(groups)-1]
 }
 
+// quantTokenMatches reports whether pref appears in base as a whole token
+// rather than as a substring of a larger alphanumeric run. Both arguments
+// must already be lowercased.
+//
+// A plain strings.Contains is wrong here: `f16` is a substring of `bf16`, so
+// asking for the `F16` quant used to wrongly select a `BF16` file (#10559).
+// Only the OUTER edges of the matched preference must hit a boundary — a
+// non-alphanumeric char (or the start/end of base). Separators inside the
+// preference itself (e.g. `ud-q4_k_xl`) are intentionally left untouched.
+func quantTokenMatches(base, pref string) bool {
+	if pref == "" {
+		return false
+	}
+	for start := strings.Index(base, pref); start != -1; {
+		end := start + len(pref)
+		leftOK := start == 0 || !isAlphaNum(base[start-1])
+		rightOK := end == len(base) || !isAlphaNum(base[end])
+		if leftOK && rightOK {
+			return true
+		}
+		next := strings.Index(base[start+1:], pref)
+		if next == -1 {
+			break
+		}
+		start += next + 1
+	}
+	return false
+}
+
+func isAlphaNum(b byte) bool {
+	return (b >= 'a' && b <= 'z') || (b >= '0' && b <= '9')
+}
+
 // maybeApplyMTPDefaults parses the picked GGUF header (range-fetched over
 // HTTP for HF/URL imports) and, if the file declares a Multi-Token Prediction
 // head, appends the auto-MTP option keys to modelConfig.Options. Failures

core/gallery/importers/llama-cpp_test.go

@@ -374,6 +374,104 @@ var _ = Describe("LlamaCPPImporter", func() {
 		})
 	})
 
+	Context("quant token boundary matching", func() {
+		// Regression for #10559: the quant preference must match as a whole
+		// token, not as a substring. Asking for `F16` used to select a
+		// `BF16` mmproj because strings.Contains("...bf16.gguf", "f16") is
+		// true — the leading `b` was ignored.
+
+		const repoBase = "https://huggingface.co/acme/example-GGUF/resolve/main/"
+
+		hfFile := func(path, sha string) hfapi.ModelFile {
+			return hfapi.ModelFile{
+				Path:   path,
+				SHA256: sha,
+				URL:    repoBase + path,
+			}
+		}
+
+		withHF := func(preferences string, files ...hfapi.ModelFile) Details {
+			d := Details{
+				URI: "https://huggingface.co/acme/example-GGUF",
+				HuggingFace: &hfapi.ModelDetails{
+					ModelID: "acme/example-GGUF",
+					Files:   files,
+				},
+			}
+			if preferences != "" {
+				d.Preferences = json.RawMessage(preferences)
+			}
+			return d
+		}
+
+		It("selects the F16 mmproj over BF16 (BF16 listed first)", func() {
+			details := withHF(`{"name":"VL","mmproj_quantizations":"F16"}`,
+				hfFile("model-Q4_K_M.gguf", "model"),
+				hfFile("mmproj-x-BF16.gguf", "bf16"),
+				hfFile("mmproj-x-F16.gguf", "f16"),
+			)
+
+			modelConfig, err := importer.Import(details)
+
+			Expect(err).ToNot(HaveOccurred())
+			Expect(modelConfig.ConfigFile).To(ContainSubstring("mmproj: llama-cpp/mmproj/VL/mmproj-x-F16.gguf"), fmt.Sprintf("%+v", modelConfig))
+			Expect(modelConfig.ConfigFile).ToNot(ContainSubstring("BF16"), fmt.Sprintf("%+v", modelConfig))
+		})
+
+		It("selects the F16 mmproj over BF16 (F16 listed first)", func() {
+			details := withHF(`{"name":"VL","mmproj_quantizations":"F16"}`,
+				hfFile("model-Q4_K_M.gguf", "model"),
+				hfFile("mmproj-x-F16.gguf", "f16"),
+				hfFile("mmproj-x-BF16.gguf", "bf16"),
+			)
+
+			modelConfig, err := importer.Import(details)
+
+			Expect(err).ToNot(HaveOccurred())
+			Expect(modelConfig.ConfigFile).To(ContainSubstring("mmproj: llama-cpp/mmproj/VL/mmproj-x-F16.gguf"), fmt.Sprintf("%+v", modelConfig))
+			Expect(modelConfig.ConfigFile).ToNot(ContainSubstring("BF16"), fmt.Sprintf("%+v", modelConfig))
+		})
+
+		It("selects BF16 when BF16 is the requested mmproj quant", func() {
+			details := withHF(`{"name":"VL","mmproj_quantizations":"BF16"}`,
+				hfFile("model-Q4_K_M.gguf", "model"),
+				hfFile("mmproj-x-F16.gguf", "f16"),
+				hfFile("mmproj-x-BF16.gguf", "bf16"),
+			)
+
+			modelConfig, err := importer.Import(details)
+
+			Expect(err).ToNot(HaveOccurred())
+			Expect(modelConfig.ConfigFile).To(ContainSubstring("mmproj: llama-cpp/mmproj/VL/mmproj-x-BF16.gguf"), fmt.Sprintf("%+v", modelConfig))
+		})
+
+		It("still matches a normal model quant with internal separators", func() {
+			// ud-q4_k_xl contains `-`/`_` internally; only the outer edges
+			// must hit a token boundary.
+			details := withHF(`{"name":"M","quantizations":"ud-q4_k_xl"}`,
+				hfFile("model-UD-Q4_K_XL.gguf", "xl"),
+				hfFile("model-Q3_K_M.gguf", "q3"),
+			)
+
+			modelConfig, err := importer.Import(details)
+
+			Expect(err).ToNot(HaveOccurred())
+			Expect(modelConfig.ConfigFile).To(ContainSubstring("model: llama-cpp/models/M/model-UD-Q4_K_XL.gguf"), fmt.Sprintf("%+v", modelConfig))
+		})
+
+		It("falls back to the last group when no preference matches", func() {
+			details := withHF(`{"name":"M","quantizations":"Q2_K"}`,
+				hfFile("model-Q8_0.gguf", "q8"),
+				hfFile("model-Q3_K_M.gguf", "q3"),
+			)
+
+			modelConfig, err := importer.Import(details)
+
+			Expect(err).ToNot(HaveOccurred())
+			Expect(modelConfig.ConfigFile).To(ContainSubstring("model: llama-cpp/models/M/model-Q3_K_M.gguf"), fmt.Sprintf("%+v", modelConfig))
+		})
+	})
+
 	Context("AdditionalBackends", func() {
 		It("advertises ik-llama-cpp and turboquant as drop-in replacements", func() {
 			entries := importer.AdditionalBackends()

core/http/app.go

@@ -23,8 +23,10 @@ import (
 
 	"github.com/mudler/LocalAI/core/application"
 	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/distributed"
 	"github.com/mudler/LocalAI/core/services/finetune"
 	"github.com/mudler/LocalAI/core/services/galleryop"
+	"github.com/mudler/LocalAI/core/services/messaging"
 	"github.com/mudler/LocalAI/core/services/nodes"
 	"github.com/mudler/LocalAI/core/services/quantization"
 
@@ -400,25 +402,45 @@ func API(application *application.Application) (*echo.Echo, error) {
 	routes.RegisterAgentPoolRoutes(e, application, agentsMw, skillsMw, collectionsMw)
 	// Fine-tuning routes
 	fineTuningMw := auth.RequireFeature(application.AuthDB(), auth.FeatureFineTuning)
+	// In distributed mode pass the shared NATS client + PostgreSQL store so
+	// fine-tune jobs stay consistent across replicas (the SyncedMap broadcasts
+	// mutations and hydrates from the DB); standalone passes nil for both.
+	var ftNats messaging.MessagingClient
+	var ftStore *distributed.FineTuneStore
+	if d := application.Distributed(); d != nil {
+		ftNats = d.Nats
+		if d.DistStores != nil && d.DistStores.FineTune != nil {
+			ftStore = d.DistStores.FineTune
+		}
+	}
 	ftService := finetune.NewFineTuneService(
 		application.ApplicationConfig(),
 		application.ModelLoader(),
 		application.ModelConfigLoader(),
+		ftNats,
+		ftStore,
 	)
-	if d := application.Distributed(); d != nil {
-		ftService.SetNATSClient(d.Nats)
-		if d.DistStores != nil && d.DistStores.FineTune != nil {
-			ftService.SetFineTuneStore(d.DistStores.FineTune)
-		}
-	}
 	routes.RegisterFineTuningRoutes(e, ftService, application.ApplicationConfig(), fineTuningMw)
 
 	// Quantization routes
 	quantizationMw := auth.RequireFeature(application.AuthDB(), auth.FeatureQuantization)
+	// In distributed mode pass the shared NATS client + PostgreSQL store so
+	// quantization jobs stay consistent across replicas (the SyncedMap broadcasts
+	// mutations and hydrates from the DB); standalone passes nil for both.
+	var quantNats messaging.MessagingClient
+	var quantStore *distributed.QuantStore
+	if d := application.Distributed(); d != nil {
+		quantNats = d.Nats
+		if d.DistStores != nil && d.DistStores.Quant != nil {
+			quantStore = d.DistStores.Quant
+		}
+	}
 	qService := quantization.NewQuantizationService(
 		application.ApplicationConfig(),
 		application.ModelLoader(),
 		application.ModelConfigLoader(),
+		quantNats,
+		quantStore,
 	)
 	routes.RegisterQuantizationRoutes(e, qService, application.ApplicationConfig(), quantizationMw)
 

core/http/endpoints/localai/config_meta.go

@@ -155,7 +155,7 @@ func AutocompleteEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, a
 // @Param name path string true "Model name"
 // @Success 200 {object} map[string]any "success message"
 // @Router /api/models/config-json/{name} [patch]
-func PatchConfigEndpoint(cl *config.ModelConfigLoader, _ *model.ModelLoader, appConfig *config.ApplicationConfig) echo.HandlerFunc {
+func PatchConfigEndpoint(cl *config.ModelConfigLoader, _ *model.ModelLoader, gs *galleryop.GalleryService, appConfig *config.ApplicationConfig) echo.HandlerFunc {
 	svc := modeladmin.NewConfigService(cl, appConfig)
 	return func(c echo.Context) error {
 		modelName := c.Param("name")
@@ -173,6 +173,14 @@ func PatchConfigEndpoint(cl *config.ModelConfigLoader, _ *model.ModelLoader, app
 		if _, err := svc.PatchConfig(c.Request().Context(), modelName, patchMap); err != nil {
 			return c.JSON(httpStatusForModelAdminError(err), map[string]any{"error": err.Error()})
 		}
+
+		// Patch rewrites the config on disk and reloads only the local loader;
+		// tell peers to refresh so the change is consistent across replicas.
+		// No-op in standalone mode.
+		if gs != nil {
+			gs.BroadcastModelsChanged(modelName, "install")
+		}
+
 		return c.JSON(http.StatusOK, map[string]any{
 			"success": true,
 			"message": fmt.Sprintf("Model '%s' updated successfully", modelName),

core/http/endpoints/localai/config_meta_test.go

@@ -45,7 +45,7 @@ var _ = Describe("Config Metadata Endpoints", func() {
 		app = echo.New()
 		app.GET("/api/models/config-metadata", ConfigMetadataEndpoint())
 		app.GET("/api/models/config-metadata/autocomplete/:provider", AutocompleteEndpoint(configLoader, modelLoader, appConfig))
-		app.PATCH("/api/models/config-json/:name", PatchConfigEndpoint(configLoader, modelLoader, appConfig))
+		app.PATCH("/api/models/config-json/:name", PatchConfigEndpoint(configLoader, modelLoader, nil, appConfig))
 	})
 
 	AfterEach(func() {

core/http/endpoints/localai/edit_model.go

@@ -10,6 +10,7 @@ import (
 	"github.com/labstack/echo/v4"
 	"github.com/mudler/LocalAI/core/config"
 	httpUtils "github.com/mudler/LocalAI/core/http/middleware"
+	"github.com/mudler/LocalAI/core/services/galleryop"
 	"github.com/mudler/LocalAI/core/services/modeladmin"
 	"github.com/mudler/LocalAI/internal"
 	"github.com/mudler/LocalAI/pkg/model"
@@ -55,7 +56,7 @@ func GetEditModelPage(cl *config.ModelConfigLoader, appConfig *config.Applicatio
 }
 
 // EditModelEndpoint handles updating existing model configurations
-func EditModelEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) echo.HandlerFunc {
+func EditModelEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, gs *galleryop.GalleryService, appConfig *config.ApplicationConfig) echo.HandlerFunc {
 	svc := modeladmin.NewConfigService(cl, appConfig)
 	return func(c echo.Context) error {
 		modelName := c.Param("name")
@@ -70,6 +71,17 @@ func EditModelEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, appC
 		if err != nil {
 			return c.JSON(httpStatusForModelAdminError(err), ModelResponse{Success: false, Error: err.Error()})
 		}
+
+		// Tell peer replicas to refresh their in-memory config: this endpoint
+		// only reloaded the local loader. A rename is a delete of the old name
+		// plus an install of the new one. No-op in standalone mode.
+		if gs != nil {
+			if result.Renamed {
+				gs.BroadcastModelsChanged(result.OldName, "delete")
+			}
+			gs.BroadcastModelsChanged(result.NewName, "install")
+		}
+
 		msg := fmt.Sprintf("Model '%s' updated successfully. Model has been reloaded with new configuration.", result.NewName)
 		if result.Renamed {
 			msg = fmt.Sprintf("Model '%s' renamed to '%s' and updated successfully.", result.OldName, result.NewName)

core/http/endpoints/localai/edit_model_test.go

@@ -56,7 +56,7 @@ var _ = Describe("Edit Model test", func() {
 			app := echo.New()
 			// Set up a simple renderer for the test
 			app.Renderer = &testRenderer{}
-			app.POST("/import-model", ImportModelEndpoint(modelConfigLoader, applicationConfig))
+			app.POST("/import-model", ImportModelEndpoint(modelConfigLoader, nil, applicationConfig))
 			app.GET("/edit-model/:name", GetEditModelPage(modelConfigLoader, applicationConfig))
 
 			requestBody := bytes.NewBufferString(`{"name": "foo", "backend": "foo", "model": "foo"}`)
@@ -106,7 +106,7 @@ var _ = Describe("Edit Model test", func() {
 			Expect(exists).To(BeTrue())
 
 			app := echo.New()
-			app.POST("/models/edit/:name", EditModelEndpoint(modelConfigLoader, modelLoader, applicationConfig))
+			app.POST("/models/edit/:name", EditModelEndpoint(modelConfigLoader, modelLoader, nil, applicationConfig))
 
 			newYAML := "name: newname\nbackend: llama\nmodel: foo\n"
 			req := httptest.NewRequest("POST", "/models/edit/oldname", bytes.NewBufferString(newYAML))
@@ -163,7 +163,7 @@ var _ = Describe("Edit Model test", func() {
 			Expect(modelConfigLoader.LoadModelConfigsFromPath(tempDir)).To(Succeed())
 
 			app := echo.New()
-			app.POST("/models/edit/:name", EditModelEndpoint(modelConfigLoader, modelLoader, applicationConfig))
+			app.POST("/models/edit/:name", EditModelEndpoint(modelConfigLoader, modelLoader, nil, applicationConfig))
 
 			req := httptest.NewRequest(
 				"POST",
@@ -204,7 +204,7 @@ var _ = Describe("Edit Model test", func() {
 			Expect(modelConfigLoader.LoadModelConfigsFromPath(tempDir)).To(Succeed())
 
 			app := echo.New()
-			app.POST("/models/edit/:name", EditModelEndpoint(modelConfigLoader, modelLoader, applicationConfig))
+			app.POST("/models/edit/:name", EditModelEndpoint(modelConfigLoader, modelLoader, nil, applicationConfig))
 
 			req := httptest.NewRequest(
 				"POST",

core/http/endpoints/localai/import_model.go

@@ -125,7 +125,7 @@ func ImportModelURIEndpoint(cl *config.ModelConfigLoader, appConfig *config.Appl
 }
 
 // ImportModelEndpoint handles creating new model configurations
-func ImportModelEndpoint(cl *config.ModelConfigLoader, appConfig *config.ApplicationConfig) echo.HandlerFunc {
+func ImportModelEndpoint(cl *config.ModelConfigLoader, gs *galleryop.GalleryService, appConfig *config.ApplicationConfig) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		// Get the raw body
 		body, err := io.ReadAll(c.Request().Body)
@@ -245,6 +245,13 @@ func ImportModelEndpoint(cl *config.ModelConfigLoader, appConfig *config.Applica
 			}
 			return c.JSON(http.StatusInternalServerError, response)
 		}
+		// Tell peer replicas to load the newly-created config from the shared
+		// models dir: this endpoint only reloaded the local loader. No-op in
+		// standalone mode.
+		if gs != nil {
+			gs.BroadcastModelsChanged(modelConfig.Name, "install")
+		}
+
 		// Return success response
 		response := ModelResponse{
 			Success:  true,

core/http/endpoints/localai/nodes.go

@@ -25,6 +25,7 @@ import (
 	"github.com/mudler/LocalAI/core/http/auth"
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/core/services/galleryop"
+	"github.com/mudler/LocalAI/core/services/messaging"
 	"github.com/mudler/LocalAI/core/services/nodes"
 	"github.com/mudler/LocalAI/core/services/nodes/prefixcache"
 	"github.com/mudler/LocalAI/pkg/httpclient"
@@ -60,7 +61,10 @@ func GetNodeEndpoint(registry *nodes.NodeRegistry) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		ctx := c.Request().Context()
 		id := c.Param("id")
-		node, err := registry.Get(ctx, id)
+		// GetWithExtras (not Get) so the response carries the node's labels,
+		// loaded-model count, and in-flight total — the bare BackendNode keeps
+		// labels in a separate table, leaving the detail view's label list empty.
+		node, err := registry.GetWithExtras(ctx, id)
 		if err != nil {
 			return c.JSON(http.StatusNotFound, nodeError(http.StatusNotFound, "node not found"))
 		}
@@ -547,12 +551,23 @@ func DeleteBackendOnNodeEndpoint(unloader nodes.NodeCommandSender) echo.HandlerF
 }
 
 // ListBackendsOnNodeEndpoint lists installed backends on a worker node via NATS.
-func ListBackendsOnNodeEndpoint(unloader nodes.NodeCommandSender) echo.HandlerFunc {
+func ListBackendsOnNodeEndpoint(unloader nodes.NodeCommandSender, registry *nodes.NodeRegistry) echo.HandlerFunc {
 	return func(c echo.Context) error {
+		nodeID := c.Param("id")
+		// Agent-type workers don't run backends and never subscribe to the
+		// nodes.<id>.backend.list NATS subject, so the request would hang
+		// until timeout with "no responders". Their backend list is simply
+		// empty. Mirror the aggregate-list guard in managers_distributed.go
+		// (skip nodes whose NodeType is set and not "backend") so the
+		// single-node and cluster-wide views stay consistent.
+		if node, err := registry.Get(c.Request().Context(), nodeID); err == nil {
+			if node.NodeType != "" && node.NodeType != nodes.NodeTypeBackend {
+				return c.JSON(http.StatusOK, []messaging.NodeBackendInfo{})
+			}
+		}
 		if unloader == nil {
 			return c.JSON(http.StatusServiceUnavailable, nodeError(http.StatusServiceUnavailable, "NATS not configured"))
 		}
-		nodeID := c.Param("id")
 		reply, err := unloader.ListBackends(nodeID)
 		if err != nil {
 			xlog.Error("Failed to list backends on node", "node", nodeID, "error", err)

core/http/endpoints/localai/nodes_backends_list_test.go

@@ -0,0 +1,103 @@
+package localai
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+
+	"github.com/labstack/echo/v4"
+	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/core/services/nodes"
+	"github.com/mudler/LocalAI/core/services/testutil"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// stubNodeCommandSender records whether ListBackends was invoked so the test can
+// assert the endpoint short-circuits (no NATS request) for agent-type nodes.
+type stubNodeCommandSender struct {
+	listBackendsCalled bool
+}
+
+func (s *stubNodeCommandSender) InstallBackend(_, _, _, _, _, _, _ string, _ int, _ string, _ func(messaging.BackendInstallProgressEvent)) (*messaging.BackendInstallReply, error) {
+	return &messaging.BackendInstallReply{}, nil
+}
+
+func (s *stubNodeCommandSender) UpgradeBackend(_, _, _, _, _, _ string, _ int, _ string, _ func(messaging.BackendInstallProgressEvent)) (*messaging.BackendUpgradeReply, error) {
+	return &messaging.BackendUpgradeReply{}, nil
+}
+
+func (s *stubNodeCommandSender) DeleteBackend(_, _ string) (*messaging.BackendDeleteReply, error) {
+	return &messaging.BackendDeleteReply{Success: true}, nil
+}
+
+func (s *stubNodeCommandSender) ListBackends(_ string) (*messaging.BackendListReply, error) {
+	s.listBackendsCalled = true
+	return &messaging.BackendListReply{Backends: []messaging.NodeBackendInfo{{Name: "llama-cpp"}}}, nil
+}
+
+func (s *stubNodeCommandSender) StopBackend(_, _ string) error { return nil }
+
+func (s *stubNodeCommandSender) UnloadModelOnNode(_, _ string) error { return nil }
+
+var _ = Describe("ListBackendsOnNodeEndpoint", func() {
+	var registry *nodes.NodeRegistry
+
+	BeforeEach(func() {
+		db := testutil.SetupTestDB()
+		var err error
+		registry, err = nodes.NewNodeRegistry(db)
+		Expect(err).ToNot(HaveOccurred())
+	})
+
+	callEndpoint := func(unloader nodes.NodeCommandSender, nodeID string) *httptest.ResponseRecorder {
+		e := echo.New()
+		req := httptest.NewRequest(http.MethodGet, "/", nil)
+		rec := httptest.NewRecorder()
+		c := e.NewContext(req, rec)
+		c.SetParamNames("id")
+		c.SetParamValues(nodeID)
+		handler := ListBackendsOnNodeEndpoint(unloader, registry)
+		Expect(handler(c)).To(Succeed())
+		return rec
+	}
+
+	It("returns an empty list for an agent node without issuing a NATS request", func() {
+		ctx := context.Background()
+		node := &nodes.BackendNode{Name: "agent-1", NodeType: nodes.NodeTypeAgent}
+		Expect(registry.Register(ctx, node, true)).To(Succeed())
+
+		stub := &stubNodeCommandSender{}
+		rec := callEndpoint(stub, node.ID)
+
+		Expect(rec.Code).To(Equal(http.StatusOK))
+		Expect(stub.listBackendsCalled).To(BeFalse(),
+			"agent workers don't subscribe to backend.list; the endpoint must not issue the doomed NATS request")
+
+		var list []messaging.NodeBackendInfo
+		Expect(json.Unmarshal(rec.Body.Bytes(), &list)).To(Succeed())
+		Expect(list).To(BeEmpty())
+		// Must be `[]`, not `null`, so the UI can render it.
+		Expect(rec.Body.String()).To(ContainSubstring("[]"))
+	})
+
+	It("consults the unloader (NATS) for a backend node", func() {
+		ctx := context.Background()
+		node := &nodes.BackendNode{Name: "backend-1", NodeType: nodes.NodeTypeBackend, Address: "10.0.0.1:50051"}
+		Expect(registry.Register(ctx, node, true)).To(Succeed())
+
+		stub := &stubNodeCommandSender{}
+		rec := callEndpoint(stub, node.ID)
+
+		Expect(rec.Code).To(Equal(http.StatusOK))
+		Expect(stub.listBackendsCalled).To(BeTrue(),
+			"backend nodes must still be queried over NATS")
+
+		var list []messaging.NodeBackendInfo
+		Expect(json.Unmarshal(rec.Body.Bytes(), &list)).To(Succeed())
+		Expect(list).To(HaveLen(1))
+		Expect(list[0].Name).To(Equal("llama-cpp"))
+	})
+})

core/http/endpoints/localai/toggle_model.go

@@ -7,6 +7,7 @@ import (
 
 	"github.com/labstack/echo/v4"
 	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/services/galleryop"
 	"github.com/mudler/LocalAI/core/services/modeladmin"
 	"github.com/mudler/LocalAI/pkg/model"
 )
@@ -24,7 +25,7 @@ import (
 // @Failure      404  {object}  ModelResponse
 // @Failure      500  {object}  ModelResponse
 // @Router       /api/models/{name}/{action} [put]
-func ToggleStateModelEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) echo.HandlerFunc {
+func ToggleStateModelEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, gs *galleryop.GalleryService, appConfig *config.ApplicationConfig) echo.HandlerFunc {
 	svc := modeladmin.NewConfigService(cl, appConfig)
 	return func(c echo.Context) error {
 		modelName := c.Param("name")
@@ -36,6 +37,14 @@ func ToggleStateModelEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoade
 		if err != nil {
 			return c.JSON(httpStatusForModelAdminError(err), ModelResponse{Success: false, Error: err.Error()})
 		}
+
+		// Enabling/disabling rewrites the config on disk and reloads only the
+		// local loader; tell peers to refresh so the model's availability is
+		// consistent across replicas. No-op in standalone mode.
+		if gs != nil {
+			gs.BroadcastModelsChanged(modelName, "install")
+		}
+
 		msg := fmt.Sprintf("Model '%s' has been %sd successfully.", modelName, action)
 		if action == modeladmin.ActionDisable {
 			msg += " The model will not be loaded on demand until re-enabled."

