Skip to content

Releases: josStorer/RWKV-Runner

v1.8.4

29 May 08:42
Compare
Choose a tag to compare

v1.8.4

  • fix f05a4ac, __init__.py is not embedded

v1.8.3

Deprecations

  • rwkv-beta is deprecated

Upgrades

Improvements

  • improve default LoRA fine-tune params

Fixes

  • fix #342, #345: cannot import name 'packaging' from 'pkg_resources'
  • fix the huge error prompt that pops up when running in webgpu mode

Install

v1.8.3

28 May 15:15
Compare
Choose a tag to compare

Deprecations

  • rwkv-beta is deprecated

Upgrades

Improvements

  • improve default LoRA fine-tune params

Fixes

  • fix #342, #345: cannot import name 'packaging' from 'pkg_resources'
  • fix the huge error prompt that pops up when running in webgpu mode

Install

v1.8.2

v1.8.1

12 May 15:45
Compare
Choose a tag to compare

Changes

Features

  • add support for dynamic state-tuned models

image

Upgrades

Improvements

  • add tps console output
  • add torch cnMirror
  • disable pre_ffn and head_qk
  • improve frontend details

Chores

  • update manifest.json and defaultModelConfigs

Install

v1.8.0

03 May 05:22
Compare
Choose a tag to compare

v1.7.9

30 Apr 15:14
Compare
Choose a tag to compare

Changes

  • bump webgpu mode ai00_server v0.4.2 (huge performance improvement)
  • upgrade to rwkv 0.8.26 (state-tuned model support)
  • update defaultConfigs and manifest.json
  • chores

Breaking Changes

  • change the default value of presystem to false

For the convenience of using the future state-tuned models, the default value of presystem has been set to false. This means that the RWKV-Runner service will no longer automatically insert recommended RWKV pre-prompts for you:

User: hi

Assistant: Hi. I am your assistant and I will provide expert full response in full details. Please feel free to ask any question and I will always answer it.

If you are using the API service and conducting a rigorous RWKV conversation, please manually send the above messages to the /chat/completions API's messages array, or manually send presystem: true to have the server automatically insert pre-prompts.

If you are using the RWKV-Runner client for chatting, you can enable Insert default system prompt at the beginning in the preset editor.

Of course, in reality, even if you do not perform the above, there is usually no significant negative impact.

If you are using the new RWKV state-tuned models, you do not need to perform the above.

The new RWKV state-tuned models can be downloaded here, they are very interesting:

If you are interested in state-tuning, please refer to: https://github.com/BlinkDL/RWKV-LM#state-tuning-tuning-the-initial-state-zero-inference-overhead

Install

v1.7.8

03 Apr 07:05
Compare
Choose a tag to compare

v1.7.7

27 Mar 02:27
Compare
Choose a tag to compare

v1.7.6

26 Mar 15:08
Compare
Choose a tag to compare

Changes

Features

  • make gate and out trainable (JL-er/RWKV-PEFT@834aea0)
  • new chat template for /chat/completions api (better system support)
  • add system role support for preset
  • proxied fetch support (for custom api url)

Improvements

  • improve preset editor
  • better compatibility for custom api (ollama etc.)
    image
  • throttling saveConfigs
  • improve error messages
  • other details

Install

v1.7.5