perf(react-ui): code-split bundle, speed up coverage suite (#10042)

* Curate the highlight.js build to ~29 languages (lib/core + the
  common set) instead of the full ~190-grammar default: -787 KB raw /
  -230 KB gz on the base bundle.
* Code-split every route via React.lazy with a per-layout <Suspense>
  in App.jsx so the sidebar stays mounted on navigation. Initial entry
  chunk drops from 3194 KB raw / 887 KB gz to 397 KB / 122 KB (-87%).
  Warm chunks on sidebar hover/focus/touch via a preload registry so
  the click finds the chunk already in flight or cached.
* Migrate Playwright coverage from istanbul (build-time counters) to
  native Chromium V8 coverage, with per-worker accumulation +
  conversion. Suite drops from 71s to 30s at 20 workers (~58%) at the
  non-instrumented floor.
* Keep the coverage gate bundling-invariant: the coverage build inlines
  dynamic imports so every shipped source file lands in the denominator
  (otherwise untested page chunks silently drop out and inflate the
  percentage). Production builds stay code-split.
* Add UI_TEST_WORKERS=N Makefile knob; tighten coverage tolerance to
  0.8pp now that jitter sits near istanbul's ~0.5pp again.

Assisted-by: Claude:claude-opus-4-7 [Claude Code]

Signed-off-by: Richard Palethorpe <io@richiejp.com>

Makefile

@@ -1313,6 +1313,13 @@ build-ui-test-server: build-mock-backend react-ui protogen-go
 test-ui-e2e: build-ui-test-server
 	cd core/http/react-ui && npm install && npx playwright install --with-deps chromium && npx playwright test
 
+## Optional Playwright worker count for the UI e2e targets below. Pass
+## UI_TEST_WORKERS=N (e.g. `make test-ui-coverage UI_TEST_WORKERS=20`) to
+## override Playwright's default (cores/2). Empty by default so Playwright
+## picks its own worker count.
+UI_TEST_WORKERS ?=
+PLAYWRIGHT_WORKERS_FLAG = $(if $(UI_TEST_WORKERS),--workers=$(UI_TEST_WORKERS),)
+
 ## Fast Playwright e2e run used by the pre-commit hook on React UI changes.
 ## Force-rebuilds the (non-instrumented) dist so the suite tests the working
 ## tree — not a stale dist the `react-ui` skip-guard would leave — re-embeds
@@ -1322,22 +1329,24 @@ test-ui-e2e: build-ui-test-server
 test-ui: build-mock-backend protogen-go
 	cd core/http/react-ui && bun install && bun run build
 	$(GOCMD) build -o tests/e2e-ui/ui-test-server ./tests/e2e-ui
-	cd core/http/react-ui && sh $(CURDIR)/scripts/ensure-playwright-browser.sh && bunx playwright test
+	cd core/http/react-ui && sh $(CURDIR)/scripts/ensure-playwright-browser.sh && bunx playwright test $(PLAYWRIGHT_WORKERS_FLAG)
 
-## React UI code coverage from the Playwright e2e suite. Builds an
-## istanbul-instrumented bundle (COVERAGE=true), re-embeds it into the
-## ui-test-server (the dist is //go:embed'ed at compile time), runs the
-## Playwright specs — which harvest window.__coverage__ via the coverage
-## fixture — and writes an nyc report to core/http/react-ui/coverage/.
-## Removes the instrumented dist afterwards so normal builds aren't served
-## instrumented assets.
+## React UI code coverage from the Playwright e2e suite. Builds a
+## NON-instrumented bundle with source maps (COVERAGE_V8=true), re-embeds it
+## into the ui-test-server (the dist is //go:embed'ed at compile time), runs the
+## Playwright specs which collect native Chromium V8 coverage (PW_V8_COVERAGE=1)
+## — far cheaper than istanbul's build-time counters (~40% faster end-to-end) —
+## convert it to istanbul via v8-to-istanbul in the coverage fixture, and write
+## an nyc report to core/http/react-ui/coverage/. Removes the dist afterwards so
+## normal builds aren't served source-mapped assets. (The legacy istanbul path
+## still exists: `bun run build:coverage` + unset PW_V8_COVERAGE.)
 test-ui-coverage: build-mock-backend protogen-go
 	trap 'rm -rf "$(CURDIR)/core/http/react-ui/dist"' EXIT; \
-	( cd core/http/react-ui && bun install && bun run build:coverage ) && \
+	( cd core/http/react-ui && bun install && bun run build:coverage-v8 ) && \
 	$(GOCMD) build -o tests/e2e-ui/ui-test-server ./tests/e2e-ui && \
 	( cd core/http/react-ui && rm -rf .nyc_output coverage && \
 	    sh $(CURDIR)/scripts/ensure-playwright-browser.sh && \
-	    bunx playwright test && bun run coverage:report )
+	    PW_V8_COVERAGE=1 bunx playwright test $(PLAYWRIGHT_WORKERS_FLAG) && bun run coverage:report )
 
 ## UI coverage baseline (committed) and the strict gate that compares against
 ## it — the React mirror of test-coverage-baseline / test-coverage-check.

backend/cpp/ds4/Makefile

@@ -1,10 +1,10 @@
 # ds4 backend Makefile.
 #
-# Upstream pin lives below as DS4_VERSION?=e8e8779b261c10f36ad6270ba732c8f0be5b62e3
+# Upstream pin lives below as DS4_VERSION?=072bc0feb187be5f374c08b16d0045e1ad7bc9bc
 # (.github/bump_deps.sh) can find and update it - matches the
 # llama-cpp / ik-llama-cpp / turboquant convention.
 
-DS4_VERSION?=e8e8779b261c10f36ad6270ba732c8f0be5b62e3
+DS4_VERSION?=072bc0feb187be5f374c08b16d0045e1ad7bc9bc
 DS4_REPO?=https://github.com/antirez/ds4
 
 CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=d2da6da05c73aeb658a3d1751f386c24e6963856
+IK_LLAMA_VERSION?=3bf7e836c2c5a895e8d12d3eb7e398ae7ab2f9ce
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=0d18aaa9d1a8af3df9abccd828e22eeaac7f840b
+LLAMA_VERSION?=aa50b2c2ae91326d5aad956ceeb015d1d48e626b
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

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?=92dc7268fc4ffb0c0cc0bd52dfcefea91326e797
+STABLEDIFFUSION_GGML_VERSION?=29ab511fc75f89fbab148665eab1a8e10a139a72
 
 CMAKE_ARGS+=-DGGML_MAX_NAME=128
 

backend/go/whisper/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=27101c01dcac1676e2b6422256233cd0f1f9ae28
+WHISPER_CPP_VERSION?=6dcdd6536456158667747f724d6bd3a2ceaa8d88
 SO_TARGET?=libgowhisper.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

core/http/endpoints/openresponses/responses.go

@@ -95,7 +95,7 @@ func ResponsesEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, eval
 