core/http/endpoints/openresponses/responses.go

@@ -3,6 +3,7 @@ package openresponses
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"time"
 
@@ -10,6 +11,7 @@ import (
 	"github.com/labstack/echo/v4"
 	"github.com/mudler/LocalAI/core/backend"
 	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/http/auth"
 	mcpTools "github.com/mudler/LocalAI/core/http/endpoints/mcp"
 	openaiEndpoint "github.com/mudler/LocalAI/core/http/endpoints/openai"
 	"github.com/mudler/LocalAI/core/http/middleware"
@@ -246,8 +248,11 @@ func ResponsesEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, eval
 			// Create cancellable context for background execution
 			bgCtx, bgCancel := context.WithCancel(context.Background())
 
-			// Store the background response
+			// Store the background response and stamp its owner before the ID
+			// is returned to the client, so later GET/cancel/resume can verify
+			// the caller owns it.
 			store.StoreBackground(responseID, input, queuedResponse, bgCancel, input.Stream)
+			store.SetOwner(responseID, ownerFromContext(c))
 
 			// Start background processing goroutine
 			go func() {
@@ -1587,6 +1592,7 @@ func handleOpenResponsesNonStream(c echo.Context, responseID string, createdAt i
 	if shouldStore {
 		store := GetGlobalStore()
 		store.Store(responseID, input, response)
+		store.SetOwner(responseID, ownerFromContext(c))
 	}
 
 	return c.JSON(200, response)
@@ -2322,6 +2328,7 @@ func handleOpenResponsesStream(c echo.Context, responseID string, createdAt int6
 		if shouldStore {
 			store := GetGlobalStore()
 			store.Store(responseID, input, responseCompleted)
+			store.SetOwner(responseID, ownerFromContext(c))
 		}
 
 		// Send [DONE]
@@ -2966,6 +2973,18 @@ func convertORToolsToOpenAIFormat(orTools []schema.ORFunctionTool) []functions.T
 	return result
 }
 
+// ownerFromContext returns the identity (user ID) of the authenticated
+// caller, or empty string when no authentication was performed (single-key /
+// no-auth deployments). It is the value stamped on a response at creation and
+// compared on read/cancel/resume to prevent one caller from accessing
+// another's response by guessing its ID.
+func ownerFromContext(c echo.Context) string {
+	if u := auth.GetUser(c); u != nil {
+		return u.ID
+	}
+	return ""
+}
+
 // GetResponseEndpoint returns a handler for GET /responses/:id
 // This endpoint is used for polling background responses or resuming streaming
 // @Summary Get a response by ID
@@ -2991,6 +3010,12 @@ func GetResponseEndpoint() func(c echo.Context) error {
 			return sendOpenResponsesError(c, 404, "not_found", fmt.Sprintf("response not found: %s", responseID), "id")
 		}
 
+		// Enforce response ownership. Return 404 (not 403) on mismatch so the
+		// existence of another caller's response is not leaked.
+		if !accessAllowed(stored, ownerFromContext(c)) {
+			return sendOpenResponsesError(c, 404, "not_found", fmt.Sprintf("response not found: %s", responseID), "id")
+		}
+
 		// Check if streaming resume is requested
 		streamParam := c.QueryParam("stream")
 		if streamParam == "true" {
@@ -3022,16 +3047,21 @@ func GetResponseEndpoint() func(c echo.Context) error {
 
 // handleStreamResume handles resuming a streaming response from a specific sequence number
 func handleStreamResume(c echo.Context, store *ResponseStore, responseID string, stored *StoredResponse, startingAfter int) error {
+	// Fetch buffered events before committing to an SSE response so an
+	// offset-lost gap can be reported as a clean HTTP status rather than a
+	// silently truncated event stream.
+	events, err := store.GetEventsAfter(responseID, startingAfter)
+	if err != nil {
+		if errors.Is(err, ErrOffsetLost) {
+			return sendOpenResponsesError(c, 409, "invalid_request_error", fmt.Sprintf("starting_after=%d is older than the oldest retained event; the resume buffer evicted those events and the stream cannot be resumed from that point", startingAfter), "starting_after")
+		}
+		return sendOpenResponsesError(c, 500, "server_error", fmt.Sprintf("failed to get events: %v", err), "")
+	}
+
 	c.Response().Header().Set("Content-Type", "text/event-stream")
 	c.Response().Header().Set("Cache-Control", "no-cache")
 	c.Response().Header().Set("Connection", "keep-alive")
 
-	// Get buffered events after the starting point
-	events, err := store.GetEventsAfter(responseID, startingAfter)
-	if err != nil {
-		return sendOpenResponsesError(c, 500, "server_error", fmt.Sprintf("failed to get events: %v", err), "")
-	}
-
 	// Send all buffered events
 	for _, event := range events {
 		fmt.Fprintf(c.Response().Writer, "event: %s\ndata: %s\n\n", event.EventType, string(event.Data))
@@ -3126,6 +3156,17 @@ func CancelResponseEndpoint() func(c echo.Context) error {
 		}
 
 		store := GetGlobalStore()
+
+		// Look up first so ownership can be checked before any mutation.
+		stored, err := store.Get(responseID)
+		if err != nil {
+			return sendOpenResponsesError(c, 404, "not_found", fmt.Sprintf("response not found: %s", responseID), "id")
+		}
+		// Return 404 (not 403) on owner mismatch so existence is not leaked.
+		if !accessAllowed(stored, ownerFromContext(c)) {
+			return sendOpenResponsesError(c, 404, "not_found", fmt.Sprintf("response not found: %s", responseID), "id")
+		}
+
 		response, err := store.Cancel(responseID)
 		if err != nil {
 			return sendOpenResponsesError(c, 404, "not_found", fmt.Sprintf("response not found: %s", responseID), "id")

core/http/endpoints/openresponses/store.go

@@ -3,6 +3,7 @@ package openresponses
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"sync"
 	"time"
@@ -11,6 +12,30 @@ import (
 	"github.com/mudler/xlog"
 )
 
+const (
+	// defaultMaxStreamEvents bounds how many resume-buffer events a single
+	// background response retains. Without a cap, a long-running or abandoned
+	// background generation grows StreamEvents without limit and can exhaust
+	// process memory. When the cap is exceeded the oldest events are evicted
+	// from the front (see AppendEvent). Mirrors llama.cpp's byte-capped slot
+	// ring used for resumable /slots state.
+	defaultMaxStreamEvents = 8192
+
+	// defaultMaxStreamBytes caps the total serialized size of retained
+	// resume-buffer events, evicting oldest-first when exceeded. This guards
+	// against a handful of very large events defeating the count cap. 0
+	// disables the byte cap (count cap still applies).
+	defaultMaxStreamBytes = 64 << 20 // 64 MiB
+)
+
+// ErrOffsetLost is returned by GetEventsAfter when the requested
+// starting_after sequence number is older than the oldest event still
+// retained in the resume buffer (i.e. the events between the requested
+// offset and the current watermark were evicted by the cap). Callers should
+// surface this to clients as a distinct error instead of silently returning
+// a truncated stream that omits the dropped events.
+var ErrOffsetLost = errors.New("resume offset lost: requested events were evicted from the buffer")
+
 // ResponseStore provides thread-safe storage for Open Responses API responses
 type ResponseStore struct {
 	mu            sync.RWMutex
@@ -18,6 +43,12 @@ type ResponseStore struct {
 	ttl           time.Duration // Time-to-live for stored responses (0 = no expiration)
 	cleanupCtx    context.Context
 	cleanupCancel context.CancelFunc
+
+	// maxStreamEvents / maxStreamBytes bound the per-response resume buffer.
+	// Set once at construction from the default constants; tests may lower
+	// them. A value <= 0 disables that particular cap.
+	maxStreamEvents int
+	maxStreamBytes  int
 }
 
 // StreamedEvent represents a buffered SSE event for streaming resume
@@ -35,6 +66,12 @@ type StoredResponse struct {
 	StoredAt  time.Time
 	ExpiresAt *time.Time // nil if no expiration
 
+	// Owner is the identity (user ID) that created this response. It is set
+	// once at creation and never mutated, so it can be read without holding
+	// mu. Empty means "no owner" (single-key / no-auth deployments), in which
+	// case ownership checks are skipped for backward compatibility.
+	Owner string
+
 	// Background execution support
 	CancelFunc    context.CancelFunc // For cancellation of background tasks
 	StreamEvents  []StreamedEvent    // Buffered events for streaming resume
@@ -42,6 +79,14 @@ type StoredResponse struct {
 	IsBackground  bool               // Was created with background=true
 	EventsChan    chan struct{}      // Signals new events for live subscribers
 	mu            sync.RWMutex       // Protect concurrent access to this response
+
+	// streamBytes tracks the total serialized size of the events currently
+	// retained in StreamEvents, used to enforce the byte cap. droppedThrough
+	// is the highest sequence number evicted from the front of the buffer
+	// (-1 = nothing evicted); it is the watermark GetEventsAfter compares
+	// against to detect a lost resume offset. Both are guarded by mu.
+	streamBytes    int
+	droppedThrough int
 }
 
 var getGlobalStore = sync.OnceValue(func() *ResponseStore {
@@ -81,8 +126,10 @@ func (s *ResponseStore) SetTTL(ttl time.Duration) {
 // If ttl is 0, responses are stored indefinitely
 func NewResponseStore(ttl time.Duration) *ResponseStore {
 	store := &ResponseStore{
-		responses: make(map[string]*StoredResponse),
-		ttl:       ttl,
+		responses:       make(map[string]*StoredResponse),
+		ttl:             ttl,
+		maxStreamEvents: defaultMaxStreamEvents,
+		maxStreamBytes:  defaultMaxStreamBytes,
 	}
 
 	// Start cleanup goroutine if TTL is set
@@ -109,11 +156,12 @@ func (s *ResponseStore) Store(responseID string, request *schema.OpenResponsesRe
 	}
 
 	stored := &StoredResponse{
-		Request:   request,
-		Response:  response,
-		Items:     items,
-		StoredAt:  time.Now(),
-		ExpiresAt: nil,
+		Request:        request,
+		Response:       response,
+		Items:          items,
+		StoredAt:       time.Now(),
+		ExpiresAt:      nil,
+		droppedThrough: -1,
 	}
 
 	// Set expiration if TTL is configured
@@ -256,16 +304,17 @@ func (s *ResponseStore) StoreBackground(responseID string, request *schema.OpenR
 	}
 
 	stored := &StoredResponse{
-		Request:       request,
-		Response:      response,
-		Items:         items,
-		StoredAt:      time.Now(),
-		ExpiresAt:     nil,
-		CancelFunc:    cancelFunc,
-		StreamEvents:  []StreamedEvent{},
-		StreamEnabled: streamEnabled,
-		IsBackground:  true,
-		EventsChan:    make(chan struct{}, 100), // Buffered channel for event notifications
+		Request:        request,
+		Response:       response,
+		Items:          items,
+		StoredAt:       time.Now(),
+		ExpiresAt:      nil,
+		CancelFunc:     cancelFunc,
+		StreamEvents:   []StreamedEvent{},
+		StreamEnabled:  streamEnabled,
+		IsBackground:   true,
+		EventsChan:     make(chan struct{}, 100), // Buffered channel for event notifications
+		droppedThrough: -1,
 	}
 
 	// Set expiration if TTL is configured
@@ -349,6 +398,25 @@ func (s *ResponseStore) AppendEvent(responseID string, event *schema.ORStreamEve
 		EventType:      event.Type,
 		Data:           data,
 	})
+	stored.streamBytes += len(data)
+
+	// Evict oldest events from the front once either cap is exceeded. The
+	// byte cap never evicts the only remaining event (a single oversized
+	// event is still served once). Each eviction advances droppedThrough so
+	// a later resume below the watermark is reported as ErrOffsetLost rather
+	// than silently skipping the dropped events.
+	for (s.maxStreamEvents > 0 && len(stored.StreamEvents) > s.maxStreamEvents) ||
+		(s.maxStreamBytes > 0 && stored.streamBytes > s.maxStreamBytes && len(stored.StreamEvents) > 1) {
+		evicted := stored.StreamEvents[0]
+		stored.streamBytes -= len(evicted.Data)
+		if evicted.SequenceNumber > stored.droppedThrough {
+			stored.droppedThrough = evicted.SequenceNumber
+		}
+		// Release the evicted payload so it can be GC'd even though the
+		// backing array element is still owned by the slice until reuse.
+		stored.StreamEvents[0].Data = nil
+		stored.StreamEvents = stored.StreamEvents[1:]
+	}
 	stored.mu.Unlock()
 
 	// Notify any subscribers of new event
@@ -374,6 +442,14 @@ func (s *ResponseStore) GetEventsAfter(responseID string, startingAfter int) ([]
 	stored.mu.RLock()
 	defer stored.mu.RUnlock()
 
+	// If the requested offset is older than the watermark, the events the
+	// client expects next (those in (startingAfter, droppedThrough]) were
+	// evicted by the cap. Signal the gap rather than returning a stream that
+	// silently skips them.
+	if startingAfter < stored.droppedThrough {
+		return nil, ErrOffsetLost
+	}
+
 	var result []StreamedEvent
 	for _, event := range stored.StreamEvents {
 		if event.SequenceNumber > startingAfter {
@@ -447,3 +523,30 @@ func (s *ResponseStore) IsStreamEnabled(responseID string) (bool, error) {
 
 	return stored.StreamEnabled, nil
 }
+
+// SetOwner records the identity that owns a stored response. It is called
+// once, right after the response is stored and before its ID is handed back
+// to any client, so no lock on the stored response is required. A no-op for
+// an empty owner or unknown response ID.
+func (s *ResponseStore) SetOwner(responseID, owner string) {
+	if owner == "" {
+		return
+	}
+
+	s.mu.RLock()
+	stored, exists := s.responses[responseID]
+	s.mu.RUnlock()
+	if !exists {
+		return
+	}
+
+	stored.Owner = owner
+}
+
+// accessAllowed reports whether a caller identified by callerID may read or
+// mutate the given stored response. An empty owner (single-key / no-auth
+// deployments) is accessible by anyone, preserving backward compatibility;
+// otherwise the caller identity must match the recorded owner.
+func accessAllowed(stored *StoredResponse, callerID string) bool {
+	return stored.Owner == "" || stored.Owner == callerID
+}

core/http/endpoints/openresponses/store_test.go

@@ -585,6 +585,86 @@ var _ = Describe("ResponseStore", func() {
 			Expect(enabled2).To(BeFalse())
 		})
 
+		It("should bound the resume buffer and evict oldest events past the cap", func() {
+			// Lower the caps so the test stays fast; production defaults are
+			// large. Same-package access to the unexported fields is fine.
+			store.maxStreamEvents = 5
+			store.maxStreamBytes = 0 // count cap only for this test
+
+			responseID := "resp_buffer_cap"
+			request := &schema.OpenResponsesRequest{Model: "test"}
+			response := &schema.ORResponseResource{
+				ID:     responseID,
+				Object: "response",
+				Status: schema.ORStatusInProgress,
+			}
+
+			_, cancel := context.WithCancel(context.Background())
+			defer cancel()
+
+			store.StoreBackground(responseID, request, response, cancel, true)
+
+			// Append well past the cap.
+			const total = 20
+			for i := range total {
+				err := store.AppendEvent(responseID, &schema.ORStreamEvent{
+					Type:           "response.output_text.delta",
+					SequenceNumber: i,
+				})
+				Expect(err).ToNot(HaveOccurred())
+			}
+
+			stored, err := store.Get(responseID)
+			Expect(err).ToNot(HaveOccurred())
+
+			// (a) Buffer length stays bounded by the cap.
+			Expect(len(stored.StreamEvents)).To(Equal(5))
+
+			// (b) Oldest events were evicted: only the last 5 sequence numbers
+			// remain (15..19).
+			Expect(stored.StreamEvents[0].SequenceNumber).To(Equal(15))
+			Expect(stored.StreamEvents[len(stored.StreamEvents)-1].SequenceNumber).To(Equal(19))
+
+			// Asking for events after the last retained seq still works.
+			retained, err := store.GetEventsAfter(responseID, 14)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(retained).To(HaveLen(5))
+
+			// (c) Asking below the dropped watermark returns ErrOffsetLost.
+			_, err = store.GetEventsAfter(responseID, 0)
+			Expect(err).To(MatchError(ErrOffsetLost))
+
+			_, err = store.GetEventsAfter(responseID, -1)
+			Expect(err).To(MatchError(ErrOffsetLost))
+		})
+
+		It("should record and enforce response ownership", func() {
+			responseID := "resp_owner_test"
+			request := &schema.OpenResponsesRequest{Model: "test"}
+			response := &schema.ORResponseResource{ID: responseID, Object: "response", Status: schema.ORStatusCompleted}
+
+			store.Store(responseID, request, response)
+			store.SetOwner(responseID, "userA")
+
+			stored, err := store.Get(responseID)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(stored.Owner).To(Equal("userA"))
+
+			// Owner matches -> allowed; different identity -> denied.
+			Expect(accessAllowed(stored, "userA")).To(BeTrue())
+			Expect(accessAllowed(stored, "userB")).To(BeFalse())
+
+			// Backward compatibility: a response with no owner is accessible
+			// by any caller (single-key / no-auth deployments).
+			noOwnerID := "resp_no_owner"
+			store.Store(noOwnerID, request, &schema.ORResponseResource{ID: noOwnerID, Object: "response"})
+			noOwner, err := store.Get(noOwnerID)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(noOwner.Owner).To(BeEmpty())
+			Expect(accessAllowed(noOwner, "anyone")).To(BeTrue())
+			Expect(accessAllowed(noOwner, "")).To(BeTrue())
+		})
+
 		It("should notify subscribers of new events", func() {
 			responseID := "resp_events_chan"
 			request := &schema.OpenResponsesRequest{Model: "test"}

core/http/routes/localai.go

@@ -72,19 +72,19 @@ func RegisterLocalAIRoutes(router *echo.Echo,
 		router.POST("/backends/upgrades/check", backendGalleryEndpointService.CheckUpgradesEndpoint(), adminMiddleware)
 		router.POST("/backends/upgrade/:name", backendGalleryEndpointService.UpgradeBackendEndpoint(), adminMiddleware)
 		// Custom model import endpoint
-		router.POST("/models/import", localai.ImportModelEndpoint(cl, appConfig), adminMiddleware)
+		router.POST("/models/import", localai.ImportModelEndpoint(cl, galleryService, appConfig), adminMiddleware)
 
 		// URI model import endpoint
 		router.POST("/models/import-uri", localai.ImportModelURIEndpoint(cl, appConfig, galleryService, opcache), adminMiddleware)
 
 		// Custom model edit endpoint
-		router.POST("/models/edit/:name", localai.EditModelEndpoint(cl, ml, appConfig), adminMiddleware)
+		router.POST("/models/edit/:name", localai.EditModelEndpoint(cl, ml, galleryService, appConfig), adminMiddleware)
 
 		// List model aliases endpoint
 		router.GET("/api/aliases", localai.ListAliasesEndpoint(cl), adminMiddleware)
 
 		// Toggle model enable/disable endpoint
-		router.PUT("/models/toggle-state/:name/:action", localai.ToggleStateModelEndpoint(cl, ml, appConfig), adminMiddleware)
+		router.PUT("/models/toggle-state/:name/:action", localai.ToggleStateModelEndpoint(cl, ml, galleryService, appConfig), adminMiddleware)
 
 		// Toggle model pinned status endpoint
 		router.PUT("/models/toggle-pinned/:name/:action", localai.TogglePinnedModelEndpoint(cl, appConfig, func() {

core/http/routes/nodes.go

@@ -88,7 +88,7 @@ func RegisterNodeAdminRoutes(e *echo.Echo, registry *nodes.NodeRegistry, unloade
 	admin.POST("/:id/approve", localai.ApproveNodeEndpoint(registry, authDB, hmacSecret, natsCfg))
 
 	// Backend management on workers
-	admin.GET("/:id/backends", localai.ListBackendsOnNodeEndpoint(unloader))
+	admin.GET("/:id/backends", localai.ListBackendsOnNodeEndpoint(unloader, registry))
 	admin.POST("/:id/backends/install", localai.InstallBackendOnNodeEndpoint(unloader, galleryService, opcache, appConfig))
 	admin.POST("/:id/backends/delete", localai.DeleteBackendOnNodeEndpoint(unloader))
 

core/http/routes/ui_api.go

@@ -922,7 +922,7 @@ func RegisterUIAPIRoutes(app *echo.Echo, cl *config.ModelConfigLoader, ml *model
 	app.GET("/api/models/config-metadata/autocomplete/:provider", localai.AutocompleteEndpoint(cl, ml, appConfig), adminMiddleware)
 
 	// PATCH config endpoint - partial update using nested JSON merge
-	app.PATCH("/api/models/config-json/:name", localai.PatchConfigEndpoint(cl, ml, appConfig), adminMiddleware)
+	app.PATCH("/api/models/config-json/:name", localai.PatchConfigEndpoint(cl, ml, galleryService, appConfig), adminMiddleware)
 
 	// VRAM estimation endpoint
 	app.POST("/api/models/vram-estimate", localai.VRAMEstimateEndpoint(cl, appConfig), adminMiddleware)

core/schema/message_test.go

@@ -68,6 +68,32 @@ var _ = Describe("LLM tests", func() {
 			Expect(protoMessages[0].Content).To(Equal("Hello World"))
 		})
 
+		// Regression for mudler/LocalAI#10524: a text part whose inner text is
+		// itself a JSON-array string (mealie sends an ingredient list) must
+		// flatten to that exact string verbatim. ToProto must NOT escape or
+		// restructure it - the C++ backend then treats it as opaque text. This
+		// pins the precise Go-side input that produced the "unsupported
+		// content[].type" gRPC error before the backend stopped re-parsing it.
+		It("flattens a JSON-array-looking text part to the verbatim string (#10524)", func() {
+			ingredients := `["1/4 cup brown sugar, packed","1 pound ground beef"]`
+			messages := Messages{
+				{
+					Role: "user",
+					Content: []any{
+						map[string]any{
+							"type": "text",
+							"text": ingredients,
+						},
+					},
+				},
+			}
+
+			protoMessages := messages.ToProto()
+
+			Expect(protoMessages).To(HaveLen(1))
+			Expect(protoMessages[0].Content).To(Equal(ingredients))
+		})
+
 		It("should convert message with tool_calls", func() {
 			messages := Messages{
 				{

core/services/agentpool/agent_jobs.go

@@ -30,6 +30,8 @@ import (
 	mcpTools "github.com/mudler/LocalAI/core/http/endpoints/mcp"
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/core/services/jobs"
+	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/core/services/syncstate"
 	"github.com/mudler/LocalAI/core/templates"
 	"github.com/mudler/LocalAI/pkg/httpclient"
 	"github.com/mudler/LocalAI/pkg/model"
@@ -43,8 +45,18 @@ type AgentJobService struct {
 	configLoader *config.ModelConfigLoader
 	evaluator    *templates.Evaluator
 
+	// tasks is the cross-replica task store: an in-memory map kept consistent
+	// across replicas via NATS, with read-through to the configured persister
+	// (file in standalone, PostgreSQL in distributed). Unlike jobs - which already
+	// converge via the dispatcher + DB read-through - tasks previously read
+	// in-memory only, so ListTasks went stale on non-originating replicas.
+	tasks *syncstate.SyncedMap[string, schema.Task]
+	// taskNats is the distributed NATS client backing the tasks SyncedMap. It is
+	// not available at construction time, so it is injected via SetTaskSyncNATS
+	// during distributed wiring; nil keeps tasks in-memory-only (standalone).
+	taskNats messaging.MessagingClient
+
 	// Storage (in-memory primary, persister for secondary persistence)
-	tasks     *xsync.SyncedMap[string, schema.Task]
 	jobs      *xsync.SyncedMap[string, schema.Job]
 	persister JobPersister
 	userID    string // Scoping: empty for global (main service), set for per-user instances
@@ -96,6 +108,31 @@ func (s *AgentJobService) SetDistributedJobStore(store *jobs.JobStore) {
 	s.persister = &dbJobPersister{store: store}
 }
 
+// SetTaskSyncNATS wires the distributed NATS client used to keep agent *tasks*
+// consistent across replicas (jobs already converge via the dispatcher + DB
+// read-through, so they are left untouched). The client is not available when the
+// service is constructed, so it is injected here during distributed wiring and the
+// tasks SyncedMap is rebuilt to pick it up. It is always called before Start /
+// hydrate, while the map is still empty, so rebuilding loses no state. Passing nil
+// (standalone) keeps the map in-memory-only with no broadcast.
+func (s *AgentJobService) SetTaskSyncNATS(nats messaging.MessagingClient) {
+	s.taskNats = nats
+	s.buildTasksMap()
+}
+
+// buildTasksMap (re)constructs the cross-replica tasks SyncedMap from the current
+// taskNats. The Store adapter reads s.persister/s.userID live, so a persister swap
+// (SetDistributedJobStore) needs no rebuild; only the NATS client, fixed at
+// New-time, forces one - hence SetTaskSyncNATS calls this.
+func (s *AgentJobService) buildTasksMap() {
+	s.tasks = syncstate.New(syncstate.Config[string, schema.Task]{
+		Name:  "agent.tasks",
+		Key:   func(t schema.Task) string { return t.ID },
+		Nats:  s.taskNats,
+		Store: &taskStoreAdapter{svc: s},
+	})
+}
+
 // Dispatcher returns the distributed dispatcher (nil if not in distributed mode).
 func (s *AgentJobService) Dispatcher() DistributedDispatcher {
 	return s.dispatcher
@@ -106,13 +143,6 @@ func (s *AgentJobService) DBStore() *jobs.JobStore {
 	return s.rawDBStore
 }
 
