Skip to content

Releases: josStorer/RWKV-Runner

v1.9.6

27 Jun 14:59
Compare
Choose a tag to compare

Changes

Llama.cpp support

  • Add llama.cpp support with pre-compiled Vulkan libraries for Windows that should work out-of-the-box with any modern GPU. Mac and Linux users still need to manually install llama-cpp-python. You can now use RWKV GGUF models as well as any other GGUF models such as DeepSeek, Qwen3, Gemma3, Phi4. You can select the llama.cpp tag in the Models page and download the required models with one click, or place downloaded GGUF models in the models directory for use.
  • The software's preset configs have been streamlined and now include some GGUF format presets. You can click the reset button to fetch the latest presets.
  • The RWKV-Runner Python server in llama.cpp mode has been optimized. After loading the model to GPU, the server process only occupies approximately 200MB of memory on Windows platform.
  • When server users call the /switch-model API to load models, you only need to pass a file path ending with .gguf to the model field to use llama.cpp mode.

Features

  • llama.cpp support
  • Add a setting to save the full rwkv-runner client state, rather than just storing necessary settings. This option is enabled by default. You can disable it and restart the software to restore the previous version's behavior
  • add a share button to save your chat screenshot

Improvements

  • reduce peak memory usage when loading rwkv7 in cuda mode
  • increase the maximum value of the top_k API parameter to 100
  • remove language tags in Models page, as all new models support global languages
  • remove useless/disabled resources
  • other small improvements
  • You can run RWKV-Runner on Windows 7 by installing the patches from the link below. Note that you still need to install Python 3.8 and dependencies manually. https://github.com/josStorer/wails/releases/tag/v2.9.2x

Fixes

  • fix the issue of failing to load the state for RWKV7
  • Fix the abnormal behavior when passing a Tool Definition array. This is a frontend only parameter construction issue.
  • fix the issue where the model list did not refresh automatically after downloading the model when using a custom model path

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.9.5

25 Jun 12:34
Compare
Choose a tag to compare

Changes

  • add torch-2.7.1+cu128 precompiled kernels
    image
  • hide unnecessary pop-up consoles on windows
  • The linux binary files released in github releases now depend on libwebkit2gtk-4.1 to support Ubuntu 24.04. This means that versions below Ubuntu 20.04 will no longer be supported for running, and users will have to build it on their own. Additionally, Windows 7 is still supported, but you need to install the KB2999226 patch.
  • add quick think support
    Image
  • fix the issue where the line breaks in the thinking content did not take effect
  • update manifest.json and defaultModelConfigs
  • bump go-webview2

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.9.4

24 Jun 17:04
Compare
Choose a tag to compare

Changes

  • Add NVIDIA hardware info display to Settings Page with PyTorch version switching capability. Auto-select optimal PyTorch version during initial setup based on detected hardware. (Currently only works on Windows)
    image
    image
  • temporarily disable the standard WebGPU strategy as it's outdated
  • improve details

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.9.3

19 May 16:40
Compare
Choose a tag to compare

Changes

  • bump rwkv pip (improve VRAM usage when using rwkv7)
  • the reasoning model renderer no longer modifies the original response's <think> tags, but only processes them during the rendering process, and fixes the issue where markdown was not correctly rendered when rendering the <think> tags in certain cases
  • update the shortcut API list and model list in the settings, add OpenRouter and DeepSeek, and update the list with the most commonly used models at present
  • update rwkv.cpp model conversion script (3e97b6f) @MollySophia
  • update manifest (add rwkv7-g1 reasoning model)
  • add make devq command to improve the startup and reload speed during project development. Requires go install github.com/josStorer/wails/v2/cmd/[email protected]

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.9.2

12 Mar 15:38
Compare
Choose a tag to compare

Changes

  • deepseek compatible prefix mode api support
  • add deepthink toggle button
  • chores

The image below shows the effect of RWKV7-G1 1.5B model trained to 16% completion. The results may not be optimal as it's mainly for demonstrating the API server and UI functionality.

image

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.9.1

09 Mar 14:15
Compare
Choose a tag to compare

Changes

  • bump webgpu(python) (rwkv7 support) (https://github.com/cryscan/web-rwkv-py)
  • bump rwkv.cpp (rwkv7 support) (https://github.com/RWKV/rwkv.cpp)
  • reasoning model renderer support (like deepseek-r1, Qwen qwq)
  • smart scroll area support, no longer locked to bottom when generating
  • precision of penalty can be set to 0.01
  • update manifest and default configs
  • update welcome message

image

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.9.0

13 Dec 06:13
Compare
Choose a tag to compare

Changes

  • fix the handling of AVOID_REPEAT_TOKENS (Chinese punctuation) that may lead to rwkv7 fp16 overflow, causing the generation to terminate
  • fix the misidentification of rwkv5 as rwkv7 (#407)
  • improve version comparison

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.8.9

11 Dec 15:41
Compare
Choose a tag to compare

Changes

  • rwkv7 support (CPU and CUDA Mode only)

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for
WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.8.8

20 Oct 07:57
Compare
Choose a tag to compare

Changes

  • potential crash fix (#396)

Note: If you encounter WebView2 crash issues, please try opening the Windows Settings, click on Apps, search for WebView2, click Modify -> Repair to update your WebView2 runtime.

Install

v1.8.7

29 Aug 07:46
Compare
Choose a tag to compare

v1.8.7

  • withdrawing the obfuscated tiny package, as it caused some exceptions in the html webui

v1.8.6

Features

  • feat(python backend): function call support (#368) Great thanks to @EliwiiKeeya

image

  • feat(ui): add navigator for web on narrow screen (#376) Great thanks to @HaloWang

Improvements

  • improve WSL installation condition detection

Chores

  • downgrade to golang1.20 for compatibility with windows7 (#377)
  • update manifest.json (hide old models and add new models)
  • html-webui.zip is now added to the release

Install