 		// Add instructions as system message if provided
 		if input.Instructions != "" {
-			messages = append([]schema.Message{{Role: "system", StringContent: input.Instructions}}, messages...)
+			messages = append([]schema.Message{{Role: "system", Content: input.Instructions, StringContent: input.Instructions}}, messages...)
 		}
 
 		// Handle tools
@@ -299,7 +299,7 @@ func convertORInputToMessages(input any, cfg *config.ModelConfig) ([]schema.Mess
 	switch v := input.(type) {
 	case string:
 		// Simple string = user message
-		return []schema.Message{{Role: "user", StringContent: v}}, nil
+		return []schema.Message{{Role: "user", Content: v, StringContent: v}}, nil
 	case []any:
 		// Array of items
 		for _, itemRaw := range v {
@@ -309,6 +309,16 @@ func convertORInputToMessages(input any, cfg *config.ModelConfig) ([]schema.Mess
 			}
 
 			itemType, _ := itemMap["type"].(string)
+			// OpenAI SDK helpers (e.g. client.responses.create(input=[{"role":...,"content":...}]))
+			// send message items without a "type" discriminator. Treat a bare {role, content}
+			// object as type:"message" so the chat-completions and responses paths agree.
+			if itemType == "" {
+				if _, hasRole := itemMap["role"].(string); hasRole {
+					if _, hasContent := itemMap["content"]; hasContent {
+						itemType = "message"
+					}
+				}
+			}
 			switch itemType {
 			case "message":
 				msg, err := convertORMessageItem(itemMap, cfg)

core/http/endpoints/openresponses/responses_convert_test.go

@@ -0,0 +1,62 @@
+package openresponses
+
+import (
+	"github.com/mudler/LocalAI/core/config"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// Regression for mudler/LocalAI#10039. convertORInputToMessages must populate
+// both Content and StringContent: the templating fallback path reads
+// StringContent, while the UseTokenizerTemplate path serialises Content via
+// Messages.ToProto(). Leaving Content nil produced an empty prompt on any model
+// without a Go-side template.chat_message block (the default for imported GGUFs).
+var _ = Describe("convertORInputToMessages", func() {
+	cfg := &config.ModelConfig{}
+
+	It("populates both Content and StringContent for plain string input", func() {
+		msgs, err := convertORInputToMessages("Hello", cfg)
+		Expect(err).NotTo(HaveOccurred())
+		Expect(msgs).To(HaveLen(1))
+		Expect(msgs[0].Role).To(Equal("user"))
+		Expect(msgs[0].StringContent).To(Equal("Hello"))
+		Expect(msgs[0].Content).To(Equal("Hello"))
+	})
+
+	It("accepts a bare {role, content} item without a type discriminator", func() {
+		// The OpenAI Python SDK helper client.responses.create(input=[{...}])
+		// sends message items with no "type" field. They must not be dropped.
+		input := []any{
+			map[string]any{"role": "user", "content": "Hi there"},
+		}
+		msgs, err := convertORInputToMessages(input, cfg)
+		Expect(err).NotTo(HaveOccurred())
+		Expect(msgs).To(HaveLen(1))
+		Expect(msgs[0].Role).To(Equal("user"))
+		Expect(msgs[0].StringContent).To(Equal("Hi there"))
+		Expect(msgs[0].Content).To(Equal("Hi there"))
+	})
+
+	It("still populates both fields for an explicit type:message item", func() {
+		input := []any{
+			map[string]any{"type": "message", "role": "user", "content": "Typed"},
+		}
+		msgs, err := convertORInputToMessages(input, cfg)
+		Expect(err).NotTo(HaveOccurred())
+		Expect(msgs).To(HaveLen(1))
+		Expect(msgs[0].StringContent).To(Equal("Typed"))
+		Expect(msgs[0].Content).To(Equal("Typed"))
+	})
+
+	It("does not treat a non-message item (no content key) as a message", func() {
+		// An item with neither a known type nor a {role, content} shape must
+		// keep falling through unchanged — no behaviour change for such inputs.
+		input := []any{
+			map[string]any{"role": "user"},
+		}
+		msgs, err := convertORInputToMessages(input, cfg)
+		Expect(err).NotTo(HaveOccurred())
+		Expect(msgs).To(BeEmpty())
+	})
+})

core/http/react-ui/bun.lock

@@ -32,6 +32,7 @@
         "yaml": "^2.8.3",
       },
       "devDependencies": {
+        "@bcoe/v8-coverage": "^1.0.2",
         "@eslint/js": "^9.27.0",
         "@playwright/test": "1.58.2",
         "@vitejs/plugin-react": "^6.0.2",
@@ -41,6 +42,7 @@
         "globals": "^16.1.0",
         "i18next-parser": "^9.4.0",
         "nyc": "^18.0.0",
+        "v8-to-istanbul": "^9.3.0",
         "vite": "^8.0.14",
         "vite-plugin-istanbul": "^9.0.0",
       },
@@ -81,6 +83,8 @@
 
     "@babel/types": ["@babel/types@7.29.0", "", { "dependencies": { "@babel/helper-string-parser": "^7.27.1", "@babel/helper-validator-identifier": "^7.28.5" } }, "sha512-LwdZHpScM4Qz8Xw2iKSzS+cfglZzJGvofQICy7W7v4caru4EaAmyUuO6BGrbyQ2mYV11W0U8j5mBhd14dd3B0A=="],
 
+    "@bcoe/v8-coverage": ["@bcoe/v8-coverage@1.0.2", "", {}, "sha512-6zABk/ECA/QYSCQ1NGiVwwbQerUCZ+TQbp64Q3AgmfNvurHH0j8TtXa1qbShXA6qqkpAj4V5W8pP6mLe1mcMqA=="],
+
     "@codemirror/autocomplete": ["@codemirror/autocomplete@6.20.1", "", { "dependencies": { "@codemirror/language": "^6.0.0", "@codemirror/state": "^6.0.0", "@codemirror/view": "^6.17.0", "@lezer/common": "^1.0.0" } }, "sha512-1cvg3Vz1dSSToCNlJfRA2WSI4ht3K+WplO0UMOgmUYPivCyy2oueZY6Lx7M9wThm7SDUBViRmuT+OG/i8+ON9A=="],
 
     "@codemirror/commands": ["@codemirror/commands@6.10.3", "", { "dependencies": { "@codemirror/language": "^6.0.0", "@codemirror/state": "^6.6.0", "@codemirror/view": "^6.27.0", "@lezer/common": "^1.1.0" } }, "sha512-JFRiqhKu+bvSkDLI+rUhJwSxQxYb759W5GBezE8Uc8mHLqC9aV/9aTC7yJSqCtB3F00pylrLCwnyS91Ap5ej4Q=="],
@@ -267,6 +271,8 @@
 