-// saveTasks persists tasks via the configured persister (file or DB).
-func (s *AgentJobService) saveTasks(task schema.Task) {
-	if err := s.persister.SaveTask(s.userID, task); err != nil {
-		xlog.Warn("Failed to persist task", "error", err, "task_id", task.ID)
-	}
-}
-
 // saveJobs persists jobs via the configured persister (file or DB).
 func (s *AgentJobService) saveJobs(job schema.Job) {
 	if err := s.persister.SaveJob(s.userID, job); err != nil {
@@ -129,18 +159,8 @@ func (s *AgentJobService) LoadFromDB() {
 
 // loadFromPersister loads tasks and jobs from the configured persister into memory.
 func (s *AgentJobService) loadFromPersister() {
-	if tasks, err := s.persister.LoadTasks(s.userID); err != nil {
+	if err := s.hydrateTasks(s.appConfig.Context); err != nil {
 		xlog.Warn("Failed to load tasks from persister", "error", err)
-	} else {
-		for _, task := range tasks {
-			s.tasks.Set(task.ID, task)
-			if task.Enabled && task.Cron != "" {
-				if err := s.ScheduleCronTask(task); err != nil {
-					xlog.Warn("Failed to schedule cron task on load", "error", err, "task_id", task.ID)
-				}
-			}
-		}
-		xlog.Info("Loaded tasks from persister", "count", len(tasks))
 	}
 
 	if loadedJobs, err := s.persister.LoadJobs(s.userID); err != nil {
@@ -153,6 +173,27 @@ func (s *AgentJobService) loadFromPersister() {
 	}
 }
 
+// hydrateTasks loads tasks into the cross-replica SyncedMap and (re)schedules
+// cron entries for enabled tasks. Hydration goes through the SyncedMap's Store
+// read-through (Start), not Set, so it neither re-persists nor re-broadcasts the
+// loaded tasks. Each service instance hydrates exactly once: the main service via
+// Start -> loadFromPersister, per-user services via LoadFromDB or LoadTasksFromFile.
+func (s *AgentJobService) hydrateTasks(ctx context.Context) error {
+	if err := s.tasks.Start(ctx); err != nil {
+		return err
+	}
+	tasks := s.tasks.List()
+	for _, task := range tasks {
+		if task.Enabled && task.Cron != "" {
+			if err := s.ScheduleCronTask(task); err != nil {
+				xlog.Warn("Failed to schedule cron task on load", "error", err, "task_id", task.ID)
+			}
+		}
+	}
+	xlog.Info("Loaded tasks from persister", "count", len(tasks))
+	return nil
+}
+
 // JobExecution represents a job to be executed
 type JobExecution struct {
 	Job    schema.Job
@@ -200,21 +241,19 @@ func NewAgentJobServiceWithPaths(
 ) *AgentJobService {
 	retentionDays := cmp.Or(appConfig.AgentJobRetentionDays, 30)
 
-	tasks := xsync.NewSyncedMap[string, schema.Task]()
 	jobsMap := xsync.NewSyncedMap[string, schema.Job]()
 
-	return &AgentJobService{
+	s := &AgentJobService{
 		appConfig:    appConfig,
 		modelLoader:  modelLoader,
 		configLoader: configLoader,
 		evaluator:    evaluator,
-		tasks:        tasks,
 		jobs:         jobsMap,
 		persister: &fileJobPersister{
-			tasks:     tasks,
 			jobs:      jobsMap,
 			tasksFile: tasksFile,
 			jobsFile:  jobsFile,
+			taskSet:   make(map[string]schema.Task),
 		},
 		jobQueue:      make(chan JobExecution, 100), // Buffer for 100 jobs
 		cancellations: xsync.NewSyncedMap[string, context.CancelFunc](),
@@ -222,25 +261,17 @@ func NewAgentJobServiceWithPaths(
 		cronEntries:   xsync.NewSyncedMap[string, cron.EntryID](),
 		retentionDays: retentionDays,
 	}
+	// Build the cross-replica tasks map standalone (nil NATS); SetTaskSyncNATS
+	// rebuilds it with the distributed client once that is available, before Start.
+	s.buildTasksMap()
+	return s
 }
 
 // LoadTasksFromFile loads tasks from the persister into the in-memory map
 // and schedules cron entries. Named "FromFile" for backward compat; in DB
 // mode it loads from the database.
 func (s *AgentJobService) LoadTasksFromFile() error {
-	tasks, err := s.persister.LoadTasks(s.userID)
-	if err != nil {
-		return err
-	}
-	for _, task := range tasks {
-		s.tasks.Set(task.ID, task)
-		if task.Enabled && task.Cron != "" {
-			if err := s.ScheduleCronTask(task); err != nil {
-				xlog.Warn("Failed to schedule cron task on load", "error", err, "task_id", task.ID)
-			}
-		}
-	}
-	return nil
+	return s.hydrateTasks(s.appConfig.Context)
 }
 
 // SaveTasksToFile flushes the current tasks map via the persister. File
@@ -293,8 +324,12 @@ func (s *AgentJobService) CreateTask(task schema.Task) (string, error) {
 		task.Enabled = true // Default to enabled
 	}
 
-	// Store task
-	s.tasks.Set(id, task)
+	// Store task: Set updates the in-memory map, write-throughs to the persister
+	// (file or DB), and broadcasts the create to peer replicas. Background ctx
+	// because CreateTask carries no request ctx (mirrors the finetune service).
+	if err := s.tasks.Set(context.Background(), task); err != nil {
+		return "", fmt.Errorf("failed to persist task: %w", err)
+	}
 
 	// Schedule cron if enabled and has cron expression
 	if task.Enabled && task.Cron != "" {
@@ -303,16 +338,15 @@ func (s *AgentJobService) CreateTask(task schema.Task) (string, error) {
 		}
 	}
 
-	s.saveTasks(task)
 	return id, nil
 }
 
 // UpdateTask updates an existing task
 func (s *AgentJobService) UpdateTask(id string, task schema.Task) error {
-	if !s.tasks.Exists(id) {
+	existing, ok := s.tasks.Get(id)
+	if !ok {
 		return fmt.Errorf("%w: %s", ErrTaskNotFound, id)
 	}
-	existing := s.tasks.Get(id)
 
 	// Preserve ID and CreatedAt
 	task.ID = id
@@ -324,8 +358,10 @@ func (s *AgentJobService) UpdateTask(id string, task schema.Task) error {
 		s.UnscheduleCronTask(id)
 	}
 
-	// Store updated task
-	s.tasks.Set(id, task)
+	// Store updated task: write-through + broadcast (see CreateTask).
+	if err := s.tasks.Set(context.Background(), task); err != nil {
+		return fmt.Errorf("failed to persist task: %w", err)
+	}
 
 	// Schedule new cron if enabled and has cron expression
 	if task.Enabled && task.Cron != "" {
@@ -334,24 +370,22 @@ func (s *AgentJobService) UpdateTask(id string, task schema.Task) error {
 		}
 	}
 
-	s.saveTasks(task)
 	return nil
 }
 
 // DeleteTask deletes a task
 func (s *AgentJobService) DeleteTask(id string) error {
-	if !s.tasks.Exists(id) {
+	if _, ok := s.tasks.Get(id); !ok {
 		return fmt.Errorf("%w: %s", ErrTaskNotFound, id)
 	}
 
 	// Unschedule cron
 	s.UnscheduleCronTask(id)
 
-	// Remove from memory
-	s.tasks.Delete(id)
-
-	if err := s.persister.DeleteTask(id); err != nil {
-		xlog.Warn("Failed to delete task from persister", "error", err, "task_id", id)
+	// Delete removes from the in-memory map, deletes from the persister, and
+	// broadcasts the removal to peer replicas.
+	if err := s.tasks.Delete(context.Background(), id); err != nil {
+		xlog.Warn("Failed to delete task from store", "error", err, "task_id", id)
 	}
 
 	return nil
@@ -359,8 +393,8 @@ func (s *AgentJobService) DeleteTask(id string) error {
 
 // GetTask retrieves a task by ID
 func (s *AgentJobService) GetTask(id string) (*schema.Task, error) {
-	task := s.tasks.Get(id)
-	if task.ID == "" {
+	task, ok := s.tasks.Get(id)
+	if !ok {
 		return nil, fmt.Errorf("%w: %s", ErrTaskNotFound, id)
 	}
 	return &task, nil
@@ -368,7 +402,7 @@ func (s *AgentJobService) GetTask(id string) (*schema.Task, error) {
 
 // ListTasks returns all tasks, sorted by creation date (newest first)
 func (s *AgentJobService) ListTasks() []schema.Task {
-	tasks := s.tasks.Values()
+	tasks := s.tasks.List()
 	// Sort by CreatedAt descending (newest first), then by Name for stability
 	slices.SortFunc(tasks, func(a, b schema.Task) int {
 		if a.CreatedAt.Equal(b.CreatedAt) {
@@ -397,8 +431,8 @@ func (s *AgentJobService) buildPrompt(templateStr string, params map[string]stri
 // ExecuteJob creates and queues a job for execution
 // multimedia can be nil for backward compatibility
 func (s *AgentJobService) ExecuteJob(taskID string, params map[string]string, triggeredBy string, multimedia *schema.MultimediaAttachment) (string, error) {
-	task := s.tasks.Get(taskID)
-	if task.ID == "" {
+	task, ok := s.tasks.Get(taskID)
+	if !ok {
 		return "", fmt.Errorf("%w: %s", ErrTaskNotFound, taskID)
 	}
 
@@ -1451,6 +1485,12 @@ func (s *AgentJobService) Stop() error {
 	if s.cronScheduler != nil {
 		s.cronScheduler.Stop()
 	}
+	// Release the tasks SyncedMap subscription / background workers.
+	if s.tasks != nil {
+		if err := s.tasks.Close(); err != nil {
+			xlog.Warn("Error closing tasks sync map", "error", err)
+		}
+	}
 	xlog.Info("AgentJobService stopped")
 	return nil
 }

core/services/agentpool/job_persister_file.go

@@ -14,24 +14,38 @@ import (
 )
 
 // fileJobPersister persists tasks and jobs to JSON files.
-// It holds references to the service's syncmaps and serializes the entire
-// map contents on each save (bulk write). Reads at runtime return nil
-// (the in-memory map is the authoritative source); LoadTasks/LoadJobs
-// are used only at startup to bootstrap the syncmaps.
+//
+// Jobs serialize the service's in-memory jobs syncmap on each save (bulk write).
+// Tasks are kept in this persister's own taskSet map instead: the tasks SyncedMap
+// calls SaveTask/DeleteTask while holding its internal lock (write-through), so
+// reading back the SyncedMap here would re-enter that lock and deadlock. The
+// self-contained taskSet, seeded by LoadTasks, lets a per-task write rewrite the
+// whole bulk file without touching the SyncedMap.
+//
+// Runtime reads (GetJob/ListJobs) return nil (the in-memory state is the
+// authoritative source); LoadTasks/LoadJobs bootstrap state at startup.
 type fileJobPersister struct {
-	tasks     *xsync.SyncedMap[string, schema.Task]
 	jobs      *xsync.SyncedMap[string, schema.Job]
 	tasksFile string
 	jobsFile  string
 	mu        sync.Mutex
+	// taskSet is the persister's own view of all tasks, seeded by LoadTasks and
+	// updated by SaveTask/DeleteTask. The bulk JSON file is rewritten from it.
+	taskSet map[string]schema.Task
 }
 
-func (p *fileJobPersister) SaveTask(_ string, _ schema.Task) error {
-	return p.saveTasksToFile()
+func (p *fileJobPersister) SaveTask(_ string, task schema.Task) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	p.taskSet[task.ID] = task
+	return p.writeTasksLocked()
 }
 
-func (p *fileJobPersister) DeleteTask(_ string) error {
-	return p.saveTasksToFile()
+func (p *fileJobPersister) DeleteTask(taskID string) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	delete(p.taskSet, taskID)
+	return p.writeTasksLocked()
 }
 
 func (p *fileJobPersister) SaveJob(_ string, _ schema.Job) error {
@@ -43,7 +57,9 @@ func (p *fileJobPersister) DeleteJob(_ string) error {
 }
 
 func (p *fileJobPersister) FlushTasks() error {
-	return p.saveTasksToFile()
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.writeTasksLocked()
 }
 
 func (p *fileJobPersister) FlushJobs() error {
@@ -83,6 +99,12 @@ func (p *fileJobPersister) LoadTasks(_ string) ([]schema.Task, error) {
 		return nil, fmt.Errorf("failed to parse tasks file: %w", err)
 	}
 
+	// Seed the in-memory set so subsequent per-task SaveTask/DeleteTask merge into
+	// (rather than overwrite) the persisted tasks when the bulk file is rewritten.
+	for _, t := range tf.Tasks {
+		p.taskSet[t.ID] = t
+	}
+
 	xlog.Info("Loaded tasks from file", "count", len(tf.Tasks))
 	return tf.Tasks, nil
 }
@@ -118,19 +140,20 @@ func (p *fileJobPersister) CleanupOldJobs(_ time.Duration) (int64, error) {
 	return 0, nil // cleanup handled via in-memory filtering
 }
 
-// saveTasksToFile serializes the entire tasks map to the JSON file.
-func (p *fileJobPersister) saveTasksToFile() error {
+// writeTasksLocked serializes the persister's task set to the JSON file. Callers
+// must hold p.mu.
+func (p *fileJobPersister) writeTasksLocked() error {
 	if p.tasksFile == "" {
 		return nil
 	}
 
-	p.mu.Lock()
-	defer p.mu.Unlock()
-
-	tf := schema.TasksFile{
-		Tasks: p.tasks.Values(),
+	tasks := make([]schema.Task, 0, len(p.taskSet))
+	for _, t := range p.taskSet {
+		tasks = append(tasks, t)
 	}
 
+	tf := schema.TasksFile{Tasks: tasks}
+
 	data, err := json.MarshalIndent(tf, "", "  ")
 	if err != nil {
 		return fmt.Errorf("failed to marshal tasks: %w", err)

core/services/agentpool/job_persister_test.go

@@ -20,28 +20,26 @@ var _ = Describe("JobPersister", func() {
 	Context("fileJobPersister", func() {
 		var (
 			p       *fileJobPersister
-			tasks   *xsync.SyncedMap[string, schema.Task]
 			jobsMap *xsync.SyncedMap[string, schema.Job]
 			tmpDir  string
 		)
 
 		BeforeEach(func() {
 			tmpDir = GinkgoT().TempDir()
-			tasks = xsync.NewSyncedMap[string, schema.Task]()
 			jobsMap = xsync.NewSyncedMap[string, schema.Job]()
 			p = &fileJobPersister{
-				tasks:     tasks,
 				jobs:      jobsMap,
 				tasksFile: filepath.Join(tmpDir, "tasks.json"),
 				jobsFile:  filepath.Join(tmpDir, "jobs.json"),
+				// taskSet is the persister's own task view (decoupled from the tasks
+				// SyncedMap to avoid re-entering its lock during write-through).
+				taskSet: make(map[string]schema.Task),
 			}
 		})
 
 		It("SaveTask writes all tasks to file", func() {
-			tasks.Set("t1", schema.Task{ID: "t1", Name: "Task One", Model: "m", Prompt: "p"})
-			tasks.Set("t2", schema.Task{ID: "t2", Name: "Task Two", Model: "m", Prompt: "p"})
-
-			Expect(p.SaveTask("", schema.Task{})).To(Succeed())
+			Expect(p.SaveTask("", schema.Task{ID: "t1", Name: "Task One", Model: "m", Prompt: "p"})).To(Succeed())
+			Expect(p.SaveTask("", schema.Task{ID: "t2", Name: "Task Two", Model: "m", Prompt: "p"})).To(Succeed())
 
 			// Verify file contents
 			data, err := os.ReadFile(p.tasksFile)
@@ -52,11 +50,9 @@ var _ = Describe("JobPersister", func() {
 		})
 
 		It("DeleteTask writes updated tasks to file", func() {
-			tasks.Set("t1", schema.Task{ID: "t1", Name: "Keep"})
-			tasks.Set("t2", schema.Task{ID: "t2", Name: "Delete"})
+			Expect(p.SaveTask("", schema.Task{ID: "t1", Name: "Keep"})).To(Succeed())
+			Expect(p.SaveTask("", schema.Task{ID: "t2", Name: "Delete"})).To(Succeed())
 
-			// Simulate deletion from memory (caller does this before calling persister)
-			tasks.Delete("t2")
 			Expect(p.DeleteTask("t2")).To(Succeed())
 
 			data, err := os.ReadFile(p.tasksFile)

core/services/agentpool/task_sync_test.go

@@ -0,0 +1,152 @@
+package agentpool
+
+// White-box tests (package agentpool) so a spec can build two AgentJobService
+// instances sharing one in-memory bus and assert that agent *tasks* converge
+// across replicas - the bug this migration fixes (ListTasks used to read
+// in-memory only, so a task created on replica A was invisible on replica B).
+// Jobs are deliberately untouched here: they already converge via the dispatcher
+// + DB read-through.
+
+import (
+	"context"
+	"time"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/core/services/syncstate"
+	"github.com/mudler/LocalAI/core/services/testutil"
+	"github.com/mudler/LocalAI/pkg/system"
+)
+
+// newTaskSyncService builds an AgentJobService wired to the given bus and a
+// throwaway data dir (so the file persister has somewhere to write). Model/config
+// loaders are nil because the task sync paths under test never touch them.
+func newTaskSyncService(bus messaging.MessagingClient) *AgentJobService {
+	tmpDir := GinkgoT().TempDir()
+	sysState := &system.SystemState{}
+	sysState.Model.ModelsPath = tmpDir
+	appConfig := config.NewApplicationConfig(
+		config.WithDynamicConfigDir(tmpDir),
+		config.WithContext(context.Background()),
+	)
+	appConfig.SystemState = sysState
+
+	svc := NewAgentJobServiceWithPaths(appConfig, nil, nil, nil,
+		// Distinct per-replica files so the file persister write-through never
+		// crosses replicas: convergence here must be proven via the bus alone.
+		tmpDir+"/tasks.json", tmpDir+"/jobs.json")
+	svc.SetTaskSyncNATS(bus)
+	return svc
+}
+
+var _ = Describe("AgentJobService task cross-replica sync", func() {
+	Describe("two replicas sharing one bus", func() {
+		var (
+			bus  *testutil.FakeBus
+			a, b *AgentJobService
+		)
+
+		BeforeEach(func() {
+			// One shared bus, two replicas: exactly the distributed topology where a
+			// round-robin request may land on a replica that did not originate the
+			// change.
+			bus = testutil.NewFakeBus()
+			a = newTaskSyncService(bus)
+			b = newTaskSyncService(bus)
+			// Start hydrates (empty here) and subscribes both replicas to deltas.
+			Expect(a.Start(context.Background())).To(Succeed())
+			Expect(b.Start(context.Background())).To(Succeed())
+		})
+
+		AfterEach(func() {
+			Expect(a.Stop()).To(Succeed())
+			Expect(b.Stop()).To(Succeed())
+		})
+
+		It("makes a task created on A visible via B's GetTask and ListTasks", func() {
+			id, err := a.CreateTask(schema.Task{Name: "Shared", Model: "m", Prompt: "p"})
+			Expect(err).NotTo(HaveOccurred())
+
+			got, err := b.GetTask(id)
+			Expect(err).NotTo(HaveOccurred(), "B must see a task A just created")
+			Expect(got.Name).To(Equal("Shared"))
+
+			listed := b.ListTasks()
+			Expect(listed).To(HaveLen(1))
+			Expect(listed[0].ID).To(Equal(id))
+		})
+
+		It("propagates a task update from A to B", func() {
+			id, err := a.CreateTask(schema.Task{Name: "Before", Model: "m", Prompt: "p"})
+			Expect(err).NotTo(HaveOccurred())
+
+			Expect(a.UpdateTask(id, schema.Task{Name: "After", Model: "m", Prompt: "p"})).To(Succeed())
+
+			got, err := b.GetTask(id)
+			Expect(err).NotTo(HaveOccurred())
+			Expect(got.Name).To(Equal("After"), "an update on A must be visible on B")
+		})
+
+		It("removes a task from B when it is deleted on A", func() {
+			id, err := a.CreateTask(schema.Task{Name: "Doomed", Model: "m", Prompt: "p"})
+			Expect(err).NotTo(HaveOccurred())
+			_, err = b.GetTask(id)
+			Expect(err).NotTo(HaveOccurred(), "precondition: B must have the task before the delete")
+
+			Expect(a.DeleteTask(id)).To(Succeed())
+
+			_, err = b.GetTask(id)
+			Expect(err).To(HaveOccurred(), "a delete on A must remove the task from B")
+			Expect(b.ListTasks()).To(BeEmpty())
+		})
+
+		It("does not re-broadcast a delta it received (echo-loop guard)", func() {
+			subject := messaging.SubjectSyncStateDelta("agent.tasks")
+
+			_, err := a.CreateTask(schema.Task{Name: "Once", Model: "m", Prompt: "p"})
+			Expect(err).NotTo(HaveOccurred())
+
+			// Exactly one publish: A's create. B applies it without re-publishing,
+			// otherwise this would be 2+ and a real bus would storm.
+			Expect(bus.PublishCount(subject)).To(Equal(1))
+		})
+	})
+
+	Describe("ListTasks ordering and scoping", func() {
+		var svc *AgentJobService
+
+		BeforeEach(func() {
+			svc = newTaskSyncService(testutil.NewFakeBus())
+			Expect(svc.Start(context.Background())).To(Succeed())
+		})
+		AfterEach(func() { Expect(svc.Stop()).To(Succeed()) })
+
+		It("sorts newest-first, breaking ties by name", func() {
+			// CreateTask stamps CreatedAt with time.Now(); space them out so ordering
+			// is deterministic rather than relying on the sub-millisecond gap.
+			oldID, err := svc.CreateTask(schema.Task{Name: "Old", Model: "m", Prompt: "p"})
+			Expect(err).NotTo(HaveOccurred())
+			time.Sleep(5 * time.Millisecond)
+			newID, err := svc.CreateTask(schema.Task{Name: "New", Model: "m", Prompt: "p"})
+			Expect(err).NotTo(HaveOccurred())
+
+			listed := svc.ListTasks()
+			Expect(listed).To(HaveLen(2))
+			Expect(listed[0].ID).To(Equal(newID), "newest first")
+			Expect(listed[1].ID).To(Equal(oldID))
+		})
+	})
+
+	Describe("compile-time adapter contract", func() {
+		It("satisfies syncstate.Store for tasks", func() {
+			// Mirrors the var assertion in task_syncstore.go; keeps the type
+			// referenced from a spec so drift surfaces here too.
+			var _ syncstate.Store[string, schema.Task] = (*taskStoreAdapter)(nil)
+			Expect(&taskStoreAdapter{}).ToNot(BeNil())
+		})
+	})
+})

core/services/agentpool/task_syncstore.go

@@ -0,0 +1,47 @@
+package agentpool
+
+import (
+	"context"
+
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/syncstate"
+)
+
+// taskStoreAdapter bridges the existing JobPersister (file- or DB-backed) to the
+// generic syncstate.Store the tasks SyncedMap consumes. Only tasks are migrated:
+// jobs already converge across replicas via the dispatcher (NATS) plus the DB
+// read-through in ListJobs/GetJob, whereas ListTasks read in-memory only and so
+// went stale on replicas that did not originate the change.
+//
+// The adapter reads svc.persister and svc.userID live (rather than capturing
+// them) because both are configured by setters - SetDistributedJobStore swaps the
+// file persister for the DB one, SetUserID scopes per-user queries - AFTER the
+// service, and thus this adapter, is constructed. Reading them at call time means
+// the SyncedMap never has to be rebuilt when the persister is swapped.
+//
+// The SyncedMap value type is schema.Task: the exact shape ListTasks returns, so
+// reads need no conversion and REST responses are provably unchanged.
+type taskStoreAdapter struct {
+	svc *AgentJobService
+}
+
+// compile-time assertion that the adapter satisfies the component's Store.
+var _ syncstate.Store[string, schema.Task] = (*taskStoreAdapter)(nil)
+
+// List hydrates the map from durable storage on Start/reconnect: the file's task
+// list (standalone) or every task row (DB / distributed).
+func (a *taskStoreAdapter) List(_ context.Context) ([]schema.Task, error) {
+	return a.svc.persister.LoadTasks(a.svc.userID)
+}
+
+// Upsert write-through persists a single task created/updated locally; the
+// SyncedMap then broadcasts the delta to peers.
+func (a *taskStoreAdapter) Upsert(_ context.Context, task schema.Task) error {
+	return a.svc.persister.SaveTask(a.svc.userID, task)
+}
+
+// Delete write-through removes a task locally; the SyncedMap then broadcasts the
+// removal to peers.
+func (a *taskStoreAdapter) Delete(_ context.Context, id string) error {
+	return a.svc.persister.DeleteTask(id)
+}

core/services/agentpool/user_services.go

@@ -7,6 +7,7 @@ import (
 	"github.com/mudler/LocalAGI/webui/collections"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/services/jobs"
+	"github.com/mudler/LocalAI/core/services/messaging"
 	"github.com/mudler/LocalAI/core/templates"
 	"github.com/mudler/LocalAI/pkg/model"
 	"github.com/mudler/xlog"
@@ -28,6 +29,9 @@ type UserServicesManager struct {
 	// Shared distributed backends (set once, inherited by per-user job services)
 	jobDispatcher DistributedDispatcher
 	jobDBStore    *jobs.JobStore
+	// jobNats keeps per-user agent tasks consistent across replicas (nil in
+	// standalone). Inherited by each per-user AgentJobService.
+	jobNats messaging.MessagingClient
 }
 
