Release KevinTex 1.2.10 with Windows startup fix

.github/workflows/linux-release.yml

@@ -40,7 +40,10 @@ jobs:
         run: |
           python3 -m pip install pillow fastapi python-multipart
           python3 -m unittest discover -v
-          python3 -m compileall -q app.py backend_gemma.py backend_mlx.py backend_vlm.py image_preprocessing.py
+          python3 -m compileall -q app.py backend_gemma.py backend_gemma_cloud.py backend_mlx.py backend_vlm.py provider_config.py image_preprocessing.py snip.py
+          bash -n packaging/build-deb.sh packaging/install-user.sh packaging/localtex-launcher run.sh
+          shellcheck packaging/build-deb.sh packaging/install-user.sh packaging/localtex-launcher packaging/macos/build.sh run.sh
+          sed -n '/<script>/,/<\/script>/p' static/index.html | sed '/<script>/d;/<\/script>/d' | node --check
 
       - name: Build Debian package
         run: |

.github/workflows/macos-release.yml

@@ -40,7 +40,7 @@ jobs:
       - name: Run tests
         run: |
           python -m unittest discover -v
-          python -m py_compile backend_mlx.py packaging/macos/desktop_macos.py
+          python -m py_compile app.py backend_gemma_cloud.py backend_mlx.py provider_config.py packaging/macos/desktop_macos.py
 
       - name: Build app, ZIP, and DMG
         env:

.github/workflows/windows-release.yml

@@ -15,6 +15,7 @@ on:
         options:
           - cpu
           - cuda
+          - vulkan
 
 permissions:
   contents: write
@@ -71,38 +72,111 @@ jobs:
           python -m pip install --upgrade pip
           python -m pip install -r requirements-windows.txt
 
-      - name: Replace runtime with CUDA wheels
-        if: env.ACCELERATION == 'cuda'
+      - name: Install Vulkan SDK
+        if: env.ACCELERATION == 'vulkan'
         run: |
-          python -m pip uninstall -y llama-cpp-python torch
-          python -m pip install torch --index-url https://download.pytorch.org/whl/cu124
-          python -m pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124
-          Set-Content -Path packaging/windows/cuda_enabled.txt -Value "CUDA build"
+          choco install vulkan-sdk -y --no-progress
+          $sdk = Get-ChildItem "C:\VulkanSDK" -Directory | Sort-Object Name -Descending | Select-Object -First 1
+          if (-not $sdk) { throw "Vulkan SDK installation was not found." }
+          "VULKAN_SDK=$($sdk.FullName)" >> $env:GITHUB_ENV
+          "$($sdk.FullName)\Bin" >> $env:GITHUB_PATH
+
+      - name: Select llama.cpp acceleration backend
+        run: |
+          & ./packaging/windows/build-llama-backend.ps1 -Acceleration $env:ACCELERATION
+
+      - name: Verify acceleration runtime
+        run: |
+          if ($env:ACCELERATION -eq "cuda") {
+            python -c "import torch; assert torch.version.cuda, 'CPU PyTorch wheel remained in CUDA build'; print('PyTorch CUDA', torch.version.cuda)"
+          }
+          @'
+          import os
+          from pathlib import Path
+
+          acceleration = os.environ["ACCELERATION"]
+          if acceleration in {"cuda", "vulkan"}:
+              import importlib.util
+
+              llama_spec = importlib.util.find_spec("llama_cpp")
+              if llama_spec is None or not llama_spec.submodule_search_locations:
+                  raise ModuleNotFoundError("llama_cpp package is missing")
+              llama_lib = Path(next(iter(llama_spec.submodule_search_locations))) / "lib"
+          if acceleration == "cuda":
+              import torch
+
+              cuda_backend = llama_lib / "ggml-cuda.dll"
+              if not cuda_backend.is_file():
+                  raise FileNotFoundError("CUDA llama.cpp backend DLL is missing")
+              imports = cuda_backend.read_bytes().lower()
+              for dependency in (b"cudart64_12.dll", b"cublas64_12.dll", b"nvcuda.dll"):
+                  if dependency not in imports:
+                      raise RuntimeError(f"{dependency.decode()} is not linked by ggml-cuda.dll")
+              torch_lib = Path(torch.__file__).parent / "lib"
+              for dependency in ("cudart64_12.dll", "cublas64_12.dll"):
+                  if not (torch_lib / dependency).is_file():
+                      raise FileNotFoundError(f"PyTorch CUDA runtime is missing {dependency}")
+              print(f"Verified CUDA backend and runtime DLLs in {llama_lib}")
+          elif acceleration == "vulkan":
+              vulkan_backend = llama_lib / "ggml-vulkan.dll"
+              if not vulkan_backend.is_file():
+                  raise FileNotFoundError("Vulkan llama.cpp backend DLL is missing")
+              if b"vulkan-1.dll" not in vulkan_backend.read_bytes().lower():
+                  raise RuntimeError("ggml-vulkan.dll is not linked to the Vulkan loader")
+              print(f"Verified Vulkan backend DLL in {llama_lib}")
+          else:
+              import llama_cpp
+              from llama_cpp import llama_cpp as native
+
+              info = native.llama_print_system_info().decode("utf-8", "replace")
+              print(info)
+          '@ | python -
 