     "@types/estree": ["@types/estree@1.0.8", "", {}, "sha512-dWHzHa2WqEXI/O1E9OjrocMTKJl2mSrEolh1Iomrv6U+JuNwaHXsXx9bLu5gG7BUWFIN0skIQJQ/L1rIex4X6w=="],
 
+    "@types/istanbul-lib-coverage": ["@types/istanbul-lib-coverage@2.0.6", "", {}, "sha512-2QF/t/auWm0lsy8XtKVPG19v3sSOQlJe/YHZgfjb/KBBHOGSV+J2q/S671rcq9uTBrLAXmZpqJiaQbMT+zNU1w=="],
+
     "@types/json-schema": ["@types/json-schema@7.0.15", "", {}, "sha512-5+fP8P8MFNC+AyZCDxrB2pkZFPGzqQWUzpSeuuVLvm8VMcorNYavBqoFcxK8bQz4Qsbn4oUEEem4wDLfcysGHA=="],
 
     "@types/minimatch": ["@types/minimatch@3.0.5", "", {}, "sha512-Klz949h02Gz2uZCMGwDUSDS1YBlTdDDgbWHi+81l29tQALUtvz4rAYi5uoVhE5Lagoq6DeqAUlbrHvW/mXDgdQ=="],
@@ -983,6 +989,8 @@
 
     "uuid": ["uuid@8.3.2", "", { "bin": { "uuid": "dist/bin/uuid" } }, "sha512-+NYs2QeMWy+GWFOEm9xnn6HCDp0l7QBD7ml8zLUmJ+93Q5NF0NocErnwkTkXVFNiX3/fpC6afS8Dhb/gz7R7eg=="],
 
+    "v8-to-istanbul": ["v8-to-istanbul@9.3.0", "", { "dependencies": { "@jridgewell/trace-mapping": "^0.3.12", "@types/istanbul-lib-coverage": "^2.0.1", "convert-source-map": "^2.0.0" } }, "sha512-kiGUalWN+rgBJ/1OHZsBtU4rXZOfj/7rKQxULKlIzwzQSvMJUUNgPwJEEh7gU6xEVxC0ahoOBvN2YI8GH6FNgA=="],
+
     "value-or-function": ["value-or-function@4.0.0", "", {}, "sha512-aeVK81SIuT6aMJfNo9Vte8Dw0/FZINGBV8BfCraGtqVxIeLAEhJyoWs8SmvRVmXfGss2PmmOwZCuBPbZR+IYWg=="],
 
     "vary": ["vary@1.1.2", "", {}, "sha512-BNGbWLfd0eUPabhkXUVm0j8uuvREyTh5ovRa/dyow/BqAbZJyC+5fU+IzQOzmAKzYqYRAISoRhdQr3eIZ/PXqg=="],
@@ -1121,6 +1129,8 @@
 
     "test-exclude/minimatch": ["minimatch@10.2.5", "", { "dependencies": { "brace-expansion": "^5.0.5" } }, "sha512-MULkVLfKGYDFYejP07QOurDLLQpcjk7Fw+7jXS2R2czRQzR56yHRveU5NDJEOviH+hETZKSkIk5c+T23GjFUMg=="],
 
+    "v8-to-istanbul/convert-source-map": ["convert-source-map@2.0.0", "", {}, "sha512-Kvp459HrV2FEJ1CAsi1Ku+MY3kasH19TFykTz2xWmMeq6bk2NU3XXvfJ+Q61m0xktWwt+1HSYf3JZsTms3aRJg=="],
+
     "vinyl-sourcemap/convert-source-map": ["convert-source-map@2.0.0", "", {}, "sha512-Kvp459HrV2FEJ1CAsi1Ku+MY3kasH19TFykTz2xWmMeq6bk2NU3XXvfJ+Q61m0xktWwt+1HSYf3JZsTms3aRJg=="],
 
     "vite-plugin-istanbul/espree": ["espree@11.2.0", "", { "dependencies": { "acorn": "^8.16.0", "acorn-jsx": "^5.3.2", "eslint-visitor-keys": "^5.0.1" } }, "sha512-7p3DrVEIopW1B1avAGLuCSh1jubc01H2JHc8B4qqGblmg5gI9yumBgACjWo4JlIc04ufug4xJ3SQI8HkS/Rgzw=="],

core/http/react-ui/coverage-baseline.txt

@@ -1 +1 @@
-30.66
+38.29

core/http/react-ui/e2e/coverage-fixtures.js

@@ -15,9 +15,41 @@ import { randomUUID } from 'node:crypto'
 import path from 'node:path'
 
 const COVERAGE_DIR = path.resolve(process.cwd(), '.nyc_output')
+const V8_COVERAGE = process.env.PW_V8_COVERAGE === '1'
+
+const withCoverage = base.extend({
+  // Worker-scoped V8 coverage accumulator: collects every test's native
+  // Chromium coverage and converts it to istanbul ONCE at worker teardown
+  // (conversion is expensive; see e2e/v8-coverage.js). null when V8 mode is off.
+  _v8acc: [
+    async ({}, use) => {
+      if (!V8_COVERAGE) {
+        await use(null)
+        return
+      }
+      const { createAccumulator } = await import('./v8-coverage.js')
+      const acc = createAccumulator()
+      await use(acc)
+      await acc.flush()
+    },
+    { scope: 'worker' },
+  ],
+
+  page: async ({ page, _v8acc }, use) => {
+    // V8 coverage path: collect native Chromium coverage (cheap), hand it to the
+    // worker accumulator on teardown. Avoids running an instrumented bundle.
+    if (V8_COVERAGE) {
+      const { startV8 } = await import('./v8-coverage.js')
+      await startV8(page)
+      await use(page)
+      try {
+        _v8acc.add(await page.coverage.stopJSCoverage())
+      } catch {
+        // page already closed — nothing to collect
+      }
+      return
+    }
 
-export const test = base.extend({
-  page: async ({ page }, use) => {
     await use(page)
 
     let coverage
@@ -37,4 +69,5 @@ export const test = base.extend({
   },
 })
 
+export const test = withCoverage
 export { expect }

core/http/react-ui/e2e/v8-coverage.js

@@ -0,0 +1,88 @@
+// V8 -> istanbul coverage harvest for the Playwright suite.
+//
+// When PW_V8_COVERAGE=1 the suite runs against a NON-instrumented build (built
+// with COVERAGE_V8=true, which only adds source maps). Chromium collects native
+// V8 coverage with near-zero runtime overhead; we convert it back to per-source
+// istanbul data via v8-to-istanbul (using the on-disk source maps), filter to
+// src/**, and write the same .nyc_output/*.json the istanbul path produced — so
+// `nyc report` and the strict baseline gate are unchanged.
+//
+// Conversion (v8-to-istanbul load() parses the large bundle source map) is the
+// expensive part, so we do NOT convert per test. Instead each worker collects
+// raw V8 coverage from every test, merges it with @bcoe/v8-coverage (which sums
+// counts and reconciles overlapping ranges correctly — applyCoverage can't be
+// called repeatedly, it pushes/overwrites), and converts ONCE at worker
+// teardown. That cuts conversions from ~152 (per test) to ~1 per worker.
+import v8toIstanbul from 'v8-to-istanbul'
+import libCoverage from 'istanbul-lib-coverage'
+import { mergeProcessCovs } from '@bcoe/v8-coverage'
+import { mkdirSync, writeFileSync, existsSync } from 'node:fs'
+import { randomUUID } from 'node:crypto'
+import path from 'node:path'
+
+const COVERAGE_DIR = path.resolve(process.cwd(), '.nyc_output')
+const DIST_ASSETS = path.resolve(process.cwd(), 'dist', 'assets')
+// Absolute app source dir. Match on this (not a bare "/src/" substring) — the
+// repo itself lives under .../go/src/..., so a substring check would collide.
+const SRC_DIR = path.resolve(process.cwd(), 'src') + path.sep
+// Only our own bundle chunks under /assets/*.js carry app source maps.
+const APP_CHUNK = /\/assets\/([^/?]+\.js)(\?|$)/
+
+export async function startV8(page) {
+  // resetOnNavigation:false so hard navigations (goto) within a test accumulate.
+  await page.coverage.startJSCoverage({ resetOnNavigation: false })
+}
+
+// One accumulator per worker (created by the worker-scoped fixture).
+export function createAccumulator() {
+  const processCovs = []
+
+  return {
+    // Called on each test teardown with that test's V8 coverage entries.
+    add(entries) {
+      const result = entries
+        .filter((e) => APP_CHUNK.test(e.url))
+        // Keep only structural fields (drop the ~1MB `source` per entry — it's
+        // re-read from disk at convert time — to bound per-worker memory).
+        .map((e) => ({ scriptId: e.scriptId || e.url, url: e.url, functions: e.functions }))
+      if (result.length) processCovs.push({ result })
+    },
+
+    // Called once at worker teardown: merge all tests' coverage, convert, write.
+    async flush() {
+      if (processCovs.length === 0) return
+      const merged = mergeProcessCovs(processCovs)
+      const map = libCoverage.createCoverageMap({})
+
+      for (const script of merged.result) {
+        const m = APP_CHUNK.exec(script.url)
+        if (!m) continue
+        const diskPath = path.join(DIST_ASSETS, m[1])
+        if (!existsSync(diskPath)) continue
+
+        // v8-to-istanbul auto-loads source + sibling .map from disk; the served
+        // bytes match dist, so the V8 ranges line up.
+        const converter = v8toIstanbul(diskPath, 0)
+        try {
+          await converter.load()
+          converter.applyCoverage(script.functions)
+          const data = converter.toIstanbul()
+          for (const [key, fileCov] of Object.entries(data)) {
+            // v8-to-istanbul keys are already absolute; keep only app sources.
+            if (!key.startsWith(SRC_DIR) || key.includes(`${path.sep}node_modules${path.sep}`)) continue
+            map.merge({ [key]: fileCov })
+          }
+        } catch {
+          // skip a chunk we couldn't convert
+        } finally {
+          converter.destroy()
+        }
+      }
+
+      const json = map.toJSON()
+      if (Object.keys(json).length === 0) return
+      mkdirSync(COVERAGE_DIR, { recursive: true })
+      writeFileSync(path.join(COVERAGE_DIR, `v8-${randomUUID()}.json`), JSON.stringify(json))
+    },
+  }
+}

core/http/react-ui/package.json

@@ -12,6 +12,7 @@
     "test:e2e": "playwright test",
     "test:e2e:ui": "playwright test --ui",
     "build:coverage": "COVERAGE=true vite build",
+    "build:coverage-v8": "COVERAGE_V8=true vite build",
     "coverage:report": "nyc report"
   },
   "dependencies": {
@@ -42,6 +43,7 @@
     "yaml": "^2.8.3"
   },
   "devDependencies": {
+    "@bcoe/v8-coverage": "^1.0.2",
     "@eslint/js": "^9.27.0",
     "@playwright/test": "1.58.2",
     "@vitejs/plugin-react": "^6.0.2",
@@ -51,6 +53,7 @@
     "globals": "^16.1.0",
     "i18next-parser": "^9.4.0",
     "nyc": "^18.0.0",
+    "v8-to-istanbul": "^9.3.0",
     "vite": "^8.0.14",
     "vite-plugin-istanbul": "^9.0.0"
   }

core/http/react-ui/src/App.jsx

@@ -1,4 +1,4 @@
-import { useState, useEffect, useRef } from 'react'
+import { useState, useEffect, useRef, Suspense } from 'react'
 import { Outlet, useLocation, useNavigate } from 'react-router-dom'
 import { useTranslation } from 'react-i18next'
 import Sidebar from './components/Sidebar'
@@ -122,7 +122,14 @@ export default function App() {
         </header>
         <div className="main-content-inner">
           <div className="page-transition" key={location.pathname}>
-            <Outlet context={{ addToast }} />
+            {/* Per-route Suspense catches React.lazy chunk loads (router.jsx)
+                here, inside the App layout. Without it, suspension would bubble
+                up to main.jsx's outer boundary and unmount the sidebar/header
+                on every navigation. fallback={null} keeps the shell stable; the
+                page-content area briefly blanks while the chunk arrives. */}
+            <Suspense fallback={null}>
+              <Outlet context={{ addToast }} />
+            </Suspense>
           </div>
         </div>
         {!isChatRoute && (

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

@@ -4,7 +4,7 @@ import { getArtifactIcon } from '../utils/artifacts'
 import { safeHref } from '../utils/url'
 import { copyToClipboard } from '../utils/clipboard'
 import DOMPurify from 'dompurify'
-import hljs from 'highlight.js'
+import hljs from '../utils/hljs'
 
 export default function CanvasPanel({ artifacts, selectedId, onSelect, onClose }) {
   const [showPreview, setShowPreview] = useState(true)

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

@@ -6,6 +6,7 @@ import LanguageSwitcher from './LanguageSwitcher'
 import { useAuth } from '../context/AuthContext'
 import { useBranding } from '../contexts/BrandingContext'
 import { apiUrl } from '../utils/basePath'
+import { preloadRoute } from '../router'
 
 const COLLAPSED_KEY = 'localai_sidebar_collapsed'
 const SECTIONS_KEY = 'localai_sidebar_sections'
@@ -85,6 +86,10 @@ const sections = [
 function NavItem({ item, onClose, collapsed }) {
   const { t } = useTranslation('nav')
   const label = t(item.labelKey)
+  // Warm the route's lazy chunk before the user clicks. Touch fires ~150ms
+  // before the synthetic click on mobile; mouseenter/focus cover desktop and
+  // keyboard. The underlying import() is memoised so multiple triggers are free.
+  const preload = () => preloadRoute(item.path)
   return (
     <NavLink
       to={item.path}
@@ -93,6 +98,9 @@ function NavItem({ item, onClose, collapsed }) {
         `nav-item ${isActive ? 'active' : ''}`
       }
       onClick={onClose}
+      onMouseEnter={preload}
+      onFocus={preload}
+      onTouchStart={preload}
       title={collapsed ? label : undefined}
     >
       <i className={`${item.icon} nav-icon`} />
@@ -296,6 +304,9 @@ export default function Sidebar({ isOpen, onClose }) {
               <button
                 className="sidebar-user-link"
                 onClick={() => { navigate('/app/account'); onClose?.() }}
+                onMouseEnter={() => preloadRoute('/app/account')}
+                onFocus={() => preloadRoute('/app/account')}
+                onTouchStart={() => preloadRoute('/app/account')}
                 title={t('accountSettings')}
               >
                 {user.avatarUrl ? (

core/http/react-ui/src/router.jsx

@@ -1,54 +1,81 @@
+import { lazy } from 'react'
 import { createBrowserRouter, Navigate, useParams } from 'react-router-dom'
 import { routerBasename } from './utils/basePath'
 import App from './App'
-import Home from './pages/Home'
-import Chat from './pages/Chat'
-import Models from './pages/Models'
-import Manage from './pages/Manage'
-import ImageGen from './pages/ImageGen'
-import VideoGen from './pages/VideoGen'
-import TTS from './pages/TTS'
-import Sound from './pages/Sound'
-import AudioTransform from './pages/AudioTransform'
-import Talk from './pages/Talk'
-import Backends from './pages/Backends'
-import Settings from './pages/Settings'
-import Traces from './pages/Traces'
-import P2P from './pages/P2P'
-import Agents from './pages/Agents'
-import AgentCreate from './pages/AgentCreate'
-import AgentChat from './pages/AgentChat'
-import AgentStatus from './pages/AgentStatus'
-import Collections from './pages/Collections'
-import CollectionDetails from './pages/CollectionDetails'
-import Skills from './pages/Skills'
-import SkillEdit from './pages/SkillEdit'
-import AgentJobs from './pages/AgentJobs'
-import AgentTaskDetails from './pages/AgentTaskDetails'
-import AgentJobDetails from './pages/AgentJobDetails'
-import ModelEditor from './pages/ModelEditor'
-// PipelineEditor removed — the Model Editor with templates handles all model types
-import ImportModel from './pages/ImportModel'
-import BackendLogs from './pages/BackendLogs'
-import Explorer from './pages/Explorer'
-import Login from './pages/Login'
-import FineTune from './pages/FineTune'
-import Quantize from './pages/Quantize'
-import Studio from './pages/Studio'
-import FaceRecognition from './pages/FaceRecognition'
-import VoiceRecognition from './pages/VoiceRecognition'
-import Nodes from './pages/Nodes'
-import NodeBackendLogs from './pages/NodeBackendLogs'
-import NotFound from './pages/NotFound'
-import Usage from './pages/Usage'
-import Users from './pages/Users'
-import Middleware from './pages/Middleware'
-import Account from './pages/Account'
 import RequireAdmin from './components/RequireAdmin'
 import RequireAuth from './components/RequireAuth'
 import RequireAuthEnabled from './components/RequireAuthEnabled'
 import RequireFeature from './components/RequireFeature'
 
+// Pages are code-split: each becomes its own chunk loaded on demand, so a route
+// no longer drags every other page (and its heavy deps — CodeMirror, the MCP
+// SDK, yaml, marked) into the initial bundle. The <Suspense> boundary in
+// App.jsx (around <Outlet/>) shows nothing while a chunk loads, keeping the
+// sidebar/header mounted.
+//
+// `page(key, loader)` registers the dynamic import under a route-segment key
+// (the first segment after /app/) so a NavLink can warm the chunk on hover via
+// `preloadRoute('/app/chat')`. Dynamic import() is memoised by the module
+// loader, so a preloaded chunk is reused — not re-fetched — when the user
+// actually navigates. Pages with `key: null` aren't sidebar-reachable; they
+// still code-split, they just won't be preloaded from the nav.
+const preloaders = {}
+function page(key, loader) {
+  if (key !== null) preloaders[key] = loader
+  return lazy(loader)
+}
+
+export function preloadRoute(path) {
+  if (!path) return
+  const m = path.match(/^\/app(?:\/([^/?#]*))?/)
+  if (!m) return
+  preloaders[m[1] ?? '']?.().catch(() => { /* network blip — real click will retry */ })
+}
+
+const Home = page('', () => import('./pages/Home'))
+const Chat = page('chat', () => import('./pages/Chat'))
+const Models = page('models', () => import('./pages/Models'))
+const Manage = page('manage', () => import('./pages/Manage'))
+const ImageGen = page('image', () => import('./pages/ImageGen'))
+const VideoGen = page('video', () => import('./pages/VideoGen'))
+const TTS = page('tts', () => import('./pages/TTS'))
+const Sound = page('sound', () => import('./pages/Sound'))
+const AudioTransform = page('transform', () => import('./pages/AudioTransform'))
+const Talk = page('talk', () => import('./pages/Talk'))
+const Backends = page('backends', () => import('./pages/Backends'))
+const Settings = page('settings', () => import('./pages/Settings'))
+const Traces = page('traces', () => import('./pages/Traces'))
+const P2P = page('p2p', () => import('./pages/P2P'))
+const Agents = page('agents', () => import('./pages/Agents'))
+const AgentCreate = page(null, () => import('./pages/AgentCreate'))
+const AgentChat = page(null, () => import('./pages/AgentChat'))
+const AgentStatus = page(null, () => import('./pages/AgentStatus'))
+const Collections = page('collections', () => import('./pages/Collections'))
+const CollectionDetails = page(null, () => import('./pages/CollectionDetails'))
+const Skills = page('skills', () => import('./pages/Skills'))
+const SkillEdit = page(null, () => import('./pages/SkillEdit'))
+const AgentJobs = page('agent-jobs', () => import('./pages/AgentJobs'))
+const AgentTaskDetails = page(null, () => import('./pages/AgentTaskDetails'))
+const AgentJobDetails = page(null, () => import('./pages/AgentJobDetails'))
+const ModelEditor = page(null, () => import('./pages/ModelEditor'))
+// PipelineEditor removed — the Model Editor with templates handles all model types
+const ImportModel = page(null, () => import('./pages/ImportModel'))
+const BackendLogs = page(null, () => import('./pages/BackendLogs'))
+const Explorer = page(null, () => import('./pages/Explorer'))
+const Login = page(null, () => import('./pages/Login'))
+const FineTune = page('fine-tune', () => import('./pages/FineTune'))
+const Quantize = page('quantize', () => import('./pages/Quantize'))
+const Studio = page('studio', () => import('./pages/Studio'))
+const FaceRecognition = page('face', () => import('./pages/FaceRecognition'))
+const VoiceRecognition = page('voice', () => import('./pages/VoiceRecognition'))
+const Nodes = page('nodes', () => import('./pages/Nodes'))
+const NodeBackendLogs = page(null, () => import('./pages/NodeBackendLogs'))
+const NotFound = page(null, () => import('./pages/NotFound'))
+const Usage = page('usage', () => import('./pages/Usage'))
+const Users = page('users', () => import('./pages/Users'))
+const Middleware = page('middleware', () => import('./pages/Middleware'))
+const Account = page('account', () => import('./pages/Account'))
+
 function BrowseRedirect() {
   const { '*': splat } = useParams()
   return <Navigate to={`/app/${splat || ''}`} replace />

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

@@ -1,6 +1,6 @@
 import { Marked } from 'marked'
 import DOMPurify from 'dompurify'
-import hljs from 'highlight.js'
+import hljs from './hljs'
 import { apiUrl } from './basePath'
 
 const FENCE_REGEX = /```(\w*)\n([\s\S]*?)```/g
@@ -119,12 +119,17 @@ export function getArtifactIcon(type, language) {
 const artifactMarked = new Marked({
   renderer: {
     code({ text, lang }) {
-      // Will be overridden per-call
+      // Match markdown.js's fallback: when the language is unknown (not in
+      // the curated hljs set, see utils/hljs.js), use highlightAuto so the
+      // block still picks up theme colors — otherwise the same fenced block
+      // would render differently in chat (auto-highlighted) vs artifact card
+      // (plain text).
       if (lang && hljs.getLanguage(lang)) {
         const highlighted = hljs.highlight(text, { language: lang }).value
         return `<pre><code class="hljs language-${lang}">${highlighted}</code></pre>`
       }
-      return `<pre><code>${text.replace(/</g, '&lt;').replace(/>/g, '&gt;')}</code></pre>`
+      const highlighted = hljs.highlightAuto(text).value
+      return `<pre><code class="hljs">${highlighted}</code></pre>`
     },
   },
   breaks: true,

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

@@ -0,0 +1,53 @@
+// Curated highlight.js build.
+//
+// `import hljs from 'highlight.js'` pulls in the full bundle — ~190 language
+// grammars, ~893 KB raw / ~294 KB gzip, the single biggest item in the app
+// bundle (measured). We render code blocks from chat/markdown/canvas only, and
+// only ever for a handful of common languages, so we import the lightweight
+// core and register just the grammars below. `highlightAuto` still works — it
+// auto-detects among the registered set, which covers what an LLM realistically
+// emits. Import hljs from THIS module, never directly from 'highlight.js'.
+import hljs from 'highlight.js/lib/core'
+
+import bash from 'highlight.js/lib/languages/bash'
+import c from 'highlight.js/lib/languages/c'
+import cpp from 'highlight.js/lib/languages/cpp'
+import csharp from 'highlight.js/lib/languages/csharp'
+import css from 'highlight.js/lib/languages/css'
+import diff from 'highlight.js/lib/languages/diff'
+import dockerfile from 'highlight.js/lib/languages/dockerfile'
+import go from 'highlight.js/lib/languages/go'
+import ini from 'highlight.js/lib/languages/ini'
+import java from 'highlight.js/lib/languages/java'
+import javascript from 'highlight.js/lib/languages/javascript'
+import json from 'highlight.js/lib/languages/json'
+import kotlin from 'highlight.js/lib/languages/kotlin'
+import lua from 'highlight.js/lib/languages/lua'
+import makefile from 'highlight.js/lib/languages/makefile'
+import markdown from 'highlight.js/lib/languages/markdown'
+import php from 'highlight.js/lib/languages/php'
+import plaintext from 'highlight.js/lib/languages/plaintext'
+import powershell from 'highlight.js/lib/languages/powershell'
+import python from 'highlight.js/lib/languages/python'
+import ruby from 'highlight.js/lib/languages/ruby'
+import rust from 'highlight.js/lib/languages/rust'
+import scss from 'highlight.js/lib/languages/scss'
+import shell from 'highlight.js/lib/languages/shell'
+import sql from 'highlight.js/lib/languages/sql'
+import swift from 'highlight.js/lib/languages/swift'
+import typescript from 'highlight.js/lib/languages/typescript'
+import xml from 'highlight.js/lib/languages/xml'
+import yaml from 'highlight.js/lib/languages/yaml'
+
+// Each grammar registers its own aliases (e.g. js→javascript, ts→typescript,
+// yml→yaml, html→xml, sh→bash, py→python), so hljs.getLanguage('js') resolves.
+const languages = {
+  bash, c, cpp, csharp, css, diff, dockerfile, go, ini, java, javascript,
+  json, kotlin, lua, makefile, markdown, php, plaintext, powershell, python,
+  ruby, rust, scss, shell, sql, swift, typescript, xml, yaml,
+}
+for (const [name, lang] of Object.entries(languages)) {
+  hljs.registerLanguage(name, lang)
+}
+
+export default hljs

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

@@ -1,6 +1,6 @@
 import { marked } from 'marked'
 import DOMPurify from 'dompurify'
-import hljs from 'highlight.js'
+import hljs from './hljs'
 
 marked.setOptions({
   highlight(code, lang) {

core/http/react-ui/vite.config.js

@@ -9,6 +9,11 @@ const backendUrl = process.env.LOCALAI_URL || 'http://localhost:8080'
 // fixture (e2e/coverage-fixtures.js). Off by default so normal/dev/prod builds
 // carry no instrumentation overhead.
 const coverage = process.env.COVERAGE === 'true'
+// COVERAGE_V8=true produces a NON-instrumented build with source maps, so the
+// Playwright coverage fixture can collect Chromium V8 coverage (near-zero
+// runtime overhead, unlike istanbul's build-time counters) and map it back to
+// source via v8-to-istanbul. Mutually exclusive with COVERAGE.
+const coverageV8 = process.env.COVERAGE_V8 === 'true'
 
 export default defineConfig({
   plugins: [
@@ -50,5 +55,20 @@ export default defineConfig({
   build: {
     outDir: 'dist',
     assetsDir: 'assets',
+    // Source maps are needed only to map V8 coverage back to original sources.
+    sourcemap: coverageV8,
+    rollupOptions: {
+      output: {
+        // The coverage build inlines all dynamic imports into a single chunk.
+        // The app is route-code-split (router.jsx uses React.lazy), so a normal
+        // build emits ~50 lazy chunks. V8 coverage only sees chunks a test
+        // actually loaded, so untested pages would silently drop out of the
+        // denominator and inflate the percentage. Bundling everything into one
+        // chunk for the coverage build keeps the denominator complete and the
+        // measurement invariant to how production is split. Production builds
+        // (COVERAGE_V8 unset) keep code-splitting for fast first paint.
+        inlineDynamicImports: coverageV8,
+      },
+    },
   },
 })

core/schema/message_test.go

@@ -332,5 +332,41 @@ var _ = Describe("LLM tests", func() {
 			// Should only extract text parts
 			Expect(protoMessages[0].Content).To(Equal("Hello"))
 		})
+
+		// Regression for mudler/LocalAI#10039: ToProto is the path taken by
+		// UseTokenizerTemplate backends (e.g. imported GGUFs, where the backend
+		// applies the GGUF's jinja template to the raw messages). It reads
+		// Content, not StringContent — so a message that only populated
+		// StringContent (the shape /v1/responses produced before the fix)
+		// reached the backend with empty content. These two cases pin that
+		// contract: Content is authoritative, and producers must set it.
+		It("emits empty content when only StringContent is set (Content nil)", func() {
+			messages := Messages{
+				{
+					Role:          "user",
+					StringContent: "Hello",
+				},
+			}
+
+			protoMessages := messages.ToProto()
+
+			Expect(protoMessages).To(HaveLen(1))
+			Expect(protoMessages[0].Content).To(BeEmpty())
+		})
+
+		It("carries Content through to proto regardless of StringContent", func() {
+			messages := Messages{
+				{
+					Role:          "user",
+					Content:       "Hello",
+					StringContent: "Hello",
+				},
+			}
+
+			protoMessages := messages.ToProto()
+
+			Expect(protoMessages).To(HaveLen(1))
+			Expect(protoMessages[0].Content).To(Equal("Hello"))
+		})
 	})
 })

core/templates/evaluator.go

@@ -111,7 +111,11 @@ func (e *Evaluator) TemplateMessages(input schema.OpenAIRequest, messages []sche
 			}
 		}
 		r := config.Roles[role]
-		contentExists := i.Content != nil && i.StringContent != ""
+		// Treat StringContent as the source of truth — every downstream fallback branch in this
+		// function reads StringContent, not Content. Gating on both with && silently drops
+		// messages that have StringContent set but Content nil (e.g. /v1/responses string-input
+		// before mudler/LocalAI#10039 fix).
+		contentExists := i.StringContent != ""
 
 		fcall := i.FunctionCall
 		if len(i.ToolCalls) > 0 {

core/templates/evaluator_test.go

@@ -218,4 +218,41 @@ var _ = Describe("Templates", func() {
 			})
 		}
 	})
+	// Regression test for mudler/LocalAI#10039: when a model has no Go-side
+	// TemplateConfig.ChatMessage block (e.g. backends that rely on the GGUF's
+	// jinja template), TemplateMessages falls through to the role-prefix path.
+	// That path must still render messages whose StringContent is populated but
+	// Content (any) is nil — which is the shape /v1/responses produced before
+	// the fix to convertORInputToMessages.
+	Context("fallback path with StringContent-only message (no ChatMessage template)", func() {
+		var evaluator *Evaluator
+		BeforeEach(func() {
+			evaluator = NewEvaluator("")
+		})
+		It("renders the role prefix and content when only StringContent is set", func() {
+			cfg := &config.ModelConfig{
+				TemplateConfig: config.TemplateConfig{},
+				Roles:          map[string]string{"user": "USER: "},
+			}
+			messages := []schema.Message{
+				{
+					Role:          "user",
+					StringContent: "hello",
+					// Content intentionally left nil — reproduces /v1/responses string-input.
+				},
+			}
+			templated := evaluator.TemplateMessages(schema.OpenAIRequest{}, messages, cfg, []functions.Function{}, false)
+			Expect(templated).To(Equal("USER: hello"), templated)
+		})
+		It("renders content even with no role mapping", func() {
+			cfg := &config.ModelConfig{
+				TemplateConfig: config.TemplateConfig{},
+			}
+			messages := []schema.Message{
+				{Role: "user", StringContent: "hello"},
+			}
+			templated := evaluator.TemplateMessages(schema.OpenAIRequest{}, messages, cfg, []functions.Function{}, false)
+			Expect(templated).To(Equal("hello"), templated)
+		})
+	})
 })

gallery/index.yaml

@@ -3,35 +3,7 @@
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
     - https://huggingface.co/Jackrong/Qwopus3.5-9B-Coder-MTP-GGUF
-  description: |
-    # 🌟 Qwopus3.5-9B-v3.5
-
-    ## 💡 Model Overview & v3.5 Design
-
-    Qwopus3.5-9B-v3.5 is a **data-scaled continuation** of the Qwopus3.5-9B-v3 model.
-
-    The training data in v3.5 is expanded to cover a broader range of domains, including mathematics, programming, puzzle-solving, multilingual dialogue, instruction-following, multi-turn interactions, and STEM-related tasks.
-
-    Qwopus3.5-9B-v3.5 is a reasoning-enhanced model based on **Qwen3.5-9B**, designed for:
-
-      - 🧩 Structured reasoning
-      - 🔧 Tool-augmented workflows
-      - 🔁 Multi-step agentic tasks
-      - ⚡ Token-efficient inference
-
-    Compared with Qwopus3.5-9B-v3, **3.5 version does not introduce a new architecture, RL stage, or template redesign**.
-
-    This version is trained with approximately **2× more SFT data**.
-
-    ## 🎯 Motivation & Generalization Insight
-
-    The motivation behind v3.5 comes from a simple observation:
-
-    > This work is motivated by the hypothesis that scaling high-quality SFT data may further enhance the generalization ability of large language models.
-
-    In earlier Qwopus3.5 experiments, structured reasoning was observed to improve both **accuracy and efficiency**:
-
-    ...
+  description: "# \U0001F31F Qwopus3.5-9B-v3.5\n\n## \U0001F4A1 Model Overview & v3.5 Design\n\nQwopus3.5-9B-v3.5 is a **data-scaled continuation** of the Qwopus3.5-9B-v3 model.\n\nThe training data in v3.5 is expanded to cover a broader range of domains, including mathematics, programming, puzzle-solving, multilingual dialogue, instruction-following, multi-turn interactions, and STEM-related tasks.\n\nQwopus3.5-9B-v3.5 is a reasoning-enhanced model based on **Qwen3.5-9B**, designed for:\n\n  - \U0001F9E9 Structured reasoning\n  - \U0001F527 Tool-augmented workflows\n  - \U0001F501 Multi-step agentic tasks\n  - ⚡ Token-efficient inference\n\nCompared with Qwopus3.5-9B-v3, **3.5 version does not introduce a new architecture, RL stage, or template redesign**.\n\nThis version is trained with approximately **2× more SFT data**.\n\n## \U0001F3AF Motivation & Generalization Insight\n\nThe motivation behind v3.5 comes from a simple observation:\n\n> This work is motivated by the hypothesis that scaling high-quality SFT data may further enhance the generalization ability of large language models.\n\nIn earlier Qwopus3.5 experiments, structured reasoning was observed to improve both **accuracy and efficiency**:\n\n...\n"
   license: "apache-2.0"
   tags:
     - llm
@@ -67,26 +39,7 @@
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:
     - https://huggingface.co/Jackrong/Qwopus3.6-27B-v2-MTP-GGUF
-  description: |
-    🪐 Qwopus3.6-27B-v2-MTP
-    MTP Release
-
-    Multi-Token Prediction reasoning model fine-tuned from Qwen3.6-27B
-
-    🧬 Trace Inversion & Negentropy
-    🧠 27B Parameters
-    ⚡ Speculative Decoding
-    🛠️ Coding / DevOps / Math
-
-    💡 What is Qwopus3.6-27B-v2-MTP?
-    🪐 Qwopus3.6-27B-v2-MTP is a speed-oriented reasoning release built on top of Qwen3.6-27B. It keeps the Qwopus line's focus on reconstructed reasoning traces, coding discipline, DevOps procedures, and mathematical derivations, while adding Multi-Token Prediction for faster generation. The goal is simple: preserve the depth and structure of a 27B reasoning model while making real interactive use noticeably faster.
-
-    ⚡ MTP DecodingAuxiliary future-token prediction improves throughput on long reasoning, code, math, and strict-format prompts.
-    🧩 Structured ReasoningInherits the Qwopus training recipe built around reconstructed step-by-step reasoning trajectories.
-    🧪 GB10 TestedValidated on a 30-question local benchmark across Logic, Coding, DevOps, Math, and Edge tasks.
-    🚀 Practical SpeedDesigned for workflows where strong answers matter, but waiting several extra minutes per task does not.
-
-    ...
+  description: "\U0001FA90 Qwopus3.6-27B-v2-MTP\nMTP Release\n\nMulti-Token Prediction reasoning model fine-tuned from Qwen3.6-27B\n\n\U0001F9EC Trace Inversion & Negentropy\n\U0001F9E0 27B Parameters\n⚡ Speculative Decoding\n\U0001F6E0️ Coding / DevOps / Math\n\n\U0001F4A1 What is Qwopus3.6-27B-v2-MTP?\n\U0001FA90 Qwopus3.6-27B-v2-MTP is a speed-oriented reasoning release built on top of Qwen3.6-27B. It keeps the Qwopus line's focus on reconstructed reasoning traces, coding discipline, DevOps procedures, and mathematical derivations, while adding Multi-Token Prediction for faster generation. The goal is simple: preserve the depth and structure of a 27B reasoning model while making real interactive use noticeably faster.\n\n⚡ MTP DecodingAuxiliary future-token prediction improves throughput on long reasoning, code, math, and strict-format prompts.\n\U0001F9E9 Structured ReasoningInherits the Qwopus training recipe built around reconstructed step-by-step reasoning trajectories.\n\U0001F9EA GB10 TestedValidated on a 30-question local benchmark across Logic, Coding, DevOps, Math, and Edge tasks.\n\U0001F680 Practical SpeedDesigned for workflows where strong answers matter, but waiting several extra minutes per task does not.\n\n...\n"
   license: "apache-2.0"
   tags:
     - llm
@@ -6208,6 +6161,7 @@
   files:
     - filename: rfdetr-nano-q8_0.gguf
       uri: huggingface://mudler/rfdetr-cpp-nano/rfdetr-nano-q8_0.gguf
+      sha256: 940084c60a780f1a19a51458ae3a601454b3b843675fa0713ff43ae5bccc0d9b
 - name: rfdetr-cpp-base
   url: github:mudler/LocalAI/gallery/virtual.yaml@master
   urls:
@@ -6233,6 +6187,7 @@
   files:
     - filename: rfdetr-base-f16.gguf
       uri: huggingface://mudler/rfdetr-cpp-base/rfdetr-base-f16.gguf
+      sha256: 8a68b21a90478564bcbb758557069a618d96e25e7c358207fd85ba45b90faf52
 - name: rfdetr-cpp-small
   url: github:mudler/LocalAI/gallery/virtual.yaml@master
   urls:

scripts/ui-coverage-check.sh

@@ -4,13 +4,16 @@
 #
 # Compares the total line coverage in an nyc coverage-summary.json against a
 # committed baseline and fails (exit 1) if it dropped by more than
-# UI_COVERAGE_TOLERANCE percentage points (default 1.0). The React UI e2e suite
+# UI_COVERAGE_TOLERANCE percentage points (default 0.8). The React UI e2e suite
 # drives the real app, so a removed feature or deleted spec shows up as a
 # coverage drop here.
 #
 # UI e2e line coverage is NOT deterministic: async/debounced paths (e.g. the
-# VRAM estimate's 500ms debounce) mean identical specs vary ~0.5pp run-to-run.
-# The tolerance absorbs that jitter; keep it just above the observed wobble.
+# VRAM estimate's 500ms debounce) mean identical specs vary run-to-run. With the
+# V8 path's single-chunk coverage build (vite.config.js inlineDynamicImports)
+# the observed wobble is ~0.5pp, similar to the old istanbul path. The tolerance
+# absorbs that jitter — keep it just above the observed wobble so a real ~1pp
+# regression still trips the gate.
 # (The Go gate carries a smaller tolerance for the same reason — its e2e slice.)
 #
 # When coverage rises meaningfully, regenerate and commit the baseline with:
@@ -19,7 +22,7 @@ set -eu
 
 summary="${1:?usage: ui-coverage-check.sh SUMMARY_JSON BASELINE_FILE}"
 baseline_file="${2:?usage: ui-coverage-check.sh SUMMARY_JSON BASELINE_FILE}"
-tolerance="${UI_COVERAGE_TOLERANCE:-1.0}"
+tolerance="${UI_COVERAGE_TOLERANCE:-0.8}"
 
 if [ ! -f "$summary" ]; then
 	echo "ui-coverage-check: coverage summary not found: $summary" >&2