 // NewUserServicesManager creates a new UserServicesManager.
@@ -162,6 +166,10 @@ func (m *UserServicesManager) GetJobs(userID string) (*AgentJobService, error) {
 	if m.jobDispatcher != nil {
 		svc.SetDistributedBackends(m.jobDispatcher)
 	}
+	// Inherit the NATS client so per-user tasks broadcast across replicas. Must be
+	// set before the hydrate below (LoadFromDB / LoadTasksFromFile) so the tasks
+	// SyncedMap is rebuilt with the client while it is still empty.
+	svc.SetTaskSyncNATS(m.jobNats)
 	if m.jobDBStore != nil {
 		svc.SetDistributedJobStore(m.jobDBStore)
 		// Load tasks/jobs from DB immediately (per-user services skip Start())
@@ -189,6 +197,12 @@ func (m *UserServicesManager) SetJobDBStore(s *jobs.JobStore) {
 	m.jobDBStore = s
 }
 
+// SetJobSyncNATS sets the NATS client used to keep per-user agent tasks consistent
+// across replicas.
+func (m *UserServicesManager) SetJobSyncNATS(nats messaging.MessagingClient) {
+	m.jobNats = nats
+}
+
 // ListAllUserIDs returns all user IDs that have scoped data directories.
 func (m *UserServicesManager) ListAllUserIDs() ([]string, error) {
 	return m.storage.ListUserDirs()

core/services/distributed/finetune.go

@@ -8,6 +8,7 @@ import (
 	"github.com/google/uuid"
 	"github.com/mudler/LocalAI/core/services/advisorylock"
 	"gorm.io/gorm"
+	"gorm.io/gorm/clause"
 )
 
 // FineTuneJobRecord tracks fine-tune jobs in PostgreSQL.
@@ -80,6 +81,34 @@ func (s *FineTuneStore) List(userID string) ([]FineTuneJobRecord, error) {
 	return jobs, q.Find(&jobs).Error
 }
 
+// ListAll returns every fine-tune job across all users. The SyncedMap that backs
+// FineTuneService is a single global map (the REST API filters by user at read
+// time), so hydrate needs the full set rather than the per-user List above.
+func (s *FineTuneStore) ListAll() ([]FineTuneJobRecord, error) {
+	var jobs []FineTuneJobRecord
+	return jobs, s.db.Order("created_at DESC").Find(&jobs).Error
+}
+
+// Upsert idempotently inserts or fully replaces a job row by primary key. The
+// SyncedMap write-through path issues a single Set per mutation regardless of
+// whether the job already exists, so it needs one create-or-update primitive
+// (Create alone fails on a duplicate key, UpdateStatus alone misses new rows and
+// only touches a few columns).
+func (s *FineTuneStore) Upsert(job *FineTuneJobRecord) error {
+	if job.ID == "" {
+		job.ID = uuid.New().String()
+	}
+	now := time.Now()
+	if job.CreatedAt.IsZero() {
+		job.CreatedAt = now
+	}
+	job.UpdatedAt = now
+	return s.db.Clauses(clause.OnConflict{
+		Columns:   []clause.Column{{Name: "id"}},
+		UpdateAll: true,
+	}).Create(job).Error
+}
+
 // UpdateStatus updates the status and message of a fine-tune job.
 func (s *FineTuneStore) UpdateStatus(id, status, message string) error {
 	return s.db.Model(&FineTuneJobRecord{}).Where("id = ?", id).Updates(map[string]any{

core/services/distributed/finetune_suite_test.go

@@ -0,0 +1,13 @@
+package distributed_test
+
+import (
+	"testing"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func TestDistributed(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "Distributed Suite")
+}

core/services/distributed/finetune_test.go

@@ -0,0 +1,61 @@
+package distributed_test
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/testutil"
+)
+
+var _ = Describe("FineTuneStore", func() {
+	var store *distributed.FineTuneStore
+
+	BeforeEach(func() {
+		db := testutil.SetupTestDB()
+		var err error
+		store, err = distributed.NewFineTuneStore(db)
+		Expect(err).ToNot(HaveOccurred())
+	})
+
+	Describe("ListAll", func() {
+		It("returns jobs across all users (unlike per-user List)", func() {
+			Expect(store.Create(&distributed.FineTuneJobRecord{ID: "j1", UserID: "u1", Status: "queued"})).To(Succeed())
+			Expect(store.Create(&distributed.FineTuneJobRecord{ID: "j2", UserID: "u2", Status: "queued"})).To(Succeed())
+
+			all, err := store.ListAll()
+			Expect(err).ToNot(HaveOccurred())
+			Expect(all).To(HaveLen(2))
+
+			perUser, err := store.List("u1")
+			Expect(err).ToNot(HaveOccurred())
+			Expect(perUser).To(HaveLen(1), "List stays per-user")
+		})
+	})
+
+	Describe("Upsert", func() {
+		It("inserts a new row", func() {
+			Expect(store.Upsert(&distributed.FineTuneJobRecord{ID: "up-1", UserID: "u1", Status: "queued"})).To(Succeed())
+
+			got, err := store.Get("up-1")
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got.Status).To(Equal("queued"))
+		})
+
+		It("idempotently updates an existing row on a repeated key", func() {
+			Expect(store.Upsert(&distributed.FineTuneJobRecord{ID: "up-2", UserID: "u1", Status: "queued"})).To(Succeed())
+			// Second Upsert with the same primary key must update, not error on a
+			// duplicate-key violation (this is the SyncedMap write-through contract).
+			Expect(store.Upsert(&distributed.FineTuneJobRecord{ID: "up-2", UserID: "u1", Status: "completed", Message: "done"})).To(Succeed())
+
+			got, err := store.Get("up-2")
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got.Status).To(Equal("completed"))
+			Expect(got.Message).To(Equal("done"))
+
+			all, err := store.ListAll()
+			Expect(err).ToNot(HaveOccurred())
+			Expect(all).To(HaveLen(1), "upsert must not create a duplicate")
+		})
+	})
+})

core/services/distributed/init.go

@@ -11,6 +11,7 @@ import (
 type Stores struct {
 	Gallery  *GalleryStore
 	FineTune *FineTuneStore
+	Quant    *QuantStore
 	Skills   *SkillStore
 }
 
@@ -26,15 +27,21 @@ func InitStores(db *gorm.DB) (*Stores, error) {
 		return nil, fmt.Errorf("fine-tune store: %w", err)
 	}
 
+	quant, err := NewQuantStore(db)
+	if err != nil {
+		return nil, fmt.Errorf("quantization store: %w", err)
+	}
+
 	skills, err := NewSkillStore(db)
 	if err != nil {
 		return nil, fmt.Errorf("skills store: %w", err)
 	}
 
-	xlog.Info("Distributed stores initialized (Gallery, FineTune, Skills)")
+	xlog.Info("Distributed stores initialized (Gallery, FineTune, Quant, Skills)")
 	return &Stores{
 		Gallery:  gallery,
 		FineTune: ft,
+		Quant:    quant,
 		Skills:   skills,
 	}, nil
 }

core/services/distributed/quant.go

@@ -0,0 +1,105 @@
+package distributed
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"github.com/google/uuid"
+	"github.com/mudler/LocalAI/core/services/advisorylock"
+	"gorm.io/gorm"
+	"gorm.io/gorm/clause"
+)
+
+// QuantJobRecord tracks quantization jobs in PostgreSQL. The columns mirror the
+// API shape (schema.QuantizationJob); the structured Config and ExtraOptions are
+// serialized into JSON text columns so a record fully reconstructs the job.
+type QuantJobRecord struct {
+	ID               string    `gorm:"primaryKey;size:36" json:"id"`
+	UserID           string    `gorm:"index;size:36" json:"user_id,omitempty"`
+	Model            string    `gorm:"size:255" json:"model"`
+	Backend          string    `gorm:"size:64" json:"backend"`
+	ModelID          string    `gorm:"size:255" json:"model_id,omitempty"`
+	QuantizationType string    `gorm:"size:32" json:"quantization_type"`
+	Status           string    `gorm:"index;size:32;default:queued" json:"status"` // queued, downloading, converting, quantizing, completed, failed, stopped
+	Message          string    `gorm:"type:text" json:"message,omitempty"`
+	OutputDir        string    `gorm:"size:512" json:"output_dir,omitempty"`
+	OutputFile       string    `gorm:"size:512" json:"output_file,omitempty"`
+	ConfigJSON       string    `gorm:"column:config;type:text" json:"-"`
+	ExtraOptsJSON    string    `gorm:"column:extra_options;type:text" json:"-"`
+	ImportStatus     string    `gorm:"size:32" json:"import_status,omitempty"`
+	ImportMessage    string    `gorm:"type:text" json:"import_message,omitempty"`
+	ImportModelName  string    `gorm:"size:255" json:"import_model_name,omitempty"`
+	CreatedAt        time.Time `json:"created_at"`
+	UpdatedAt        time.Time `json:"updated_at"`
+}
+
+func (QuantJobRecord) TableName() string { return "quantization_jobs" }
+
+// QuantStore manages quantization job state in PostgreSQL.
+type QuantStore struct {
+	db *gorm.DB
+}
+
+// NewQuantStore creates a new QuantStore and auto-migrates.
+// Uses a PostgreSQL advisory lock to prevent concurrent migration races
+// when multiple instances (frontend + workers) start at the same time.
+func NewQuantStore(db *gorm.DB) (*QuantStore, error) {
+	if err := advisorylock.WithLockCtx(context.Background(), db, advisorylock.KeySchemaMigrate, func() error {
+		return db.AutoMigrate(&QuantJobRecord{})
+	}); err != nil {
+		return nil, fmt.Errorf("migrating quantization_jobs: %w", err)
+	}
+	return &QuantStore{db: db}, nil
+}
+
+// Create stores a new quantization job.
+func (s *QuantStore) Create(job *QuantJobRecord) error {
+	if job.ID == "" {
+		job.ID = uuid.New().String()
+	}
+	job.CreatedAt = time.Now()
+	job.UpdatedAt = job.CreatedAt
+	return s.db.Create(job).Error
+}
+
+// Get retrieves a quantization job by ID.
+func (s *QuantStore) Get(id string) (*QuantJobRecord, error) {
+	var job QuantJobRecord
+	if err := s.db.First(&job, "id = ?", id).Error; err != nil {
+		return nil, err
+	}
+	return &job, nil
+}
+
+// ListAll returns every quantization job across all users. The SyncedMap that
+// backs QuantizationService is a single global map (the REST API filters by user
+// at read time), so hydrate needs the full set.
+func (s *QuantStore) ListAll() ([]QuantJobRecord, error) {
+	var jobs []QuantJobRecord
+	return jobs, s.db.Order("created_at DESC").Find(&jobs).Error
+}
+
+// Upsert idempotently inserts or fully replaces a job row by primary key. The
+// SyncedMap write-through path issues a single Set per mutation regardless of
+// whether the job already exists, so it needs one create-or-update primitive
+// (Create alone fails on a duplicate key).
+func (s *QuantStore) Upsert(job *QuantJobRecord) error {
+	if job.ID == "" {
+		job.ID = uuid.New().String()
+	}
+	now := time.Now()
+	if job.CreatedAt.IsZero() {
+		job.CreatedAt = now
+	}
+	job.UpdatedAt = now
+	return s.db.Clauses(clause.OnConflict{
+		Columns:   []clause.Column{{Name: "id"}},
+		UpdateAll: true,
+	}).Create(job).Error
+}
+
+// Delete removes a quantization job.
+func (s *QuantStore) Delete(id string) error {
+	return s.db.Where("id = ?", id).Delete(&QuantJobRecord{}).Error
+}

core/services/distributed/quant_test.go

@@ -0,0 +1,57 @@
+package distributed_test
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/testutil"
+)
+
+var _ = Describe("QuantStore", func() {
+	var store *distributed.QuantStore
+
+	BeforeEach(func() {
+		db := testutil.SetupTestDB()
+		var err error
+		store, err = distributed.NewQuantStore(db)
+		Expect(err).ToNot(HaveOccurred())
+	})
+
+	Describe("ListAll", func() {
+		It("returns jobs across all users", func() {
+			Expect(store.Create(&distributed.QuantJobRecord{ID: "j1", UserID: "u1", Status: "queued"})).To(Succeed())
+			Expect(store.Create(&distributed.QuantJobRecord{ID: "j2", UserID: "u2", Status: "queued"})).To(Succeed())
+
+			all, err := store.ListAll()
+			Expect(err).ToNot(HaveOccurred())
+			Expect(all).To(HaveLen(2))
+		})
+	})
+
+	Describe("Upsert", func() {
+		It("inserts a new row", func() {
+			Expect(store.Upsert(&distributed.QuantJobRecord{ID: "up-1", UserID: "u1", Status: "queued"})).To(Succeed())
+
+			got, err := store.Get("up-1")
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got.Status).To(Equal("queued"))
+		})
+
+		It("idempotently updates an existing row on a repeated key", func() {
+			Expect(store.Upsert(&distributed.QuantJobRecord{ID: "up-2", UserID: "u1", Status: "queued"})).To(Succeed())
+			// Second Upsert with the same primary key must update, not error on a
+			// duplicate-key violation (this is the SyncedMap write-through contract).
+			Expect(store.Upsert(&distributed.QuantJobRecord{ID: "up-2", UserID: "u1", Status: "completed", Message: "done"})).To(Succeed())
+
+			got, err := store.Get("up-2")
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got.Status).To(Equal("completed"))
+			Expect(got.Message).To(Equal("done"))
+
+			all, err := store.ListAll()
+			Expect(err).ToNot(HaveOccurred())
+			Expect(all).To(HaveLen(1), "upsert must not create a duplicate")
+		})
+	})
+})

core/services/finetune/finetune_suite_test.go

@@ -0,0 +1,13 @@
+package finetune
+
+import (
+	"testing"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func TestFinetune(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "Finetune Suite")
+}

core/services/finetune/service.go

@@ -19,6 +19,7 @@ import (
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/core/services/distributed"
 	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/core/services/syncstate"
 	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
 	"github.com/mudler/LocalAI/pkg/model"
 	"github.com/mudler/LocalAI/pkg/utils"
@@ -32,44 +33,63 @@ type FineTuneService struct {
 	modelLoader  *model.ModelLoader
 	configLoader *config.ModelConfigLoader
 
-	mu   sync.Mutex
-	jobs map[string]*schema.FineTuneJob
+	// mu serializes the read-modify-write of job values. The SyncedMap guards its
+	// own map structure, but a job is a pointer mutated in place (e.g. the export
+	// goroutine), so the service still needs a lock to keep those field updates
+	// and the subsequent Set atomic with respect to readers.
+	mu sync.Mutex
 
-	// Distributed mode (nil when not in distributed mode)
-	natsClient    messaging.Publisher
-	fineTuneStore *distributed.FineTuneStore
+	// jobs is the cross-replica job store: an in-memory map kept consistent across
+	// replicas via NATS, optionally read-through to PostgreSQL in distributed mode.
+	jobs *syncstate.SyncedMap[string, *schema.FineTuneJob]
 }
 
-// SetNATSClient sets the NATS client for distributed progress publishing.
-func (s *FineTuneService) SetNATSClient(nc messaging.Publisher) {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	s.natsClient = nc
-}
-
-// SetFineTuneStore sets the PostgreSQL fine-tune store for distributed persistence.
-func (s *FineTuneService) SetFineTuneStore(store *distributed.FineTuneStore) {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	s.fineTuneStore = store
-}
-
-// NewFineTuneService creates a new FineTuneService.
+// NewFineTuneService creates a new FineTuneService. In distributed mode pass the
+// shared NATS client and PostgreSQL store so jobs stay consistent across
+// replicas; pass nil for both in standalone mode, where the disk Loader hydrates
+// the map and there is nothing to broadcast.
 func NewFineTuneService(
 	appConfig *config.ApplicationConfig,
 	modelLoader *model.ModelLoader,
 	configLoader *config.ModelConfigLoader,
+	nats messaging.MessagingClient,
+	store *distributed.FineTuneStore,
 ) *FineTuneService {
 	s := &FineTuneService{
 		appConfig:    appConfig,
 		modelLoader:  modelLoader,
 		configLoader: configLoader,
-		jobs:         make(map[string]*schema.FineTuneJob),
 	}
-	s.loadAllJobs()
+
+	// Only attach a Store interface when a concrete store exists, otherwise the
+	// SyncedMap would see a non-nil interface wrapping a nil pointer and try to
+	// hydrate/write through a nil DB.
+	var syncStore syncstate.Store[string, *schema.FineTuneJob]
+	if store != nil {
+		syncStore = &fineTuneStoreAdapter{store: store}
+	}
+
+	s.jobs = syncstate.New(syncstate.Config[string, *schema.FineTuneJob]{
+		Name:   "finetune.jobs",
+		Key:    func(j *schema.FineTuneJob) string { return j.ID },
+		Nats:   nats,
+		Store:  syncStore,
+		Loader: s.loadJobsFromDisk, // ignored when Store is set (distributed mode)
+	})
+
+	// Hydrate + subscribe. A hydrate failure must not take the server down: log
+	// and continue degraded (standalone), mirroring the OpCache wiring.
+	if err := s.jobs.Start(appConfig.Context); err != nil {
+		xlog.Warn("FineTune SyncedMap start failed; running degraded", "error", err)
+	}
 	return s
 }
 
+// Close releases the SyncedMap subscription and background workers.
+func (s *FineTuneService) Close() error {
+	return s.jobs.Close()
+}
+
 // fineTuneBaseDir returns the base directory for fine-tune job data.
 func (s *FineTuneService) fineTuneBaseDir() string {
 	return filepath.Join(s.appConfig.DataPath, "fine-tune")
@@ -100,15 +120,18 @@ func (s *FineTuneService) saveJobState(job *schema.FineTuneJob) {
 	}
 }
 
-// loadAllJobs scans the fine-tune directory for persisted jobs and loads them.
-func (s *FineTuneService) loadAllJobs() {
+// loadJobsFromDisk scans the fine-tune directory for persisted jobs and returns
+// them. It is the SyncedMap Loader used in standalone mode (no DB); the returned
+// slice hydrates the map on Start.
+func (s *FineTuneService) loadJobsFromDisk(_ context.Context) ([]*schema.FineTuneJob, error) {
 	baseDir := s.fineTuneBaseDir()
 	entries, err := os.ReadDir(baseDir)
 	if err != nil {
-		// Directory doesn't exist yet — that's fine
-		return
+		// Directory doesn't exist yet — that's fine, start empty.
+		return nil, nil
 	}
 
+	var jobs []*schema.FineTuneJob
 	for _, entry := range entries {
 		if !entry.IsDir() {
 			continue
@@ -137,12 +160,13 @@ func (s *FineTuneService) loadAllJobs() {
 			job.ExportMessage = "Server restarted while export was running"
 		}
 
-		s.jobs[job.ID] = &job
+		jobs = append(jobs, &job)
 	}
 
-	if len(s.jobs) > 0 {
-		xlog.Info("Loaded persisted fine-tune jobs", "count", len(s.jobs))
+	if len(jobs) > 0 {
+		xlog.Info("Loaded persisted fine-tune jobs", "count", len(jobs))
 	}
+	return jobs, nil
 }
 
 // StartJob starts a new fine-tuning job.
@@ -236,27 +260,13 @@ func (s *FineTuneService) StartJob(ctx context.Context, userID string, req schem
 		CreatedAt:      time.Now().UTC().Format(time.RFC3339),
 		Config:         &req,
 	}
-	s.jobs[jobID] = job
-	s.saveJobState(job)
-
-	// Persist to PostgreSQL in distributed mode
-	if s.fineTuneStore != nil {
-		configJSON, _ := json.Marshal(req)
-		extraJSON, _ := json.Marshal(req.ExtraOptions)
-		s.fineTuneStore.Create(&distributed.FineTuneJobRecord{
-			ID:             jobID,
-			UserID:         userID,
-			Model:          req.Model,
-			Backend:        backendName,
-			ModelID:        modelID,
-			TrainingType:   req.TrainingType,
-			TrainingMethod: req.TrainingMethod,
-			Status:         "queued",
-			OutputDir:      outputDir,
-			ConfigJSON:     string(configJSON),
-			ExtraOptsJSON:  string(extraJSON),
-		})
+	// Set write-through persists to PostgreSQL (distributed) and broadcasts to
+	// peer replicas; the disk state.json is written separately for restart
+	// recovery / standalone hydrate.
+	if err := s.jobs.Set(ctx, job); err != nil {
+		return nil, fmt.Errorf("failed to persist job: %w", err)
 	}
+	s.saveJobState(job)
 
 	return &schema.FineTuneJobResponse{
 		ID:      jobID,
@@ -270,7 +280,7 @@ func (s *FineTuneService) GetJob(userID, jobID string) (*schema.FineTuneJob, err
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		return nil, fmt.Errorf("job not found: %s", jobID)
 	}
@@ -286,7 +296,7 @@ func (s *FineTuneService) ListJobs(userID string) []*schema.FineTuneJob {
 	defer s.mu.Unlock()
 
 	var result []*schema.FineTuneJob
-	for _, job := range s.jobs {
+	for _, job := range s.jobs.List() {
 		if userID == "" || job.UserID == userID {
 			result = append(result, job)
 		}
@@ -302,7 +312,7 @@ func (s *FineTuneService) ListJobs(userID string) []*schema.FineTuneJob {
 // StopJob stops a running fine-tuning job.
 func (s *FineTuneService) StopJob(ctx context.Context, userID, jobID string, saveCheckpoint bool) error {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return fmt.Errorf("job not found: %s", jobID)
@@ -323,10 +333,10 @@ func (s *FineTuneService) StopJob(ctx context.Context, userID, jobID string, sav
 	s.mu.Lock()
 	job.Status = "stopped"
 	job.Message = "Training stopped by user"
-	s.saveJobState(job)
-	if s.fineTuneStore != nil {
-		s.fineTuneStore.UpdateStatus(jobID, "stopped", "Training stopped by user")
+	if err := s.jobs.Set(ctx, job); err != nil {
+		xlog.Warn("Failed to persist stopped job", "job_id", jobID, "error", err)
 	}
+	s.saveJobState(job)
 	s.mu.Unlock()
 
 	return nil
@@ -335,7 +345,7 @@ func (s *FineTuneService) StopJob(ctx context.Context, userID, jobID string, sav
 // DeleteJob removes a fine-tuning job and its associated data from disk.
 func (s *FineTuneService) DeleteJob(userID, jobID string) error {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return fmt.Errorf("job not found: %s", jobID)
@@ -360,9 +370,10 @@ func (s *FineTuneService) DeleteJob(userID, jobID string) error {
 	}
 
 	exportModelName := job.ExportModelName
-	delete(s.jobs, jobID)
-	if s.fineTuneStore != nil {
-		s.fineTuneStore.Delete(jobID)
+	// Delete write-through removes the DB row (distributed) and broadcasts the
+	// removal to peer replicas. DeleteJob has no ctx, so use Background.
+	if err := s.jobs.Delete(context.Background(), jobID); err != nil {
+		xlog.Warn("Failed to delete job from store", "job_id", jobID, "error", err)
 	}
 	s.mu.Unlock()
 
@@ -398,7 +409,7 @@ func (s *FineTuneService) DeleteJob(userID, jobID string) error {
 // StreamProgress opens a gRPC progress stream and calls the callback for each update.
 func (s *FineTuneService) StreamProgress(ctx context.Context, userID, jobID string, callback func(event *schema.FineTuneProgressEvent)) error {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return fmt.Errorf("job not found: %s", jobID)
@@ -427,7 +438,7 @@ func (s *FineTuneService) StreamProgress(ctx context.Context, userID, jobID stri
 	}, func(update *pb.FineTuneProgressUpdate) {
 		// Update job status and persist
 		s.mu.Lock()
-		if j, ok := s.jobs[jobID]; ok {
+		if j, ok := s.jobs.Get(jobID); ok {
 			// Don't let progress updates overwrite terminal states
 			isTerminal := j.Status == "stopped" || j.Status == "completed" || j.Status == "failed"
 			if !isTerminal {
@@ -436,10 +447,10 @@ func (s *FineTuneService) StreamProgress(ctx context.Context, userID, jobID stri
 			if update.Message != "" {
 				j.Message = update.Message
 			}
-			s.saveJobState(j)
-			if s.fineTuneStore != nil {
-				s.fineTuneStore.UpdateStatus(jobID, j.Status, j.Message)
+			if err := s.jobs.Set(ctx, j); err != nil {
+				xlog.Warn("Failed to persist progress update", "job_id", jobID, "error", err)
 			}
+			s.saveJobState(j)
 		}
 		s.mu.Unlock()
 