       - name: Run source checks
         run: |
           python -m py_compile desktop.py packaging/windows/generate_icon.py packaging/windows/generate_version_info.py
-          python -c "import ast, pathlib; [ast.parse(pathlib.Path(p).read_text(encoding='utf-8')) for p in ('app.py', 'backend_gemma.py', 'backend_vlm.py', 'image_preprocessing.py', 'snip.py')]"
+          $tokens = $null
+          $parseErrors = $null
+          $null = [System.Management.Automation.Language.Parser]::ParseFile("packaging/windows/build-llama-backend.ps1", [ref]$tokens, [ref]$parseErrors)
+          if ($parseErrors) { throw ($parseErrors | Out-String) }
+          python -c "import ast, pathlib; [ast.parse(pathlib.Path(p).read_text(encoding='utf-8')) for p in ('app.py', 'backend_gemma.py', 'backend_gemma_cloud.py', 'backend_mlx.py', 'backend_vlm.py', 'provider_config.py', 'image_preprocessing.py', 'snip.py')]"
+          python -m unittest discover -v
           python packaging/windows/generate_version_info.py $env:RELEASE_TAG
           python packaging/windows/generate_icon.py
 
       - name: Build native desktop application
         run: |
           pyinstaller --clean --noconfirm packaging/windows/kevintex.spec
+          Copy-Item -Force "packaging/windows/KevinTex.exe.config" "dist/KevinTex/KevinTex.exe.config"
 
       - name: Smoke-test packaged application
         run: |
+          if (-not (Test-Path "dist/KevinTex/KevinTex.exe.config")) {
+            throw "Packaged .NET application configuration is missing."
+          }
           $process = Start-Process -FilePath "dist/KevinTex/KevinTex.exe" -ArgumentList "--smoke-test" -PassThru -Wait
           if ($process.ExitCode -ne 0) {
             throw "Packaged application smoke test failed with exit code $($process.ExitCode)."
           }
           if (-not (Test-Path "dist/KevinTex/_internal/static/index.html")) {
             throw "Packaged frontend data is missing."
           }
+          if ($env:ACCELERATION -eq "cuda") {
+            foreach ($dependency in @("ggml-cuda.dll", "cudart64_12.dll", "cublas64_12.dll")) {
+              if (-not (Get-ChildItem "dist/KevinTex" -Recurse -Filter $dependency -File)) {
+                throw "Packaged CUDA build is missing $dependency."
+              }
+            }
+          }
+          if ($env:ACCELERATION -eq "vulkan") {
+            if (-not (Get-ChildItem "dist/KevinTex" -Recurse -Filter "ggml-vulkan.dll" -File)) {
+              throw "Packaged Vulkan build is missing ggml-vulkan.dll."
+            }
+          }
 
       - name: Create versioned archive
         run: |
-          $suffix = if ($env:ACCELERATION -eq "cuda") { "-cuda" } else { "" }
+          $suffix = if ($env:ACCELERATION -eq "cpu") { "" } else { "-$env:ACCELERATION" }
           if ($env:ACCELERATION -eq "cuda") {
             # GitHub release assets are capped below 2 GiB. The CUDA runtime is
             # large, so publish standard 7-Zip multi-volume files that extract
@@ -112,7 +186,7 @@ jobs:
             if ($LASTEXITCODE -ne 0) { throw "7-Zip packaging failed." }
             "ARCHIVE_GLOB=$archive.*" >> $env:GITHUB_ENV
           } else {
-            $archive = "KevinTex-$env:VERSION-windows-x64.zip"
+            $archive = "KevinTex-$env:VERSION-windows-x64$suffix.zip"
             Compress-Archive -Path "dist/KevinTex/*" -DestinationPath $archive
             "ARCHIVE_GLOB=$archive" >> $env:GITHUB_ENV
           }

README.md

@@ -2,20 +2,23 @@
   <img src="assets/kevintex-banner.png" alt="KevinTex — Screenshot to LaTeX" width="900">
 </p>
 
-# KevinTex — Snip & Get, fully offline
+# KevinTex — Snip & Get
 
 A local, free, unlimited formula-image → LaTeX converter (a self-hosted
 alternative to SimpleTex). Paste, drop, or **snip** a screenshot of a math
 formula and get editable Markdown + LaTeX with a live rendered preview.
-Recognition runs entirely on your machine using the quantized
+Recognition defaults to the quantized
 [Gemma 4 E2B-it](https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF)
-vision-language model through llama.cpp — no cloud APIs, accounts, or quotas.
+vision-language model through llama.cpp. On first launch, you can either
+download those local weights or use hosted Gemma 4 through a Google AI Studio
+API key.
 
