common: mark --split-mode tensor as experimental (#21684)

* refactor: Build improvements

* chore: Formatting + package lock update

common/arg.cpp

@@ -2353,7 +2353,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "- none: use one GPU only\n"
         "- layer (default): split layers and KV across GPUs (pipelined)\n"
         "- row: split weight across GPUs by rows (parallelized)\n"
-        "- tensor: split weights and KV across GPUs (parallelized)",
+        "- tensor: split weights and KV across GPUs (parallelized, EXPERIMENTAL)",
         [](common_params & params, const std::string & value) {
             if (value == "none") {
                 params.split_mode = LLAMA_SPLIT_MODE_NONE;

common/chat.cpp

@@ -1083,7 +1083,9 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
 
     data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
     data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
-    data.supports_thinking = true;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "<|channel>thought";
+    data.thinking_end_tag   = "<channel|>";
 
     data.preserved_tokens = {
         "<|channel>",
@@ -1102,9 +1104,9 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
         auto start = p.rule("start", p.prefix(inputs.generation_prompt, "<|channel>"));
 
         if (extract_reasoning) {
-            p.rule("thought", p.literal("<|channel>thought\n") + p.reasoning(p.until("<channel|>")) + p.literal("<channel|>"));
+            p.rule("thought", p.literal("<|channel>thought") + p.space() + p.reasoning(p.until("<channel|>")) + p.literal("<channel|>"));
         } else {
-            p.rule("thought", p.content(p.literal("<|channel>thought\n") + p.until("<channel|>") + p.literal("<channel|>")));
+            p.rule("thought", p.content(p.literal("<|channel>thought") + p.space() + p.until("<channel|>") + p.literal("<channel|>")));
         }
 
         auto thought = (p.peek(p.literal("<|channel>")) + p.ref("thought")) | p.negate(p.literal("<|channel>"));

docs/backend/OPENVINO.md

@@ -3,7 +3,7 @@
 > [!NOTE]
 > Performance and memory optimizations, accuracy validation, broader quantization coverage, broader operator and model support are work in progress.
 
-[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware, including CPUs, GPUs, and NPUs, in the cloud, on-premises, and on the edge. [OpenVINO backend for llama.cpp](../../src/ggml-openvino) enables hardware-accelerated inference on **Intel® CPUs, GPUs, and NPUs** while remaining compatible with the existing **GGUF model ecosystem**. The backend translates GGML compute graphs into OpenVINO graphs and leverages graph compilation, kernel fusion, and device-specific optimizations to improve inference performance on supported Intel hardware.
+[OpenVINO](https://docs.openvino.ai/) is an open-source toolkit for optimizing and deploying high-performance AI inference, specifically designed for Intel hardware, including CPUs, GPUs, and NPUs, in the cloud, on-premises, and on the edge. [OpenVINO backend for llama.cpp](../../ggml/src/ggml-openvino) enables hardware-accelerated inference on **Intel® CPUs, GPUs, and NPUs** while remaining compatible with the existing **GGUF model ecosystem**. The backend translates GGML compute graphs into OpenVINO graphs and leverages graph compilation, kernel fusion, and device-specific optimizations to improve inference performance on supported Intel hardware.
 
 The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a translation layer for core GGML operations. The OpenVINO backend replaces the standard GGML graph execution path with Intel's OpenVINO inference engine. This approach allows the same GGUF model file to run on Intel CPUs, Intel GPUs (integrated and discrete), and Intel NPUs without changes to the model or the rest of the llama.cpp stack. When a `ggml_cgraph` is dispatched to OpenVINO backend, it:
 

tests/test-chat.cpp

@@ -1988,6 +1988,13 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect(message_assist_thoughts)
             .run();
 
+        // Empty reasoning (budget=0: sampler forces end tag before newline)
+        tst.test(
+                "<|channel>thought<channel|>Hello, world!\nWhat's up?")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .expect(simple_assist_msg("Hello, world!\nWhat's up?", ""))
+            .run();
+
         // Reasoning and content with reasoning_format = none
         tst.test(
                 "<|channel>thought\nI'm\nthinking<channel|>Hello, world!\nWhat's up?")

tools/server/public/index.html

@@ -1,5 +1,5 @@
 <!--
-  This is a single file build of the frontend.
+  This is a static build of the frontend.
   It is automatically generated by the build process.
   Do not edit this file directly.
   To make changes, refer to the "Web UI" section in the README.
@@ -18,7 +18,7 @@
 		<div style="display: contents">
 			<script>
 				{
-					__sveltekit_1ao0o9h = {
+					__sveltekit__ = {
 						base: new URL('.', location).pathname.slice(0, -1)
 					};
 

tools/server/webui/package-lock.json

@@ -1,11 +1,11 @@
 {
-	"name": "webui",
+	"name": "llama-server-webui",
 	"version": "1.0.0",
 	"lockfileVersion": 3,
 	"requires": true,
 	"packages": {
 		"": {
-			"name": "webui",
+			"name": "llama-server-webui",
 			"version": "1.0.0",
 			"dependencies": {
 				"@modelcontextprotocol/sdk": "^1.25.1",