@@ -474,7 +485,7 @@ func (s *FineTuneService) StreamProgress(ctx context.Context, userID, jobID stri
 // ListCheckpoints lists checkpoints for a job.
 func (s *FineTuneService) ListCheckpoints(ctx context.Context, userID, jobID string) ([]*pb.CheckpointInfo, error) {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return nil, fmt.Errorf("job not found: %s", jobID)
@@ -520,7 +531,7 @@ func sanitizeModelName(s string) string {
 // ExportModel starts an async model export from a checkpoint and returns the intended model name immediately.
 func (s *FineTuneService) ExportModel(ctx context.Context, userID, jobID string, req schema.ExportRequest) (string, error) {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return "", fmt.Errorf("job not found: %s", jobID)
@@ -572,6 +583,9 @@ func (s *FineTuneService) ExportModel(ctx context.Context, userID, jobID string,
 	job.ExportStatus = "exporting"
 	job.ExportMessage = ""
 	job.ExportModelName = ""
+	if err := s.jobs.Set(ctx, job); err != nil {
+		xlog.Warn("Failed to persist export start", "job_id", jobID, "error", err)
+	}
 	s.saveJobState(job)
 	s.mu.Unlock()
 
@@ -662,24 +676,30 @@ func (s *FineTuneService) ExportModel(ctx context.Context, userID, jobID string,
 
 		xlog.Info("Model exported and registered", "job_id", jobID, "model_name", modelName, "format", req.ExportFormat)
 
+		// Runs after the HTTP request returns, so use Background rather than the
+		// (now likely cancelled) request ctx for the write-through.
 		s.mu.Lock()
 		job.ExportStatus = "completed"
 		job.ExportModelName = modelName
 		job.ExportMessage = ""
-		s.saveJobState(job)
-		if s.fineTuneStore != nil {
-			s.fineTuneStore.UpdateExportStatus(jobID, "completed", "", modelName)
+		if err := s.jobs.Set(context.Background(), job); err != nil {
+			xlog.Warn("Failed to persist export completion", "job_id", jobID, "error", err)
 		}
+		s.saveJobState(job)
 		s.mu.Unlock()
 	}()
 
 	return modelName, nil
 }
 
-// setExportMessage updates the export message and persists the job state.
+// setExportMessage updates the export message and persists the job state. Called
+// from the background export goroutine, so it uses Background for write-through.
 func (s *FineTuneService) setExportMessage(job *schema.FineTuneJob, msg string) {
 	s.mu.Lock()
 	job.ExportMessage = msg
+	if err := s.jobs.Set(context.Background(), job); err != nil {
+		xlog.Warn("Failed to persist export message", "job_id", job.ID, "error", err)
+	}
 	s.saveJobState(job)
 	s.mu.Unlock()
 }
@@ -687,7 +707,7 @@ func (s *FineTuneService) setExportMessage(job *schema.FineTuneJob, msg string)
 // GetExportedModelPath returns the path to the exported model directory and its name.
 func (s *FineTuneService) GetExportedModelPath(userID, jobID string) (string, string, error) {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return "", "", fmt.Errorf("job not found: %s", jobID)
@@ -723,10 +743,10 @@ func (s *FineTuneService) setExportFailed(job *schema.FineTuneJob, message strin
 	s.mu.Lock()
 	job.ExportStatus = "failed"
 	job.ExportMessage = message
-	s.saveJobState(job)
-	if s.fineTuneStore != nil {
-		s.fineTuneStore.UpdateExportStatus(job.ID, "failed", message, "")
+	if err := s.jobs.Set(context.Background(), job); err != nil {
+		xlog.Warn("Failed to persist export failure", "job_id", job.ID, "error", err)
 	}
+	s.saveJobState(job)
 	s.mu.Unlock()
 }
 

core/services/finetune/service_test.go

@@ -0,0 +1,185 @@
+package finetune
+
+// White-box tests (package finetune) so a spec can drive the service's internal
+// SyncedMap the same way StartJob does (via jobs.Set) without standing up a
+// training backend, then assert the cross-replica reads (GetJob/ListJobs) and
+// the adapter conversions that keep REST responses byte-for-byte unchanged.
+
+import (
+	"context"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/testutil"
+)
+
+// newTestService builds a standalone FineTuneService wired to the given bus. The
+// model/config loaders are nil because the read/sync paths under test never touch
+// them; the data dir is a throwaway temp dir so the disk Loader finds nothing.
+func newTestService(bus *testutil.FakeBus) *FineTuneService {
+	appConfig := &config.ApplicationConfig{
+		Context:  context.Background(),
+		DataPath: GinkgoT().TempDir(),
+	}
+	return NewFineTuneService(appConfig, nil, nil, bus, nil)
+}
+
+var _ = Describe("FineTuneService", func() {
+	ctx := context.Background()
+
+	Describe("cross-replica job visibility", func() {
+		var (
+			bus  *testutil.FakeBus
+			a, b *FineTuneService
+		)
+
+		BeforeEach(func() {
+			// One shared bus, two replicas: exactly the distributed topology where
+			// a round-robin request may land on a replica that did not originate
+			// the change.
+			bus = testutil.NewFakeBus()
+			a = newTestService(bus)
+			b = newTestService(bus)
+		})
+
+		AfterEach(func() {
+			Expect(a.Close()).To(Succeed())
+			Expect(b.Close()).To(Succeed())
+		})
+
+		It("makes a job created on A visible via B's GetJob and ListJobs", func() {
+			job := &schema.FineTuneJob{ID: "job-1", UserID: "user-1", Status: "queued", CreatedAt: "2026-06-27T10:00:00Z"}
+			// StartJob persists via jobs.Set; drive that directly to avoid a backend.
+			Expect(a.jobs.Set(ctx, job)).To(Succeed())
+
+			got, err := b.GetJob("user-1", "job-1")
+			Expect(err).ToNot(HaveOccurred(), "B must see a job A just created")
+			Expect(got.Status).To(Equal("queued"))
+
+			listed := b.ListJobs("user-1")
+			Expect(listed).To(HaveLen(1))
+			Expect(listed[0].ID).To(Equal("job-1"))
+		})
+
+		It("removes a job from B when it is deleted on A", func() {
+			job := &schema.FineTuneJob{ID: "job-2", UserID: "user-1", Status: "completed", CreatedAt: "2026-06-27T10:00:00Z"}
+			Expect(a.jobs.Set(ctx, job)).To(Succeed())
+			_, err := b.GetJob("user-1", "job-2")
+			Expect(err).ToNot(HaveOccurred(), "precondition: B must have the job before the delete")
+
+			Expect(a.jobs.Delete(ctx, "job-2")).To(Succeed())
+
+			_, err = b.GetJob("user-1", "job-2")
+			Expect(err).To(HaveOccurred(), "a delete on A must remove the job from B")
+		})
+
+		It("propagates a status update from A to B", func() {
+			job := &schema.FineTuneJob{ID: "job-3", UserID: "user-1", Status: "training", CreatedAt: "2026-06-27T10:00:00Z"}
+			Expect(a.jobs.Set(ctx, job)).To(Succeed())
+
+			updated := &schema.FineTuneJob{ID: "job-3", UserID: "user-1", Status: "completed", CreatedAt: "2026-06-27T10:00:00Z"}
+			Expect(a.jobs.Set(ctx, updated)).To(Succeed())
+
+			got, err := b.GetJob("user-1", "job-3")
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got.Status).To(Equal("completed"))
+		})
+	})
+
+	Describe("ListJobs", func() {
+		var svc *FineTuneService
+
+		BeforeEach(func() {
+			svc = newTestService(testutil.NewFakeBus())
+		})
+		AfterEach(func() { Expect(svc.Close()).To(Succeed()) })
+
+		It("filters by user and sorts newest-first", func() {
+			Expect(svc.jobs.Set(ctx, &schema.FineTuneJob{ID: "old", UserID: "u1", CreatedAt: "2026-06-25T10:00:00Z"})).To(Succeed())
+			Expect(svc.jobs.Set(ctx, &schema.FineTuneJob{ID: "new", UserID: "u1", CreatedAt: "2026-06-27T10:00:00Z"})).To(Succeed())
+			Expect(svc.jobs.Set(ctx, &schema.FineTuneJob{ID: "other", UserID: "u2", CreatedAt: "2026-06-26T10:00:00Z"})).To(Succeed())
+
+			jobs := svc.ListJobs("u1")
+			Expect(jobs).To(HaveLen(2), "only u1's jobs")
+			Expect(jobs[0].ID).To(Equal("new"), "newest first")
+			Expect(jobs[1].ID).To(Equal("old"))
+		})
+
+		It("returns every user's jobs when the userID filter is empty", func() {
+			Expect(svc.jobs.Set(ctx, &schema.FineTuneJob{ID: "a", UserID: "u1", CreatedAt: "2026-06-25T10:00:00Z"})).To(Succeed())
+			Expect(svc.jobs.Set(ctx, &schema.FineTuneJob{ID: "b", UserID: "u2", CreatedAt: "2026-06-26T10:00:00Z"})).To(Succeed())
+
+			Expect(svc.ListJobs("")).To(HaveLen(2))
+		})
+
+		It("rejects GetJob for a job owned by another user", func() {
+			Expect(svc.jobs.Set(ctx, &schema.FineTuneJob{ID: "x", UserID: "owner", CreatedAt: "2026-06-25T10:00:00Z"})).To(Succeed())
+
+			_, err := svc.GetJob("intruder", "x")
+			Expect(err).To(HaveOccurred(), "a different user must not read someone else's job")
+		})
+	})
+
+	Describe("store adapter conversion", func() {
+		// The SyncedMap value type is *schema.FineTuneJob (the exact REST shape).
+		// These specs prove the DB adapter round-trips it losslessly, so hydrate
+		// and write-through in distributed mode keep responses unchanged.
+		It("round-trips a job through jobToRecord/recordToJob preserving the API shape", func() {
+			original := &schema.FineTuneJob{
+				ID:              "rt-1",
+				UserID:          "user-1",
+				Model:           "base-model",
+				Backend:         "trl",
+				ModelID:         "trl-finetune-rt-1",
+				TrainingType:    "lora",
+				TrainingMethod:  "sft",
+				Status:          "completed",
+				Message:         "done",
+				OutputDir:       "/data/fine-tune/rt-1",
+				ExtraOptions:    map[string]string{"hf_token": "secret"},
+				CreatedAt:       "2026-06-27T10:00:00Z",
+				ExportStatus:    "completed",
+				ExportMessage:   "",
+				ExportModelName: "base-model-ft-rt-1",
+				Config:          &schema.FineTuneJobRequest{Model: "base-model", Backend: "trl", DatasetSource: "data.jsonl"},
+			}
+
+			rec := jobToRecord(original)
+			Expect(rec.ID).To(Equal("rt-1"))
+			Expect(rec.ConfigJSON).ToNot(BeEmpty(), "structured config must serialize into the JSON column")
+			Expect(rec.ExtraOptsJSON).ToNot(BeEmpty())
+
+			back := recordToJob(rec)
+			Expect(back.ID).To(Equal(original.ID))
+			Expect(back.UserID).To(Equal(original.UserID))
+			Expect(back.Model).To(Equal(original.Model))
+			Expect(back.Backend).To(Equal(original.Backend))
+			Expect(back.ModelID).To(Equal(original.ModelID))
+			Expect(back.TrainingType).To(Equal(original.TrainingType))
+			Expect(back.TrainingMethod).To(Equal(original.TrainingMethod))
+			Expect(back.Status).To(Equal(original.Status))
+			Expect(back.Message).To(Equal(original.Message))
+			Expect(back.OutputDir).To(Equal(original.OutputDir))
+			Expect(back.ExportStatus).To(Equal(original.ExportStatus))
+			Expect(back.ExportModelName).To(Equal(original.ExportModelName))
+			Expect(back.CreatedAt).To(Equal(original.CreatedAt))
+			Expect(back.ExtraOptions).To(Equal(original.ExtraOptions))
+			Expect(back.Config).ToNot(BeNil())
+			Expect(back.Config.DatasetSource).To(Equal("data.jsonl"))
+		})
+	})
+
+	Describe("compile-time adapter contract", func() {
+		It("satisfies syncstate.Store for *distributed.FineTuneStore", func() {
+			// Guards against drift between the adapter and the component interface;
+			// the var assertion in syncstore.go covers it at build time, this keeps
+			// the type referenced from a spec too.
+			var _ *distributed.FineTuneStore
+			Expect(&fineTuneStoreAdapter{}).ToNot(BeNil())
+		})
+	})
+})

core/services/finetune/syncstore.go

@@ -0,0 +1,114 @@
+package finetune
+
+import (
+	"context"
+	"encoding/json"
+	"time"
+
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/syncstate"
+)
+
+// fineTuneStoreAdapter bridges the distributed PostgreSQL FineTuneStore to the
+// generic syncstate.Store the SyncedMap consumes. It is only wired in distributed
+// mode; standalone leaves Store nil and hydrates from disk via a Loader instead.
+//
+// The SyncedMap value type is *schema.FineTuneJob (the exact shape the REST API
+// returns) so reads need no conversion and the response JSON is provably
+// unchanged. The adapter is the single place that translates between that API
+// shape and the DB FineTuneJobRecord.
+type fineTuneStoreAdapter struct {
+	store *distributed.FineTuneStore
+}
+
+// compile-time assertion that the adapter satisfies the component's Store.
+var _ syncstate.Store[string, *schema.FineTuneJob] = (*fineTuneStoreAdapter)(nil)
+
+func (a *fineTuneStoreAdapter) List(_ context.Context) ([]*schema.FineTuneJob, error) {
+	records, err := a.store.ListAll()
+	if err != nil {
+		return nil, err
+	}
+	jobs := make([]*schema.FineTuneJob, 0, len(records))
+	for i := range records {
+		jobs = append(jobs, recordToJob(&records[i]))
+	}
+	return jobs, nil
+}
+
+func (a *fineTuneStoreAdapter) Upsert(_ context.Context, job *schema.FineTuneJob) error {
+	return a.store.Upsert(jobToRecord(job))
+}
+
+func (a *fineTuneStoreAdapter) Delete(_ context.Context, id string) error {
+	return a.store.Delete(id)
+}
+
+// recordToJob maps a persisted DB record back to the API shape, reconstructing
+// the structured Config / ExtraOptions from their JSON columns.
+func recordToJob(r *distributed.FineTuneJobRecord) *schema.FineTuneJob {
+	job := &schema.FineTuneJob{
+		ID:              r.ID,
+		UserID:          r.UserID,
+		Model:           r.Model,
+		Backend:         r.Backend,
+		ModelID:         r.ModelID,
+		TrainingType:    r.TrainingType,
+		TrainingMethod:  r.TrainingMethod,
+		Status:          r.Status,
+		Message:         r.Message,
+		OutputDir:       r.OutputDir,
+		ExportStatus:    r.ExportStatus,
+		ExportMessage:   r.ExportMessage,
+		ExportModelName: r.ExportModelName,
+		CreatedAt:       r.CreatedAt.UTC().Format(time.RFC3339),
+	}
+	if r.ExtraOptsJSON != "" {
+		// Best-effort: a malformed column must not drop the whole job from the API.
+		_ = json.Unmarshal([]byte(r.ExtraOptsJSON), &job.ExtraOptions)
+	}
+	if r.ConfigJSON != "" {
+		var cfg schema.FineTuneJobRequest
+		if err := json.Unmarshal([]byte(r.ConfigJSON), &cfg); err == nil {
+			job.Config = &cfg
+		}
+	}
+	return job
+}
+
+// jobToRecord maps the API shape to a DB record for write-through, serializing
+// the structured Config / ExtraOptions into their JSON columns. CreatedAt is
+// parsed back from the RFC3339 string the service stamps; an unparseable value
+// is left zero so FineTuneStore.Upsert stamps "now".
+func jobToRecord(job *schema.FineTuneJob) *distributed.FineTuneJobRecord {
+	rec := &distributed.FineTuneJobRecord{
+		ID:              job.ID,
+		UserID:          job.UserID,
+		Model:           job.Model,
+		Backend:         job.Backend,
+		ModelID:         job.ModelID,
+		TrainingType:    job.TrainingType,
+		TrainingMethod:  job.TrainingMethod,
+		Status:          job.Status,
+		Message:         job.Message,
+		OutputDir:       job.OutputDir,
+		ExportStatus:    job.ExportStatus,
+		ExportMessage:   job.ExportMessage,
+		ExportModelName: job.ExportModelName,
+	}
+	if job.Config != nil {
+		if data, err := json.Marshal(job.Config); err == nil {
+			rec.ConfigJSON = string(data)
+		}
+	}
+	if job.ExtraOptions != nil {
+		if data, err := json.Marshal(job.ExtraOptions); err == nil {
+			rec.ExtraOptsJSON = string(data)
+		}
+	}
+	if t, err := time.Parse(time.RFC3339, job.CreatedAt); err == nil {
+		rec.CreatedAt = t
+	}
+	return rec
+}

core/services/galleryop/distributed_sync_test.go

@@ -404,6 +404,36 @@ var _ = Describe("GalleryService cache invalidation broadcasts", func() {
 			Element: "x", Op: "install",
 		})).To(Succeed())
 	})
+
+	It("BroadcastModelsChanged delivers the element and op to a peer's OnModelsChanged", func() {
+		var (
+			mu   sync.Mutex
+			seen []messaging.CacheInvalidateEvent
+		)
+		svcB.OnModelsChanged = func(evt messaging.CacheInvalidateEvent) {
+			mu.Lock()
+			seen = append(seen, evt)
+			mu.Unlock()
+		}
+		Expect(svcA.SubscribeBroadcasts()).To(Succeed())
+		Expect(svcB.SubscribeBroadcasts()).To(Succeed())
+
+		// An admin edit on replica A must reach replica B over the same subject
+		// the gallery path uses, so B refreshes its in-memory config loader.
+		svcA.BroadcastModelsChanged("my-alias", "install")
+
+		mu.Lock()
+		defer mu.Unlock()
+		Expect(seen).To(ContainElement(messaging.CacheInvalidateEvent{
+			Element: "my-alias", Op: "install",
+		}))
+	})
+
+	It("BroadcastModelsChanged is a no-op when NATS is not wired (standalone)", func() {
+		standalone := galleryop.NewGalleryService(&config.ApplicationConfig{}, nil)
+		// No SetNATSClient: must not panic and must simply do nothing.
+		Expect(func() { standalone.BroadcastModelsChanged("x", "delete") }).ToNot(Panic())
+	})
 })
 
 var _ = Describe("GalleryService PostgreSQL hydration", func() {

core/services/galleryop/service.go

@@ -201,6 +201,24 @@ func (g *GalleryService) publishCacheInvalidate(subject string, evt messaging.Ca
 	}
 }
 
+// BroadcastModelsChanged notifies peer replicas that a model config was
+// created, edited, or removed out-of-band of the gallery install/delete
+// channel (e.g. the admin /models/edit, /models/import and
+// /models/toggle-state endpoints, which write the YAML and reload only the
+// local in-memory loader). Peers receive it via OnModelsChanged and refresh
+// their own ModelConfigLoader so a request load-balanced to any replica sees
+// the same config. No-op in standalone mode (no NATS client).
+//
+// op is "install" for a create/edit (the element must be (re)loaded from
+// disk) or "delete" for a removal (the element must be pruned from memory,
+// which a reload-from-path cannot do because the loader is additive).
+func (g *GalleryService) BroadcastModelsChanged(element, op string) {
+	g.publishCacheInvalidate(messaging.SubjectCacheInvalidateModels, messaging.CacheInvalidateEvent{
+		Element: element,
+		Op:      op,
+	})
+}
+
 // mergeStatus is the broadcast-side merge: it updates the in-memory map from
 // a peer's GalleryProgressEvent without re-publishing to NATS or re-writing
 // to PostgreSQL. UpdateStatus is the local-write entry point and does both;

core/services/messaging/client.go

@@ -22,6 +22,14 @@ const subscribeConfirmTimeout = 5 * time.Second
 type Client struct {
 	conn *nats.Conn
 	mu   sync.RWMutex
+
+	// reconnectCbs are invoked after the underlying connection is
+	// re-established. nats.go transparently resubscribes existing
+	// subscriptions on reconnect, but it cannot know that a consumer kept
+	// derived in-memory state (e.g. syncstate.SyncedMap) that may have drifted
+	// while the link was down — these callbacks let such consumers re-hydrate.
+	cbMu         sync.Mutex
+	reconnectCbs []func()
 }
 
 // New creates a new NATS client with auto-reconnect.
@@ -31,6 +39,10 @@ func New(url string, opts ...Option) (*Client, error) {
 		o(&cfg)
 	}
 
+	// Allocate the client up front so the reconnect handler closure can reach
+	// it; conn is populated after nats.Connect succeeds below.
+	c := &Client{}
+
 	natsOpts := []nats.Option{
 		nats.RetryOnFailedConnect(true),
 		nats.MaxReconnects(-1),
@@ -41,6 +53,7 @@ func New(url string, opts ...Option) (*Client, error) {
 		}),
 		nats.ReconnectHandler(func(_ *nats.Conn) {
 			xlog.Info("NATS reconnected")
+			c.runReconnectCallbacks()
 		}),
 		nats.ClosedHandler(func(_ *nats.Conn) {
 			xlog.Info("NATS connection closed")
@@ -103,7 +116,33 @@ func New(url string, opts ...Option) (*Client, error) {
 		return nil, fmt.Errorf("connecting to NATS at %s: %w", sanitize.URL(url), err)
 	}
 
-	return &Client{conn: nc}, nil
+	c.conn = nc
+	return c, nil
+}
+
+// OnReconnect registers a callback invoked after the NATS connection is
+// re-established. It is consumed via an optional interface type-assertion
+// (interface{ OnReconnect(func()) }) rather than being added to MessagingClient,
+// so the messaging abstraction stays minimal and standalone/test clients are not
+// forced to implement reconnect semantics. A nil callback is ignored.
+func (c *Client) OnReconnect(cb func()) {
+	if cb == nil {
+		return
+	}
+	c.cbMu.Lock()
+	c.reconnectCbs = append(c.reconnectCbs, cb)
+	c.cbMu.Unlock()
+}
+
+// runReconnectCallbacks invokes registered reconnect callbacks. It copies the
+// slice under the lock so a callback that (re)registers cannot deadlock.
+func (c *Client) runReconnectCallbacks() {
+	c.cbMu.Lock()
+	cbs := append([]func(){}, c.reconnectCbs...)
+	c.cbMu.Unlock()
+	for _, cb := range cbs {
+		cb()
+	}
 }
 
 // Publish marshals data as JSON and publishes it to the given subject.

core/services/messaging/subjects.go

@@ -380,6 +380,20 @@ func SubjectCacheInvalidateCollection(name string) string {
 	return "cache.invalidate.collections." + sanitizeSubjectToken(name)
 }
 