 ## Features
 
 - **Snip button** (the computer-with-+ icon) — click it, drag a screen region,
-  and the capture is converted instantly. Fully self-contained (Pillow +
-  tkinter region selector); no `gnome-screenshot`/`flameshot`/portal needed.
+  and the capture is converted instantly. It uses Pillow + tkinter directly,
+  with optional `gnome-screenshot`, `grim`, `spectacle`, or `ffmpeg` fallbacks
+  when the desktop session blocks Pillow access.
 - **Paste (Ctrl+V), drag-and-drop, or browse** for a formula image (PNG/JPG/BMP/WEBP)
 - **Text sharpener** with persisted Off / Auto / Strong modes — conservative
   contrast normalization, smart upscaling, scan denoising, and edge sharpening
@@ -25,7 +28,8 @@ vision-language model through llama.cpp — no cloud APIs, accounts, or quotas.
   fraction/root/integral/matrix starters. Drawings go directly to local Gemma.
 - **Voice-to-LaTeX** — dictate a formula through the microphone; Gemma turns
   spoken math into editable Markdown + LaTeX entirely on-device through MLX on
-  Apple Silicon and llama.cpp on Windows/Linux.
+  Apple Silicon and llama.cpp on Windows/Linux, or through the audio-capable
+  Gemini model when Google AI Studio is selected.
 - **Image preparation tools** — rotate before OCR, invert dark screenshots, and
   compare the original with the exact processed image sent to recognition.
 - **Live KaTeX preview** (bundled locally — the app works with no internet at all)
@@ -37,22 +41,31 @@ vision-language model through llama.cpp — no cloud APIs, accounts, or quotas.
 - **Recognize again** using the current source image and active Thinking setting
 - **Download `.tex`** output and keyboard shortcuts (`Ctrl/Cmd+K`, `Ctrl/Cmd+S`)
 - **Native desktop builds** for Ubuntu, Windows, and Apple Silicon macOS
-- **Hardware acceleration** with CUDA on Linux/Windows and MLX/Metal on Apple Silicon
+- **Hardware acceleration** with NVIDIA CUDA, AMD ROCm/HIP, Intel/AMD
+  Vulkan, Intel SYCL, and Apple MLX/Metal; optimized CPU builds remain available
 - **Switchable backend** — Gemma is the default; `LOCALTEX_BACKEND=lfm-vl` or
   `LOCALTEX_BACKEND=pix2tex` selects an alternative local backend. The macOS
   application selects `LOCALTEX_BACKEND=mlx` automatically.
+- **Local or cloud Gemma** — choose **Download model weights** or **Use Google
+  AI Studio API** on first launch. Change the choice later in Settings. The
+  hosted option uses `gemma-4-26b-a4b-it` for images and `gemini-3.5-flash`
+  for voice. Cloud images and voice recordings are sent to Google, and usage
+  follows your AI Studio quota/billing.
 
 ## Install
 
 Prebuilt packages are attached to each [GitHub release](https://github.com/EV3KevinDEV/KevinTex/releases):
 
-- **Ubuntu:** `kevintex_1.2.4_all.deb`
-- **Windows:** portable `KevinTex-1.2.4-windows-x64.zip` containing `KevinTex.exe`
-- **Apple Silicon macOS:** `KevinTex-1.2.4-macOS-arm64.dmg` or `.zip`
+- **Ubuntu:** `kevintex_1.2.10_all.deb`
+- **Windows:** portable `KevinTex-1.2.10-windows-x64.zip` containing `KevinTex.exe`
+- **Windows GPU:** `KevinTex-1.2.10-windows-x64-cuda.7z.001` or
+  `KevinTex-1.2.10-windows-x64-vulkan.zip`
+- **Apple Silicon macOS:** `KevinTex-1.2.10-macOS-arm64.dmg` or `.zip`
 
-The Windows and macOS applications open in a native window. Model weights are
-not bundled; they download to the current user's application-data directory on
-first launch. The macOS build requires an M-series Mac and uses MLX/Metal.
+The Windows and macOS applications open in a native window. Local model weights
+are not bundled; if you choose the local provider, they download to the current
+user's application-data directory. The macOS build requires an M-series Mac
+and uses MLX/Metal for its local provider.
 
