Skip to content

Releases: josStorer/RWKV-Runner

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

14 Mar 05:41
Compare
Choose a tag to compare

v1.7.4

13 Mar 13:47
Compare
Choose a tag to compare

Changes

Features

Improvements

  • improve markdown rendering
  • improve theme
  • improve usability
  • for Chinese users, replace Tsinghua pip mirrors with Alibaba Cloud to avoid 403 http error

Install

v1.7.3

11 Mar 11:15
Compare
Choose a tag to compare

Changes

Features

Fixes

  • fix a generation exception caused by potentially dangerous regex being passed into the stop array
  • fix max_tokens parameter of Chat page not being passed to backend
  • fix the issue where penalty_decay and global_penalty are not being passed to the backend default config when running the model through client

Improvements

  • prevent 'torch' has no attribute 'cuda' error in torch_gc, so user can use CPU or WebGPU (#302)

Chores

  • bump dependencies
  • add pre-release workflow
  • dep_check.py now ignores GPUtil

Install

v1.7.2

02 Mar 11:54
Compare
Choose a tag to compare

v1.7.1

01 Mar 08:11
Compare
Choose a tag to compare

Changes

This version includes important bug fixes, it is strongly recommended to upgrade to this version.

Upgrades

Features

  • allow setting quantizedLayers of WebGPU mode

Improvements

  • improve occurrence[token] condition
  • disable AVOID_PENALTY_TOKENS when generating (still enabled when preprocessing)
  • enable useHfMirror by default for chinese users

Fixes

  • fix the issue where state cache could be modified leading to inconsistent hit results
  • fix convert_safetensors.py for rwkv6
  • add python3-dev to lora fine-tune dependencies (this may previously lead to the error of v5 fine-tune)

Chores

  • hide MPS and CUDA-Beta Options
  • update manifest

Install

v1.7.0

21 Feb 16:18
Compare
Choose a tag to compare

Changes

Upgrades

  • rwkv 0.8.25 (fix v6 fp16i8)

Fixes

  • fix /update-config can make the default value of unclearly specified fields invalid by passing in None fields

Chores

  • change AVOID_PENALTY to \n only
  • chores

Install

v1.6.9

05 Feb 04:52
Compare
Choose a tag to compare

Changes

Upgrades

Features

  • add markdown renderer switch
  • allow loading conversation
  • allow setting history message number
  • expose penalty_decay, top_k
  • add AVOID_PENALTY_TOKENS
  • add parse_api_log.py, this script can extract
    formatted data from api.log

Improvements

  • improve macos experience
  • improve fine-tune performance
  • add better custom tokenizer support and tokenizer-midipiano.json
  • improve path processing
  • add EOS state cache point
  • reduce package size

Fixes

  • fix WSL2 WindowsOptionalFeature: Microsoft-Windows-Subsystem-Linux -> VirtualMachinePlatform
  • fix finetune errorsMap ($modelInfo)

Chores

  • update defaultPresets
  • update defaultModelConfigs
  • rename manifest tag "Main" -> "Official"
  • update manifest.json
  • update Related Repositories
  • other minor changes

Install

v1.6.8

05 Jan 06:02
Compare
Choose a tag to compare

v1.6.7

29 Dec 04:33
Compare
Choose a tag to compare