+// SyncedMap State Sync (Pub/Sub — broadcast to all frontends)
+//
+// The reusable syncstate.SyncedMap component publishes a {op,key,value} delta on
+// this subject whenever a replica mutates a piece of cross-replica in-memory
+// state. Peers subscribe and apply the delta to their own map, so a round-robin
+// API request that lands on a replica which did not originate the change still
+// sees it. Convergence on (re)connect is done by re-hydrating from the durable
+// source, so no request/reply snapshot subject is needed here.
+func SubjectSyncStateDelta(name string) string {
+	return subjectSyncStatePrefix + sanitizeSubjectToken(name) + ".delta"
+}
+
+const subjectSyncStatePrefix = "state."
+
 // Prefix-Cache Routing Sync (Pub/Sub - broadcast to all frontends)
 //
 // Frontends share prefix-cache observations so a request routed to any replica

core/services/modeladmin/remote_sync.go

@@ -0,0 +1,53 @@
+package modeladmin
+
+import (
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/pkg/model"
+
+	"github.com/mudler/xlog"
+)
+
+// opDelete is the CacheInvalidateEvent.Op value the gallery delete path and the
+// admin delete endpoint use; a delete must prune (a reload-from-path cannot).
+const opDelete = "delete"
+
+// ApplyRemoteChange refreshes this replica's in-memory model state from a peer
+// replica's model-config change broadcast (messaging.CacheInvalidateEvent on
+// SubjectCacheInvalidateModels). It is the subscriber-side counterpart to
+// GalleryService.BroadcastModelsChanged.
+//
+// The op matters because LoadModelConfigsFromPath is additive: it loads every
+// YAML on disk into the loader but never removes an entry whose file is gone.
+// So a delete cannot be propagated by a plain reload - the deleted element must
+// be explicitly pruned. Specifically:
+//
+//   - op == "delete" with a named element: prune that element from the loader.
+//   - otherwise: reload all configs from disk (picks up creates and edits).
+//
+// In both cases, when an element is named, any running instance on this replica
+// is shut down (best-effort) so the next request rebuilds it from the new
+// config instead of serving the stale one - mirroring what the originating
+// replica does on a local edit/delete.
+//
+// ml may be nil (no running instances to shut down). modelsPath and opts are
+// forwarded to LoadModelConfigsFromPath.
+func ApplyRemoteChange(cl *config.ModelConfigLoader, ml *model.ModelLoader, modelsPath string, evt messaging.CacheInvalidateEvent, opts ...config.ConfigLoaderOption) error {
+	if evt.Op == opDelete && evt.Element != "" {
+		cl.RemoveModelConfig(evt.Element)
+	} else if err := cl.LoadModelConfigsFromPath(modelsPath, opts...); err != nil {
+		return err
+	}
+
+	// Drop any running instance of the affected model so the next request
+	// rebuilds it from the refreshed config instead of serving the stale one.
+	// Best-effort: the model may not be loaded on this replica, which surfaces
+	// as a benign error here.
+	if ml != nil && evt.Element != "" {
+		if err := ml.ShutdownModel(evt.Element); err != nil {
+			xlog.Debug("ApplyRemoteChange: could not shut down model instance (likely not loaded)",
+				"model", evt.Element, "error", err)
+		}
+	}
+	return nil
+}

core/services/modeladmin/remote_sync_test.go

@@ -0,0 +1,80 @@
+package modeladmin
+
+import (
+	"os"
+	"path/filepath"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+	"gopkg.in/yaml.v3"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/services/messaging"
+)
+
+var _ = Describe("ApplyRemoteChange", func() {
+	var (
+		dir    string
+		loader *config.ModelConfigLoader
+	)
+
+	BeforeEach(func() {
+		dir = GinkgoT().TempDir()
+		loader = config.NewModelConfigLoader(dir)
+	})
+
+	writeYAML := func(name string, body map[string]any) {
+		body["name"] = name
+		data, err := yaml.Marshal(body)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(os.WriteFile(filepath.Join(dir, name+".yaml"), data, 0644)).To(Succeed())
+	}
+
+	It("loads a peer-created config from disk on an install event", func() {
+		// Peer wrote the YAML to the shared models dir; this replica has not
+		// loaded it yet (empty in-memory loader).
+		writeYAML("peer-alias", map[string]any{"alias": "qwen"})
+		_, ok := loader.GetModelConfig("peer-alias")
+		Expect(ok).To(BeFalse(), "precondition: not yet in memory")
+
+		err := ApplyRemoteChange(loader, nil, dir, messaging.CacheInvalidateEvent{
+			Element: "peer-alias", Op: "install",
+		})
+		Expect(err).ToNot(HaveOccurred())
+
+		_, ok = loader.GetModelConfig("peer-alias")
+		Expect(ok).To(BeTrue(), "install event must reload the new config from disk")
+	})
+
+	It("prunes a peer-deleted config that a reload-from-path cannot drop", func() {
+		// Model is present in memory (loaded earlier) but its file is now gone
+		// from the shared dir. LoadModelConfigsFromPath is additive, so only an
+		// explicit prune can remove it - this is the cross-replica delete bug.
+		writeYAML("doomed", map[string]any{"alias": "qwen"})
+		Expect(loader.LoadModelConfigsFromPath(dir)).To(Succeed())
+		_, ok := loader.GetModelConfig("doomed")
+		Expect(ok).To(BeTrue(), "precondition: in memory")
+		Expect(os.Remove(filepath.Join(dir, "doomed.yaml"))).To(Succeed())
+
+		err := ApplyRemoteChange(loader, nil, dir, messaging.CacheInvalidateEvent{
+			Element: "doomed", Op: "delete",
+		})
+		Expect(err).ToNot(HaveOccurred())
+
+		_, ok = loader.GetModelConfig("doomed")
+		Expect(ok).To(BeFalse(), "delete event must prune the element from memory")
+	})
+
+	It("does a full reload when no element is named", func() {
+		writeYAML("m1", map[string]any{"alias": "qwen"})
+		writeYAML("m2", map[string]any{"alias": "qwen"})
+
+		err := ApplyRemoteChange(loader, nil, dir, messaging.CacheInvalidateEvent{})
+		Expect(err).ToNot(HaveOccurred())
+
+		_, ok1 := loader.GetModelConfig("m1")
+		_, ok2 := loader.GetModelConfig("m2")
+		Expect(ok1).To(BeTrue())
+		Expect(ok2).To(BeTrue())
+	})
+})

core/services/nodes/registry.go

@@ -673,6 +673,49 @@ func (r *NodeRegistry) Get(ctx context.Context, nodeID string) (*BackendNode, er
 	return &node, nil
 }
 
+// GetWithExtras returns a single node enriched with the same computed fields as
+// ListWithExtras (labels, loaded-model count, in-flight total). The plain Get
+// returns a bare BackendNode whose Labels live in a separate table, so the node
+// detail view needs this to show a node's existing labels and live counts.
+func (r *NodeRegistry) GetWithExtras(ctx context.Context, nodeID string) (*NodeWithExtras, error) {
+	node, err := r.Get(ctx, nodeID)
+	if err != nil {
+		return nil, err
+	}
+
+	labels := make(map[string]string)
+	nodeLabels, err := r.GetNodeLabels(ctx, nodeID)
+	if err != nil {
+		xlog.Warn("GetWithExtras: failed to get labels", "node", nodeID, "error", err)
+	} else {
+		for _, l := range nodeLabels {
+			labels[l.Key] = l.Value
+		}
+	}
+
+	var modelCount int64
+	if err := r.db.WithContext(ctx).Model(&NodeModel{}).
+		Where("node_id = ? AND state = ?", nodeID, "loaded").
+		Count(&modelCount).Error; err != nil {
+		xlog.Warn("GetWithExtras: failed to get model count", "node", nodeID, "error", err)
+	}
+
+	var inFlight struct{ Total int }
+	if err := r.db.WithContext(ctx).Model(&NodeModel{}).
+		Select("COALESCE(SUM(in_flight), 0) as total").
+		Where("node_id = ? AND state IN ?", nodeID, []string{"loaded", "unloading"}).
+		Scan(&inFlight).Error; err != nil {
+		xlog.Warn("GetWithExtras: failed to get in-flight count", "node", nodeID, "error", err)
+	}
+
+	return &NodeWithExtras{
+		BackendNode:   *node,
+		ModelCount:    int(modelCount),
+		InFlightCount: inFlight.Total,
+		Labels:        labels,
+	}, nil
+}
+
 // GetByName returns a single node by name.
 func (r *NodeRegistry) GetByName(ctx context.Context, name string) (*BackendNode, error) {
 	var node BackendNode

core/services/nodes/registry_test.go

@@ -646,6 +646,38 @@ var _ = Describe("NodeRegistry", func() {
 		})
 	})
 
+	Describe("GetWithExtras", func() {
+		It("returns the node enriched with its labels map", func() {
+			node := makeNode("extras-node", "10.0.0.80:50051", 8_000_000_000)
+			Expect(registry.Register(context.Background(), node, true)).To(Succeed())
+			Expect(registry.SetNodeLabel(context.Background(), node.ID, "env", "prod")).To(Succeed())
+			Expect(registry.SetNodeLabel(context.Background(), node.ID, "region", "us-east")).To(Succeed())
+
+			got, err := registry.GetWithExtras(context.Background(), node.ID)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got).ToNot(BeNil())
+			Expect(got.ID).To(Equal(node.ID))
+			Expect(got.Name).To(Equal("extras-node"))
+			Expect(got.Labels).To(Equal(map[string]string{"env": "prod", "region": "us-east"}))
+		})
+
+		It("returns an empty (non-nil) labels map when the node has none", func() {
+			node := makeNode("extras-no-labels", "10.0.0.81:50051", 8_000_000_000)
+			Expect(registry.Register(context.Background(), node, true)).To(Succeed())
+
+			got, err := registry.GetWithExtras(context.Background(), node.ID)
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got).ToNot(BeNil())
+			Expect(got.Labels).ToNot(BeNil())
+			Expect(got.Labels).To(BeEmpty())
+		})
+
+		It("returns an error for an unknown node", func() {
+			_, err := registry.GetWithExtras(context.Background(), "does-not-exist")
+			Expect(err).To(HaveOccurred())
+		})
+	})
+
 	Describe("FindNodesBySelector", func() {
 		It("returns nodes matching all labels in selector", func() {
 			n1 := makeNode("sel-match", "10.0.0.80:50051", 8_000_000_000)

core/services/nodes/router.go

@@ -63,6 +63,11 @@ type SmartRouterOptions struct {
 	// The reconciler reads the same instance to autoscale a saturated cache-warm
 	// replica. nil disables recording (the disabled path stays a no-op).
 	Pressure *prefixcache.Pressure
+	// SharedModels asserts that every node mounts the same models directory at
+	// the same path. When true, stageModelFiles skips all uploading and leaves
+	// the absolute model paths untouched so the worker loads them directly from
+	// the shared volume (#10556). See config.DistributedConfig.SharedModels.
+	SharedModels bool
 }
 
 // SmartRouter routes inference requests to the best available backend node.
@@ -93,6 +98,9 @@ type SmartRouter struct {
 	// per-request routing doesn't stall behind a busy backend's serialized
 	// HealthCheck/Predict. See probe_cache.go for the rationale.
 	probeCache *probeCache
+	// sharedModels skips file staging when all nodes mount the same models
+	// directory at the same path (see SmartRouterOptions.SharedModels).
+	sharedModels bool
 }
 
 // probeCacheTTL is how long a successful gRPC HealthCheck on a backend is
@@ -122,6 +130,7 @@ func NewSmartRouter(registry ModelRouter, opts SmartRouterOptions) *SmartRouter
 		prefixProvider:   opts.PrefixProvider,
 		prefixConfig:     opts.PrefixConfig,
 		pressure:         opts.Pressure,
+		sharedModels:     opts.SharedModels,
 	}
 }
 
@@ -947,6 +956,19 @@ func (r *SmartRouter) buildClientForAddr(node *BackendNode, addr string, paralle
 // simply remove the {ModelsPath}/{trackingKey}/ directory.
 func (r *SmartRouter) stageModelFiles(ctx context.Context, node *BackendNode, opts *pb.ModelOptions, trackingKey string) (*pb.ModelOptions, error) {
 	opts = proto.Clone(opts).(*pb.ModelOptions)
+
+	// Shared-models mode: every node mounts the same models directory at the
+	// same path, so the frontend's absolute model paths are already valid on the
+	// worker. Staging would only re-upload files that already exist on the shared
+	// volume (under a tracking-key subdir the probe never reuses), re-downloading
+	// the model on every load (#10556). Return the clone untouched: no upload, no
+	// path rewrite, no staging tracker.
+	if r.sharedModels {
+		xlog.Info("Skipping model file staging: shared-models mode is on (LOCALAI_DISTRIBUTED_SHARED_MODELS); worker loads directly from the shared volume",
+			"node", node.Name, "modelFile", opts.ModelFile, "trackingKey", trackingKey)
+		return opts, nil
+	}
+
 	xlog.Info("Staging model files for remote node", "node", node.Name, "modelFile", opts.ModelFile, "trackingKey", trackingKey)
 
 	// Derive the frontend models directory from ModelFile and Model.

core/services/nodes/router_sharedmodels_test.go

@@ -0,0 +1,85 @@
+package nodes
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+// These tests cover shared-models mode (LOCALAI_DISTRIBUTED_SHARED_MODELS): when
+// every node mounts the same models directory at the same path, the router must
+// NOT stage model files to workers. The canonical absolute path is already valid
+// on the worker, so staging would only re-download what is already present
+// (#10556).
+var _ = Describe("stageModelFiles shared-models mode", func() {
+	var (
+		stager  *fakeFileStager
+		node    *BackendNode
+		tmp     string
+		gguf    string
+		modelID = "ornith-1.0-35b"
+	)
+
+	BeforeEach(func() {
+		stager = &fakeFileStager{}
+		node = &BackendNode{ID: "node-1", Name: "node-1", Address: "10.0.0.1:50051"}
+		tmp = GinkgoT().TempDir()
+
+		modelDir := filepath.Join(tmp, "models", "llama-cpp", "models")
+		Expect(os.MkdirAll(modelDir, 0o755)).To(Succeed())
+		gguf = filepath.Join(modelDir, "ornith.gguf")
+		Expect(os.WriteFile(gguf, []byte("weights"), 0o644)).To(Succeed())
+	})
+
+	It("does not stage and keeps the canonical absolute ModelFile when shared-models is enabled", func() {
+		router := &SmartRouter{
+			fileStager:     stager,
+			stagingTracker: NewStagingTracker(),
+			sharedModels:   true,
+		}
+
+		opts := &pb.ModelOptions{
+			Model:     "llama-cpp/models/ornith.gguf",
+			ModelFile: gguf,
+		}
+
+		staged, err := router.stageModelFiles(context.Background(), node, opts, modelID)
+		Expect(err).ToNot(HaveOccurred())
+
+		// The file stager must never be touched: no upload, no re-download.
+		Expect(stager.ensureCalls).To(BeEmpty())
+		// The worker loads directly from the shared volume, so the path is unchanged.
+		Expect(staged.ModelFile).To(Equal(gguf))
+	})
+
+	It("stages files (existing behavior) when shared-models is disabled", func() {
+		router := &SmartRouter{
+			fileStager:     stager,
+			stagingTracker: NewStagingTracker(),
+			sharedModels:   false,
+		}
+
+		opts := &pb.ModelOptions{
+			Model:     "llama-cpp/models/ornith.gguf",
+			ModelFile: gguf,
+		}
+
+		staged, err := router.stageModelFiles(context.Background(), node, opts, modelID)
+		Expect(err).ToNot(HaveOccurred())
+
+		// Default mode uploads the model file to the worker.
+		Expect(stager.ensureCalls).ToNot(BeEmpty())
+		stagedLocals := make([]string, 0, len(stager.ensureCalls))
+		for _, c := range stager.ensureCalls {
+			stagedLocals = append(stagedLocals, c.localPath)
+		}
+		Expect(stagedLocals).To(ContainElement(gguf))
+		// ModelFile is rewritten to the remote (tracking-key namespaced) path.
+		Expect(staged.ModelFile).ToNot(Equal(gguf))
+	})
+})

core/services/quantization/quantization_suite_test.go

@@ -0,0 +1,13 @@
+package quantization
+
+import (
+	"testing"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func TestQuantization(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "Quantization Suite")
+}

core/services/quantization/service.go

@@ -17,6 +17,9 @@ import (
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/gallery/importers"
 	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/core/services/syncstate"
 	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
 	"github.com/mudler/LocalAI/pkg/model"
 	"github.com/mudler/LocalAI/pkg/utils"
@@ -30,26 +33,63 @@ type QuantizationService struct {
 	modelLoader  *model.ModelLoader
 	configLoader *config.ModelConfigLoader
 
-	mu   sync.Mutex
-	jobs map[string]*schema.QuantizationJob
+	// mu serializes the read-modify-write of job values. The SyncedMap guards its
+	// own map structure, but a job is a pointer mutated in place (e.g. the import
+	// goroutine), so the service still needs a lock to keep those field updates and
+	// the subsequent Set atomic with respect to readers.
+	mu sync.Mutex
+
+	// jobs is the cross-replica job store: an in-memory map kept consistent across
+	// replicas via NATS, optionally read-through to PostgreSQL in distributed mode.
+	jobs *syncstate.SyncedMap[string, *schema.QuantizationJob]
 }
 
-// NewQuantizationService creates a new QuantizationService.
+// NewQuantizationService creates a new QuantizationService. In distributed mode
+// pass the shared NATS client and PostgreSQL store so jobs stay consistent across
+// replicas; pass nil for both in standalone mode, where the disk Loader hydrates
+// the map and there is nothing to broadcast.
 func NewQuantizationService(
 	appConfig *config.ApplicationConfig,
 	modelLoader *model.ModelLoader,
 	configLoader *config.ModelConfigLoader,
+	nats messaging.MessagingClient,
+	store *distributed.QuantStore,
 ) *QuantizationService {
 	s := &QuantizationService{
 		appConfig:    appConfig,
 		modelLoader:  modelLoader,
 		configLoader: configLoader,
-		jobs:         make(map[string]*schema.QuantizationJob),
 	}
-	s.loadAllJobs()
+
+	// Only attach a Store interface when a concrete store exists, otherwise the
+	// SyncedMap would see a non-nil interface wrapping a nil pointer and try to
+	// hydrate/write through a nil DB.
+	var syncStore syncstate.Store[string, *schema.QuantizationJob]
+	if store != nil {
+		syncStore = &quantStoreAdapter{store: store}
+	}
+
+	s.jobs = syncstate.New(syncstate.Config[string, *schema.QuantizationJob]{
+		Name:   "quant.jobs",
+		Key:    func(j *schema.QuantizationJob) string { return j.ID },
+		Nats:   nats,
+		Store:  syncStore,
+		Loader: s.loadJobsFromDisk, // ignored when Store is set (distributed mode)
+	})
+
+	// Hydrate + subscribe. A hydrate failure must not take the server down: log and
+	// continue degraded (standalone), mirroring the FineTune/OpCache wiring.
+	if err := s.jobs.Start(appConfig.Context); err != nil {
+		xlog.Warn("Quantization SyncedMap start failed; running degraded", "error", err)
+	}
 	return s
 }
 
+// Close releases the SyncedMap subscription and background workers.
+func (s *QuantizationService) Close() error {
+	return s.jobs.Close()
+}
+
 // quantizationBaseDir returns the base directory for quantization job data.
 func (s *QuantizationService) quantizationBaseDir() string {
 	return filepath.Join(s.appConfig.DataPath, "quantization")
@@ -80,15 +120,18 @@ func (s *QuantizationService) saveJobState(job *schema.QuantizationJob) {
 	}
 }
 
-// loadAllJobs scans the quantization directory for persisted jobs and loads them.
-func (s *QuantizationService) loadAllJobs() {
+// loadJobsFromDisk scans the quantization directory for persisted jobs and
+// returns them. It is the SyncedMap Loader used in standalone mode (no DB); the
+// returned slice hydrates the map on Start.
+func (s *QuantizationService) loadJobsFromDisk(_ context.Context) ([]*schema.QuantizationJob, error) {
 	baseDir := s.quantizationBaseDir()
 	entries, err := os.ReadDir(baseDir)
 	if err != nil {
-		// Directory doesn't exist yet — that's fine
-		return
+		// Directory doesn't exist yet — that's fine, start empty.
+		return nil, nil
 	}
 
+	var jobs []*schema.QuantizationJob
 	for _, entry := range entries {
 		if !entry.IsDir() {
 			continue
@@ -117,12 +160,13 @@ func (s *QuantizationService) loadAllJobs() {
 			job.ImportMessage = "Server restarted while import was running"
 		}
 
-		s.jobs[job.ID] = &job
+		jobs = append(jobs, &job)
 	}
 
-	if len(s.jobs) > 0 {
-		xlog.Info("Loaded persisted quantization jobs", "count", len(s.jobs))
+	if len(jobs) > 0 {
+		xlog.Info("Loaded persisted quantization jobs", "count", len(jobs))
 	}
+	return jobs, nil
 }
 
 // StartJob starts a new quantization job.
@@ -188,7 +232,12 @@ func (s *QuantizationService) StartJob(ctx context.Context, userID string, req s
 		CreatedAt:        time.Now().UTC().Format(time.RFC3339),
 		Config:           &req,
 	}
-	s.jobs[jobID] = job
+	// Set write-through persists to PostgreSQL (distributed) and broadcasts to
+	// peer replicas; the disk state.json is written separately for restart
+	// recovery / standalone hydrate.
+	if err := s.jobs.Set(ctx, job); err != nil {
+		return nil, fmt.Errorf("failed to persist job: %w", err)
+	}
 	s.saveJobState(job)
 
 	return &schema.QuantizationJobResponse{
@@ -203,7 +252,7 @@ func (s *QuantizationService) GetJob(userID, jobID string) (*schema.Quantization
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		return nil, fmt.Errorf("job not found: %s", jobID)
 	}
@@ -219,7 +268,7 @@ func (s *QuantizationService) ListJobs(userID string) []*schema.QuantizationJob
 	defer s.mu.Unlock()
 
 	var result []*schema.QuantizationJob
-	for _, job := range s.jobs {
+	for _, job := range s.jobs.List() {
 		if userID == "" || job.UserID == userID {
 			result = append(result, job)
 		}
@@ -235,7 +284,7 @@ func (s *QuantizationService) ListJobs(userID string) []*schema.QuantizationJob
 // StopJob stops a running quantization job.
 func (s *QuantizationService) StopJob(ctx context.Context, userID, jobID string) error {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return fmt.Errorf("job not found: %s", jobID)
@@ -256,6 +305,9 @@ func (s *QuantizationService) StopJob(ctx context.Context, userID, jobID string)
 	s.mu.Lock()
 	job.Status = "stopped"
 	job.Message = "Quantization stopped by user"
+	if err := s.jobs.Set(ctx, job); err != nil {
+		xlog.Warn("Failed to persist stopped job", "job_id", jobID, "error", err)
+	}
 	s.saveJobState(job)
 	s.mu.Unlock()
 
@@ -265,7 +317,7 @@ func (s *QuantizationService) StopJob(ctx context.Context, userID, jobID string)
 // DeleteJob removes a quantization job and its associated data from disk.
 func (s *QuantizationService) DeleteJob(userID, jobID string) error {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return fmt.Errorf("job not found: %s", jobID)
@@ -289,7 +341,11 @@ func (s *QuantizationService) DeleteJob(userID, jobID string) error {
 	}
 
 	importModelName := job.ImportModelName
-	delete(s.jobs, jobID)
+	// Delete write-through removes the DB row (distributed) and broadcasts the
+	// removal to peer replicas. DeleteJob has no ctx, so use Background.
+	if err := s.jobs.Delete(context.Background(), jobID); err != nil {
+		xlog.Warn("Failed to delete job from store", "job_id", jobID, "error", err)
+	}
 	s.mu.Unlock()
 
 	// Remove job directory (state.json, output files)
@@ -324,7 +380,7 @@ func (s *QuantizationService) DeleteJob(userID, jobID string) error {
 // StreamProgress opens a gRPC progress stream and calls the callback for each update.
 func (s *QuantizationService) StreamProgress(ctx context.Context, userID, jobID string, callback func(event *schema.QuantizationProgressEvent)) error {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return fmt.Errorf("job not found: %s", jobID)
@@ -353,7 +409,7 @@ func (s *QuantizationService) StreamProgress(ctx context.Context, userID, jobID
 	}, func(update *pb.QuantizationProgressUpdate) {
 		// Update job status and persist
 		s.mu.Lock()
-		if j, ok := s.jobs[jobID]; ok {
+		if j, ok := s.jobs.Get(jobID); ok {
 			// Don't let progress updates overwrite terminal states
 			isTerminal := j.Status == "stopped" || j.Status == "completed" || j.Status == "failed"
 			if !isTerminal {
@@ -365,6 +421,9 @@ func (s *QuantizationService) StreamProgress(ctx context.Context, userID, jobID
 			if update.OutputFile != "" {
 				j.OutputFile = update.OutputFile
 			}
+			if err := s.jobs.Set(ctx, j); err != nil {
+				xlog.Warn("Failed to persist progress update", "job_id", jobID, "error", err)
+			}
 			s.saveJobState(j)
 		}
 		s.mu.Unlock()
@@ -399,7 +458,7 @@ func sanitizeQuantModelName(s string) string {
 // ImportModel imports a quantized model into LocalAI asynchronously.
 func (s *QuantizationService) ImportModel(ctx context.Context, userID, jobID string, req schema.QuantizationImportRequest) (string, error) {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return "", fmt.Errorf("job not found: %s", jobID)
@@ -459,6 +518,9 @@ func (s *QuantizationService) ImportModel(ctx context.Context, userID, jobID str
 	job.ImportStatus = "importing"
 	job.ImportMessage = ""
 	job.ImportModelName = ""
+	if err := s.jobs.Set(ctx, job); err != nil {
+		xlog.Warn("Failed to persist import start", "job_id", jobID, "error", err)
+	}
 	s.saveJobState(job)
 	s.mu.Unlock()
 
@@ -514,10 +576,15 @@ func (s *QuantizationService) ImportModel(ctx context.Context, userID, jobID str
 
 		xlog.Info("Quantized model imported and registered", "job_id", jobID, "model_name", modelName)
 
+		// Runs after the HTTP request returns, so use Background rather than the
+		// (now likely cancelled) request ctx for the write-through.
 		s.mu.Lock()
 		job.ImportStatus = "completed"
 		job.ImportModelName = modelName
 		job.ImportMessage = ""
+		if err := s.jobs.Set(context.Background(), job); err != nil {
+			xlog.Warn("Failed to persist import completion", "job_id", jobID, "error", err)
+		}
 		s.saveJobState(job)
 		s.mu.Unlock()
 	}()
@@ -525,10 +592,14 @@ func (s *QuantizationService) ImportModel(ctx context.Context, userID, jobID str
 	return modelName, nil
 }
 