 ### Ubuntu desktop app
 
@@ -62,7 +75,7 @@ Two options:
 
 ```bash
 packaging/build-deb.sh
-sudo apt install ./dist/kevintex_1.2.4_all.deb
+sudo apt install ./dist/kevintex_1.2.10_all.deb
 ```
 
 **Current user only (no root):**
@@ -74,9 +87,36 @@ packaging/install-user.sh
 Either way you get a **KevinTex** entry in the applications menu with its own
 icon. Launching it starts the local server and opens the app in its own
 native window (Chrome/Chromium app mode; falls back to your browser).
-Closing the window stops the server. On a machine without the model
-environment, the first launch shows a one-time setup dialog that downloads
-PyTorch, llama.cpp, and the Gemma GGUF weights.
+Closing the window stops the server. On first launch, the app shows a provider
+setup screen. The local choice downloads PyTorch/llama.cpp model weights; the
+AI Studio choice asks for a key from
+<https://aistudio.google.com/app/apikey> and does not download model weights.
+
+### Linux and Windows hardware support
+
+| Hardware | Recommended backend | Linux | Windows |
+| --- | --- | --- | --- |
+| Intel or AMD CPU | CPU | Automatic fallback | Standard ZIP |
+| NVIDIA GPU | CUDA 12.4 | Automatic when the driver is detected | CUDA archive |
+| AMD GPU | ROCm/HIP | Automatic when ROCm is installed | Local HIP SDK build |
+| Intel or AMD GPU | Vulkan | Automatic with Vulkan build tools | Vulkan ZIP |
+| Intel GPU | SYCL/oneAPI | Explicit local build | Explicit local build |
+
+The Linux launcher auto-detects a usable backend and stores the choice under
+`~/.local/share/localtex/acceleration`. Override it for the next launch with:
+
+```bash
+KEVINTEX_ACCELERATION=cpu kevintex
+KEVINTEX_ACCELERATION=rocm kevintex
+KEVINTEX_ACCELERATION=vulkan kevintex
+KEVINTEX_ACCELERATION=sycl kevintex
+```
+
+ROCm requires the AMD ROCm SDK and access to `/dev/kfd`. Vulkan source builds
+need `build-essential cmake libvulkan-dev glslc spirv-headers`. SYCL requires
+Intel oneAPI (`icx` and `icpx`). Changing the override rebuilds only the
+llama.cpp Python backend; downloaded model weights remain untouched. Hardware
+and driver support still follows each vendor's compatibility list.
 
 The app window's **Snip** button (computer-with-+ icon) opens a fullscreen
 region selector: drag a box around a formula and it's converted immediately —
@@ -89,9 +129,8 @@ tool required.
 ./run.sh
 ```
 
-Then open http://127.0.0.1:8321 (the script opens it for you). The first
-launch downloads the Gemma weights (one time); after that it
-is fully offline.
+Then open http://127.0.0.1:8321 (the script opens it for you). The first launch
+asks whether to download the local Gemma weights or configure Google AI Studio.
 
 ## Manual setup (if moving to another machine)
 
@@ -124,9 +163,11 @@ The same fields are accepted by `POST /api/convert`; `/api/snip` accepts them as
 query parameters. Conversion responses include the applied preprocessing
 metadata.
 
-`POST /api/voice` accepts a 16 kHz mono WAV file in the `audio` multipart field
-and an optional `thinking` boolean. It is available with the default `gemma`
-backend and with `LOCALTEX_BACKEND=mlx`.
+`POST /api/voice` accepts a short mono WAV file in the `audio` multipart field
+and an optional `thinking` boolean. It is available with local Gemma backends
+and the hosted AI Studio provider. Cloud voice requests use Gemini's documented
+[inline audio input](https://ai.google.dev/gemini-api/docs/generate-content/audio)
+and are limited to 14 MiB to stay below its encoded request limit.
 
 ## Performance and resource controls
 
@@ -139,5 +180,14 @@ The following optional environment variables tune local inference:
 - `LOCALTEX_N_GPU_LAYERS` — llama.cpp GPU offload (`-1` means all)
 - `LOCALTEX_N_BATCH` — llama.cpp prompt batch size
 - `LOCALTEX_N_THREADS` — CPU inference threads
+- `KEVINTEX_ACCELERATION=auto|cpu|cuda|rocm|vulkan|sycl` — select the Linux
+  llama.cpp runtime (`auto` is the default)
 - `LOCALTEX_MAX_UPLOAD_BYTES` — upload cap (default 20 MiB)
 - `LOCALTEX_INFERENCE_QUEUE_TIMEOUT` — wait before returning busy
+- `LOCALTEX_PROVIDER=local|cloud` — optionally select the Gemma provider before startup
+- `GEMINI_API_KEY` — optionally provide the Google AI Studio key through the environment
+- `LOCALTEX_GEMINI_AUDIO_MODEL` — override the cloud voice model (default `gemini-3.5-flash`)
+
+The Settings UI stores a key in the per-user `provider.json` file with
+user-only permissions where supported. The key is not stored in browser
+`localStorage` or returned by the provider status API.