tools/server/webui/package.json

@@ -1,5 +1,5 @@
 {
-	"name": "webui",
+	"name": "llama-server-webui",
 	"private": true,
 	"version": "1.0.0",
 	"type": "module",

tools/server/webui/scripts/vite-plugin-llama-cpp-build.ts

@@ -0,0 +1,84 @@
+import { readFileSync, writeFileSync, existsSync, readdirSync, copyFileSync } from 'fs';
+import { resolve } from 'path';
+import type { Plugin } from 'vite';
+
+const GUIDE_FOR_FRONTEND = `
+<!--
+  This is a static build of the frontend.
+  It is automatically generated by the build process.
+  Do not edit this file directly.
+  To make changes, refer to the "Web UI" section in the README.
+-->
+`.trim();
+
+export function llamaCppBuildPlugin(): Plugin {
+	return {
+		name: 'llamacpp:build',
+		apply: 'build',
+		closeBundle() {
+			// Ensure the SvelteKit adapter has finished writing to ../public
+			setTimeout(() => {
+				try {
+					const indexPath = resolve('../public/index.html');
+					if (!existsSync(indexPath)) return;
+
+					let content = readFileSync(indexPath, 'utf-8');
+
+					const faviconPath = resolve('static/favicon.svg');
+
+					if (existsSync(faviconPath)) {
+						const faviconContent = readFileSync(faviconPath, 'utf-8');
+						const faviconBase64 = Buffer.from(faviconContent).toString('base64');
+						const faviconDataUrl = `data:image/svg+xml;base64,${faviconBase64}`;
+
+						content = content.replace(/href="[^"]*favicon\.svg"/g, `href="${faviconDataUrl}"`);
+
+						console.log('✓ Inlined favicon.svg as base64 data URL');
+					}
+
+					content = content.replace(/\r/g, '');
+					content = GUIDE_FOR_FRONTEND + '\n' + content;
+					content = content.replace(/\/_app\/immutable\/bundle\.[^"]+\.js/g, './bundle.js');
+					content = content.replace(
+						/\/_app\/immutable\/assets\/bundle\.[^"]+\.css/g,
+						'./bundle.css'
+					);
+					content = content.replace(/__sveltekit_[a-z0-9]+/g, '__sveltekit__');
+
+					writeFileSync(indexPath, content, 'utf-8');
+					console.log('✓ Updated index.html');
+
+					// Copy bundle.*.js -> ../public/bundle.js
+					const immutableDir = resolve('../public/_app/immutable');
+					const bundleDir = resolve('../public/_app/immutable/assets');
+
+					if (existsSync(immutableDir)) {
+						const jsFiles = readdirSync(immutableDir).filter((f) => f.match(/^bundle\..+\.js$/));
+
+						if (jsFiles.length > 0) {
+							copyFileSync(resolve(immutableDir, jsFiles[0]), resolve('../public/bundle.js'));
+							// Normalize __sveltekit_<hash> to __sveltekit__ in bundle.js
+							const bundleJsPath = resolve('../public/bundle.js');
+							let bundleJs = readFileSync(bundleJsPath, 'utf-8');
+							bundleJs = bundleJs.replace(/__sveltekit_[a-z0-9]+/g, '__sveltekit__');
+							writeFileSync(bundleJsPath, bundleJs, 'utf-8');
+							console.log(`✓ Copied ${jsFiles[0]} -> bundle.js`);
+						}
+					}
+
+					// Copy bundle.*.css -> ../public/bundle.css
+					if (existsSync(bundleDir)) {
+						const cssFiles = readdirSync(bundleDir).filter((f) => f.match(/^bundle\..+\.css$/));
+
+						if (cssFiles.length > 0) {
+							copyFileSync(resolve(bundleDir, cssFiles[0]), resolve('../public/bundle.css'));
+							console.log(`✓ Copied ${cssFiles[0]} -> bundle.css`);
+						}
+					}
+				} catch (error) {
+					console.error('Failed to update index.html:', error);
+				}
+			}, 100);
+		}
+	};
+}

tools/server/webui/src/lib/components/app/chat/ChatForm/ChatFormHelperText.svelte

@@ -22,7 +22,8 @@
 			</p>
 		{:else}
 			<p class="text-xs text-muted-foreground">
-				Press <kbd class="rounded bg-muted px-1 py-0.5 font-mono text-xs">{modKey} + Enter</kbd> to send,
+				Press <kbd class="rounded bg-muted px-1 py-0.5 font-mono text-xs">{modKey} + Enter</kbd> to
+				send,
 				<kbd class="rounded bg-muted px-1 py-0.5 font-mono text-xs">Enter</kbd> for new line
 			</p>
 		{/if}

tools/server/webui/svelte.config.js

@@ -25,6 +25,9 @@ const config = {
 		},
 		alias: {
 			$styles: 'src/styles'
+		},
+		version: {
+			name: 'llama-server-webui'
 		}
 	},
 

tools/server/webui/vite.config.ts

@@ -1,108 +1,33 @@
 import tailwindcss from '@tailwindcss/vite';
 import { sveltekit } from '@sveltejs/kit/vite';
-import { readFileSync, writeFileSync, existsSync, readdirSync, copyFileSync } from 'fs';
 import { dirname, resolve } from 'path';
 import { fileURLToPath } from 'url';
 
 import { defineConfig, searchForWorkspaceRoot } from 'vite';
 import devtoolsJson from 'vite-plugin-devtools-json';
 import { storybookTest } from '@storybook/addon-vitest/vitest-plugin';
+import { llamaCppBuildPlugin } from './scripts/vite-plugin-llama-cpp-build';
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 
-const GUIDE_FOR_FRONTEND = `
-<!--
-  This is a single file build of the frontend.
-  It is automatically generated by the build process.
-  Do not edit this file directly.
-  To make changes, refer to the "Web UI" section in the README.
--->
-`.trim();
-
-/**
- * the maximum size of an embedded asset in bytes,
- * e.g. maximum size of embedded font (see node_modules/katex/dist/fonts/*.woff2)
- */
-const MAX_ASSET_SIZE = 32000;
-
-/** public/index.html minified flag */
-const ENABLE_JS_MINIFICATION = true;
-
-function llamaCppBuildPlugin() {
-	return {
-		name: 'llamacpp:build',
-		apply: 'build' as const,
-		closeBundle() {
-			// Ensure the SvelteKit adapter has finished writing to ../public
-			setTimeout(() => {
-				try {
-					const indexPath = resolve('../public/index.html');
-
-					if (!existsSync(indexPath)) {
-						return;
-					}
-
-					let content = readFileSync(indexPath, 'utf-8');
-
-					const faviconPath = resolve('static/favicon.svg');
-					if (existsSync(faviconPath)) {
-						const faviconContent = readFileSync(faviconPath, 'utf-8');
-						const faviconBase64 = Buffer.from(faviconContent).toString('base64');
-						const faviconDataUrl = `data:image/svg+xml;base64,${faviconBase64}`;
-
-						content = content.replace(/href="[^"]*favicon\.svg"/g, `href="${faviconDataUrl}"`);
-
-						console.log('✓ Inlined favicon.svg as base64 data URL');
-					}
-
-					content = content.replace(/\r/g, '');
-					content = GUIDE_FOR_FRONTEND + '\n' + content;
-					content = content.replace(/\/_app\/immutable\/bundle\.[^"]+\.js/g, './bundle.js');
-					content = content.replace(
-						/\/_app\/immutable\/assets\/bundle\.[^"]+\.css/g,
-						'./bundle.css'
-					);
-
-					writeFileSync(indexPath, content, 'utf-8');
-					console.log('✓ Updated index.html');
-
-					// Copy bundle.*.js -> ../public/bundle.js
-					const immutableDir = resolve('../public/_app/immutable');
-					const bundleDir = resolve('../public/_app/immutable/assets');
-					if (existsSync(immutableDir)) {
-						const jsFiles = readdirSync(immutableDir).filter((f) => f.match(/^bundle\..+\.js$/));
-						if (jsFiles.length > 0) {
-							copyFileSync(resolve(immutableDir, jsFiles[0]), resolve('../public/bundle.js'));
-							console.log(`✓ Copied ${jsFiles[0]} -> bundle.js`);
-						}
-					}
-					// Copy bundle.*.css -> ../public/bundle.css
-					if (existsSync(bundleDir)) {
-						const cssFiles = readdirSync(bundleDir).filter((f) => f.match(/^bundle\..+\.css$/));
-						if (cssFiles.length > 0) {
-							copyFileSync(resolve(bundleDir, cssFiles[0]), resolve('../public/bundle.css'));
-							console.log(`✓ Copied ${cssFiles[0]} -> bundle.css`);
-						}
-					}
-				} catch (error) {
-					console.error('Failed to update index.html:', error);
-				}
-			}, 100);
-		}
-	};
-}
-
 export default defineConfig({
 	resolve: {
 		alias: {
 			'katex-fonts': resolve('node_modules/katex/dist/fonts')
 		}
 	},
+
 	build: {
-		assetsInlineLimit: MAX_ASSET_SIZE,
+		assetsInlineLimit: 32000,
 		chunkSizeWarningLimit: 3072,
-		minify: ENABLE_JS_MINIFICATION
+		minify: true
 	},
+
+	esbuild: {
+		lineLimit: 500,
+		minifyIdentifiers: false
+	},
+
 	css: {
 		preprocessorOptions: {
 			scss: {
@@ -114,7 +39,9 @@ export default defineConfig({
 			}
 		}
 	},
+
 	plugins: [tailwindcss(), sveltekit(), devtoolsJson(), llamaCppBuildPlugin()],
+
 	test: {
 		projects: [
 			{
@@ -131,6 +58,7 @@ export default defineConfig({
 					setupFiles: ['./vitest-setup-client.ts']
 				}
 			},
+
 			{
 				extends: './vite.config.ts',
 				test: {
@@ -139,6 +67,7 @@ export default defineConfig({
 					include: ['tests/unit/**/*.{test,spec}.{js,ts}']
 				}
 			},
+
 			{
 				extends: './vite.config.ts',
 				test: {