-// setImportMessage updates the import message and persists the job state.
+// setImportMessage updates the import message and persists the job state. Called
+// from the background import goroutine, so it uses Background for write-through.
 func (s *QuantizationService) setImportMessage(job *schema.QuantizationJob, msg string) {
 	s.mu.Lock()
 	job.ImportMessage = msg
+	if err := s.jobs.Set(context.Background(), job); err != nil {
+		xlog.Warn("Failed to persist import message", "job_id", job.ID, "error", err)
+	}
 	s.saveJobState(job)
 	s.mu.Unlock()
 }
@@ -539,6 +610,9 @@ func (s *QuantizationService) setImportFailed(job *schema.QuantizationJob, messa
 	s.mu.Lock()
 	job.ImportStatus = "failed"
 	job.ImportMessage = message
+	if err := s.jobs.Set(context.Background(), job); err != nil {
+		xlog.Warn("Failed to persist import failure", "job_id", job.ID, "error", err)
+	}
 	s.saveJobState(job)
 	s.mu.Unlock()
 }
@@ -546,7 +620,7 @@ func (s *QuantizationService) setImportFailed(job *schema.QuantizationJob, messa
 // GetOutputPath returns the path to the quantized model file and a download name.
 func (s *QuantizationService) GetOutputPath(userID, jobID string) (string, string, error) {
 	s.mu.Lock()
-	job, ok := s.jobs[jobID]
+	job, ok := s.jobs.Get(jobID)
 	if !ok {
 		s.mu.Unlock()
 		return "", "", fmt.Errorf("job not found: %s", jobID)

core/services/quantization/service_test.go

@@ -0,0 +1,187 @@
+package quantization
+
+// White-box tests (package quantization) so a spec can drive the service's
+// internal SyncedMap the same way StartJob does (via jobs.Set) without standing
+// up a quantization backend, then assert the cross-replica reads
+// (GetJob/ListJobs) and the adapter conversions that keep REST responses
+// byte-for-byte unchanged.
+
+import (
+	"context"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/testutil"
+)
+
+// newTestService builds a standalone QuantizationService wired to the given bus.
+// The model/config loaders are nil because the read/sync paths under test never
+// touch them; the data dir is a throwaway temp dir so the disk Loader finds
+// nothing.
+func newTestService(bus *testutil.FakeBus) *QuantizationService {
+	appConfig := &config.ApplicationConfig{
+		Context:  context.Background(),
+		DataPath: GinkgoT().TempDir(),
+	}
+	return NewQuantizationService(appConfig, nil, nil, bus, nil)
+}
+
+var _ = Describe("QuantizationService", func() {
+	ctx := context.Background()
+
+	Describe("cross-replica job visibility", func() {
+		var (
+			bus  *testutil.FakeBus
+			a, b *QuantizationService
+		)
+
+		BeforeEach(func() {
+			// One shared bus, two replicas: exactly the distributed topology where a
+			// round-robin request may land on a replica that did not originate the
+			// change.
+			bus = testutil.NewFakeBus()
+			a = newTestService(bus)
+			b = newTestService(bus)
+		})
+
+		AfterEach(func() {
+			Expect(a.Close()).To(Succeed())
+			Expect(b.Close()).To(Succeed())
+		})
+
+		It("makes a job created on A visible via B's GetJob and ListJobs", func() {
+			job := &schema.QuantizationJob{ID: "job-1", UserID: "user-1", Status: "queued", CreatedAt: "2026-06-27T10:00:00Z"}
+			// StartJob persists via jobs.Set; drive that directly to avoid a backend.
+			Expect(a.jobs.Set(ctx, job)).To(Succeed())
+
+			got, err := b.GetJob("user-1", "job-1")
+			Expect(err).ToNot(HaveOccurred(), "B must see a job A just created")
+			Expect(got.Status).To(Equal("queued"))
+
+			listed := b.ListJobs("user-1")
+			Expect(listed).To(HaveLen(1))
+			Expect(listed[0].ID).To(Equal("job-1"))
+		})
+
+		It("removes a job from B when it is deleted on A", func() {
+			job := &schema.QuantizationJob{ID: "job-2", UserID: "user-1", Status: "completed", CreatedAt: "2026-06-27T10:00:00Z"}
+			Expect(a.jobs.Set(ctx, job)).To(Succeed())
+			_, err := b.GetJob("user-1", "job-2")
+			Expect(err).ToNot(HaveOccurred(), "precondition: B must have the job before the delete")
+
+			Expect(a.jobs.Delete(ctx, "job-2")).To(Succeed())
+
+			_, err = b.GetJob("user-1", "job-2")
+			Expect(err).To(HaveOccurred(), "a delete on A must remove the job from B")
+		})
+
+		It("propagates a status update from A to B", func() {
+			job := &schema.QuantizationJob{ID: "job-3", UserID: "user-1", Status: "quantizing", CreatedAt: "2026-06-27T10:00:00Z"}
+			Expect(a.jobs.Set(ctx, job)).To(Succeed())
+
+			updated := &schema.QuantizationJob{ID: "job-3", UserID: "user-1", Status: "completed", CreatedAt: "2026-06-27T10:00:00Z"}
+			Expect(a.jobs.Set(ctx, updated)).To(Succeed())
+
+			got, err := b.GetJob("user-1", "job-3")
+			Expect(err).ToNot(HaveOccurred())
+			Expect(got.Status).To(Equal("completed"))
+		})
+	})
+
+	Describe("ListJobs", func() {
+		var svc *QuantizationService
+
+		BeforeEach(func() {
+			svc = newTestService(testutil.NewFakeBus())
+		})
+		AfterEach(func() { Expect(svc.Close()).To(Succeed()) })
+
+		It("filters by user and sorts newest-first", func() {
+			Expect(svc.jobs.Set(ctx, &schema.QuantizationJob{ID: "old", UserID: "u1", CreatedAt: "2026-06-25T10:00:00Z"})).To(Succeed())
+			Expect(svc.jobs.Set(ctx, &schema.QuantizationJob{ID: "new", UserID: "u1", CreatedAt: "2026-06-27T10:00:00Z"})).To(Succeed())
+			Expect(svc.jobs.Set(ctx, &schema.QuantizationJob{ID: "other", UserID: "u2", CreatedAt: "2026-06-26T10:00:00Z"})).To(Succeed())
+
+			jobs := svc.ListJobs("u1")
+			Expect(jobs).To(HaveLen(2), "only u1's jobs")
+			Expect(jobs[0].ID).To(Equal("new"), "newest first")
+			Expect(jobs[1].ID).To(Equal("old"))
+		})
+
+		It("returns every user's jobs when the userID filter is empty", func() {
+			Expect(svc.jobs.Set(ctx, &schema.QuantizationJob{ID: "a", UserID: "u1", CreatedAt: "2026-06-25T10:00:00Z"})).To(Succeed())
+			Expect(svc.jobs.Set(ctx, &schema.QuantizationJob{ID: "b", UserID: "u2", CreatedAt: "2026-06-26T10:00:00Z"})).To(Succeed())
+
+			Expect(svc.ListJobs("")).To(HaveLen(2))
+		})
+
+		It("rejects GetJob for a job owned by another user", func() {
+			Expect(svc.jobs.Set(ctx, &schema.QuantizationJob{ID: "x", UserID: "owner", CreatedAt: "2026-06-25T10:00:00Z"})).To(Succeed())
+
+			_, err := svc.GetJob("intruder", "x")
+			Expect(err).To(HaveOccurred(), "a different user must not read someone else's job")
+		})
+	})
+
+	Describe("store adapter conversion", func() {
+		// The SyncedMap value type is *schema.QuantizationJob (the exact REST shape).
+		// These specs prove the DB adapter round-trips it losslessly, so hydrate and
+		// write-through in distributed mode keep responses unchanged.
+		It("round-trips a job through jobToRecord/recordToJob preserving the API shape", func() {
+			original := &schema.QuantizationJob{
+				ID:               "rt-1",
+				UserID:           "user-1",
+				Model:            "base-model",
+				Backend:          "llama-cpp-quantization",
+				ModelID:          "llama-cpp-quantization-quantize-rt-1",
+				QuantizationType: "q4_k_m",
+				Status:           "completed",
+				Message:          "done",
+				OutputDir:        "/data/quantization/rt-1",
+				OutputFile:       "/data/quantization/rt-1/model.gguf",
+				ExtraOptions:     map[string]string{"hf_token": "secret"},
+				CreatedAt:        "2026-06-27T10:00:00Z",
+				ImportStatus:     "completed",
+				ImportMessage:    "",
+				ImportModelName:  "base-model-q4_k_m-rt-1",
+				Config:           &schema.QuantizationJobRequest{Model: "base-model", Backend: "llama-cpp-quantization", QuantizationType: "q4_k_m"},
+			}
+
+			rec := jobToRecord(original)
+			Expect(rec.ID).To(Equal("rt-1"))
+			Expect(rec.ConfigJSON).ToNot(BeEmpty(), "structured config must serialize into the JSON column")
+			Expect(rec.ExtraOptsJSON).ToNot(BeEmpty())
+
+			back := recordToJob(rec)
+			Expect(back.ID).To(Equal(original.ID))
+			Expect(back.UserID).To(Equal(original.UserID))
+			Expect(back.Model).To(Equal(original.Model))
+			Expect(back.Backend).To(Equal(original.Backend))
+			Expect(back.ModelID).To(Equal(original.ModelID))
+			Expect(back.QuantizationType).To(Equal(original.QuantizationType))
+			Expect(back.Status).To(Equal(original.Status))
+			Expect(back.Message).To(Equal(original.Message))
+			Expect(back.OutputDir).To(Equal(original.OutputDir))
+			Expect(back.OutputFile).To(Equal(original.OutputFile))
+			Expect(back.ImportStatus).To(Equal(original.ImportStatus))
+			Expect(back.ImportModelName).To(Equal(original.ImportModelName))
+			Expect(back.CreatedAt).To(Equal(original.CreatedAt))
+			Expect(back.ExtraOptions).To(Equal(original.ExtraOptions))
+			Expect(back.Config).ToNot(BeNil())
+			Expect(back.Config.QuantizationType).To(Equal("q4_k_m"))
+		})
+	})
+
+	Describe("compile-time adapter contract", func() {
+		It("satisfies syncstate.Store for *distributed.QuantStore", func() {
+			// Guards against drift between the adapter and the component interface;
+			// the var assertion in syncstore.go covers it at build time, this keeps
+			// the type referenced from a spec too.
+			var _ *distributed.QuantStore
+			Expect(&quantStoreAdapter{}).ToNot(BeNil())
+		})
+	})
+})

core/services/quantization/syncstore.go

@@ -0,0 +1,114 @@
+package quantization
+
+import (
+	"context"
+	"encoding/json"
+	"time"
+
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/core/services/distributed"
+	"github.com/mudler/LocalAI/core/services/syncstate"
+)
+
+// quantStoreAdapter bridges the distributed PostgreSQL QuantStore to the generic
+// syncstate.Store the SyncedMap consumes. It is only wired in distributed mode;
+// standalone leaves Store nil and hydrates from disk via a Loader instead.
+//
+// The SyncedMap value type is *schema.QuantizationJob (the exact shape the REST
+// API returns) so reads need no conversion and the response JSON is provably
+// unchanged. The adapter is the single place that translates between that API
+// shape and the DB QuantJobRecord.
+type quantStoreAdapter struct {
+	store *distributed.QuantStore
+}
+
+// compile-time assertion that the adapter satisfies the component's Store.
+var _ syncstate.Store[string, *schema.QuantizationJob] = (*quantStoreAdapter)(nil)
+
+func (a *quantStoreAdapter) List(_ context.Context) ([]*schema.QuantizationJob, error) {
+	records, err := a.store.ListAll()
+	if err != nil {
+		return nil, err
+	}
+	jobs := make([]*schema.QuantizationJob, 0, len(records))
+	for i := range records {
+		jobs = append(jobs, recordToJob(&records[i]))
+	}
+	return jobs, nil
+}
+
+func (a *quantStoreAdapter) Upsert(_ context.Context, job *schema.QuantizationJob) error {
+	return a.store.Upsert(jobToRecord(job))
+}
+
+func (a *quantStoreAdapter) Delete(_ context.Context, id string) error {
+	return a.store.Delete(id)
+}
+
+// recordToJob maps a persisted DB record back to the API shape, reconstructing
+// the structured Config / ExtraOptions from their JSON columns.
+func recordToJob(r *distributed.QuantJobRecord) *schema.QuantizationJob {
+	job := &schema.QuantizationJob{
+		ID:               r.ID,
+		UserID:           r.UserID,
+		Model:            r.Model,
+		Backend:          r.Backend,
+		ModelID:          r.ModelID,
+		QuantizationType: r.QuantizationType,
+		Status:           r.Status,
+		Message:          r.Message,
+		OutputDir:        r.OutputDir,
+		OutputFile:       r.OutputFile,
+		ImportStatus:     r.ImportStatus,
+		ImportMessage:    r.ImportMessage,
+		ImportModelName:  r.ImportModelName,
+		CreatedAt:        r.CreatedAt.UTC().Format(time.RFC3339),
+	}
+	if r.ExtraOptsJSON != "" {
+		// Best-effort: a malformed column must not drop the whole job from the API.
+		_ = json.Unmarshal([]byte(r.ExtraOptsJSON), &job.ExtraOptions)
+	}
+	if r.ConfigJSON != "" {
+		var cfg schema.QuantizationJobRequest
+		if err := json.Unmarshal([]byte(r.ConfigJSON), &cfg); err == nil {
+			job.Config = &cfg
+		}
+	}
+	return job
+}
+
+// jobToRecord maps the API shape to a DB record for write-through, serializing
+// the structured Config / ExtraOptions into their JSON columns. CreatedAt is
+// parsed back from the RFC3339 string the service stamps; an unparseable value is
+// left zero so QuantStore.Upsert stamps "now".
+func jobToRecord(job *schema.QuantizationJob) *distributed.QuantJobRecord {
+	rec := &distributed.QuantJobRecord{
+		ID:               job.ID,
+		UserID:           job.UserID,
+		Model:            job.Model,
+		Backend:          job.Backend,
+		ModelID:          job.ModelID,
+		QuantizationType: job.QuantizationType,
+		Status:           job.Status,
+		Message:          job.Message,
+		OutputDir:        job.OutputDir,
+		OutputFile:       job.OutputFile,
+		ImportStatus:     job.ImportStatus,
+		ImportMessage:    job.ImportMessage,
+		ImportModelName:  job.ImportModelName,
+	}
+	if job.Config != nil {
+		if data, err := json.Marshal(job.Config); err == nil {
+			rec.ConfigJSON = string(data)
+		}
+	}
+	if job.ExtraOptions != nil {
+		if data, err := json.Marshal(job.ExtraOptions); err == nil {
+			rec.ExtraOptsJSON = string(data)
+		}
+	}
+	if t, err := time.Parse(time.RFC3339, job.CreatedAt); err == nil {
+		rec.CreatedAt = t
+	}
+	return rec
+}

core/services/syncstate/syncstate.go

@@ -0,0 +1,289 @@
+// Package syncstate provides SyncedMap, a reusable cross-replica in-memory map.
+//
+// LocalAI in distributed mode runs multiple frontend replicas behind a
+// round-robin load balancer. Several features keep process-local in-memory state
+// that is surfaced to the HTTP/UI API; without cross-replica sync a poll that
+// lands on a replica which did not originate a change sees stale or missing data.
+// SyncedMap collapses the three legs each feature otherwise hand-wires - an
+// in-memory map, a NATS broadcast/apply path, and optional durable read-through -
+// into one well-tested component so cross-replica consistency is a configuration
+// choice rather than a bespoke re-implementation.
+package syncstate
+
+import (
+	"context"
+	"sync"
+	"time"
+
+	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/xlog"
+)
+
+// Op values carried on the wire and passed to OnApply.
+const (
+	opSet    = "set"
+	opDelete = "delete"
+)
+
+// Store is optional durable backing for a SyncedMap. In distributed mode it is a
+// single shared DB, so the apply path (a delta received from a peer) updates
+// memory only and never re-writes the Store.
+type Store[K comparable, V any] interface {
+	List(ctx context.Context) ([]V, error)
+	Upsert(ctx context.Context, v V) error
+	Delete(ctx context.Context, k K) error
+}
+
+// Config configures a SyncedMap.
+type Config[K comparable, V any] struct {
+	Name      string                                 // subject namespace, e.g. "finetune.jobs"
+	Key       func(V) K                              // extract the key from a value
+	Nats      messaging.MessagingClient              // nil => standalone: in-memory only, no broadcast/subscribe
+	Store     Store[K, V]                            // optional read-through persistence
+	Loader    func(ctx context.Context) ([]V, error) // source when there is no Store (e.g. disk reload)
+	OnApply   func(op string, k K, v V)              // optional hook after an applied change (e.g. ShutdownModel)
+	Reconcile time.Duration                          // optional periodic re-hydrate; 0 = off
+}
+
+// delta is the JSON wire envelope broadcast on every local mutation. Value is
+// omitempty so a delete carries only op+key.
+type delta[K comparable, V any] struct {
+	Op    string `json:"op"`
+	Key   K      `json:"key"`
+	Value V      `json:"value,omitempty"`
+}
+
+// SyncedMap is a cross-replica in-memory map. A local write (Set/Delete) updates
+// memory, the optional durable Store, then broadcasts a delta to peers. A peer's
+// delta updates memory only and fires OnApply - it never re-broadcasts and never
+// writes the Store. That structural split is the echo-loop guard (same pattern as
+// galleryop.mergeStatus / OpCache.applyStart): receiving your own broadcast just
+// re-applies an idempotent value to memory, so there is no storm and no
+// double-write.
+type SyncedMap[K comparable, V any] struct {
+	cfg Config[K, V]
+
+	mu   sync.RWMutex
+	data map[K]V
+
+	sub Subscription
+
+	// lifeCtx outlives Start's argument: a reconnect callback or reconcile tick
+	// can fire long after Start returns, so they must not be tied to a ctx the
+	// caller may cancel. Close cancels it.
+	lifeCtx context.Context
+	cancel  context.CancelFunc
+	wg      sync.WaitGroup
+}
+
+// Subscription is the subset of messaging.Subscription the component holds onto.
+type Subscription = messaging.Subscription
+
+// New constructs a SyncedMap. Call Start to hydrate and begin syncing.
+func New[K comparable, V any](cfg Config[K, V]) *SyncedMap[K, V] {
+	return &SyncedMap[K, V]{cfg: cfg, data: make(map[K]V)}
+}
+
+func (m *SyncedMap[K, V]) subject() string {
+	return messaging.SubjectSyncStateDelta(m.cfg.Name)
+}
+
+// Start hydrates from the source, subscribes for peer deltas, registers a
+// reconnect re-hydrate (when the client supports it), and starts the optional
+// reconcile ticker.
+func (m *SyncedMap[K, V]) Start(ctx context.Context) error {
+	if err := m.hydrate(ctx); err != nil {
+		return err
+	}
+
+	// The cancel func is stored on the struct and invoked in Close (covered by
+	// tests); lifeCtx must outlive Start to drive the reconnect/reconcile
+	// goroutines, so it cannot be cancelled or deferred within this scope.
+	m.lifeCtx, m.cancel = context.WithCancel(context.Background()) // #nosec G118 -- cancel is invoked in Close()
+
+	if m.cfg.Nats != nil {
+		sub, err := messaging.SubscribeJSON(m.cfg.Nats, m.subject(), m.apply)
+		if err != nil {
+			return err
+		}
+		m.sub = sub
+
+		// nats.go transparently resubscribes on reconnect, but it cannot know we
+		// kept derived in-memory state that may have drifted while the link was
+		// down, so re-hydrate from the durable source. Detected via an optional
+		// interface so MessagingClient itself stays minimal; standalone/test
+		// clients without the method simply fall back to the reconcile ticker.
+		if r, ok := m.cfg.Nats.(interface{ OnReconnect(func()) }); ok {
+			r.OnReconnect(func() {
+				if err := m.hydrate(m.lifeCtx); err != nil {
+					xlog.Warn("syncstate: reconnect re-hydrate failed", "name", m.cfg.Name, "error", err)
+				}
+			})
+		}
+	}
+
+	if m.cfg.Reconcile > 0 {
+		m.wg.Add(1)
+		go m.reconcileLoop()
+	}
+	return nil
+}
+
+// Close unsubscribes and stops the reconcile ticker.
+func (m *SyncedMap[K, V]) Close() error {
+	if m.cancel != nil {
+		m.cancel()
+	}
+	m.wg.Wait()
+	if m.sub != nil {
+		return m.sub.Unsubscribe()
+	}
+	return nil
+}
+
+// Set updates the value locally, writes through the Store, then broadcasts.
+// Per the data-flow contract the Store write happens under the lock so memory and
+// durable state move together; the broadcast is best-effort after unlocking.
+func (m *SyncedMap[K, V]) Set(ctx context.Context, v V) error {
+	k := m.cfg.Key(v)
+	m.mu.Lock()
+	m.data[k] = v
+	if m.cfg.Store != nil {
+		if err := m.cfg.Store.Upsert(ctx, v); err != nil {
+			m.mu.Unlock()
+			return err
+		}
+	}
+	m.mu.Unlock()
+	m.publish(opSet, k, v)
+	return nil
+}
+
+// Delete removes the key locally, deletes it from the Store, then broadcasts.
+func (m *SyncedMap[K, V]) Delete(ctx context.Context, k K) error {
+	m.mu.Lock()
+	delete(m.data, k)
+	if m.cfg.Store != nil {
+		if err := m.cfg.Store.Delete(ctx, k); err != nil {
+			m.mu.Unlock()
+			return err
+		}
+	}
+	m.mu.Unlock()
+	var zero V
+	m.publish(opDelete, k, zero)
+	return nil
+}
+
+// Get returns the value for k and whether it was present.
+func (m *SyncedMap[K, V]) Get(k K) (V, bool) {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	v, ok := m.data[k]
+	return v, ok
+}
+
+// List returns a snapshot slice of all values.
+func (m *SyncedMap[K, V]) List() []V {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	out := make([]V, 0, len(m.data))
+	for _, v := range m.data {
+		out = append(out, v)
+	}
+	return out
+}
+
+// Snapshot returns a copy of the underlying map.
+func (m *SyncedMap[K, V]) Snapshot() map[K]V {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	out := make(map[K]V, len(m.data))
+	for k, v := range m.data {
+		out[k] = v
+	}
+	return out
+}
+
+// publish broadcasts a delta. Standalone (nil Nats) is a strict no-op.
+func (m *SyncedMap[K, V]) publish(op string, k K, v V) {
+	if m.cfg.Nats == nil {
+		return
+	}
+	if err := m.cfg.Nats.Publish(m.subject(), delta[K, V]{Op: op, Key: k, Value: v}); err != nil {
+		xlog.Warn("syncstate: failed to broadcast delta", "name", m.cfg.Name, "op", op, "error", err)
+	}
+}
+
+// apply handles a peer's delta: memory-only update plus OnApply. It deliberately
+// never writes the Store nor re-publishes - that is the echo-loop guard.
+func (m *SyncedMap[K, V]) apply(d delta[K, V]) {
+	switch d.Op {
+	case opSet:
+		m.mu.Lock()
+		m.data[d.Key] = d.Value
+		m.mu.Unlock()
+	case opDelete:
+		m.mu.Lock()
+		delete(m.data, d.Key)
+		m.mu.Unlock()
+	default:
+		xlog.Warn("syncstate: ignoring delta with unknown op", "name", m.cfg.Name, "op", d.Op)
+		return
+	}
+	if m.cfg.OnApply != nil {
+		m.cfg.OnApply(d.Op, d.Key, d.Value)
+	}
+}
+
+// hydrate replaces the whole map from the durable source: Store if present, else
+// Loader. With neither, a late joiner starts empty and catches up via deltas
+// (acceptable only for ephemeral state).
+func (m *SyncedMap[K, V]) hydrate(ctx context.Context) error {
+	var (
+		vals []V
+		err  error
+	)
+	switch {
+	case m.cfg.Store != nil:
+		vals, err = m.cfg.Store.List(ctx)
+	case m.cfg.Loader != nil:
+		vals, err = m.cfg.Loader(ctx)
+	default:
+		return nil
+	}
+	if err != nil {
+		return err
+	}
+	m.replaceAll(vals)
+	return nil
+}
+
+// replaceAll atomically swaps the map contents for the given values, keyed via
+// cfg.Key.
+func (m *SyncedMap[K, V]) replaceAll(vals []V) {
+	next := make(map[K]V, len(vals))
+	for _, v := range vals {
+		next[m.cfg.Key(v)] = v
+	}
+	m.mu.Lock()
+	m.data = next
+	m.mu.Unlock()
+}
+
+// reconcileLoop periodically re-hydrates to repair silent drift (missed deltas).
+func (m *SyncedMap[K, V]) reconcileLoop() {
+	defer m.wg.Done()
+	t := time.NewTicker(m.cfg.Reconcile)
+	defer t.Stop()
+	for {
+		select {
+		case <-m.lifeCtx.Done():
+			return
+		case <-t.C:
+			if err := m.hydrate(m.lifeCtx); err != nil {
+				xlog.Warn("syncstate: reconcile re-hydrate failed", "name", m.cfg.Name, "error", err)
+			}
+		}
+	}
+}

core/services/syncstate/syncstate_suite_test.go

@@ -0,0 +1,13 @@
+package syncstate_test
+
+import (
+	"testing"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func TestSyncstate(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "Syncstate Suite")
+}

