server: remove loading.html (#25500)

* server: remove loading.html

* apply ui changes

.github/workflows/ui-publish.yml

@@ -73,4 +73,3 @@ jobs:
           hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/index.html --yes 2>/dev/null || true
           hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/bundle.js --yes 2>/dev/null || true
           hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/bundle.css --yes 2>/dev/null || true
-          hf buckets rm ggml-org/${{ env.HF_BUCKET_NAME }}/loading.html --yes 2>/dev/null || true

common/arg.cpp

@@ -3036,7 +3036,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--tools"}, "TOOL1,TOOL2,...",
         "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
         "specify \"all\" to enable all tools\n"
-        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, apply_diff, get_datetime",
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime",
         [](common_params & params, const std::string & value) {
             params.server_tools = parse_csv_row(value);
         }

ggml/CMakeLists.txt

@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 15)
-set(GGML_VERSION_PATCH 3)
+set(GGML_VERSION_MINOR 16)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -263,13 +263,13 @@ void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const uint8x16_t raw16 = vcombine_u8(raw, raw);
 
             // First 16 elements: replicate bytes 0-3, shift, mask, subtract 1
-            uint8x16_t bytes0 = vqtbl1q_u8(raw16, idx_lo);
+            uint8x16_t bytes0 = ggml_vqtbl1q_u8(raw16, idx_lo);
             int8x16_t qv0 = vsubq_s8(
                 vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes0, shifts), mask2)),
                 one);
 
             // Second 16 elements: replicate bytes 4-7, shift, mask, subtract 1
-            uint8x16_t bytes1 = vqtbl1q_u8(raw16, idx_hi);
+            uint8x16_t bytes1 = ggml_vqtbl1q_u8(raw16, idx_hi);
             int8x16_t qv1 = vsubq_s8(
                 vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes1, shifts), mask2)),
                 one);

scripts/sync-ggml.last

@@ -1 +1 @@
-eced84c86f8b012c752c016f7fe789adea168e1e
+eaa0a74fa768bb72da623a61d9da3d436053ea91

tools/cli/cli-context.cpp

@@ -153,9 +153,19 @@ bool cli_context::init() {
 
     if (use_external_server) {
         spinner.reset();
-        if (!list_and_ask_models()) {
+        try {
+            if (!list_and_ask_models()) {
+                return false;
+            }
+        } catch (const json::parse_error & e) {
+            ui::show_error(e.what());
+            ui::show_message("This might be caused by an incorrect server-base endpoint URL");
+            return false;
+        } catch (const std::exception & e) {
+            ui::show_error(e.what());
             return false;
         }
+
         // restore the spinner for the next step
         spinner.emplace("Waiting for server...");
     }

tools/server/server-http.cpp

@@ -175,6 +175,15 @@ bool server_http_context::init(const common_params & params) {
     // Middlewares
     //
 
+    // Frontend paths - all embedded UI assets
+    static const std::unordered_set<std::string> frontend_paths = []() {
+        std::unordered_set<std::string> paths { "/" };
+        for (const llama_ui_asset & a : llama_ui_get_assets()) {
+            paths.insert("/" + a.name);
+        }
+        return paths;
+    }();
+
     // Public endpoints - API routes plus all embedded UI assets
     static const std::unordered_set<std::string> get_public_endpoints = []() {
         std::unordered_set<std::string> endpoints {
@@ -182,11 +191,8 @@ bool server_http_context::init(const common_params & params) {
             "/v1/health",
             "/models",
             "/v1/models",
-            "/",
         };
-        for (const llama_ui_asset & a : llama_ui_get_assets()) {
-            endpoints.insert("/" + a.name);
-        }
+        endpoints.insert(frontend_paths.begin(), frontend_paths.end());
         return endpoints;
     }();
 
@@ -239,18 +245,9 @@ bool server_http_context::init(const common_params & params) {
 
     auto middleware_server_state = [this](const httplib::Request & req, httplib::Response & res) {
         if (!is_ready.load()) {
-#if defined(LLAMA_UI_HAS_ASSETS)
-            if (const auto tmp = string_split<std::string>(req.path, '.');
-                req.path == "/" || (!tmp.empty() && tmp.back() == "html")) {
-                if (const llama_ui_asset * a = llama_ui_find_asset("loading.html")) {
-                    res.status = 503;
-                    res.set_content(reinterpret_cast<const char*>(a->data), a->size, "text/html; charset=utf-8");
-                    return false;
-                }
+            if (frontend_paths.count(req.path)) {
+                return true; // frontend asset, allow it to load and show "loading"
             }
-#else
-            (void)req;
-#endif
             // no endpoints are allowed to be accessed when the server is not ready
             // this is to prevent any data races or inconsistent states
             res.status = 503;

tools/server/server-tools.h

@@ -9,10 +9,10 @@ struct server_tool {
     bool permission_write = false;
 
     virtual ~server_tool() = default;
-    virtual json get_definition() = 0;
-    virtual json invoke(json params) = 0;
+    virtual json get_definition() const = 0;
+    virtual json invoke(json params) const = 0;
 
-    json to_json();
+    json to_json() const;
 };
 
 struct server_tools {

tools/server/tests/unit/test_tools_builtin.py

@@ -0,0 +1,125 @@
+import os
+
+import pytest
+from utils import *
+
+server: ServerProcess
+
+# project root, used as the search directory for grep_search/file_glob_search
+PROJECT_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", ".."))
+
+# marker for the grep_search test to find in this file
+GREP_MARKER = "llama_cpp_test_tools_builtin_marker_grep_search"
+
+
+@pytest.fixture(autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.router()
+    server.server_tools = "all"
+
+
+def call_tool(name: str, params: dict) -> dict:
+    res = server.make_request("POST", "/tools", data={"tool": name, "params": params})
+    assert res.status_code == 200, res.body
+    assert "error" not in res.body, res.body
+    return res.body
+
+
+def call_tool_expect_error(name: str, params: dict) -> str:
+    res = server.make_request("POST", "/tools", data={"tool": name, "params": params})
+    assert res.status_code == 200, res.body
+    assert "error" in res.body, res.body
+    return res.body["error"]
+
+
+def test_tools_builtin_grep_search():
+    global server
+    server.start()
+
+    res = call_tool("grep_search", {
+        "path": PROJECT_ROOT,
+        "pattern": GREP_MARKER,
+        "include": "test_tools_builtin.py",  # bare pattern -> matches basename at any depth
+    })
+    text = res["plain_text_response"]
+    assert "test_tools_builtin.py" in text
+    assert GREP_MARKER in text
+    assert "Total matches: 1" in text
+
+
+def test_tools_builtin_read_file():
+    global server
+    server.start()
+
+    this_file = os.path.join(PROJECT_ROOT, "tools", "server", "tests", "unit", "test_tools_builtin.py")
+    res = call_tool("read_file", {"path": this_file})
+    text = res["plain_text_response"]
+    assert GREP_MARKER in text
+    assert "def test_tools_builtin_read_file" in text
+
+
+def test_tools_builtin_write_then_edit_file():
+    global server
+    server.start()
+
+    log_path = os.path.join(PROJECT_ROOT, "test.log")
+    try:
+        write_res = call_tool("write_file", {"path": log_path, "content": "line1\nline2\nline3\n"})
+        assert write_res["result"] == "file written successfully"
+
+        read_before = call_tool("read_file", {"path": log_path})
+        assert read_before["plain_text_response"] == "line1\nline2\nline3\n"
+
+        edit_res = call_tool("edit_file", {
+            "path": log_path,
+            "edits": [
+                {"old_text": "line2", "new_text": "line2-edited"},
+                {"old_text": "line3\n", "new_text": "line3\nline4\n"},
+            ],
+        })
+        assert edit_res["result"] == "file edited successfully"
+        assert edit_res["edits_applied"] == 2
+
+        read_after = call_tool("read_file", {"path": log_path})
+        assert read_after["plain_text_response"] == "line1\nline2-edited\nline3\nline4\n"
+    finally:
+        if os.path.exists(log_path):
+            os.remove(log_path)
+
+
+def test_tools_builtin_edit_file_rejects_non_unique_old_text():
+    global server
+    server.start()
+
+    log_path = os.path.join(PROJECT_ROOT, "test.log")
+    try:
+        call_tool("write_file", {"path": log_path, "content": "dup\ndup\n"})
+        err = call_tool_expect_error("edit_file", {
+            "path": log_path,
+            "edits": [{"old_text": "dup", "new_text": "changed"}],
+        })
+        assert "unique" in err
+    finally:
+        if os.path.exists(log_path):
+            os.remove(log_path)
+
+
+def test_tools_builtin_edit_file_rejects_overlapping_edits():
+    global server
+    server.start()
+
+    log_path = os.path.join(PROJECT_ROOT, "test.log")
+    try:
+        call_tool("write_file", {"path": log_path, "content": "line1\nline2\n"})
+        err = call_tool_expect_error("edit_file", {
+            "path": log_path,
+            "edits": [
+                {"old_text": "line1\nline2", "new_text": "a"},
+                {"old_text": "line2", "new_text": "b"},
+            ],
+        })
+        assert "overlap" in err
+    finally:
+        if os.path.exists(log_path):
+            os.remove(log_path)

tools/server/tests/utils.py

@@ -113,6 +113,7 @@ class ServerProcess:
     ui_mcp_proxy: bool = False
     backend_sampling: bool = False
     gcp_compat: bool = False
+    server_tools: str | None = None
 
     # session variables
     process: subprocess.Popen | None = None
@@ -256,6 +257,8 @@ class ServerProcess:
             server_args.append("--no-cache-idle-slots")
         if self.ui_mcp_proxy:
             server_args.append("--ui-mcp-proxy")
+        if self.server_tools:
+            server_args.extend(["--tools", self.server_tools])
         if self.backend_sampling:
             server_args.append("--backend_sampling")
         if self.gcp_compat:

tools/ui/embed.cpp

@@ -187,7 +187,6 @@ int main(int argc, char ** argv) {
         struct required_check { const char * label; match_fn match; bool found; };
         required_check checks[] = {
             { "index.html",           exact("index.html"),           false },
-            { "loading.html",         exact("loading.html"),         false },
             { "manifest.webmanifest", exact("manifest.webmanifest"), false },
             { "sw.js",                exact("sw.js"),                false },
             { "build.json",           exact("build.json"),           false },

tools/ui/src/lib/components/app/chat/ChatScreen/ChatScreenServerError.svelte

@@ -1,10 +1,11 @@
 <script lang="ts">
-	import { AlertTriangle, RefreshCw } from '@lucide/svelte';
+	import { AlertTriangle, Loader2, RefreshCw } from '@lucide/svelte';
 	import { fadeInView } from '$lib/actions/fade-in-view.svelte';
 	import * as Alert from '$lib/components/ui/alert';
-	import { serverError, serverLoading, serverStore } from '$lib/stores/server.svelte';
+	import { serverError, serverLoading, serverStatus, serverStore } from '$lib/stores/server.svelte';
 
 	let hasError = $derived(!!serverError());
+	let isLoadingModel = $derived(serverStatus() === 503);
 </script>
 
 {#if hasError}
@@ -12,23 +13,31 @@
 		class="pointer-events-auto mx-auto mb-4 max-w-[48rem] px-1"
 		use:fadeInView={{ y: 10, duration: 250 }}
 	>
-		<Alert.Root variant="destructive">
-			<AlertTriangle class="h-4 w-4" />
+		<Alert.Root variant={isLoadingModel ? 'default' : 'destructive'}>
+			{#if isLoadingModel}
+				<Loader2 class="h-4 w-4 animate-spin" />
+			{:else}
+				<AlertTriangle class="h-4 w-4" />
+			{/if}
 
 			<Alert.Title class="flex items-center justify-between">
-				<span>Server unavailable</span>
+				<span>{isLoadingModel ? 'Loading model' : 'Server unavailable'}</span>
 
-				<button
-					onclick={() => serverStore.fetch()}
-					disabled={serverLoading()}
-					class="flex items-center gap-1.5 rounded-lg bg-destructive/20 px-2 py-1 text-xs font-medium hover:bg-destructive/30 disabled:opacity-50"
-				>
-					<RefreshCw class="h-3 w-3 {serverLoading() ? 'animate-spin' : ''}" />
-					{serverLoading() ? 'Retrying...' : 'Retry'}
-				</button>
+				{#if !isLoadingModel}
+					<button
+						onclick={() => serverStore.fetch()}
+						disabled={serverLoading()}
+						class="flex items-center gap-1.5 rounded-lg bg-destructive/20 px-2 py-1 text-xs font-medium hover:bg-destructive/30 disabled:opacity-50"
+					>
+						<RefreshCw class="h-3 w-3 {serverLoading() ? 'animate-spin' : ''}" />
+						{serverLoading() ? 'Retrying...' : 'Retry'}
+					</button>
+				{/if}
 			</Alert.Title>
 
-			<Alert.Description>{serverError()}</Alert.Description>
+			{#if !isLoadingModel}
+				<Alert.Description>{serverError()}</Alert.Description>
+			{/if}
 		</Alert.Root>
 	</div>
 {/if}

tools/ui/src/lib/components/ui/tooltip/tooltip-content.svelte

@@ -5,7 +5,7 @@
 	let {
 		ref = $bindable(null),
 		class: className,
-		sideOffset = 0,
+		sideOffset = 4,
 		side = 'top',
 		children,
 		arrowClasses,

tools/ui/src/lib/constants/pwa.ts

@@ -258,12 +258,6 @@ export const GLOB_PATTERNS: string[] = [
 	'**/*.{js,css,html,ico,svg,png,webp,woff,woff2,json,webmanifest}'
 ];
 
-// loading.html is the model loading page served by llama-server itself.
-// The SvelteKit PWA manifest transform strips the html extension from every
-// precache entry to match clean URLs, but loading.html is a plain static asset
-// with no clean URL, so static servers answer 404 and the SW install fails.
-export const GLOB_IGNORES: string[] = ['**/loading.html'];
-
 export const SW_CONFIG = {
 	CHECK_INTERVAL_MS: 60000,
 	UPDATE_FETCH_OPTIONS: {
@@ -317,7 +311,6 @@ export const SVELTEKIT_PWA_OPTIONS: SvelteKitPWAOptions = {
 		// Uses '**/' because SvelteKit outputs files under _app/immutable/
 		// subdirectories.
 		globPatterns: GLOB_PATTERNS,
-		globIgnores: GLOB_IGNORES,
 		maximumFileSizeToCacheInBytes: CACHE_SETTINGS.MAX_FILE_SIZE_BYTES,
 
 		// Prevent @vite-pwa/sveltekit from auto-adding a NavigationRoute by

tools/ui/src/lib/stores/conversations.svelte.ts

@@ -1115,21 +1115,18 @@ class ConversationsStore {
 	}
 
 	/**
-	 * Downloads a conversation as JSON file.
+	 * Downloads a single conversation as a JSONL file, serializing the full message tree.
 	 * @param convId - The conversation ID to download
 	 */
 	async downloadConversation(convId: string): Promise<void> {
-		let conversation: DatabaseConversation | null;
-		let messages: DatabaseMessage[];
+		const conversation =
+			this.activeConversation?.id === convId
+				? this.activeConversation
+				: await DatabaseService.getConversation(convId);
 
-		if (this.activeConversation?.id === convId) {
-			conversation = this.activeConversation;
-			messages = this.activeMessages;
-		} else {
-			conversation = await DatabaseService.getConversation(convId);
-			if (!conversation) return;
-			messages = await DatabaseService.getConversationMessages(convId);
-		}
+		if (!conversation) return;
+
+		const messages = await DatabaseService.getConversationMessages(convId);
 
 		this.downloadConversationFile({ conv: conversation, messages });
 	}

tools/ui/src/lib/stores/server.svelte.ts

@@ -1,5 +1,8 @@
 import { PropsService } from '$lib/services/props.service';
 import { ServerRole } from '$lib/enums';
+import { ApiError } from '$lib/utils/api-fetch';
+
+const LOADING_RETRY_INTERVAL_MS = 1000;
 
 /**
  * serverStore - Server connection state, configuration, and role detection
@@ -29,8 +32,10 @@ class ServerStore {
 	props = $state<ApiLlamaCppServerProps | null>(null);
 	loading = $state(false);
 	error = $state<string | null>(null);
+	status = $state<number | null>(null);
 	role = $state<ServerRole | null>(null);
 	private fetchPromise: Promise<void> | null = null;
+	private retryTimer: ReturnType<typeof setTimeout> | null = null;
 
 	/**
 	 *
@@ -70,23 +75,43 @@ class ServerStore {
 	 *
 	 */
 
-	async fetch(): Promise<void> {
+	/**
+	 * @param background - Set by the automatic "still loading" poll. Skips the
+	 * `loading` flag flip so the UI doesn't bounce between the full loading
+	 * splash and the chat screen every retry tick.
+	 */
+	async fetch({ background = false }: { background?: boolean } = {}): Promise<void> {
 		if (this.fetchPromise) return this.fetchPromise;
 
-		this.loading = true;
-		this.error = null;
+		this.clearRetryTimer();
+		if (!background) {
+			this.loading = true;
+		}
+		// Don't clear an existing "still loading" error before a retry -
+		// doing so would unmount/remount the error banner every second.
+		if (this.status !== 503) {
+			this.error = null;
+		}
 
 		const fetchPromise = (async () => {
 			try {
 				const props = await PropsService.fetch();
 				this.props = props;
 				this.error = null;
+				this.status = null;
 				this.detectRole(props);
 			} catch (error: unknown) {
 				this.error = error instanceof Error ? error.message : String(error);
+				this.status = error instanceof ApiError ? error.status : null;
 				console.error('Error fetching server properties:', error);
+
+				if (this.status === 503) {
+					this.scheduleRetry();
+				}
 			} finally {
-				this.loading = false;
+				if (!background) {
+					this.loading = false;
+				}
 				this.fetchPromise = null;
 			}
 		})();
@@ -96,13 +121,30 @@ class ServerStore {
 	}
 
 	clear(): void {
+		this.clearRetryTimer();
 		this.props = null;
 		this.error = null;
+		this.status = null;
 		this.loading = false;
 		this.role = null;
 		this.fetchPromise = null;
 	}
 
+	private scheduleRetry(): void {
+		if (this.retryTimer) return;
+		this.retryTimer = setTimeout(() => {
+			this.retryTimer = null;
+			this.fetch({ background: true });
+		}, LOADING_RETRY_INTERVAL_MS);
+	}
+
+	private clearRetryTimer(): void {
+		if (this.retryTimer) {
+			clearTimeout(this.retryTimer);
+			this.retryTimer = null;
+		}
+	}
+
 	/**
 	 *
 	 *
@@ -125,6 +167,7 @@ export const serverStore = new ServerStore();
 export const serverProps = () => serverStore.props;
 export const serverLoading = () => serverStore.loading;
 export const serverError = () => serverStore.error;
+export const serverStatus = () => serverStore.status;
 export const serverRole = () => serverStore.role;
 export const defaultParams = () => serverStore.defaultParams;
 export const contextSize = () => serverStore.contextSize;

tools/ui/src/lib/utils/api-fetch.ts

@@ -12,6 +12,21 @@ import { ERROR_MESSAGES, HTTP_CODE_TO_STRING } from '$lib/constants/error';
  * - Base path resolution
  */
 
+/**
+ * Error thrown when an API request fails, carrying the HTTP status code
+ * so callers can distinguish e.g. a 503 "still loading" response from a
+ * genuine failure.
+ */
+export class ApiError extends Error {
+	status: number;
+
+	constructor(message: string, status: number) {
+		super(message);
+		this.name = 'ApiError';
+		this.status = status;
+	}
+}
+
 export interface ApiFetchOptions extends Omit<RequestInit, 'headers'> {
 	/**
 	 * Use auth-only headers (no Content-Type).
@@ -67,7 +82,7 @@ export async function apiFetch<T>(path: string, options: ApiFetchOptions = {}):
 
 	if (!response.ok) {
 		const errorMessage = await parseErrorMessage(response);
-		throw new Error(errorMessage);
+		throw new ApiError(errorMessage, response.status);
 	}
 
 	return response.json() as Promise<T>;
@@ -119,7 +134,7 @@ export async function apiFetchWithParams<T>(
 
 	if (!response.ok) {
 		const errorMessage = await parseErrorMessage(response);
-		throw new Error(errorMessage);
+		throw new ApiError(errorMessage, response.status);
 	}
 
 	return response.json() as Promise<T>;

tools/ui/static/loading.html

@@ -1,12 +0,0 @@
-<!DOCTYPE html>
-<html>
-    <head>
-        <meta http-equiv="refresh" content="5">
-    </head>
-    <body>
-        <div id="loading">
-            The model is loading. Please wait.<br/>
-            The user interface will appear soon.
-        </div>
-    </body>
-</html>

tools/ui/tests/unit/pwa.spec.ts

@@ -189,9 +189,5 @@ describe('PWA Build Output', () => {
 			expect(existsSync(resolve(DIST_DIR, 'pwa-192x192.png'))).toBeTruthy();
 			expect(existsSync(resolve(DIST_DIR, 'pwa-512x512.png'))).toBeTruthy();
 		});
-
-		it('has loading.html fallback page', () => {
-			expect(existsSync(resolve(DIST_DIR, 'loading.html'))).toBeTruthy();
-		});
 	});
 });