core/services/syncstate/syncstate_test.go

@@ -0,0 +1,291 @@
+package syncstate_test
+
+import (
+	"context"
+	"sync"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	"github.com/mudler/LocalAI/core/services/messaging"
+	"github.com/mudler/LocalAI/core/services/syncstate"
+	"github.com/mudler/LocalAI/core/services/testutil"
+)
+
+// job is a minimal JSON-serializable value stand-in for the real cross-replica
+// records (finetune/quant/agent jobs) the component is built for.
+type job struct {
+	ID     string `json:"id"`
+	Status string `json:"status"`
+}
+
+func jobKey(j *job) string { return j.ID }
+
+const stateName = "test.jobs"
+
+func deltaSubject() string { return messaging.SubjectSyncStateDelta(stateName) }
+
+// fakeStore is an in-memory Store that records call counts so specs can assert
+// the write-through-vs-apply split (local writes hit the Store; applied deltas
+// must not).
+type fakeStore struct {
+	mu          sync.Mutex
+	data        map[string]*job
+	upsertCalls int
+	deleteCalls int
+	listCalls   int
+}
+
+func newFakeStore(seed ...*job) *fakeStore {
+	s := &fakeStore{data: map[string]*job{}}
+	for _, j := range seed {
+		s.data[j.ID] = j
+	}
+	return s
+}
+
+func (s *fakeStore) List(_ context.Context) ([]*job, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.listCalls++
+	out := make([]*job, 0, len(s.data))
+	for _, j := range s.data {
+		out = append(out, j)
+	}
+	return out, nil
+}
+
+func (s *fakeStore) Upsert(_ context.Context, j *job) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.upsertCalls++
+	s.data[j.ID] = j
+	return nil
+}
+
+func (s *fakeStore) Delete(_ context.Context, k string) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.deleteCalls++
+	delete(s.data, k)
+	return nil
+}
+
+// add simulates a peer replica writing to the shared DB out-of-band (e.g. while
+// this replica was partitioned), so a re-hydrate can be observed to pick it up.
+func (s *fakeStore) add(j *job) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.data[j.ID] = j
+}
+
+func (s *fakeStore) counts() (upsert, del, list int) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.upsertCalls, s.deleteCalls, s.listCalls
+}
+
+var _ = Describe("SyncedMap", func() {
+	ctx := context.Background()
+
+	Describe("cross-replica delta propagation", func() {
+		var (
+			bus  *testutil.FakeBus
+			a, b *syncstate.SyncedMap[string, *job]
+		)
+
+		BeforeEach(func() {
+			bus = testutil.NewFakeBus()
+			a = syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus})
+			b = syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus})
+			Expect(a.Start(ctx)).To(Succeed())
+			Expect(b.Start(ctx)).To(Succeed())
+		})
+
+		AfterEach(func() {
+			Expect(a.Close()).To(Succeed())
+			Expect(b.Close()).To(Succeed())
+		})
+
+		It("propagates a Set on A to B", func() {
+			Expect(a.Set(ctx, &job{ID: "1", Status: "running"})).To(Succeed())
+
+			got, ok := b.Get("1")
+			Expect(ok).To(BeTrue(), "replica B should see the value A just set")
+			Expect(got.Status).To(Equal("running"))
+		})
+
+		It("prunes a Delete on A from B", func() {
+			Expect(a.Set(ctx, &job{ID: "1", Status: "running"})).To(Succeed())
+			_, present := b.Get("1")
+			Expect(present).To(BeTrue(), "precondition: B must have the value before the delete")
+
+			Expect(a.Delete(ctx, "1")).To(Succeed())
+
+			_, ok := b.Get("1")
+			Expect(ok).To(BeFalse(), "a delete on A must remove the key from B")
+		})
+	})
+
+	Describe("hydration", func() {
+		It("hydrates on Start from a preloaded Store", func() {
+			store := newFakeStore(&job{ID: "x", Status: "done"})
+			m := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Store: store})
+			Expect(m.Start(ctx)).To(Succeed())
+
+			got, ok := m.Get("x")
+			Expect(ok).To(BeTrue(), "Start must populate the map from the Store")
+			Expect(got.Status).To(Equal("done"))
+		})
+
+		It("uses the Loader when Store is nil", func() {
+			m := syncstate.New(syncstate.Config[string, *job]{
+				Name: stateName,
+				Key:  jobKey,
+				Loader: func(_ context.Context) ([]*job, error) {
+					return []*job{{ID: "l", Status: "loaded"}}, nil
+				},
+			})
+			Expect(m.Start(ctx)).To(Succeed())
+
+			got, ok := m.Get("l")
+			Expect(ok).To(BeTrue(), "Loader output must hydrate the map when there is no Store")
+			Expect(got.Status).To(Equal("loaded"))
+		})
+	})
+
+	Describe("echo-loop guard", func() {
+		It("applies its own broadcast once and does not re-publish", func() {
+			bus := testutil.NewFakeBus()
+			a := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus})
+			b := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus})
+			Expect(a.Start(ctx)).To(Succeed())
+			Expect(b.Start(ctx)).To(Succeed())
+			defer func() {
+				Expect(a.Close()).To(Succeed())
+				Expect(b.Close()).To(Succeed())
+			}()
+
+			Expect(a.Set(ctx, &job{ID: "e", Status: "running"})).To(Succeed())
+
+			// One local write must produce exactly one broadcast: A and B both
+			// receive it and apply to memory, but the apply path never re-publishes.
+			Expect(bus.PublishCount(deltaSubject())).To(Equal(1),
+				"the apply path must not re-broadcast, otherwise replicas storm")
+			Expect(a.List()).To(HaveLen(1), "A must not double-store its own echo")
+			_, ok := b.Get("e")
+			Expect(ok).To(BeTrue())
+		})
+	})
+
+	Describe("Store write-through vs apply", func() {
+		It("writes the Store on local Set/Delete but not on an applied delta", func() {
+			bus := testutil.NewFakeBus()
+			storeA := newFakeStore()
+			storeB := newFakeStore()
+			a := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus, Store: storeA})
+			b := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus, Store: storeB})
+			Expect(a.Start(ctx)).To(Succeed())
+			Expect(b.Start(ctx)).To(Succeed())
+			defer func() {
+				Expect(a.Close()).To(Succeed())
+				Expect(b.Close()).To(Succeed())
+			}()
+
+			Expect(a.Set(ctx, &job{ID: "w", Status: "running"})).To(Succeed())
+
+			upA, _, _ := storeA.counts()
+			upB, _, _ := storeB.counts()
+			Expect(upA).To(Equal(1), "local Set must write through to its own Store")
+			Expect(upB).To(Equal(0), "the apply path must never write the peer's Store")
+
+			Expect(a.Delete(ctx, "w")).To(Succeed())
+			_, delA, _ := storeA.counts()
+			_, delB, _ := storeB.counts()
+			Expect(delA).To(Equal(1), "local Delete must delete from its own Store")
+			Expect(delB).To(Equal(0), "the apply path must never delete from the peer's Store")
+		})
+	})
+
+	Describe("OnApply hook", func() {
+		It("fires with the correct op and key on an applied delta", func() {
+			bus := testutil.NewFakeBus()
+			var (
+				mu   sync.Mutex
+				ops  []string
+				keys []string
+			)
+			a := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus})
+			b := syncstate.New(syncstate.Config[string, *job]{
+				Name: stateName, Key: jobKey, Nats: bus,
+				OnApply: func(op string, k string, _ *job) {
+					mu.Lock()
+					ops = append(ops, op)
+					keys = append(keys, k)
+					mu.Unlock()
+				},
+			})
+			Expect(a.Start(ctx)).To(Succeed())
+			Expect(b.Start(ctx)).To(Succeed())
+			defer func() {
+				Expect(a.Close()).To(Succeed())
+				Expect(b.Close()).To(Succeed())
+			}()
+
+			Expect(a.Set(ctx, &job{ID: "o", Status: "running"})).To(Succeed())
+			Expect(a.Delete(ctx, "o")).To(Succeed())
+
+			mu.Lock()
+			defer mu.Unlock()
+			Expect(ops).To(Equal([]string{"set", "delete"}))
+			Expect(keys).To(Equal([]string{"o", "o"}))
+		})
+	})
+
+	Describe("standalone (nil Nats)", func() {
+		It("works in-memory with no panic and nothing to broadcast", func() {
+			m := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey})
+			Expect(m.Start(ctx)).To(Succeed())
+			defer func() { Expect(m.Close()).To(Succeed()) }()
+
+			Expect(func() {
+				Expect(m.Set(ctx, &job{ID: "s", Status: "running"})).To(Succeed())
+			}).ToNot(Panic())
+
+			got, ok := m.Get("s")
+			Expect(ok).To(BeTrue())
+			Expect(got.Status).To(Equal("running"))
+			Expect(m.List()).To(HaveLen(1))
+			Expect(m.Snapshot()).To(HaveKey("s"))
+
+			Expect(m.Delete(ctx, "s")).To(Succeed())
+			_, ok = m.Get("s")
+			Expect(ok).To(BeFalse())
+		})
+	})
+
+	Describe("reconnect re-hydrate", func() {
+		It("re-reads the source when the messaging client reconnects", func() {
+			bus := testutil.NewFakeBus()
+			store := newFakeStore(&job{ID: "init", Status: "running"})
+			m := syncstate.New(syncstate.Config[string, *job]{Name: stateName, Key: jobKey, Nats: bus, Store: store})
+			Expect(m.Start(ctx)).To(Succeed())
+			defer func() { Expect(m.Close()).To(Succeed()) }()
+
+			_, ok := m.Get("init")
+			Expect(ok).To(BeTrue())
+
+			// A peer writes to the shared DB while we are unaware (no delta seen).
+			store.add(&job{ID: "late", Status: "running"})
+			_, ok = m.Get("late")
+			Expect(ok).To(BeFalse(), "the new row should not appear before a re-hydrate")
+
+			bus.TriggerReconnect()
+
+			_, ok = m.Get("late")
+			Expect(ok).To(BeTrue(), "reconnect must re-hydrate from the source and pick up drift")
+			_, _, list := store.counts()
+			Expect(list).To(Equal(2), "exactly one Start hydrate plus one reconnect re-hydrate")
+		})
+	})
+})

core/services/testutil/fakebus.go

@@ -0,0 +1,160 @@
+package testutil
+
+import (
+	"encoding/json"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/mudler/LocalAI/core/services/messaging"
+)
+
+// FakeBus is an in-memory messaging.MessagingClient that delivers each published
+// message synchronously to every registered subscriber whose subject filter
+// matches, including NATS-style wildcard subjects (`*` matches exactly one
+// token).
+//
+// Synchronous delivery keeps specs deterministic: the moment Publish returns,
+// every matching subscriber's handler has already run, so the spec body can read
+// the resulting state without polling. It is the shared test double for every
+// cross-replica-sync adopter (gallery, syncstate, ...) so they exercise the same
+// delivery semantics. It deliberately depends only on the standard library and
+// the messaging package — no test framework — so it is importable anywhere.
+type FakeBus struct {
+	mu   sync.Mutex
+	subs []fakeBusSub
+	// publishCounts records how many messages were published per subject, so a
+	// spec can assert the echo-loop guard (an applied delta must not re-publish).
+	publishCounts map[string]int
+
+	// reconnectCbs back the optional OnReconnect/TriggerReconnect pair, letting a
+	// spec exercise the component's reconnect re-hydrate path without a real
+	// NATS server.
+	reconnectCbs []func()
+}
+
+type fakeBusSub struct {
+	subject string
+	handler func([]byte)
+}
+
+// NewFakeBus returns a ready-to-use in-memory bus.
+func NewFakeBus() *FakeBus {
+	return &FakeBus{publishCounts: map[string]int{}}
+}
+
+// subjectMatches reports whether a subscription filter matches a concrete
+// subject, honoring the single-token `*` wildcard used by NATS.
+func subjectMatches(filter, subject string) bool {
+	if filter == subject {
+		return true
+	}
+	fp := strings.Split(filter, ".")
+	sp := strings.Split(subject, ".")
+	if len(fp) != len(sp) {
+		return false
+	}
+	for i := range fp {
+		if fp[i] == "*" {
+			continue
+		}
+		if fp[i] != sp[i] {
+			return false
+		}
+	}
+	return true
+}
+
+// Publish marshals data as JSON and delivers it synchronously to every matching
+// subscriber.
+func (b *FakeBus) Publish(subject string, data any) error {
+	payload, err := json.Marshal(data)
+	if err != nil {
+		return err
+	}
+	b.mu.Lock()
+	b.publishCounts[subject]++
+	subs := append([]fakeBusSub(nil), b.subs...)
+	b.mu.Unlock()
+	for _, s := range subs {
+		if subjectMatches(s.subject, subject) {
+			s.handler(payload)
+		}
+	}
+	return nil
+}
+
+// PublishCount returns how many messages were published on the exact subject.
+func (b *FakeBus) PublishCount(subject string) int {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.publishCounts[subject]
+}
+
+type fakeBusSubscription struct {
+	bus    *FakeBus
+	subRef fakeBusSub
+}
+
+func (s *fakeBusSubscription) Unsubscribe() error {
+	s.bus.mu.Lock()
+	defer s.bus.mu.Unlock()
+	for i, candidate := range s.bus.subs {
+		if candidate.subject == s.subRef.subject {
+			s.bus.subs = append(s.bus.subs[:i], s.bus.subs[i+1:]...)
+			return nil
+		}
+	}
+	return nil
+}
+
+func (b *FakeBus) Subscribe(subject string, handler func([]byte)) (messaging.Subscription, error) {
+	sub := fakeBusSub{subject: subject, handler: handler}
+	b.mu.Lock()
+	b.subs = append(b.subs, sub)
+	b.mu.Unlock()
+	return &fakeBusSubscription{bus: b, subRef: sub}, nil
+}
+
+func (b *FakeBus) QueueSubscribe(subject, _ string, handler func([]byte)) (messaging.Subscription, error) {
+	return b.Subscribe(subject, handler)
+}
+
+func (b *FakeBus) QueueSubscribeReply(string, string, func([]byte, func([]byte))) (messaging.Subscription, error) {
+	return &fakeBusSubscription{bus: b}, nil
+}
+
+func (b *FakeBus) SubscribeReply(string, func([]byte, func([]byte))) (messaging.Subscription, error) {
+	return &fakeBusSubscription{bus: b}, nil
+}
+
+func (b *FakeBus) Request(string, []byte, time.Duration) ([]byte, error) {
+	return nil, nil
+}
+
+func (b *FakeBus) IsConnected() bool { return true }
+func (b *FakeBus) Close()            {}
+
+// OnReconnect mirrors *messaging.Client.OnReconnect so a spec can drive the
+// component's reconnect re-hydrate path. The component detects this method via an
+// optional interface assertion; implementing it here keeps the fake a faithful
+// stand-in for the concrete client.
+func (b *FakeBus) OnReconnect(cb func()) {
+	if cb == nil {
+		return
+	}
+	b.mu.Lock()
+	b.reconnectCbs = append(b.reconnectCbs, cb)
+	b.mu.Unlock()
+}
+
+// TriggerReconnect runs every registered reconnect callback, simulating a NATS
+// reconnect event.
+func (b *FakeBus) TriggerReconnect() {
+	b.mu.Lock()
+	cbs := append([]func(){}, b.reconnectCbs...)
+	b.mu.Unlock()
+	for _, cb := range cbs {
+		cb()
+	}
+}

docker-compose.distributed.yaml

@@ -57,6 +57,11 @@ services:
       LOCALAI_AGENT_POOL_VECTOR_ENGINE: "postgres"
       LOCALAI_AGENT_POOL_DATABASE_URL: "postgresql://localai:localai@postgres:5432/localai?sslmode=disable"
       LOCALAI_REGISTRATION_TOKEN: "changeme"  # Change this in production!
+      # Shared-models mode (optional): set when every node mounts the SAME
+      # models directory at the SAME path (see "Shared Volume Mode" below).
+      # The router then skips gRPC file staging and workers load models
+      # directly from the shared volume instead of re-downloading them.
+      # LOCALAI_DISTRIBUTED_SHARED_MODELS: "true"
       # Auth (required for distributed mode — must use PostgreSQL)
       LOCALAI_AUTH: "true"
       LOCALAI_AUTH_DATABASE_URL: "postgresql://localai:localai@postgres:5432/localai?sslmode=disable"
@@ -157,8 +162,11 @@ services:
   # Then add to the volumes section:
   #   shared_models:
   #
-  # With shared volumes, model files are already available on the backend —
-  # gRPC file staging becomes a no-op (paths match).
+  # With shared volumes the model files are already present on every worker at
+  # the same path. Set LOCALAI_DISTRIBUTED_SHARED_MODELS=true on the frontend
+  # (see its environment above) so the router skips gRPC file staging and the
+  # worker loads the model directly from the shared path instead of
+  # re-downloading it into a per-model subdirectory.
 
   # --- Adding More Workers ---
   # Copy the worker-1 service above and change:

docs/content/features/distributed-mode.md

@@ -67,6 +67,7 @@ The frontend is a standard LocalAI instance with distributed mode enabled. These
 | `--registration-require-auth` | `LOCALAI_REGISTRATION_REQUIRE_AUTH` | `false` | Fail startup when distributed mode is enabled but the registration token is empty (node endpoints and worker file-transfer would otherwise be unauthenticated) |
 | `--distributed-require-auth` | `LOCALAI_DISTRIBUTED_REQUIRE_AUTH` | `false` | **Umbrella switch.** Implies both `--nats-require-auth` and `--registration-require-auth` — one knob to lock down the NATS bus *and* the registration/file-transfer layer. Set this in production instead of the two granular flags. |
 | `--auto-approve-nodes` | `LOCALAI_AUTO_APPROVE_NODES` | `false` | Auto-approve new worker nodes (skip admin approval) |
+| `--distributed-shared-models` | `LOCALAI_DISTRIBUTED_SHARED_MODELS` | `false` | Assert that every node mounts the **same** models directory at the **same** path (a shared volume). When `true`, the router skips file staging entirely and workers load models directly from the shared path instead of re-downloading them. See [Shared models directory](#shared-models-directory). |
 | `--auth` | `LOCALAI_AUTH` | `false` | **Must be `true`** for distributed mode |
 | `--auth-database-url` | `LOCALAI_AUTH_DATABASE_URL` | *(required)* | PostgreSQL connection URL |
 | `--backend-install-timeout` | `LOCALAI_NATS_BACKEND_INSTALL_TIMEOUT` | `15m` | How long the frontend waits for a worker to acknowledge a backend install before considering the request stalled. Raise it when workers pull large backend images over slow links. If a worker takes longer than this, the operation shows as "still installing in background" in the admin UI and clears once the worker finishes. |
@@ -133,6 +134,14 @@ When S3 is not configured, model files are transferred directly from the fronten
 
 For high-throughput or very large model files, S3 can be more efficient since it avoids streaming through the frontend.
 
+### Shared models directory
+
+If every node (frontend and workers) mounts the **same** models directory at the **same** path - for example a shared volume or network filesystem, as shown in the "Shared Volume Mode" section of `docker-compose.distributed.yaml` - the model files are already present on each worker at their canonical path. In that case staging is wasted work: it copies files that already exist into a per-model subdirectory the worker then loads from, which shows up as a re-download of a model you already have.
+
+Set `LOCALAI_DISTRIBUTED_SHARED_MODELS=true` (or `--distributed-shared-models`) on the frontend to skip staging entirely. The router then leaves the model's absolute paths untouched and the worker loads them directly from the shared volume.
+
+This flag is a contract you assert: all nodes must mount identical paths. Leave it off (the default) when workers have independent models directories - the frontend stages files to them over HTTP (or S3) as described above.
+
 {{% notice warning %}}
 The worker HTTP file transfer server is authenticated by `LOCALAI_REGISTRATION_TOKEN`. If the token is **empty**, the server **fails open** — anyone who can reach the port gets read/write access to the worker's models/staging/data directories (a remote model-poisoning / exfiltration vector). The worker logs a loud warning at startup in this case. Always set `LOCALAI_REGISTRATION_TOKEN` in distributed mode, and set `LOCALAI_DISTRIBUTED_REQUIRE_AUTH=true` (frontend **and** workers) to make a missing token *or* missing NATS credentials a hard startup error rather than a silent fail-open. Firewall the file-transfer port (gRPC base − 1) so only the frontend can reach it.
 {{% /notice %}}

docs/content/installation/macos.md

@@ -22,13 +22,16 @@ Download the latest DMG from GitHub releases:
 3. Drag the LocalAI application to your Applications folder
 4. Launch LocalAI from your Applications folder
 
-## Known Issues
+## Verification
 
-> **Note**: The DMGs are not signed by Apple and may show as quarantined.
->
-> **Workaround**: See [this issue](https://github.com/mudler/LocalAI/issues/6268) for details on how to bypass the quarantine.
->
-> **Fix tracking**: The signing issue is being tracked in [this issue](https://github.com/mudler/LocalAI/issues/6244).
+The `LocalAI.dmg` (and the app inside it) and the `local-ai` server binary are
+signed with an Apple Developer ID and notarized by Apple, so they launch with no
+quarantine prompt or workaround. To inspect the signature yourself:
+
+```bash
+spctl --assess --type open --context context:primary-signature -v /Applications/LocalAI.app
+codesign --verify --deep --strict --verbose=2 /Applications/LocalAI.app
+```
 
 ## Next Steps
 

docs/data/version.json

@@ -1,3 +1,3 @@
 {
-  "version": "v4.5.0"
+  "version": "v4.5.5"
 }

gallery/index.yaml

@@ -1,4 +1,106 @@
 ---
+- name: "qwen-agentworld-35b-a3b"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF
+  description: |
+    # Qwen-AgentWorld-35B-A3B
+
+    📑 Technical Report |
+    📖 Blog |
+    🤗 Hugging Face |
+    🤖 ModelScope |
+    💻 GitHub |
+    🖥️ Demo
+
+    > [!Note]
+    > This repository contains the model weights and configuration files for **Qwen-AgentWorld-35B-A3B**, a native language world model trained for agentic environment simulation.
+    >
+    > These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, etc.
+
+    **Qwen-AgentWorld** is the first language world model to cover seven agent interaction domains within a single model. It simulates agentic environments via long chain-of-thought reasoning, predicting the next environment state given an agent's action and interaction history. Trained through a three-stage pipeline — CPT injects environment knowledge, SFT activates next-state-prediction reasoning, RL sharpens simulation fidelity — Qwen-AgentWorld is a **native world model**: environment modeling is the training objective from the CPT stage onward, not a post-hoc add-on.
+
+    ## Highlights
+
+    ...
+  license: "apache-2.0"
+  tags:
+    - llm
+    - gguf
+    - qwen
+  icon: https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwen-AgentWorld/logo.png
+  overrides:
+    backend: llama-cpp
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    options:
+      - use_jinja:true
+    parameters:
+      model: llama-cpp/models/Qwen-AgentWorld-35B-A3B-GGUF/Qwen-AgentWorld-35B-A3B-UD-Q4_K_M.gguf
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/Qwen-AgentWorld-35B-A3B-GGUF/Qwen-AgentWorld-35B-A3B-UD-Q4_K_M.gguf
+      sha256: e7a8eafdd8013443b6bcc4b6fb47b2d2025f772d359650b9ceb7d75971e22cad
+      uri: https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF/resolve/main/Qwen-AgentWorld-35B-A3B-UD-Q4_K_M.gguf
+- name: "ornith-1.0-9b"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B-GGUF
+  description: |
+    [](https://deep-reinforce.com/ornith.html)
+
+    # Ornith-1.0-9B-GGUF
+
+    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
+    parameters:
+      model: llama-cpp/models/Ornith-1.0-9B-GGUF/ornith-1.0-9b-Q4_K_M.gguf
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/Ornith-1.0-9B-GGUF/ornith-1.0-9b-Q4_K_M.gguf
+      sha256: 5720d1f671b4996481274fffe01868c3c36e87c135cc8538471cc7bd6087b106
+      uri: https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B-GGUF/resolve/main/ornith-1.0-9b-Q4_K_M.gguf
 - name: "ornith-1.0-35b"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
@@ -477,6 +579,10 @@
   icon: https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.6/Figures/qwen3.6_35b_a3b_score.png
   overrides:
     backend: llama-cpp
+    # NVFP4 GGUFs use a quant type the GGUF metadata parser cannot read, so
+    # context size cannot be auto-derived; set it explicitly (the model trains
+    # to 262144, 32768 is a safe default operators can raise).
+    context_size: 32768
     function:
       automatic_tool_parsing_fallback: true
       grammar:
@@ -509,6 +615,9 @@
     - gguf
   overrides:
     backend: llama-cpp
+    # NVFP4 GGUFs use a quant type the GGUF metadata parser cannot read, so
+    # context size cannot be auto-derived; set it explicitly.
+    context_size: 32768
     function:
       automatic_tool_parsing_fallback: true
       grammar:
@@ -536,6 +645,9 @@
   icon: https://cdn-uploads.huggingface.co/production/uploads/66309bd090589b7c65950665/sGQKmrMc6L6guMoaB5_Y2.png
   overrides:
     backend: llama-cpp
+    # NVFP4 GGUFs use a quant type the GGUF metadata parser cannot read, so
+    # context size cannot be auto-derived; set it explicitly.
+    context_size: 32768
     function:
       automatic_tool_parsing_fallback: true
       grammar:
@@ -586,6 +698,10 @@
   icon: https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.6/Figures/qwen3.6_27b_score.png
   overrides:
     backend: llama-cpp
+    # NVFP4 GGUFs use a quant type the GGUF metadata parser cannot read, so
+    # context size cannot be auto-derived; set it explicitly (the model trains
+    # to 262144, 32768 is a safe default operators can raise).
+    context_size: 32768
     function:
       automatic_tool_parsing_fallback: true
       grammar: