release v1.4.8

completion page: add format content button
improve overflow
2023-10-03 15:05:13 +08:00 · 2023-10-03 14:54:36 +08:00 · 2023-10-03 14:27:44 +08:00 · 2023-10-03 14:25:31 +08:00 · 2023-10-03 14:04:46 +08:00 · 2023-10-03 13:50:58 +08:00
128 changed files with 105492 additions and 2171 deletions
--- a/.gitattributes
+++ b/.gitattributes
@@ -2,5 +2,9 @@ backend-python/rwkv_pip/** linguist-vendored
 backend-python/wkv_cuda_utils/** linguist-vendored
 backend-python/get-pip.py linguist-vendored
 backend-python/convert_model.py linguist-vendored
+backend-python/convert_safetensors.py linguist-vendored
+backend-python/utils/midi.py linguist-vendored
 build/** linguist-vendored
+finetune/lora/** linguist-vendored
+finetune/json2binidx_tool/** linguist-vendored
 frontend/wailsjs/** linguist-generated
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -11,7 +11,7 @@ env:

 jobs:
  create-draft:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-22.04
    steps:
      - run: echo "VERSION=${GITHUB_REF_NAME#v}" >> $GITHUB_ENV
      - uses: actions/checkout@v3
@@ -35,7 +35,7 @@ jobs:
          gh release create ${{github.ref_name}} -d -F CURRENT_CHANGE.md -t ${{github.ref_name}}

  windows:
-    runs-on: windows-latest
+    runs-on: windows-2022
    needs: create-draft
    steps:
      - uses: actions/checkout@v3
@@ -48,26 +48,39 @@ jobs:
        id: cp310
        with:
          python-version: '3.10'
+      - uses: actions-rs/toolchain@v1
+        with:
+          toolchain: stable
+          override: true
+          target: wasm32-unknown-unknown
      - uses: crazy-max/ghaction-chocolatey@v2
        with:
          args: install upx
      - run: |
+          Start-BitsTransfer https://github.com/josStorer/LibreHardwareMonitor.Console/releases/download/v0.1.0/LibreHardwareMonitor.Console.zip ./LibreHardwareMonitor.Console.zip
+          Expand-Archive ./LibreHardwareMonitor.Console.zip -DestinationPath ./components/LibreHardwareMonitor.Console
          Start-BitsTransfer https://www.python.org/ftp/python/3.10.11/python-3.10.11-embed-amd64.zip ./python-3.10.11-embed-amd64.zip
          Expand-Archive ./python-3.10.11-embed-amd64.zip -DestinationPath ./py310
          $content=Get-Content "./py310/python310._pth"; $content | ForEach-Object {if ($_.ReadCount -eq 3) {"Lib\\site-packages"} else {$_}} | Set-Content ./py310/python310._pth
          ./py310/python ./backend-python/get-pip.py
-          ./py310/python -m pip install Cython
+          ./py310/python -m pip install Cython==0.29.36
          Copy-Item -Path "${{ steps.cp310.outputs.python-path }}/../include" -Destination "py310/include" -Recurse
          Copy-Item -Path "${{ steps.cp310.outputs.python-path }}/../libs" -Destination "py310/libs" -Recurse
-          ./py310/python -m pip install cyac
+          ./py310/python -m pip install cyac==1.7
+          git clone https://github.com/josStorer/ai00_rwkv_server --depth=1
+          cd ai00_rwkv_server
+          cargo build --release
+          mv ./target/release/ai00_server.exe ../backend-rust/webgpu_server.exe
+          cd ..
          go install github.com/wailsapp/wails/v2/cmd/wails@latest
+          (Get-Content -Path ./backend-golang/app.go) -replace "//go:custom_build windows ", "" | Set-Content -Path ./backend-golang/app.go
          make
          Rename-Item -Path "build/bin/RWKV-Runner.exe" -NewName "RWKV-Runner_windows_x64.exe"

      - run: gh release upload ${{github.ref_name}} build/bin/RWKV-Runner_windows_x64.exe

  linux:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-20.04
    needs: create-draft
    steps:
      - uses: actions/checkout@v3
@@ -76,11 +89,31 @@ jobs:
      - uses: actions/setup-go@v4
        with:
          go-version: '1.20.5'
+      - uses: actions-rs/toolchain@v1
+        with:
+          toolchain: stable
+          override: true
+          target: wasm32-unknown-unknown
      - run: |
          sudo apt-get update
          sudo apt-get install upx
          sudo apt-get install build-essential libgtk-3-dev libwebkit2gtk-4.0-dev
+          git clone https://github.com/josStorer/ai00_rwkv_server --depth=1
+          cd ai00_rwkv_server
+          sudo apt-get install libudev-dev
+          sudo apt-get install libasound2-dev
+          rustup target add x86_64-unknown-linux-gnu
+          cargo build --release --target x86_64-unknown-linux-gnu
+          mv ./target/x86_64-unknown-linux-gnu/release/ai00_server ../backend-rust/webgpu_server
+          cd ..
          go install github.com/wailsapp/wails/v2/cmd/wails@latest
+          rm ./backend-python/rwkv_pip/wkv_cuda.pyd
+          rm ./backend-python/rwkv_pip/rwkv5.pyd
+          rm ./backend-python/rwkv_pip/beta/wkv_cuda.pyd
+          rm ./backend-python/get-pip.py
+          sed -i '1,2d' ./backend-golang/wsl_not_windows.go
+          rm ./backend-golang/wsl.go
+          mv ./backend-golang/wsl_not_windows.go ./backend-golang/wsl.go
          make
          mv build/bin/RWKV-Runner build/bin/RWKV-Runner_linux_x64

@@ -96,8 +129,25 @@ jobs:
      - uses: actions/setup-go@v4
        with:
          go-version: '1.20.5'
+      - uses: actions-rs/toolchain@v1
+        with:
+          toolchain: stable
+          override: true
+          target: wasm32-unknown-unknown
      - run: |
+          git clone https://github.com/josStorer/ai00_rwkv_server --depth=1
+          cd ai00_rwkv_server
+          cargo build --release
+          mv ./target/release/ai00_server ../backend-rust/webgpu_server
+          cd ..
          go install github.com/wailsapp/wails/v2/cmd/wails@latest
+          rm ./backend-python/rwkv_pip/wkv_cuda.pyd
+          rm ./backend-python/rwkv_pip/rwkv5.pyd
+          rm ./backend-python/rwkv_pip/beta/wkv_cuda.pyd
+          rm ./backend-python/get-pip.py
+          sed -i '' '1,2d' ./backend-golang/wsl_not_windows.go
+          rm ./backend-golang/wsl.go
+          mv ./backend-golang/wsl_not_windows.go ./backend-golang/wsl.go
          make
          cp build/darwin/Readme_Install.txt build/bin/Readme_Install.txt
          cp build/bin/RWKV-Runner.app/Contents/MacOS/RWKV-Runner build/bin/RWKV-Runner_darwin_universal
@@ -106,7 +156,7 @@ jobs:
      - run: gh release upload ${{github.ref_name}} build/bin/RWKV-Runner_macos_universal.zip build/bin/RWKV-Runner_darwin_universal

  publish-release:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-22.04
    needs: [ windows, linux, macos ]
    steps:
      - uses: actions/checkout@v3
--- a/.gitignore
+++ b/.gitignore
@@ -5,9 +5,12 @@ __pycache__
 .idea
 .vs
 *.pth
+*.st
+*.safetensors
 *.bin
 /config.json
 /cache.json
+/presets.json
 /frontend/stats.html
 /frontend/package.json.md5
 /py310
@@ -19,4 +22,8 @@ __pycache__
 *.old
 .DS_Store
 *.log.*
-*.log
+*.log
+train_log.txt
+finetune/json2binidx_tool/data
+/wsl.state
+/components
--- a/CURRENT_CHANGE.md
+++ b/CURRENT_CHANGE.md
@@ -1,8 +1,8 @@
 ## Changes

- add DPI Scaling setting
- allow custom api url, key, model
- add chat and completion error messages
+- latest rwkv-5.2 is now supported (with pre-compiled kernel for windows)
+- completion page: add format content button
+- chore

 ## Install

--- a/README.md
+++ b/README.md
@@ -13,7 +13,7 @@ compatible with the OpenAI API, which means that every ChatGPT client is an RWKV
 [![license][license-image]][license-url]
 [![release][release-image]][release-url]

-English | [简体中文](README_ZH.md)
+English | [简体中文](README_ZH.md) | [日本語](README_JA.md)

 ### Install

@@ -49,7 +49,7 @@ English | [简体中文](README_ZH.md)

 #### Default configs has enabled custom CUDA kernel acceleration, which is much faster and consumes much less VRAM. If you encounter possible compatibility issues, go to the Configs page and turn off `Use Custom CUDA kernel to Accelerate`.

-#### If Windows Defender claims this is a virus, you can try downloading [v1.0.8](https://github.com/josStorer/RWKV-Runner/releases/tag/v1.0.8)/[v1.0.9](https://github.com/josStorer/RWKV-Runner/releases/tag/v1.0.9) and letting it update automatically to the latest version, or add it to the trusted list.
+#### If Windows Defender claims this is a virus, you can try downloading [v1.3.7_win.zip](https://github.com/josStorer/RWKV-Runner/releases/download/v1.3.7/RWKV-Runner_win.zip) and letting it update automatically to the latest version, or add it to the trusted list (`Windows Security` -> `Virus & threat protection` -> `Manage settings` -> `Exclusions` -> `Add or remove exclusions` -> `Add an exclusion` -> `Folder` -> `RWKV-Runner`).

 #### For different tasks, adjusting API parameters can achieve better results. For example, for translation tasks, you can try setting Temperature to 1 and Top_P to 0.3.

@@ -64,6 +64,8 @@ English | [简体中文](README_ZH.md)
 - Easy-to-understand and operate parameter configuration
 - Built-in model conversion tool
 - Built-in download management and remote model inspection
+- Built-in one-click LoRA Finetune
+- Can also be used as an OpenAI ChatGPT and GPT-Playground client
 - Multilingual localization
 - Theme switching
 - Automatic updates
@@ -89,6 +91,9 @@ body.json:

 ## Embeddings API Example

+Note: v1.4.0 has improved the quality of embeddings API. The generated results are not compatible
+with previous versions. If you are using embeddings API to generate knowledge bases or similar, please regenerate.
+
 If you are using langchain, just use `OpenAIEmbeddings(openai_api_base="http://127.0.0.1:8000", openai_api_key="sk-")`

 ```python
@@ -126,46 +131,49 @@ for i in np.argsort(embeddings_cos_sim)[::-1]:
    print(f"{embeddings_cos_sim[i]:.10f} - {values[i]}")
 ```

-## Todo
-
- [ ] Model training functionality
- [x] CUDA operator int8 acceleration
- [x] macOS support
- [x] Linux support
- [ ] Local State Cache DB
-
 ## Related Repositories:

+- RWKV-4-World: https://huggingface.co/BlinkDL/rwkv-4-world/tree/main
 - RWKV-4-Raven: https://huggingface.co/BlinkDL/rwkv-4-raven/tree/main
 - ChatRWKV: https://github.com/BlinkDL/ChatRWKV
 - RWKV-LM: https://github.com/BlinkDL/RWKV-LM
+- RWKV-LM-LoRA: https://github.com/Blealtan/RWKV-LM-LoRA
+- MIDI-LLM-tokenizer: https://github.com/briansemrau/MIDI-LLM-tokenizer

 ## Preview

 ### Homepage

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/60efbb65-29e3-4346-a597-5bdcd099251c)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/d7f24d80-f382-428d-8b28-edf87e1549e2)

 ### Chat

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/6cde9c45-51bb-4dee-b1fe-746862448520)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/80009872-528f-4932-aeb2-f724fa892e7c)

 ### Completion

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/52f47f92-d21d-4cd7-b04e-d6f9af937a97)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/bf49de8e-3b89-4543-b1ef-7cd4b19a1836)
+
+### Composition
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/e8ad908d-3fd2-4e92-bcdb-96815cb836ee)

 ### Configuration

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/93270a68-9d6d-4247-b6a3-e543c65a876b)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/48befdc6-e03c-4851-9bee-22f77ee2640e)

 ### Model Management

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/6f96fdd3-fdf5-4b78-af80-2afbd1ad173b)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/367fe4f8-cc12-475f-9371-3cf62cdbf293)

 ### Download Management

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/6982e7ee-bace-4a88-bb47-92379185bf9d)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c8153cf9-c8cb-4618-8268-60c82a5be539)
+
+### LoRA Finetune
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/4715045a-683e-4d2a-9b0e-090c7a5df63f)

 ### Settings

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b3b2ab46-344c-4f04-b066-1503f776eeb9)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/1067e635-8c07-4217-86a8-e48a5fcbb075)
--- a/README_JA.md
+++ b/README_JA.md
@@ -0,0 +1,180 @@
+<p align="center">
+    <img src="https://github.com/josStorer/RWKV-Runner/assets/13366013/d24834b0-265d-45f5-93c0-fac1e19562af">
+</p>
+
+<h1 align="center">RWKV Runner</h1>
+
+<div align="center">
+
+このプロジェクトは、すべてを自動化することで、大規模な言語モデルを使用する際の障壁をなくすことを目的としています。必要なのは、
+わずか数メガバイトの軽量な実行プログラムだけです。さらに、このプロジェクトは OpenAI API と互換性のあるインターフェイスを提供しており、
+すべての ChatGPT クライアントは RWKV クライアントであることを意味します。
+
+[![license][license-image]][license-url]
+[![release][release-image]][release-url]
+
+[English](README.md) | [简体中文](README_ZH.md) | 日本語
+
+### インストール
+
+[![Windows][Windows-image]][Windows-url]
+[![MacOS][MacOS-image]][MacOS-url]
+[![Linux][Linux-image]][Linux-url]
+
+[FAQs](https://github.com/josStorer/RWKV-Runner/wiki/FAQs) | [プレビュー](#Preview) | [ダウンロード][download-url] | [サーバーデプロイ例](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples)
+
+[license-image]: http://img.shields.io/badge/license-MIT-blue.svg
+
+[license-url]: https://github.com/josStorer/RWKV-Runner/blob/master/LICENSE
+
+[release-image]: https://img.shields.io/github/release/josStorer/RWKV-Runner.svg
+
+[release-url]: https://github.com/josStorer/RWKV-Runner/releases/latest
+
+[download-url]: https://github.com/josStorer/RWKV-Runner/releases
+
+[Windows-image]: https://img.shields.io/badge/-Windows-blue?logo=windows
+
+[Windows-url]: https://github.com/josStorer/RWKV-Runner/blob/master/build/windows/Readme_Install.txt
+
+[MacOS-image]: https://img.shields.io/badge/-MacOS-black?logo=apple
+
+[MacOS-url]: https://github.com/josStorer/RWKV-Runner/blob/master/build/darwin/Readme_Install.txt
+
+[Linux-image]: https://img.shields.io/badge/-Linux-black?logo=linux
+
+[Linux-url]: https://github.com/josStorer/RWKV-Runner/blob/master/build/linux/Readme_Install.txt
+
+</div>
+
+#### デフォルトの設定はカスタム CUDA カーネルアクセラレーションを有効にしています。互換性の問題が発生する可能性がある場合は、コンフィグページに移動し、`Use Custom CUDA kernel to Accelerate` をオフにしてください。
+
+#### Windows Defender がこれをウイルスだと主張する場合は、[v1.3.7_win.zip](https://github.com/josStorer/RWKV-Runner/releases/download/v1.3.7/RWKV-Runner_win.zip) をダウンロードして最新版に自動更新させるか、信頼済みリストに追加してみてください (`Windows Security` -> `Virus & threat protection` -> `Manage settings` -> `Exclusions` -> `Add or remove exclusions` -> `Add an exclusion` -> `Folder` -> `RWKV-Runner`)。
+
+#### 異なるタスクについては、API パラメータを調整することで、より良い結果を得ることができます。例えば、翻訳タスクの場合、Temperature を 1 に、Top_P を 0.3 に設定してみてください。
+
+## 特徴
+
+- RWKV モデル管理とワンクリック起動
+- OpenAI API と完全に互換性があり、すべての ChatGPT クライアントを RWKV クライアントにします。モデル起動後、
+  http://127.0.0.1:8000/docs を開いて詳細をご覧ください。
+- 依存関係の自動インストールにより、軽量な実行プログラムのみを必要とします
+- 2G から 32G の VRAM のコンフィグが含まれており、ほとんどのコンピュータで動作します
+- ユーザーフレンドリーなチャットと完成インタラクションインターフェースを搭載
+- 分かりやすく操作しやすいパラメータ設定
+- 内蔵モデル変換ツール
+- ダウンロード管理とリモートモデル検査機能内蔵
+- 内蔵のLoRA微調整機能を搭載しています
+- このプログラムは、OpenAI ChatGPTとGPT Playgroundのクライアントとしても使用できます
+- 多言語ローカライズ
+- テーマ切り替え
+- 自動アップデート
+
+## API 同時実行ストレステスト
+
+```bash
+ab -p body.json -T application/json -c 20 -n 100 -l http://127.0.0.1:8000/chat/completions
+```
+
+body.json:
+
+```json
+{
+  "messages": [
+    {
+      "role": "user",
+      "content": "Hello"
+    }
+  ]
+}
+```
+
+## 埋め込み API の例
+
+注意: v1.4.0 では、埋め込み API の品質が向上しました。生成される結果は、以前のバージョンとは互換性がありません。
+もし、embeddings API を使って知識ベースなどを生成している場合は、再生成してください。
+
+LangChain を使用している場合は、`OpenAIEmbeddings(openai_api_base="http://127.0.0.1:8000", openai_api_key="sk-")`
+を使用してください
+
+```python
+import numpy as np
+import requests
+
+
+def cosine_similarity(a, b):
+    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))
+
+
+values = [
+    "I am a girl",
+    "我是个女孩",
+    "私は女の子です",
+    "广东人爱吃福建人",
+    "我是个人类",
+    "I am a human",
+    "that dog is so cute",
+    "私はねこむすめです、にゃん♪",
+    "宇宙级特大事件！号外号外！"
+]
+
+embeddings = []
+for v in values:
+    r = requests.post("http://127.0.0.1:8000/embeddings", json={"input": v})
+    embedding = r.json()["data"][0]["embedding"]
+    embeddings.append(embedding)
+
+compared_embedding = embeddings[0]
+
+embeddings_cos_sim = [cosine_similarity(compared_embedding, e) for e in embeddings]
+
+for i in np.argsort(embeddings_cos_sim)[::-1]:
+    print(f"{embeddings_cos_sim[i]:.10f} - {values[i]}")
+```
+
+## 関連リポジトリ:
+
+- RWKV-4-World: https://huggingface.co/BlinkDL/rwkv-4-world/tree/main
+- RWKV-4-Raven: https://huggingface.co/BlinkDL/rwkv-4-raven/tree/main
+- ChatRWKV: https://github.com/BlinkDL/ChatRWKV
+- RWKV-LM: https://github.com/BlinkDL/RWKV-LM
+- RWKV-LM-LoRA: https://github.com/Blealtan/RWKV-LM-LoRA
+- MIDI-LLM-tokenizer: https://github.com/briansemrau/MIDI-LLM-tokenizer
+
+## プレビュー
+
+### ホームページ
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/d7f24d80-f382-428d-8b28-edf87e1549e2)
+
+### チャット
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/80009872-528f-4932-aeb2-f724fa892e7c)
+
+### 補完
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/bf49de8e-3b89-4543-b1ef-7cd4b19a1836)
+
+### 作曲
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/e8ad908d-3fd2-4e92-bcdb-96815cb836ee)
+
+### コンフィグ
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/48befdc6-e03c-4851-9bee-22f77ee2640e)
+
+### モデル管理
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/367fe4f8-cc12-475f-9371-3cf62cdbf293)
+
+### ダウンロード管理
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c8153cf9-c8cb-4618-8268-60c82a5be539)
+
+### LoRA Finetune
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/4715045a-683e-4d2a-9b0e-090c7a5df63f)
+
+### 設定
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/1067e635-8c07-4217-86a8-e48a5fcbb075)
--- a/README_ZH.md
+++ b/README_ZH.md
@@ -12,7 +12,7 @@ API兼容的接口，这意味着一切ChatGPT客户端都是RWKV客户端。
 [![license][license-image]][license-url]
 [![release][release-image]][release-url]

-[English](README.md) | 简体中文
+[English](README.md) | 简体中文 | [日本語](README_JA.md)

 ### 安装

@@ -20,7 +20,7 @@ API兼容的接口，这意味着一切ChatGPT客户端都是RWKV客户端。
 [![MacOS][MacOS-image]][MacOS-url]
 [![Linux][Linux-image]][Linux-url]

-[视频演示](https://www.bilibili.com/video/BV1hM4y1v76R) | [疑难解答](https://www.bilibili.com/read/cv23921171) | [预览](#Preview) | [下载][download-url] | [懒人包](https://pan.baidu.com/s/1wchIUHgne3gncIiLIeKBEQ?pwd=1111) | [服务器部署示例](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples)
+[视频演示](https://www.bilibili.com/video/BV1hM4y1v76R) | [疑难解答](https://www.bilibili.com/read/cv23921171) | [预览](#Preview) | [下载][download-url] | [懒人包](https://pan.baidu.com/s/1zdzZ_a0uM3gDqi6pXIZVAA?pwd=1111) | [服务器部署示例](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples)

 [license-image]: http://img.shields.io/badge/license-MIT-blue.svg

@@ -46,11 +46,9 @@ API兼容的接口，这意味着一切ChatGPT客户端都是RWKV客户端。

 </div>

-#### 注意 目前RWKV中文模型质量一般，推荐使用英文模型或World(100+ 语言)体验实际RWKV能力
-
 #### 预设配置已经开启自定义CUDA算子加速，速度更快，且显存消耗更少。如果你遇到可能的兼容性问题，前往配置页面，关闭`使用自定义CUDA算子加速`

-#### 如果Windows Defender说这是一个病毒，你可以尝试下载[v1.0.8](https://github.com/josStorer/RWKV-Runner/releases/tag/v1.0.8)/[v1.0.9](https://github.com/josStorer/RWKV-Runner/releases/tag/v1.0.9)然后让其自动更新到最新版，或添加信任
+#### 如果Windows Defender说这是一个病毒，你可以尝试下载[v1.3.7_win.zip](https://github.com/josStorer/RWKV-Runner/releases/download/v1.3.7/RWKV-Runner_win.zip)，然后让其自动更新到最新版，或添加信任 (`Windows Security` -> `Virus & threat protection` -> `Manage settings` -> `Exclusions` -> `Add or remove exclusions` -> `Add an exclusion` -> `Folder` -> `RWKV-Runner`)

 #### 对于不同的任务，调整API参数会获得更好的效果，例如对于翻译任务，你可以尝试设置Temperature为1，Top_P为0.3

@@ -60,10 +58,12 @@ API兼容的接口，这意味着一切ChatGPT客户端都是RWKV客户端。
 - 与OpenAI API完全兼容，一切ChatGPT客户端，都是RWKV客户端。启动模型后，打开 http://127.0.0.1:8000/docs 查看详细内容
 - 全自动依赖安装，你只需要一个轻巧的可执行程序
 - 预设了2G至32G显存的配置，几乎在各种电脑上工作良好
- 自带用户友好的聊天和补全交互页面
+- 自带用户友好的聊天和续写交互页面
 - 易于理解和操作的参数配置
 - 内置模型转换工具
 - 内置下载管理和远程模型检视
+- 内置一键LoRA微调
+- 也可用作 OpenAI ChatGPT 和 GPT Playground 客户端
 - 多语言本地化
 - 主题切换
 - 自动更新
@@ -89,6 +89,8 @@ body.json:

 ## Embeddings API 示例

+注意: 1.4.0 版本对embeddings API质量进行了改善，生成结果与之前的版本不兼容，如果你正在使用此API生成知识库等，请重新生成
+
 如果你在用langchain, 直接使用 `OpenAIEmbeddings(openai_api_base="http://127.0.0.1:8000", openai_api_key="sk-")`

 ```python
@@ -126,46 +128,49 @@ for i in np.argsort(embeddings_cos_sim)[::-1]:
    print(f"{embeddings_cos_sim[i]:.10f} - {values[i]}")
 ```

-## Todo
-
- [ ] 模型训练功能
- [x] CUDA算子int8提速
- [x] macOS支持
- [x] linux支持
- [ ] 本地状态缓存数据库
-
 ## 相关仓库:

+- RWKV-4-World: https://huggingface.co/BlinkDL/rwkv-4-world/tree/main
 - RWKV-4-Raven: https://huggingface.co/BlinkDL/rwkv-4-raven/tree/main
 - ChatRWKV: https://github.com/BlinkDL/ChatRWKV
 - RWKV-LM: https://github.com/BlinkDL/RWKV-LM
+- RWKV-LM-LoRA: https://github.com/Blealtan/RWKV-LM-LoRA
+- MIDI-LLM-tokenizer: https://github.com/briansemrau/MIDI-LLM-tokenizer

 ## Preview

 ### 主页

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/9d25380a-a17b-443f-b823-86c754ebebf0)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/ff2b1eef-dd3b-4cbf-98fb-b5a1ecee43e1)

 ### 聊天

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/0e66d5fa-f34a-409f-9cd4-d880815733f3)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/9570e73b-dca2-4316-9e92-09961f3c48c4)

-### 补全
+### 续写

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/d4178ee9-a188-4878-9777-25c916872c29)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/69f9ba7a-2fe8-4a5e-94cb-aa655aa409e2)
+
+### 作曲
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/95b34893-80c2-4706-87f9-bc141032ed4b)

 ### 配置

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/ad9921fc-7248-40a3-9e18-03445b86e4bf)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/59460f69-b172-4c7a-86cb-573262543076)

 ### 模型管理

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/7c36f15f-3e77-49cd-a16d-99a29f870bdf)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/551121ee-1bfe-421b-a9d1-24125126ab4b)

 ### 下载管理

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/32fde30b-11dd-43b9-9667-ad6975be2106)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/cc076038-2a91-4d36-bd39-266020e8ea87)
+
+### LoRA微调
+
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/31939b8f-9546-4f44-b434-295b492ec625)

 ### 设置

-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/e8a0f746-9da7-48e3-b3fc-e1453ac50de2)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/9652d7cc-ac33-4587-a8fb-03e5a6f5ea77)
--- a/assets/default_sound_font.sf2
+++ b/assets/default_sound_font.sf2
--- a/assets/sound-font/sound_fetch.py
+++ b/assets/sound-font/sound_fetch.py
@@ -0,0 +1,116 @@
+# https://github.com/magenta/magenta-js/issues/164
+
+import json
+import os
+import urllib.request
+
+
+def get_pitches_array(min_pitch, max_pitch):
+    return list(range(min_pitch, max_pitch + 1))
+
+
+base_url = 'https://storage.googleapis.com/magentadata/js/soundfonts'
+soundfont_path = 'sgm_plus'
+soundfont_json_url = f"{base_url}/{soundfont_path}/soundfont.json"
+
+# Download soundfont.json
+soundfont_json = ""
+
+if not os.path.exists('soundfont.json'):
+    try:
+        with urllib.request.urlopen(soundfont_json_url) as response:
+            soundfont_json = response.read()
+
+        # Save soundfont.json
+        with open('soundfont.json', 'wb') as file:
+            file.write(soundfont_json)
+
+    except:
+        print("Failed to download soundfont.json")
+
+else:
+    # If file exists, get it from the file system
+    with open('soundfont.json', 'rb') as file:
+        soundfont_json = file.read()
+
+# Parse soundfont.json
+soundfont_data = json.loads(soundfont_json)
+
+if soundfont_data is not None:
+
+    # Iterate over each instrument
+    for instrument_id, instrument_name in soundfont_data['instruments'].items():
+
+        if not os.path.isdir(instrument_name):
+
+            # Create instrument directory if it doesn't exist
+            os.makedirs(instrument_name)
+
+        instrument_json = ""
+
+        instrument_path = f"{soundfont_path}/{instrument_name}"
+
+        if not os.path.exists(f"{instrument_name}/instrument.json"):
+
+            # Download instrument.json
+            instrument_json_url = f"{base_url}/{instrument_path}/instrument.json"
+
+            try:
+                with urllib.request.urlopen(instrument_json_url) as response:
+                    instrument_json = response.read()
+
+                # Save instrument.json
+                with open(f"{instrument_name}/instrument.json", 'wb') as file:
+                    file.write(instrument_json)
+
+            except:
+                print(f"Failed to download {instrument_name}/instrument.json")
+
+        else:
+
+            # If file exists, get it from the file system
+            with open(f"{instrument_name}/instrument.json", 'rb') as file:
+                instrument_json = file.read()
+
+        # Parse instrument.json
+        instrument_data = json.loads(instrument_json)
+
+        if instrument_data is not None:
+            # Iterate over each pitch and velocity
+            for velocity in instrument_data['velocities']:
+
+                pitches = get_pitches_array(instrument_data['minPitch'], instrument_data['maxPitch'])
+
+                for pitch in pitches:
+
+                    # Create the file name
+                    file_name = f'p{pitch}_v{velocity}.mp3'
+
+                    # Check if the file already exists
+                    if os.path.exists(f"{instrument_name}/{file_name}"):
+                        pass
+                        #print(f"Skipping {instrument_name}/{file_name} - File already exists")
+
+                    else:
+
+                        # Download pitch/velocity file
+                        file_url = f"{base_url}/{instrument_path}/{file_name}"
+
+                        try:
+                            with urllib.request.urlopen(file_url) as response:
+                                file_contents = response.read()
+
+                            # Save pitch/velocity file
+                            with open(f"{instrument_name}/{file_name}", 'wb') as file:
+                                file.write(file_contents)
+
+                            print(f"Downloaded {instrument_name}/{file_name}")
+
+                        except:
+                            print(f"Failed to download {instrument_name}/{file_name}")
+
+        else:
+            print(f"Failed to parse instrument.json for {instrument_name}")
+
+else:
+    print('Failed to parse soundfont.json')
--- a/assets/sound-font/soundfont.json
+++ b/assets/sound-font/soundfont.json
@@ -0,0 +1,134 @@
+{
+  "name": "sgm_plus",
+  "instruments": {
+    "0": "acoustic_grand_piano",
+    "1": "bright_acoustic_piano",
+    "2": "electric_grand_piano",
+    "3": "honkytonk_piano",
+    "4": "electric_piano_1",
+    "5": "electric_piano_2",
+    "6": "harpsichord",
+    "7": "clavichord",
+    "8": "celesta",
+    "9": "glockenspiel",
+    "10": "music_box",
+    "11": "vibraphone",
+    "12": "marimba",
+    "13": "xylophone",
+    "14": "tubular_bells",
+    "15": "dulcimer",
+    "16": "drawbar_organ",
+    "17": "percussive_organ",
+    "18": "rock_organ",
+    "19": "church_organ",
+    "20": "reed_organ",
+    "21": "accordion",
+    "22": "harmonica",
+    "23": "tango_accordion",
+    "24": "acoustic_guitar_nylon",
+    "25": "acoustic_guitar_steel",
+    "26": "electric_guitar_jazz",
+    "27": "electric_guitar_clean",
+    "28": "electric_guitar_muted",
+    "29": "overdriven_guitar",
+    "30": "distortion_guitar",
+    "31": "guitar_harmonics",
+    "32": "acoustic_bass",
+    "33": "electric_bass_finger",
+    "34": "electric_bass_pick",
+    "35": "fretless_bass",
+    "36": "slap_bass_1",
+    "37": "slap_bass_2",
+    "38": "synth_bass_1",
+    "39": "synth_bass_2",
+    "40": "violin",
+    "41": "viola",
+    "42": "cello",
+    "43": "contrabass",
+    "44": "tremolo_strings",
+    "45": "pizzicato_strings",
+    "46": "orchestral_harp",
+    "47": "timpani",
+    "48": "string_ensemble_1",
+    "49": "string_ensemble_2",
+    "50": "synthstrings_1",
+    "51": "synthstrings_2",
+    "52": "choir_aahs",
+    "53": "voice_oohs",
+    "54": "synth_voice",
+    "55": "orchestra_hit",
+    "56": "trumpet",
+    "57": "trombone",
+    "58": "tuba",
+    "59": "muted_trumpet",
+    "60": "french_horn",
+    "61": "brass_section",
+    "62": "synthbrass_1",
+    "63": "synthbrass_2",
+    "64": "soprano_sax",
+    "65": "alto_sax",
+    "66": "tenor_sax",
+    "67": "baritone_sax",
+    "68": "oboe",
+    "69": "english_horn",
+    "70": "bassoon",
+    "71": "clarinet",
+    "72": "piccolo",
+    "73": "flute",
+    "74": "recorder",
+    "75": "pan_flute",
+    "76": "blown_bottle",
+    "77": "shakuhachi",
+    "78": "whistle",
+    "79": "ocarina",
+    "80": "lead_1_square",
+    "81": "lead_2_sawtooth",
+    "82": "lead_3_calliope",
+    "83": "lead_4_chiff",
+    "84": "lead_5_charang",
+    "85": "lead_6_voice",
+    "86": "lead_7_fifths",
+    "87": "lead_8_bass_lead",
+    "88": "pad_1_new_age",
+    "89": "pad_2_warm",
+    "90": "pad_3_polysynth",
+    "91": "pad_4_choir",
+    "92": "pad_5_bowed",
+    "93": "pad_6_metallic",
+    "94": "pad_7_halo",
+    "95": "pad_8_sweep",
+    "96": "fx_1_rain",
+    "97": "fx_2_soundtrack",
+    "98": "fx_3_crystal",
+    "99": "fx_4_atmosphere",
+    "100": "fx_5_brightness",
+    "101": "fx_6_goblins",
+    "102": "fx_7_echoes",
+    "103": "fx_8_scifi",
+    "104": "sitar",
+    "105": "banjo",
+    "106": "shamisen",
+    "107": "koto",
+    "108": "kalimba",
+    "109": "bag_pipe",
+    "110": "fiddle",
+    "111": "shanai",
+    "112": "tinkle_bell",
+    "113": "agogo",
+    "114": "steel_drums",
+    "115": "woodblock",
+    "116": "taiko_drum",
+    "117": "melodic_tom",
+    "118": "synth_drum",
+    "119": "reverse_cymbal",
+    "120": "guitar_fret_noise",
+    "121": "breath_noise",
+    "122": "seashore",
+    "123": "bird_tweet",
+    "124": "telephone_ring",
+    "125": "helicopter",
+    "126": "applause",
+    "127": "gunshot",
+    "drums": "percussion"
+  }
+}
--- a/assets/soundfont_builder.rb
+++ b/assets/soundfont_builder.rb
@@ -0,0 +1,469 @@
+#!/usr/bin/env ruby
+#
+# JavaScript Soundfont Builder for MIDI.js
+# Author: 0xFE <mohit@muthanna.com>
+# edited by Valentijn Nieman <valentijnnieman@gmail.com>
+#
+# Requires:
+#
+#   FluidSynth
+#   Lame
+#   Ruby Gems: midilib parallel
+#
+#   $ brew install fluidsynth lame (on OSX)
+#   $ gem install midilib parallel
+#
+# You'll need to download a GM soundbank to generate audio.
+#
+# Usage:
+#
+# 1) Install the above dependencies.
+# 2) Edit BUILD_DIR, SOUNDFONT, and INSTRUMENTS as required.
+# 3) Run without any argument.
+
+require 'base64'
+require 'digest/sha1'
+require 'etc'
+require 'fileutils'
+require 'midilib'
+require 'parallel'
+require 'zlib'
+require 'json'
+
+include FileUtils
+
+BUILD_DIR = "./sound-font" # Output path
+SOUNDFONT = "./default_sound_font.sf2" # Soundfont file path
+
+# This script will generate MIDI.js-compatible instrument JS files for
+# all instruments in the below array. Add or remove as necessary.
+INSTRUMENTS = [
+  0,
+  1,
+  2,
+  3,
+  4,
+  5,
+  6,
+  7,
+  8,
+  9,
+  10,
+  11,
+  12,
+  13,
+  14,
+  15,
+  16,
+  17,
+  18,
+  19,
+  20,
+  21,
+  22,
+  23,
+  24,
+  25,
+  26,
+  27,
+  28,
+  29,
+  30,
+  31,
+  32,
+  33,
+  34,
+  35,
+  36,
+  37,
+  38,
+  39,
+  40,
+  41,
+  42,
+  43,
+  44,
+  45,
+  46,
+  47,
+  48,
+  49,
+  50,
+  51,
+  52,
+  53,
+  54,
+  55,
+  56,
+  57,
+  58,
+  59,
+  60,
+  61,
+  62,
+  63,
+  64,
+  65,
+  66,
+  67,
+  68,
+  69,
+  70,
+  71,
+  72,
+  73,
+  74,
+  75,
+  76,
+  77,
+  78,
+  79,
+  80,
+  81,
+  82,
+  83,
+  84,
+  85,
+  86,
+  87,
+  88,
+  89,
+  90,
+  91,
+  92,
+  93,
+  94,
+  95,
+  96,
+  97,
+  98,
+  99,
+  100,
+  101,
+  102,
+  103,
+  104,
+  105,
+  106,
+  107,
+  108,
+  109,
+  110,
+  111,
+  112,
+  113,
+  114,
+  115,
+  116,
+  117,
+  118,
+  119,
+  120,
+  121,
+  122,
+  123,
+  124,
+  125,
+  126,
+  127
+]
+
+# It was found that midilib uses names that are incompatible with MIDI.js
+# For example, midilib uses "SynthBrass 1" -> https://github.com/jimm/midilib/blob/6c8e481ae72cd9f00a38eb3700ddfca6b549f153/lib/midilib/consts.rb#L280
+# and the MIDI association uses "SynthBrass 1" -> https://www.midi.org/specifications-old/item/gm-level-1-sound-set
+# but the MIDI.js calls this "Synth Brass 1" -> https://github.com/mudcube/MIDI.js/blob/a8a84257afa70721ae462448048a87301fc1554a/js/midi/gm.js#L44
+# there are others like "Bag pipe" vs "Bagpipe", etc.
+# here, we use the MIDI.js definitions because that is how most users will interact with the generated soundfonts.
+MIDIJS_PATCH_NAMES = [
+  "Acoustic Grand Piano",
+  "Bright Acoustic Piano",
+  "Electric Grand Piano",
+  "Honky-tonk Piano",
+  "Electric Piano 1",
+  "Electric Piano 2",
+  "Harpsichord",
+  "Clavinet",
+  "Celesta",
+  "Glockenspiel",
+  "Music Box",
+  "Vibraphone",
+  "Marimba",
+  "Xylophone",
+  "Tubular Bells",
+  "Dulcimer",
+  "Drawbar Organ",
+  "Percussive Organ",
+  "Rock Organ",
+  "Church Organ",
+  "Reed Organ",
+  "Accordion",
+  "Harmonica",
+  "Tango Accordion",
+  "Acoustic Guitar (nylon)",
+  "Acoustic Guitar (steel)",
+  "Electric Guitar (jazz)",
+  "Electric Guitar (clean)",
+  "Electric Guitar (muted)",
+  "Overdriven Guitar",
+  "Distortion Guitar",
+  "Guitar Harmonics",
+  "Acoustic Bass",
+  "Electric Bass (finger)",
+  "Electric Bass (pick)",
+  "Fretless Bass",
+  "Slap Bass 1",
+  "Slap Bass 2",
+  "Synth Bass 1",
+  "Synth Bass 2",
+  "Violin",
+  "Viola",
+  "Cello",
+  "Contrabass",
+  "Tremolo Strings",
+  "Pizzicato Strings",
+  "Orchestral Harp",
+  "Timpani",
+  "String Ensemble 1",
+  "String Ensemble 2",
+  "Synth Strings 1",
+  "Synth Strings 2",
+  "Choir Aahs",
+  "Voice Oohs",
+  "Synth Choir",
+  "Orchestra Hit",
+  "Trumpet",
+  "Trombone",
+  "Tuba",
+  "Muted Trumpet",
+  "French Horn",
+  "Brass Section",
+  "Synth Brass 1",
+  "Synth Brass 2",
+  "Soprano Sax",
+  "Alto Sax",
+  "Tenor Sax",
+  "Baritone Sax",
+  "Oboe",
+  "English Horn",
+  "Bassoon",
+  "Clarinet",
+  "Piccolo",
+  "Flute",
+  "Recorder",
+  "Pan Flute",
+  "Blown Bottle",
+  "Shakuhachi",
+  "Whistle",
+  "Ocarina",
+  "Lead 1 (square)",
+  "Lead 2 (sawtooth)",
+  "Lead 3 (calliope)",
+  "Lead 4 (chiff)",
+  "Lead 5 (charang)",
+  "Lead 6 (voice)",
+  "Lead 7 (fifths)",
+  "Lead 8 (bass + lead)",
+  "Pad 1 (new age)",
+  "Pad 2 (warm)",
+  "Pad 3 (polysynth)",
+  "Pad 4 (choir)",
+  "Pad 5 (bowed)",
+  "Pad 6 (metallic)",
+  "Pad 7 (halo)",
+  "Pad 8 (sweep)",
+  "FX 1 (rain)",
+  "FX 2 (soundtrack)",
+  "FX 3 (crystal)",
+  "FX 4 (atmosphere)",
+  "FX 5 (brightness)",
+  "FX 6 (goblins)",
+  "FX 7 (echoes)",
+  "FX 8 (sci-fi)",
+  "Sitar",
+  "Banjo",
+  "Shamisen",
+  "Koto",
+  "Kalimba",
+  "Bagpipe",
+  "Fiddle",
+  "Shanai",
+  "Tinkle Bell",
+  "Agogo",
+  "Steel Drums",
+  "Woodblock",
+  "Taiko Drum",
+  "Melodic Tom",
+  "Synth Drum",
+  "Reverse Cymbal",
+  "Guitar Fret Noise",
+  "Breath Noise",
+  "Seashore",
+  "Bird Tweet",
+  "Telephone Ring",
+  "Helicopter",
+  "Applause",
+  "Gunshot"
+]
+
+# The encoders and tools are expected in your PATH. You can supply alternate
+# paths by changing the constants below.
+LAME = "lame" # `which lame`.chomp
+FLUIDSYNTH = "fluidsynth" # `which fluidsynth`.chomp
+
+puts "Building the following instruments using font: " + SOUNDFONT
+
+# Display instrument names.
+INSTRUMENTS.each do |i|
+  puts "    #{i}: " + MIDIJS_PATCH_NAMES[i]
+end
+
+puts
+puts "Using MP3 encoder: " + LAME
+puts "Using FluidSynth encoder: " + FLUIDSYNTH
+puts
+puts "Sending output to: " + BUILD_DIR
+puts
+
+raise "Can't find soundfont: #{SOUNDFONT}" unless File.exist? SOUNDFONT
+raise "Can't find 'lame' command" if LAME.empty?
+raise "Can't find 'fluidsynth' command" if FLUIDSYNTH.empty?
+raise "Output directory does not exist: #{BUILD_DIR}" unless File.exist?(BUILD_DIR)
+
+puts "Hit return to begin."
+$stdin.readline
+
+NOTES = {
+  "C"  => 0,
+  "Db" => 1,
+  "D"  => 2,
+  "Eb" => 3,
+  "E"  => 4,
+  "F"  => 5,
+  "Gb" => 6,
+  "G"  => 7,
+  "Ab" => 8,
+  "A"  => 9,
+  "Bb" => 10,
+  "B"  => 11
+}
+
+MIDI_C0 = 12
+VELOCITY = 100
+DURATION = Integer(3000)
+TEMP_FILE = "#{BUILD_DIR}/%s%stemp.midi"
+FLUIDSYNTH_RAW = "%s.wav"
+
+def deflate(string, level)
+  z = Zlib::Deflate.new(level)
+  dst = z.deflate(string, Zlib::FINISH)
+  z.close
+  dst
+end
+
+def note_to_int(note, octave)
+  value = NOTES[note]
+  increment = MIDI_C0 * octave
+  return value + increment
+end
+
+def int_to_note(value)
+  raise "Bad Value" if value < MIDI_C0
+  reverse_notes = NOTES.invert
+  value -= MIDI_C0
+  octave = value / 12
+  note = value % 12
+  return { key: reverse_notes[note],
+           octave: octave }
+end
+
+# Run a quick table validation
+MIDI_C0.upto(100) do |x|
+  note = int_to_note x
+  #raise "Broken table" unless note_to_int(note[:key], note[:octave]) == x
+end
+
+def generate_midi(program, note_value, file)
+  include MIDI
+  seq = Sequence.new()
+  track = Track.new(seq)
+
+  seq.tracks << track
+  track.events << ProgramChange.new(0, Integer(program))
+  track.events << NoteOn.new(0, note_value, VELOCITY, 0) # channel, note, velocity, delta
+  track.events << NoteOff.new(0, note_value, VELOCITY, DURATION)
+
+  File.open(file, 'wb') { | file | seq.write(file) }
+end
+
+def run_command(cmd)
+  puts "Running: " + cmd
+  `#{cmd}`
+end
+
+def midi_to_audio(source, target)
+  run_command "#{FLUIDSYNTH} -C no -R no -g 0.5 -F #{target} #{SOUNDFONT} #{source}"
+  run_command "#{LAME} -v -b 8 -B 64 #{target}"
+  rm target
+end
+
+def open_js_file(instrument_key, type)
+  js_file = File.open("#{BUILD_DIR}/#{instrument_key}-#{type}.js", "w")
+  js_file.write(
+"""
+if (typeof(MIDI) === 'undefined') var MIDI = {};
+if (typeof(MIDI.Soundfont) === 'undefined') MIDI.Soundfont = {};
+MIDI.Soundfont.#{instrument_key} = {
+""")
+  return js_file
+end
+
+def close_js_file(file)
+  file.write("\n}\n")
+  file.close
+end
+
+def base64js(note, file, type)
+  output = '"' + note + '": '
+  output += '"' + "data:audio/#{type};base64,"
+  output += Base64.strict_encode64(File.read(file)) + '"'
+  return output
+end
+
+def generate_audio(program)
+  instrument = MIDIJS_PATCH_NAMES[program]
+  instrument_key = instrument.downcase.gsub(/[^a-z0-9 ]/, "").gsub(/[ ]/, "_")
+
+  puts "Generating audio for: " + instrument + "(#{instrument_key})"
+
+  mkdir_p "#{BUILD_DIR}/#{instrument_key}"
+
+
+  note_to_int("A", 0).upto(note_to_int("C", 8)) do |note_value|
+    output_name = "p#{note_value}_v#{VELOCITY}"
+    output_path_prefix = BUILD_DIR + "/#{instrument_key}" + output_name
+
+    puts "Generating: #{output_name}"
+    temp_file_specific = TEMP_FILE % [output_name, instrument_key]
+    generate_midi(program, note_value, temp_file_specific)
+    midi_to_audio(temp_file_specific, output_path_prefix + ".wav")
+
+    mv output_path_prefix + ".mp3", "#{BUILD_DIR}/#{instrument_key}/#{output_name}.mp3"
+    rm temp_file_specific
+  end
+
+  tempHash = {
+    "name" => instrument_key,
+    "minPitch" => 0,
+    "maxPitch" => 127,
+    "durationSeconds" => 3.0,
+    "releaseSeconds" => 1.0,
+    "percussive": false,
+    "velocities": [100]
+  }
+
+  File.open("#{BUILD_DIR}/#{instrument_key}/instrument.json", "w") do |f|
+    f.write(tempHash.to_json)
+  end
+end
+
+Parallel.each(INSTRUMENTS, :in_processes=>Etc.nprocessors){|i| generate_audio(i)}
--- a/backend-golang/app.go
+++ b/backend-golang/app.go
@@ -1,6 +1,7 @@
 package backend_golang

 import (
+	"bufio"
 	"context"
 	"errors"
 	"net/http"
@@ -8,7 +9,9 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"syscall"

+	"github.com/fsnotify/fsnotify"
 	"github.com/minio/selfupdate"
 	wruntime "github.com/wailsapp/wails/v2/pkg/runtime"
 )
@@ -40,7 +43,79 @@ func (a *App) OnStartup(ctx context.Context) {
 		a.cmdPrefix = "cd " + a.exDir + " && "
 	}

+	os.Chmod("./backend-rust/webgpu_server", 0777)
+	os.Mkdir(a.exDir+"models", os.ModePerm)
+	os.Mkdir(a.exDir+"lora-models", os.ModePerm)
+	os.Mkdir(a.exDir+"finetune/json2binidx_tool/data", os.ModePerm)
+	f, err := os.Create(a.exDir + "lora-models/train_log.txt")
+	if err == nil {
+		f.Close()
+	}
+
 	a.downloadLoop()
+	a.watchFs()
+	a.monitorHardware()
+}
+
+func (a *App) OnBeforeClose(ctx context.Context) bool {
+	if monitor != nil {
+		monitor.Process.Kill()
+	}
+	return false
+}
+
+func (a *App) watchFs() {
+	watcher, err := fsnotify.NewWatcher()
+	if err == nil {
+		watcher.Add("./lora-models")
+		watcher.Add("./models")
+		go func() {
+			for {
+				select {
+				case event, ok := <-watcher.Events:
+					if !ok {
+						return
+					}
+					wruntime.EventsEmit(a.ctx, "fsnotify", event.Name)
+				case _, ok := <-watcher.Errors:
+					if !ok {
+						return
+					}
+				}
+			}
+		}()
+	}
+}
+
+var monitor *exec.Cmd
+
+func (a *App) monitorHardware() {
+	if runtime.GOOS != "windows" {
+		return
+	}
+
+	monitor = exec.Command("./components/LibreHardwareMonitor.Console/LibreHardwareMonitor.Console.exe")
+	stdout, err := monitor.StdoutPipe()
+	if err != nil {
+		monitor = nil
+		return
+	}
+
+	go func() {
+		reader := bufio.NewReader(stdout)
+		for {
+			line, _, err := reader.ReadLine()
+			if err != nil {
+				wruntime.EventsEmit(a.ctx, "monitorerr", err.Error())
+				break
+			}
+			wruntime.EventsEmit(a.ctx, "monitor", string(line))
+		}
+	}()
+
+	monitor.SysProcAttr = &syscall.SysProcAttr{}
+	//go:custom_build windows monitor.SysProcAttr.HideWindow = true
+	monitor.Start()
 }

 func (a *App) UpdateApp(url string) (broken bool, err error) {
--- a/backend-golang/file.go
+++ b/backend-golang/file.go
@@ -122,6 +122,10 @@ func (a *App) CopyFile(src string, dst string) error {
 }

 func (a *App) OpenSaveFileDialog(filterPattern string, defaultFileName string, savedContent string) (string, error) {
+	return a.OpenSaveFileDialogBytes(filterPattern, defaultFileName, []byte(savedContent))
+}
+
+func (a *App) OpenSaveFileDialogBytes(filterPattern string, defaultFileName string, savedContent []byte) (string, error) {
 	path, err := wruntime.SaveFileDialog(a.ctx, wruntime.SaveDialogOptions{
 		DefaultFilename: defaultFileName,
 		Filters: []wruntime.FileFilter{{
@@ -135,7 +139,7 @@ func (a *App) OpenSaveFileDialog(filterPattern string, defaultFileName string, s
 	if path == "" {
 		return "", nil
 	}
-	if err := os.WriteFile(path, []byte(savedContent), 0644); err != nil {
+	if err := os.WriteFile(path, savedContent, 0644); err != nil {
 		return "", err
 	}
 	return path, nil
--- a/backend-golang/rwkv.go
+++ b/backend-golang/rwkv.go
@@ -1,16 +1,16 @@
 package backend_golang

 import (
+	"encoding/json"
 	"errors"
 	"os"
 	"os/exec"
-	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
 )

-func (a *App) StartServer(python string, port int, host string) (string, error) {
+func (a *App) StartServer(python string, port int, host string, rwkvBeta bool) (string, error) {
 	var err error
 	if python == "" {
 		python, err = GetPython()
@@ -18,7 +18,19 @@ func (a *App) StartServer(python string, port int, host string) (string, error)
 	if err != nil {
 		return "", err
 	}
-	return Cmd(python, "./backend-python/main.py", strconv.Itoa(port), host)
+	args := []string{python, "./backend-python/main.py"}
+	if rwkvBeta {
+		args = append(args, "--rwkv-beta")
+	}
+	args = append(args, "--port", strconv.Itoa(port), "--host", host)
+	return Cmd(args...)
+}
+
+func (a *App) StartWebGPUServer(port int, host string) (string, error) {
+	args := []string{"./backend-rust/webgpu_server"}
+	args = append(args, "-a", "0", "-t", "backend-rust/assets/rwkv_vocab_v20230424.json",
+		"--port", strconv.Itoa(port), "--ip", host)
+	return Cmd(args...)
 }

 func (a *App) ConvertModel(python string, modelPath string, strategy string, outPath string) (string, error) {
@@ -32,6 +44,82 @@ func (a *App) ConvertModel(python string, modelPath string, strategy string, out
 	return Cmd(python, "./backend-python/convert_model.py", "--in", modelPath, "--out", outPath, "--strategy", strategy)
 }

+func (a *App) ConvertSafetensors(python string, modelPath string, outPath string) (string, error) {
+	var err error
+	if python == "" {
+		python, err = GetPython()
+	}
+	if err != nil {
+		return "", err
+	}
+	return Cmd(python, "./backend-python/convert_safetensors.py", "--input", modelPath, "--output", outPath)
+}
+
+func (a *App) ConvertData(python string, input string, outputPrefix string, vocab string) (string, error) {
+	var err error
+	if python == "" {
+		python, err = GetPython()
+	}
+	if err != nil {
+		return "", err
+	}
+	tokenizerType := "HFTokenizer"
+	if strings.Contains(vocab, "rwkv_vocab_v20230424") {
+		tokenizerType = "RWKVTokenizer"
+	}
+
+	input = strings.TrimSuffix(input, "/")
+	if fi, err := os.Stat(input); err == nil && fi.IsDir() {
+		files, err := os.ReadDir(input)
+		if err != nil {
+			return "", err
+		}
+		jsonlFile, err := os.Create(outputPrefix + ".jsonl")
+		if err != nil {
+			return "", err
+		}
+		defer jsonlFile.Close()
+		for _, file := range files {
+			if file.IsDir() || !strings.HasSuffix(file.Name(), ".txt") {
+				continue
+			}
+			textContent, err := os.ReadFile(input + "/" + file.Name())
+			if err != nil {
+				return "", err
+			}
+			textJson, err := json.Marshal(map[string]string{"text": strings.ReplaceAll(strings.ReplaceAll(string(textContent), "\r\n", "\n"), "\r", "\n")})
+			if err != nil {
+				return "", err
+			}
+			if _, err := jsonlFile.WriteString(string(textJson) + "\n"); err != nil {
+				return "", err
+			}
+		}
+		input = outputPrefix + ".jsonl"
+	} else if err != nil {
+		return "", err
+	}
+
+	return Cmd(python, "./finetune/json2binidx_tool/tools/preprocess_data.py", "--input", input, "--output-prefix", outputPrefix, "--vocab", vocab,
+		"--tokenizer-type", tokenizerType, "--dataset-impl", "mmap", "--append-eod")
+}
+
+func (a *App) MergeLora(python string, useGpu bool, loraAlpha int, baseModel string, loraPath string, outputPath string) (string, error) {
+	var err error
+	if python == "" {
+		python, err = GetPython()
+	}
+	if err != nil {
+		return "", err
+	}
+	args := []string{python, "./finetune/lora/merge_lora.py"}
+	if useGpu {
+		args = append(args, "--use-gpu")
+	}
+	args = append(args, strconv.Itoa(loraAlpha), baseModel, loraPath, outputPath)
+	return Cmd(args...)
+}
+
 func (a *App) DepCheck(python string) error {
 	var err error
 	if python == "" {
@@ -52,11 +140,7 @@ func (a *App) InstallPyDep(python string, cnMirror bool) (string, error) {
 	if python == "" {
 		python, err = GetPython()
 		if runtime.GOOS == "windows" {
-			python, err = filepath.Abs(python)
-			if err != nil {
-				return "", err
-			}
-			python = `"` + python + `"`
+			python = `"%CD%/` + python + `"`
 		}
 	}
 	if err != nil {
@@ -71,7 +155,6 @@ func (a *App) InstallPyDep(python string, cnMirror bool) (string, error) {
 			"exit"
 		if !cnMirror {
 			installScript = strings.Replace(installScript, " -i https://pypi.tuna.tsinghua.edu.cn/simple", "", -1)
-			installScript = strings.Replace(installScript, "requirements.txt", "requirements_versions.txt", -1)
 		}
 		err = os.WriteFile("./install-py-dep.bat", []byte(installScript), 0644)
 		if err != nil {
@@ -86,3 +169,11 @@ func (a *App) InstallPyDep(python string, cnMirror bool) (string, error) {
 		return Cmd(python, "-m", "pip", "install", "-r", "./backend-python/requirements_without_cyac.txt")
 	}
 }
+
+func (a *App) GetPyError() string {
+	content, err := os.ReadFile("./error.txt")
+	if err != nil {
+		return ""
+	}
+	return string(content)
+}
--- a/backend-golang/wsl.go
+++ b/backend-golang/wsl.go
@@ -0,0 +1,181 @@
+//go:build windows
+
+package backend_golang
+
+import (
+	"bufio"
+	"context"
+	"errors"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"time"
+
+	su "github.com/nyaosorg/go-windows-su"
+	wsl "github.com/ubuntu/gowsl"
+	wruntime "github.com/wailsapp/wails/v2/pkg/runtime"
+)
+
+var distro *wsl.Distro
+var stdin io.WriteCloser
+var cmd *exec.Cmd
+
+func isWslRunning() (bool, error) {
+	if distro == nil {
+		return false, nil
+	}
+	state, err := distro.State()
+	if err != nil {
+		return false, err
+	}
+	if state != wsl.Running {
+		distro = nil
+		return false, nil
+	}
+	return true, nil
+}
+
+func (a *App) WslStart() error {
+	running, err := isWslRunning()
+	if err != nil {
+		return err
+	}
+	if running {
+		return nil
+	}
+	distros, err := wsl.RegisteredDistros(context.Background())
+	if err != nil {
+		return err
+	}
+	for _, d := range distros {
+		if strings.Contains(d.Name(), "Ubuntu") {
+			distro = &d
+			break
+		}
+	}
+	if distro == nil {
+		return errors.New("ubuntu not found")
+	}
+
+	cmd = exec.Command("wsl", "-d", distro.Name(), "-u", "root")
+
+	stdin, err = cmd.StdinPipe()
+	if err != nil {
+		return err
+	}
+
+	stdout, err := cmd.StdoutPipe()
+	cmd.Stderr = cmd.Stdout
+	if err != nil {
+		// stdin.Close()
+		stdin = nil
+		return err
+	}
+
+	go func() {
+		reader := bufio.NewReader(stdout)
+		for {
+			if stdin == nil {
+				break
+			}
+			line, _, err := reader.ReadLine()
+			if err != nil {
+				wruntime.EventsEmit(a.ctx, "wslerr", err.Error())
+				break
+			}
+			wruntime.EventsEmit(a.ctx, "wsl", string(line))
+		}
+		// stdout.Close()
+	}()
+
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (a *App) WslCommand(command string) error {
+	running, err := isWslRunning()
+	if err != nil {
+		return err
+	}
+	if !running {
+		return errors.New("wsl not running")
+	}
+	_, err = stdin.Write([]byte(command + "\n"))
+	if err != nil {
+		return err
+	}
+	return nil
+}
+
+func (a *App) WslStop() error {
+	running, err := isWslRunning()
+	if err != nil {
+		return err
+	}
+	if !running {
+		return errors.New("wsl not running")
+	}
+	if cmd != nil {
+		err = cmd.Process.Kill()
+		cmd = nil
+	}
+	// stdin.Close()
+	stdin = nil
+	distro = nil
+	if err != nil {
+		return err
+	}
+	return nil
+}
+
+func (a *App) WslIsEnabled() error {
+	ex, err := os.Executable()
+	if err != nil {
+		return err
+	}
+	exDir := filepath.Dir(ex)
+
+	data, err := os.ReadFile(exDir + "/wsl.state")
+	if err == nil {
+		if strings.Contains(string(data), "Enabled") {
+			return nil
+		}
+	}
+
+	cmd := `-Command (Get-WindowsOptionalFeature -Online -FeatureName Microsoft-Windows-Subsystem-Linux).State | Out-File -Encoding utf8 -FilePath ` + exDir + "/wsl.state"
+	_, err = su.ShellExecute(su.RUNAS, "powershell", cmd, exDir)
+	if err != nil {
+		return err
+	}
+	time.Sleep(2 * time.Second)
+	data, err = os.ReadFile(exDir + "/wsl.state")
+	if err != nil {
+		return err
+	}
+	if strings.Contains(string(data), "Enabled") {
+		return nil
+	} else {
+		return errors.New("wsl is not enabled")
+	}
+}
+
+func (a *App) WslEnable(forceMode bool) error {
+	cmd := `/online /enable-feature /featurename:Microsoft-Windows-Subsystem-Linux`
+	_, err := su.ShellExecute(su.RUNAS, "dism", cmd, `C:\`)
+	if err != nil {
+		return err
+	}
+	if forceMode {
+		os.WriteFile("./wsl.state", []byte("Enabled"), 0644)
+	}
+	return nil
+}
+
+func (a *App) WslInstallUbuntu() error {
+	_, err := Cmd("ms-windows-store://pdp/?ProductId=9PN20MSR04DW")
+	return err
+}
--- a/backend-golang/wsl_not_windows.go
+++ b/backend-golang/wsl_not_windows.go
@@ -0,0 +1,31 @@
+//go:build darwin || linux
+
+package backend_golang
+
+import (
+	"errors"
+)
+
+func (a *App) WslStart() error {
+	return errors.New("wsl not supported")
+}
+
+func (a *App) WslCommand(command string) error {
+	return errors.New("wsl not supported")
+}
+
+func (a *App) WslStop() error {
+	return errors.New("wsl not supported")
+}
+
+func (a *App) WslIsEnabled() error {
+	return errors.New("wsl not supported")
+}
+
+func (a *App) WslEnable(forceMode bool) error {
+	return errors.New("wsl not supported")
+}
+
+func (a *App) WslInstallUbuntu() error {
+	return errors.New("wsl not supported")
+}
--- a/backend-python/convert_model.py
+++ b/backend-python/convert_model.py
@@ -219,13 +219,17 @@ def get_args():
    return p.parse_args()


-args = get_args()
-if not args.quiet:
-    print(f"** {args}")
+try:
+    args = get_args()
+    if not args.quiet:
+        print(f"** {args}")

-RWKV(
-    getattr(args, "in"),
-    args.strategy,
-    verbose=not args.quiet,
-    convert_and_save_and_exit=args.out,
-)
+    RWKV(
+        getattr(args, "in"),
+        args.strategy,
+        verbose=not args.quiet,
+        convert_and_save_and_exit=args.out,
+    )
+except Exception as e:
+    with open("error.txt", "w") as f:
+        f.write(str(e))
--- a/backend-python/convert_safetensors.py
+++ b/backend-python/convert_safetensors.py
@@ -0,0 +1,53 @@
+import json
+import os
+import sys
+import copy
+import torch
+from safetensors.torch import load_file, save_file
+
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--input", type=str, help="Path to input pth model")
+parser.add_argument(
+    "--output",
+    type=str,
+    default="./converted.st",
+    help="Path to output safetensors model",
+)
+args = parser.parse_args()
+
+
+def convert_file(
+    pt_filename: str,
+    sf_filename: str,
+):
+    loaded = torch.load(pt_filename, map_location="cpu")
+    if "state_dict" in loaded:
+        loaded = loaded["state_dict"]
+
+    loaded = {k: v.clone().half() for k, v in loaded.items()}
+    for k, v in loaded.items():
+        print(f"{k}\t{v.shape}\t{v.dtype}")
+
+    # For tensors to be contiguous
+    loaded = {k: v.contiguous() for k, v in loaded.items()}
+
+    dirname = os.path.dirname(sf_filename)
+    os.makedirs(dirname, exist_ok=True)
+    save_file(loaded, sf_filename, metadata={"format": "pt"})
+    reloaded = load_file(sf_filename)
+    for k in loaded:
+        pt_tensor = loaded[k]
+        sf_tensor = reloaded[k]
+        if not torch.equal(pt_tensor, sf_tensor):
+            raise RuntimeError(f"The output tensors do not match for key {k}")
+
+
+if __name__ == "__main__":
+    try:
+        convert_file(args.input, args.output)
+        print(f"Saved to {args.output}")
+    except Exception as e:
+        with open("error.txt", "w") as f:
+            f.write(str(e))
--- a/backend-python/dep_check.py
+++ b/backend-python/dep_check.py
@@ -1,8 +1,16 @@
+import safetensors
+import midi2audio
+import mido
+import lm_dataformat
+import ftfy
+import tqdm
 import tiktoken
 import GPUtil

 import torch
 import rwkv
+import numpy
+import tokenizers
 import fastapi
 import uvicorn
 import sse_starlette
--- a/backend-python/global_var.py
+++ b/backend-python/global_var.py
@@ -1,5 +1,6 @@
 from enum import Enum, auto

+Args = "args"
 Model = "model"
 Model_Status = "model_status"
 Model_Config = "model_config"
--- a/backend-python/main.py
+++ b/backend-python/main.py
@@ -1,5 +1,11 @@
+import time
+
+start_time = time.time()
+
 import os
 import sys
+import argparse
+from typing import Sequence

 sys.path.append(os.path.dirname(os.path.realpath(__file__)))

@@ -12,7 +18,7 @@ from utils.rwkv import *
 from utils.torch import *
 from utils.ngrok import *
 from utils.log import log_middleware
-from routes import completion, config, state_cache
+from routes import completion, config, state_cache, midi, misc
 import global_var

 app = FastAPI(dependencies=[Depends(log_middleware)])
@@ -27,12 +33,19 @@ app.add_middleware(

 app.include_router(completion.router)
 app.include_router(config.router)
+app.include_router(midi.router)
+app.include_router(misc.router)
 app.include_router(state_cache.router)


@app.on_event("startup")
 def init():
    global_var.init()
+    cmd_params = os.environ["RWKV_RUNNER_PARAMS"]
+    global_var.set(
+        global_var.Args, get_args(cmd_params.split(" ") if cmd_params else None)
+    )
+
    state_cache.init()

    set_torch()
@@ -41,12 +54,12 @@ def init():
        ngrok_connect()


-@app.get("/")
+@app.get("/", tags=["Root"])
 def read_root():
    return {"Hello": "World!"}


-@app.post("/exit")
+@app.post("/exit", tags=["Root"])
 def exit():
    parent_pid = os.getpid()
    parent = psutil.Process(parent_pid)
@@ -55,20 +68,34 @@ def exit():
    parent.kill()


-def debug():
-    model = RWKV(
-        model="../models/RWKV-4-Raven-7B-v11-Eng49%-Chn49%-Jpn1%-Other1%-20230430-ctx8192.pth",
-        strategy="cuda fp16",
-        tokens_path="20B_tokenizer.json",
+def get_args(args: Union[Sequence[str], None] = None):
+    parser = argparse.ArgumentParser()
+    group = parser.add_argument_group(title="server arguments")
+    group.add_argument(
+        "--port",
+        type=int,
+        default=8000,
+        help="port to run the server on (default: 8000)",
    )
-    d = model.pipeline.decode([])
-    print(d)
+    group.add_argument(
+        "--host",
+        type=str,
+        default="127.0.0.1",
+        help="host to run the server on (default: 127.0.0.1)",
+    )
+    group = parser.add_argument_group(title="mode arguments")
+    group.add_argument(
+        "--rwkv-beta",
+        action="store_true",
+        help="whether to use rwkv-beta (default: False)",
+    )
+    args = parser.parse_args(args)
+
+    return args


 if __name__ == "__main__":
-    uvicorn.run(
-        "main:app",
-        port=8000 if len(sys.argv) < 2 else int(sys.argv[1]),
-        host="127.0.0.1" if len(sys.argv) < 3 else sys.argv[2],
-    )
-    # debug()
+    args = get_args()
+    os.environ["RWKV_RUNNER_PARAMS"] = " ".join(sys.argv[1:])
+    print("--- %s seconds ---" % (time.time() - start_time))
+    uvicorn.run("main:app", port=args.port, host=args.host, workers=1)
--- a/backend-python/requirements.txt
+++ b/backend-python/requirements.txt
--- a/backend-python/requirements_versions.txt
+++ b/backend-python/requirements_versions.txt
--- a/backend-python/requirements_without_cyac.txt
+++ b/backend-python/requirements_without_cyac.txt
--- a/backend-python/routes/completion.py
+++ b/backend-python/routes/completion.py
@@ -1,12 +1,13 @@
 import asyncio
 import json
 from threading import Lock
-from typing import List
+from typing import List, Union
+from enum import Enum
 import base64

 from fastapi import APIRouter, Request, status, HTTPException
 from sse_starlette.sse import EventSourceResponse
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 import numpy as np
 import tiktoken
 from utils.rwkv import *
@@ -16,24 +17,52 @@ import global_var
 router = APIRouter()


+class Role(Enum):
+    User = "user"
+    Assistant = "assistant"
+    System = "system"
+
+
 class Message(BaseModel):
-    role: str
-    content: str
+    role: Role
+    content: str = Field(min_length=0)
+    raw: bool = Field(False, description="Whether to treat content as raw text")
+
+
+default_stop = [
+    "\n\nUser",
+    "\n\nQuestion",
+    "\n\nQ",
+    "\n\nHuman",
+    "\n\nBob",
+]


 class ChatCompletionBody(ModelConfigBody):
-    messages: List[Message]
-    model: str = "rwkv"
+    messages: Union[List[Message], None]
+    model: Union[str, None] = "rwkv"
    stream: bool = False
-    stop: str = None
+    stop: Union[str, List[str], None] = default_stop
+    user_name: Union[str, None] = Field(None, description="Internal user name")
+    assistant_name: Union[str, None] = Field(
+        None, description="Internal assistant name"
+    )
+    presystem: bool = Field(
+        True, description="Whether to insert default system prompt at the beginning"
+    )

    class Config:
        schema_extra = {
            "example": {
-                "messages": [{"role": "user", "content": "hello"}],
+                "messages": [
+                    {"role": Role.User.value, "content": "hello", "raw": False}
+                ],
                "model": "rwkv",
                "stream": False,
                "stop": None,
+                "user_name": None,
+                "assistant_name": None,
+                "presystem": True,
                "max_tokens": 1000,
                "temperature": 1.2,
                "top_p": 0.5,
@@ -44,10 +73,10 @@ class ChatCompletionBody(ModelConfigBody):


 class CompletionBody(ModelConfigBody):
-    prompt: str
-    model: str = "rwkv"
+    prompt: Union[str, List[str], None]
+    model: Union[str, None] = "rwkv"
    stream: bool = False
-    stop: str = None
+    stop: Union[str, List[str], None] = None

    class Config:
        schema_extra = {
@@ -72,12 +101,12 @@ requests_num = 0


 async def eval_rwkv(
-    model: RWKV,
+    model: AbstractRWKV,
    request: Request,
    body: ModelConfigBody,
    prompt: str,
    stream: bool,
-    stop: str,
+    stop: Union[str, List[str], None],
    chat_mode: bool,
 ):
    global requests_num
@@ -95,222 +124,235 @@ async def eval_rwkv(
            return
        await asyncio.sleep(0.1)
    else:
-        completion_lock.acquire()
-        if await request.is_disconnected():
-            completion_lock.release()
+        with completion_lock:
+            if await request.is_disconnected():
+                requests_num = requests_num - 1
+                print(f"{request.client} Stop Waiting (Lock)")
+                quick_log(
+                    request,
+                    None,
+                    "Stop Waiting (Lock). RequestsNum: " + str(requests_num),
+                )
+                return
+            set_rwkv_config(model, global_var.get(global_var.Model_Config))
+            set_rwkv_config(model, body)
+
+            response, prompt_tokens, completion_tokens = "", 0, 0
+            for response, delta, prompt_tokens, completion_tokens in model.generate(
+                prompt,
+                stop=stop,
+            ):
+                if await request.is_disconnected():
+                    break
+                if stream:
+                    yield json.dumps(
+                        {
+                            "object": "chat.completion.chunk"
+                            if chat_mode
+                            else "text_completion",
+                            # "response": response,
+                            "model": model.name,
+                            "choices": [
+                                {
+                                    "delta": {"content": delta},
+                                    "index": 0,
+                                    "finish_reason": None,
+                                }
+                                if chat_mode
+                                else {
+                                    "text": delta,
+                                    "index": 0,
+                                    "finish_reason": None,
+                                }
+                            ],
+                        }
+                    )
+            # torch_gc()
            requests_num = requests_num - 1
-            print(f"{request.client} Stop Waiting (Lock)")
+            if await request.is_disconnected():
+                print(f"{request.client} Stop Waiting")
+                quick_log(
+                    request,
+                    body,
+                    response + "\nStop Waiting. RequestsNum: " + str(requests_num),
+                )
+                return
            quick_log(
                request,
-                None,
-                "Stop Waiting (Lock). RequestsNum: " + str(requests_num),
+                body,
+                response + "\nFinished. RequestsNum: " + str(requests_num),
            )
-            return
-        set_rwkv_config(model, global_var.get(global_var.Model_Config))
-        set_rwkv_config(model, body)
-
-        response, prompt_tokens, completion_tokens = "", 0, 0
-        for response, delta, prompt_tokens, completion_tokens in model.generate(
-            prompt,
-            stop=stop,
-        ):
-            if await request.is_disconnected():
-                break
            if stream:
                yield json.dumps(
                    {
                        "object": "chat.completion.chunk"
                        if chat_mode
                        else "text_completion",
-                        "response": response,
+                        # "response": response,
                        "model": model.name,
                        "choices": [
                            {
-                                "delta": {"content": delta},
+                                "delta": {},
                                "index": 0,
-                                "finish_reason": None,
+                                "finish_reason": "stop",
                            }
                            if chat_mode
                            else {
-                                "text": delta,
+                                "text": "",
                                "index": 0,
-                                "finish_reason": None,
+                                "finish_reason": "stop",
                            }
                        ],
                    }
                )
-        # torch_gc()
-        requests_num = requests_num - 1
-        completion_lock.release()
-        if await request.is_disconnected():
-            print(f"{request.client} Stop Waiting")
-            quick_log(
-                request,
-                body,
-                response + "\nStop Waiting. RequestsNum: " + str(requests_num),
-            )
-            return
-        quick_log(
-            request,
-            body,
-            response + "\nFinished. RequestsNum: " + str(requests_num),
-        )
-        if stream:
-            yield json.dumps(
-                {
-                    "object": "chat.completion.chunk"
-                    if chat_mode
-                    else "text_completion",
-                    "response": response,
+                yield "[DONE]"
+            else:
+                yield {
+                    "object": "chat.completion" if chat_mode else "text_completion",
+                    # "response": response,
                    "model": model.name,
+                    "usage": {
+                        "prompt_tokens": prompt_tokens,
+                        "completion_tokens": completion_tokens,
+                        "total_tokens": prompt_tokens + completion_tokens,
+                    },
                    "choices": [
                        {
-                            "delta": {},
+                            "message": {
+                                "role": Role.Assistant.value,
+                                "content": response,
+                            },
                            "index": 0,
                            "finish_reason": "stop",
                        }
                        if chat_mode
                        else {
-                            "text": "",
+                            "text": response,
                            "index": 0,
                            "finish_reason": "stop",
                        }
                    ],
                }
-            )
-            yield "[DONE]"
-        else:
-            yield {
-                "object": "chat.completion" if chat_mode else "text_completion",
-                "response": response,
-                "model": model.name,
-                "usage": {
-                    "prompt_tokens": prompt_tokens,
-                    "completion_tokens": completion_tokens,
-                    "total_tokens": prompt_tokens + completion_tokens,
-                },
-                "choices": [
-                    {
-                        "message": {
-                            "role": "assistant",
-                            "content": response,
-                        },
-                        "index": 0,
-                        "finish_reason": "stop",
-                    }
-                    if chat_mode
-                    else {
-                        "text": response,
-                        "index": 0,
-                        "finish_reason": "stop",
-                    }
-                ],
-            }


-@router.post("/v1/chat/completions")
-@router.post("/chat/completions")
+@router.post("/v1/chat/completions", tags=["Completions"])
+@router.post("/chat/completions", tags=["Completions"])
 async def chat_completions(body: ChatCompletionBody, request: Request):
-    model: RWKV = global_var.get(global_var.Model)
+    model: TextRWKV = global_var.get(global_var.Model)
    if model is None:
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "model not loaded")

-    question = body.messages[-1]
-    if question.role == "user":
-        question = question.content
-    elif question.role == "system":
-        question = body.messages[-2]
-        if question.role == "user":
-            question = question.content
-        else:
-            raise HTTPException(status.HTTP_400_BAD_REQUEST, "no question found")
-    else:
-        raise HTTPException(status.HTTP_400_BAD_REQUEST, "no question found")
+    if body.messages is None or body.messages == []:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "messages not found")

    interface = model.interface
-    user = model.user
-    bot = model.bot
+    user = model.user if body.user_name is None else body.user_name
+    bot = model.bot if body.assistant_name is None else body.assistant_name

-    completion_text = (
-        f"""
+    is_raven = model.rwkv_type == RWKVType.Raven
+
+    completion_text: str = ""
+    basic_system: Union[str, None] = None
+    if body.presystem:
+        if body.messages[0].role == Role.System:
+            basic_system = body.messages[0].content
+
+        if basic_system is None:
+            completion_text = (
+                f"""
 The following is a coherent verbose detailed conversation between a girl named {bot} and her friend {user}. \
 {bot} is very intelligent, creative and friendly. \
 {bot} is unlikely to disagree with {user}, and {bot} doesn't like to ask {user} questions. \
 {bot} likes to tell {user} a lot about herself and her opinions. \
 {bot} usually gives {user} kind, helpful and informative advices.\n
 """
-        if user == "Bob"
-        else f"{user}{interface} hi\n\n{bot}{interface} 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.\n\n"
-    )
-    for message in body.messages:
-        if message.role == "system":
+                if is_raven
+                else (
+                    f"{user}{interface} hi\n\n{bot}{interface} 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.\n\n"
+                )
+            )
+        else:
+            if not body.messages[0].raw:
+                basic_system = (
+                    basic_system.replace("\r\n", "\n")
+                    .replace("\r", "\n")
+                    .replace("\n\n", "\n")
+                    .replace("\n", " ")
+                    .strip()
+                )
            completion_text = (
-                f"The following is a coherent verbose detailed conversation between a girl named {bot} and her friend {user}. "
-                if user == "Bob"
-                else f"{user}{interface} hi\n\n{bot}{interface} Hi. "
-                + message.content.replace("\\n", "\n")
-                .replace("\r\n", "\n")
-                .replace("\n\n", "\n")
-                .replace("\n", " ")
-                .strip()
-                .replace("You are", f"{bot} is" if user == "Bob" else "I am")
-                .replace("you are", f"{bot} is" if user == "Bob" else "I am")
-                .replace("You're", f"{bot} is" if user == "Bob" else "I'm")
-                .replace("you're", f"{bot} is" if user == "Bob" else "I'm")
-                .replace("You", f"{bot}" if user == "Bob" else "I")
-                .replace("you", f"{bot}" if user == "Bob" else "I")
-                .replace("Your", f"{bot}'s" if user == "Bob" else "My")
-                .replace("your", f"{bot}'s" if user == "Bob" else "my")
-                .replace("你", f"{bot}" if user == "Bob" else "我")
+                (
+                    f"The following is a coherent verbose detailed conversation between a girl named {bot} and her friend {user}. "
+                    if is_raven
+                    else f"{user}{interface} hi\n\n{bot}{interface} Hi. "
+                )
+                + basic_system.replace("You are", f"{bot} is" if is_raven else "I am")
+                .replace("you are", f"{bot} is" if is_raven else "I am")
+                .replace("You're", f"{bot} is" if is_raven else "I'm")
+                .replace("you're", f"{bot} is" if is_raven else "I'm")
+                .replace("You", f"{bot}" if is_raven else "I")
+                .replace("you", f"{bot}" if is_raven else "I")
+                .replace("Your", f"{bot}'s" if is_raven else "My")
+                .replace("your", f"{bot}'s" if is_raven else "my")
+                .replace("你", f"{bot}" if is_raven else "我")
                + "\n\n"
            )
-            break
-    for message in body.messages:
-        if message.role == "user":
-            completion_text += (
-                f"{user}{interface} "
-                + message.content.replace("\\n", "\n")
-                .replace("\r\n", "\n")
+
+    for message in body.messages[(0 if basic_system is None else 1) :]:
+        append_message: str = ""
+        if message.role == Role.User:
+            append_message = f"{user}{interface} " + message.content
+        elif message.role == Role.Assistant:
+            append_message = f"{bot}{interface} " + message.content
+        elif message.role == Role.System:
+            append_message = message.content
+        if not message.raw:
+            append_message = (
+                append_message.replace("\r\n", "\n")
+                .replace("\r", "\n")
                .replace("\n\n", "\n")
                .strip()
-                + "\n\n"
-            )
-        elif message.role == "assistant":
-            completion_text += (
-                f"{bot}{interface} "
-                + message.content.replace("\\n", "\n")
-                .replace("\r\n", "\n")
-                .replace("\n\n", "\n")
-                .strip()
-                + "\n\n"
            )
+        completion_text += append_message + "\n\n"
    completion_text += f"{bot}{interface}"

-    stop = f"\n\n{user}" if body.stop is None else body.stop
+    if type(body.stop) == str:
+        body.stop = [body.stop, f"\n\n{user}", f"\n\n{bot}"]
+    elif type(body.stop) == list:
+        body.stop.append(f"\n\n{user}")
+        body.stop.append(f"\n\n{bot}")
+    elif body.stop is None:
+        body.stop = default_stop
+
    if body.stream:
        return EventSourceResponse(
-            eval_rwkv(model, request, body, completion_text, body.stream, stop, True)
+            eval_rwkv(
+                model, request, body, completion_text, body.stream, body.stop, True
+            )
        )
    else:
        try:
            return await eval_rwkv(
-                model, request, body, completion_text, body.stream, stop, True
+                model, request, body, completion_text, body.stream, body.stop, True
            ).__anext__()
        except StopAsyncIteration:
            return None


-@router.post("/v1/completions")
-@router.post("/completions")
+@router.post("/v1/completions", tags=["Completions"])
+@router.post("/completions", tags=["Completions"])
 async def completions(body: CompletionBody, request: Request):
-    model: RWKV = global_var.get(global_var.Model)
+    model: AbstractRWKV = global_var.get(global_var.Model)
    if model is None:
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "model not loaded")

-    if body.prompt is None or body.prompt == "":
+    if body.prompt is None or body.prompt == "" or body.prompt == []:
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "prompt not found")

+    if type(body.prompt) == list:
+        body.prompt = body.prompt[0]  # TODO: support multiple prompts
+
    if body.stream:
        return EventSourceResponse(
            eval_rwkv(model, request, body, body.prompt, body.stream, body.stop, False)
@@ -325,8 +367,8 @@ async def completions(body: CompletionBody, request: Request):


 class EmbeddingsBody(BaseModel):
-    input: str | List[str] | List[List[int]]
-    model: str = "rwkv"
+    input: Union[str, List[str], List[List[int]], None]
+    model: Union[str, None] = "rwkv"
    encoding_format: str = None
    fast_mode: bool = False

@@ -345,12 +387,12 @@ def embedding_base64(embedding: List[float]) -> str:
    return base64.b64encode(np.array(embedding).astype(np.float32)).decode("utf-8")


-@router.post("/v1/embeddings")
-@router.post("/embeddings")
-@router.post("/v1/engines/text-embedding-ada-002/embeddings")
-@router.post("/engines/text-embedding-ada-002/embeddings")
+@router.post("/v1/embeddings", tags=["Embeddings"])
+@router.post("/embeddings", tags=["Embeddings"])
+@router.post("/v1/engines/text-embedding-ada-002/embeddings", tags=["Embeddings"])
+@router.post("/engines/text-embedding-ada-002/embeddings", tags=["Embeddings"])
 async def embeddings(body: EmbeddingsBody, request: Request):
-    model: RWKV = global_var.get(global_var.Model)
+    model: AbstractRWKV = global_var.get(global_var.Model)
    if model is None:
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "model not loaded")

@@ -372,81 +414,88 @@ async def embeddings(body: EmbeddingsBody, request: Request):
            return
        await asyncio.sleep(0.1)
    else:
-        completion_lock.acquire()
-        if await request.is_disconnected():
-            completion_lock.release()
-            requests_num = requests_num - 1
-            print(f"{request.client} Stop Waiting (Lock)")
-            quick_log(
-                request,
-                None,
-                "Stop Waiting (Lock). RequestsNum: " + str(requests_num),
-            )
-            return
+        with completion_lock:
+            if await request.is_disconnected():
+                requests_num = requests_num - 1
+                print(f"{request.client} Stop Waiting (Lock)")
+                quick_log(
+                    request,
+                    None,
+                    "Stop Waiting (Lock). RequestsNum: " + str(requests_num),
+                )
+                return

-        base64_format = False
-        if body.encoding_format == "base64":
-            base64_format = True
+            base64_format = False
+            if body.encoding_format == "base64":
+                base64_format = True

-        embeddings = []
-        prompt_tokens = 0
-        if type(body.input) == list:
-            if type(body.input[0]) == list:
-                encoding = tiktoken.model.encoding_for_model("text-embedding-ada-002")
-                for i in range(len(body.input)):
-                    if await request.is_disconnected():
-                        break
-                    input = encoding.decode(body.input[i])
-                    embedding, token_len = model.get_embedding(input, body.fast_mode)
-                    prompt_tokens = prompt_tokens + token_len
-                    if base64_format:
-                        embedding = embedding_base64(embedding)
-                    embeddings.append(embedding)
-            else:
-                for i in range(len(body.input)):
-                    if await request.is_disconnected():
-                        break
-                    embedding, token_len = model.get_embedding(
-                        body.input[i], body.fast_mode
+            embeddings = []
+            prompt_tokens = 0
+            if type(body.input) == list:
+                if type(body.input[0]) == list:
+                    encoding = tiktoken.model.encoding_for_model(
+                        "text-embedding-ada-002"
                    )
-                    prompt_tokens = prompt_tokens + token_len
-                    if base64_format:
-                        embedding = embedding_base64(embedding)
-                    embeddings.append(embedding)
-        else:
-            embedding, prompt_tokens = model.get_embedding(body.input, body.fast_mode)
-            if base64_format:
-                embedding = embedding_base64(embedding)
-            embeddings.append(embedding)
+                    for i in range(len(body.input)):
+                        if await request.is_disconnected():
+                            break
+                        input = encoding.decode(body.input[i])
+                        embedding, token_len = model.get_embedding(
+                            input, body.fast_mode
+                        )
+                        prompt_tokens = prompt_tokens + token_len
+                        if base64_format:
+                            embedding = embedding_base64(embedding)
+                        embeddings.append(embedding)
+                else:
+                    for i in range(len(body.input)):
+                        if await request.is_disconnected():
+                            break
+                        embedding, token_len = model.get_embedding(
+                            body.input[i], body.fast_mode
+                        )
+                        prompt_tokens = prompt_tokens + token_len
+                        if base64_format:
+                            embedding = embedding_base64(embedding)
+                        embeddings.append(embedding)
+            else:
+                embedding, prompt_tokens = model.get_embedding(
+                    body.input, body.fast_mode
+                )
+                if base64_format:
+                    embedding = embedding_base64(embedding)
+                embeddings.append(embedding)

-        requests_num = requests_num - 1
-        completion_lock.release()
-        if await request.is_disconnected():
-            print(f"{request.client} Stop Waiting")
+            requests_num = requests_num - 1
+            if await request.is_disconnected():
+                print(f"{request.client} Stop Waiting")
+                quick_log(
+                    request,
+                    None,
+                    "Stop Waiting. RequestsNum: " + str(requests_num),
+                )
+                return
            quick_log(
                request,
                None,
-                "Stop Waiting. RequestsNum: " + str(requests_num),
+                "Finished. RequestsNum: " + str(requests_num),
            )
-            return
-        quick_log(
-            request,
-            None,
-            "Finished. RequestsNum: " + str(requests_num),
-        )

-        ret_data = [
-            {
-                "object": "embedding",
-                "index": i,
-                "embedding": embedding,
+            ret_data = [
+                {
+                    "object": "embedding",
+                    "index": i,
+                    "embedding": embedding,
+                }
+                for i, embedding in enumerate(embeddings)
+            ]
+
+            return {
+                "object": "list",
+                "data": ret_data,
+                "model": model.name,
+                "usage": {
+                    "prompt_tokens": prompt_tokens,
+                    "total_tokens": prompt_tokens,
+                },
            }
-            for i, embedding in enumerate(embeddings)
-        ]
-
-        return {
-            "object": "list",
-            "data": ret_data,
-            "model": model.name,
-            "usage": {"prompt_tokens": prompt_tokens, "total_tokens": prompt_tokens},
-        }
--- a/backend-python/routes/config.py
+++ b/backend-python/routes/config.py
@@ -6,20 +6,22 @@ from pydantic import BaseModel
 from utils.rwkv import *
 from utils.torch import *
 import global_var
-import GPUtil

 router = APIRouter()


 def get_tokens_path(model_path: str):
    model_path = model_path.lower()
-    default_tokens_path = (
-        f"{pathlib.Path(__file__).parent.parent.resolve()}/rwkv_pip/20B_tokenizer.json"
-    )
+    tokenizer_dir = f"{pathlib.Path(__file__).parent.parent.resolve()}/rwkv_pip/"
+
+    default_tokens_path = tokenizer_dir + "20B_tokenizer.json"
+
    if "raven" in model_path:
        return default_tokens_path
    elif "world" in model_path:
        return "rwkv_vocab_v20230424"
+    elif "midi" in model_path:
+        return tokenizer_dir + "tokenizer-midi.json"
    else:
        return default_tokens_path

@@ -27,6 +29,7 @@ def get_tokens_path(model_path: str):
 class SwitchModelBody(BaseModel):
    model: str
    strategy: str
+    tokenizer: Union[str, None] = None
    customCuda: bool = False

    class Config:
@@ -34,12 +37,13 @@ class SwitchModelBody(BaseModel):
            "example": {
                "model": "models/RWKV-4-World-3B-v1-20230619-ctx4096.pth",
                "strategy": "cuda fp16",
+                "tokenizer": None,
                "customCuda": False,
            }
        }


-@router.post("/switch-model")
+@router.post("/switch-model", tags=["Configs"])
 def switch_model(body: SwitchModelBody, response: Response, request: Request):
    if global_var.get(global_var.Model_Status) is global_var.ModelStatus.Loading:
        response.status_code = Status.HTTP_304_NOT_MODIFIED
@@ -52,16 +56,35 @@ def switch_model(body: SwitchModelBody, response: Response, request: Request):
    if body.model == "":
        return "success"

+    if "->" in body.strategy:
+        state_cache.disable_state_cache()
+    else:
+        try:
+            state_cache.enable_state_cache()
+        except HTTPException:
+            pass
+
    os.environ["RWKV_CUDA_ON"] = "1" if body.customCuda else "0"

    global_var.set(global_var.Model_Status, global_var.ModelStatus.Loading)
+    tokenizer = (
+        get_tokens_path(body.model)
+        if body.tokenizer is None or body.tokenizer == ""
+        else body.tokenizer
+    )
    try:
        global_var.set(
            global_var.Model,
-            RWKV(
+            TextRWKV(
                model=body.model,
                strategy=body.strategy,
-                tokens_path=get_tokens_path(body.model),
+                tokens_path=tokenizer,
+            )
+            if "midi" not in body.model.lower()
+            else MusicRWKV(
+                model=body.model,
+                strategy=body.strategy,
+                tokens_path=tokenizer,
            ),
        )
    except Exception as e:
@@ -81,7 +104,7 @@ def switch_model(body: SwitchModelBody, response: Response, request: Request):
    return "success"


-@router.post("/update-config")
+@router.post("/update-config", tags=["Configs"])
 def update_config(body: ModelConfigBody):
    """
    Will not update the model config immediately, but set it when completion called to avoid modifications during generation
@@ -93,8 +116,10 @@ def update_config(body: ModelConfigBody):
    return "success"


-@router.get("/status")
+@router.get("/status", tags=["Configs"])
 def status():
+    import GPUtil
+
    gpus = GPUtil.getGPUs()
    if len(gpus) == 0:
        device_name = "CPU"
--- a/backend-python/routes/midi.py
+++ b/backend-python/routes/midi.py
@@ -0,0 +1,131 @@
+import io
+from fastapi import APIRouter, HTTPException, status
+from starlette.responses import StreamingResponse
+from pydantic import BaseModel
+from utils.midi import *
+from midi2audio import FluidSynth
+
+router = APIRouter()
+
+
+class TextToMidiBody(BaseModel):
+    text: str
+
+    class Config:
+        schema_extra = {
+            "example": {
+                "text": "p:24:a p:2a:a p:31:a p:39:a p:3b:a p:45:a b:26:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:24:0 p:2a:0 p:31:0 p:39:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:26:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:2e:a p:3b:a p:45:a b:26:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:2e:0 p:3b:0 p:45:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:2e:a p:3b:a p:45:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:2e:0 p:3b:0 p:45:0 b:26:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:26:a p:2a:a p:3b:a p:45:a t14 p:26:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a b:26:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:2a:0 p:3b:0 p:45:0 b:26:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:2d:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 b:2d:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:24:a p:2e:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:24:0 p:2e:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:26:a p:2a:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:26:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:26:a p:2e:a p:31:a p:39:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:26:0 p:2e:0 p:31:0 p:39:0 p:3b:0 p:45:0 b:21:0 t2 p:26:a p:2e:a p:31:a p:39:a p:3b:a p:45:a b:21:a t14 p:26:0 p:2e:0 p:31:0 p:39:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:24:a p:2a:a p:31:a p:39:a p:3b:a p:45:a b:1f:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:24:0 p:2a:0 p:31:0 p:39:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:1f:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:2e:a p:3b:a p:45:a b:1f:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:2e:0 p:3b:0 p:45:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:2e:a p:3b:a p:45:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:2e:0 p:3b:0 p:45:0 b:1f:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:26:a p:2a:a p:3b:a p:45:a t14 p:26:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a b:1f:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:2a:0 p:3b:0 p:45:0 b:1f:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:1f:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 b:1f:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:24:a p:2e:a p:3b:a p:45:a b:26:a g:39:a g:39:a g:3e:a g:3e:a g:42:a g:42:a pi:39:a pi:3e:a pi:42:a t14 p:24:0 p:2e:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0",
+            }
+        }
+
+
+@router.post("/text-to-midi", tags=["MIDI"])
+def text_to_midi(body: TextToMidiBody):
+    vocab_config = "backend-python/utils/midi_vocab_config.json"
+    cfg = VocabConfig.from_json(vocab_config)
+    mid = convert_str_to_midi(cfg, body.text.strip())
+    mid_data = io.BytesIO()
+    mid.save(None, mid_data)
+    mid_data.seek(0)
+
+    return StreamingResponse(mid_data, media_type="audio/midi")
+
+
+class TxtToMidiBody(BaseModel):
+    txt_path: str
+    midi_path: str
+
+    class Config:
+        schema_extra = {
+            "example": {
+                "txt_path": "midi/sample.txt",
+                "midi_path": "midi/sample.mid",
+            }
+        }
+
+
+@router.post("/txt-to-midi", tags=["MIDI"])
+def txt_to_midi(body: TxtToMidiBody):
+    if not body.midi_path.startswith("midi/"):
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "bad output path")
+
+    vocab_config = "backend-python/utils/midi_vocab_config.json"
+    cfg = VocabConfig.from_json(vocab_config)
+    with open(body.txt_path, "r") as f:
+        text = f.read()
+    text = text.strip()
+    mid = convert_str_to_midi(cfg, text)
+    mid.save(body.midi_path)
+
+    return "success"
+
+
+class MidiToWavBody(BaseModel):
+    midi_path: str
+    wav_path: str
+    sound_font_path: str = "assets/default_sound_font.sf2"
+
+    class Config:
+        schema_extra = {
+            "example": {
+                "midi_path": "midi/sample.mid",
+                "wav_path": "midi/sample.wav",
+                "sound_font_path": "assets/default_sound_font.sf2",
+            }
+        }
+
+
+@router.post("/midi-to-wav", tags=["MIDI"])
+def midi_to_wav(body: MidiToWavBody):
+    """
+    Install fluidsynth first, see more: https://github.com/FluidSynth/fluidsynth/wiki/Download#distributions
+    """
+
+    if not body.wav_path.startswith("midi/"):
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "bad output path")
+
+    fs = FluidSynth(body.sound_font_path)
+    fs.midi_to_audio(body.midi_path, body.wav_path)
+
+    return "success"
+
+
+class TextToWavBody(BaseModel):
+    text: str
+    wav_name: str
+    sound_font_path: str = "assets/default_sound_font.sf2"
+
+    class Config:
+        schema_extra = {
+            "example": {
+                "text": "p:24:a p:2a:a p:31:a p:39:a p:3b:a p:45:a b:26:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:24:0 p:2a:0 p:31:0 p:39:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:26:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:2e:a p:3b:a p:45:a b:26:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:2e:0 p:3b:0 p:45:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:2e:a p:3b:a p:45:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:2e:0 p:3b:0 p:45:0 b:26:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:26:a p:2a:a p:3b:a p:45:a t14 p:26:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a b:26:a g:3e:a g:3e:a g:42:a g:42:a g:45:a g:45:a pi:3e:a pi:42:a pi:45:a t14 p:2a:0 p:3b:0 p:45:0 b:26:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:2d:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 b:2d:0 g:3e:0 g:3e:0 g:42:0 g:42:0 g:45:0 g:45:0 pi:3e:0 pi:42:0 pi:45:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:24:a p:2e:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:24:0 p:2e:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:26:a p:2a:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:26:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:26:a p:2e:a p:31:a p:39:a p:3b:a p:45:a b:21:a g:39:a g:39:a g:3d:a g:3d:a g:40:a g:40:a pi:39:a pi:3d:a pi:40:a t14 p:26:0 p:2e:0 p:31:0 p:39:0 p:3b:0 p:45:0 b:21:0 t2 p:26:a p:2e:a p:31:a p:39:a p:3b:a p:45:a b:21:a t14 p:26:0 p:2e:0 p:31:0 p:39:0 p:3b:0 p:45:0 b:21:0 g:39:0 g:39:0 g:3d:0 g:3d:0 g:40:0 g:40:0 pi:39:0 pi:3d:0 pi:40:0 t2 p:24:a p:2a:a p:31:a p:39:a p:3b:a p:45:a b:1f:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:24:0 p:2a:0 p:31:0 p:39:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0 p:45:0 b:1f:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:2e:a p:3b:a p:45:a b:1f:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:2e:0 p:3b:0 p:45:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:2e:a p:3b:a p:45:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:2e:0 p:3b:0 p:45:0 b:1f:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:26:a p:2a:a p:3b:a p:45:a t14 p:26:0 p:2a:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a b:1f:a g:3b:a g:3b:a g:3e:a g:3e:a g:43:a g:43:a pi:3b:a pi:3e:a pi:43:a t14 p:2a:0 p:3b:0 p:45:0 b:1f:0 t2 p:24:a p:2a:a p:3b:a p:45:a b:1f:a t14 p:24:0 p:2a:0 p:3b:0 p:45:0 b:1f:0 g:3b:0 g:3b:0 g:3e:0 g:3e:0 g:43:0 g:43:0 pi:3b:0 pi:3e:0 pi:43:0 t2 p:24:a p:2e:a p:3b:a p:45:a b:26:a g:39:a g:39:a g:3e:a g:3e:a g:42:a g:42:a pi:39:a pi:3e:a pi:42:a t14 p:24:0 p:2e:0 p:3b:0 p:45:0 t2 p:2a:a p:3b:a p:45:a t14 p:2a:0 p:3b:0",
+                "wav_name": "sample",
+                "sound_font_path": "assets/default_sound_font.sf2",
+            }
+        }
+
+
+@router.post("/text-to-wav", tags=["MIDI"])
+def text_to_wav(body: TextToWavBody):
+    """
+    Install fluidsynth first, see more: https://github.com/FluidSynth/fluidsynth/wiki/Download#distributions
+    """
+
+    text = body.text.strip()
+    if not text.startswith("<start>"):
+        text = "<start> " + text
+    if not text.endswith("<end>"):
+        text = text + " <end>"
+    txt_path = f"midi/{body.wav_name}.txt"
+    midi_path = f"midi/{body.wav_name}.mid"
+    wav_path = f"midi/{body.wav_name}.wav"
+    with open(txt_path, "w") as f:
+        f.write(text)
+    txt_to_midi(TxtToMidiBody(txt_path=txt_path, midi_path=midi_path))
+    midi_to_wav(
+        MidiToWavBody(
+            midi_path=midi_path, wav_path=wav_path, sound_font_path=body.sound_font_path
+        )
+    )
+
+    return "success"
--- a/backend-python/routes/misc.py
+++ b/backend-python/routes/misc.py
@@ -0,0 +1,131 @@
+from fastapi import APIRouter, HTTPException, status
+from utils.rwkv import AbstractRWKV
+import global_var
+
+router = APIRouter()
+
+
+@router.get("/dashboard/billing/credit_grants", tags=["MISC"])
+def credit_grants():
+    return {
+        "object": "credit_summary",
+        "total_granted": 10000,
+        "total_used": 0,
+        "total_available": 10000,
+        "grants": {
+            "object": "list",
+            "data": [
+                {
+                    "object": "credit_grant",
+                    "grant_amount": 10000,
+                    "used_amount": 0,
+                    "effective_at": 1672531200,
+                    "expires_at": 33229440000,
+                }
+            ],
+        },
+    }
+
+
+fake_models = [
+    {
+        "id": "gpt-3.5-turbo",
+        "object": "model",
+        "created": 1677610602,
+        "owned_by": "openai",
+        "permission": [
+            {
+                "id": "modelperm-zy5TOjnE2zVaicIcKO9bQDgX",
+                "object": "model_permission",
+                "created": 1690864883,
+                "allow_create_engine": False,
+                "allow_sampling": True,
+                "allow_logprobs": True,
+                "allow_search_indices": False,
+                "allow_view": True,
+                "allow_fine_tuning": False,
+                "organization": "*",
+                "group": None,
+                "is_blocking": False,
+            }
+        ],
+        "root": "gpt-3.5-turbo",
+        "parent": None,
+    },
+    {
+        "id": "text-davinci-003",
+        "object": "model",
+        "created": 1669599635,
+        "owned_by": "openai-internal",
+        "permission": [
+            {
+                "id": "modelperm-a6niqBmW2JaGmo0fDO7FEt1n",
+                "object": "model_permission",
+                "created": 1690930172,
+                "allow_create_engine": False,
+                "allow_sampling": True,
+                "allow_logprobs": True,
+                "allow_search_indices": False,
+                "allow_view": True,
+                "allow_fine_tuning": False,
+                "organization": "*",
+                "group": None,
+                "is_blocking": False,
+            }
+        ],
+        "root": "text-davinci-003",
+        "parent": None,
+    },
+]
+
+
+@router.get("/v1/models", tags=["MISC"])
+@router.get("/models", tags=["MISC"])
+def models():
+    model: AbstractRWKV = global_var.get(global_var.Model)
+    model_name = model.name if model else "rwkv"
+
+    return {
+        "object": "list",
+        "data": [
+            {
+                "id": model_name,
+                "object": "model",
+                "owned_by": "rwkv",
+                "root": model_name,
+                "parent": None,
+            },
+            *fake_models,
+        ],
+    }
+
+
+@router.get("/v1/models/{model_id}", tags=["MISC"])
+@router.get("/models/{model_id}", tags=["MISC"])
+def model(model_id: str):
+    for fake_model in fake_models:
+        if fake_model["id"] == model_id:
+            return fake_model
+
+    if "rwkv" in model_id.lower():
+        model: AbstractRWKV = global_var.get(global_var.Model)
+        model_name = model.name if model else "rwkv"
+        return {
+            "id": model_name,
+            "object": "model",
+            "owned_by": "rwkv",
+            "root": model_name,
+            "parent": None,
+        }
+
+    raise HTTPException(
+        status.HTTP_404_NOT_FOUND,
+        {
+            "error": {
+                "message": f"The model '{model_id}' does not exist",
+                "type": "invalid_request_error",
+                "param": "model",
+                "code": "model_not_found",
+            }
+        },
+    )
--- a/backend-python/routes/state_cache.py
+++ b/backend-python/routes/state_cache.py
@@ -1,11 +1,9 @@
-from typing import Any, Dict, List
+from typing import Any, Dict, List, Union
 from utils.log import quick_log
 from fastapi import APIRouter, HTTPException, Request, Response, status
 from pydantic import BaseModel
 import gc
 import copy
-import sys
-import torch

 router = APIRouter()

@@ -34,19 +32,47 @@ def init():
        print("cyac not found")


+@router.post("/disable-state-cache", tags=["State Cache"])
+def disable_state_cache():
+    global trie, dtrie
+
+    trie = None
+    dtrie = {}
+    gc.collect()
+
+    return "success"
+
+
+@router.post("/enable-state-cache", tags=["State Cache"])
+def enable_state_cache():
+    global trie, dtrie
+    try:
+        import cyac
+
+        trie = cyac.Trie()
+        dtrie = {}
+        gc.collect()
+
+        return "success"
+    except ModuleNotFoundError:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "cyac not found")
+
+
 class AddStateBody(BaseModel):
    prompt: str
-    tokens: List[str]
+    tokens: List[Union[str, int]]
    state: Any
    logits: Any


-@router.post("/add-state")
+@router.post("/add-state", tags=["State Cache"])
 def add_state(body: AddStateBody):
    global trie, dtrie, loop_del_trie_id
    if trie is None:
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")

+    import torch
+
    try:
        id: int = trie.insert(body.prompt)
        device: torch.device = body.state[0].device
@@ -70,7 +96,7 @@ def add_state(body: AddStateBody):
        quick_log(
            None,
            None,
-            f"New Trie Id: {id}\nTrie Len: {len(trie)}\nTrie Buff Size: {trie.buff_size()}\nDtrie Buff Size Of Id: {_get_a_dtrie_buff_size(dtrie[id])}",
+            f"New Trie Id: {id}\nTrie Len: {len(trie)}\nTrie Buff Size: {trie.buff_size()}\nDtrie Buff Size Of Id: {__get_a_dtrie_buff_size(dtrie[id])}",
        )
        return "success"
    except Exception as e:
@@ -79,12 +105,14 @@ def add_state(body: AddStateBody):
        )


-@router.post("/reset-state")
+@router.post("/reset-state", tags=["State Cache"])
 def reset_state():
    global trie, dtrie
    if trie is None:
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")

+    import cyac
+
    trie = cyac.Trie()
    dtrie = {}
    gc.collect()
@@ -96,7 +124,7 @@ class LongestPrefixStateBody(BaseModel):
    prompt: str


-def _get_a_dtrie_buff_size(dtrie_v):
+def __get_a_dtrie_buff_size(dtrie_v):
    # print(sys.getsizeof(dtrie_v["tokens"][0]))  # str
    # print(sys.getsizeof(dtrie_v["tokens"][0]) * len(dtrie_v["tokens"]))
    # print(dtrie_v["state"][0][0].element_size())
@@ -113,12 +141,14 @@ def _get_a_dtrie_buff_size(dtrie_v):
    return 54 * len(dtrie_v["tokens"]) + 491520 + 262144 + 28  # TODO


-@router.post("/longest-prefix-state")
+@router.post("/longest-prefix-state", tags=["State Cache"])
 def longest_prefix_state(body: LongestPrefixStateBody, request: Request):
    global trie
    if trie is None:
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")

+    import torch
+
    id = -1
    try:
        for id, len in trie.prefix(body.prompt):
@@ -150,7 +180,7 @@ def longest_prefix_state(body: LongestPrefixStateBody, request: Request):
        }


-@router.post("/save-state")
+@router.post("/save-state", tags=["State Cache"])
 def save_state():
    global trie
    if trie is None:
--- a/backend-python/rwkv_pip/beta/cuda/att_one.cu
+++ b/backend-python/rwkv_pip/beta/cuda/att_one.cu
@@ -0,0 +1,124 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "element_wise.h"
+#include "util.h"
+
+// Equivalent Python code:
+// ww = t_first + k
+// p = torch.maximum(pp, ww)
+// e1 = torch.exp(pp - p)
+// e2 = torch.exp(ww - p)
+// wkv = ((e1 * aa + e2 * v) / (e1 * bb + e2)).to(dtype=x.dtype)
+// ww = t_decay + pp
+// p = torch.maximum(ww, k)
+// e1 = torch.exp(ww - p)
+// e2 = torch.exp(k - p)
+// t1 = e1 * aa + e2 * v
+// t2 = e1 * bb + e2
+// r = r * wkv
+// return t1, t2, p, r
+struct WkvForwardOne {
+  const float *t_first;
+  const float *k;
+  const float *pp;
+  const float *aa;
+  const float *bb;
+  const float *t_decay;
+  const float *v;
+  /* out */ float *t1;
+  /* out */ float *t2;
+  /* out */ float *p;
+  /* in & out */ half *r;
+
+  __device__ void operator()(int i) const {
+    float ww = t_first[i] + k[i];
+    float pp_ = pp[i];
+    float p_ = (pp_ > ww) ? pp_ : ww;
+    float e1 = expf(pp_ - p_);
+    float e2 = expf(ww - p_);
+    float aa_ = aa[i];
+    float bb_ = bb[i];
+    float v_ = v[i];
+    r[i] = __hmul(r[i], __float2half(((e1 * aa_ + e2 * v_) / (e1 * bb_ + e2))));
+    ww = t_decay[i] + pp_;
+    float k_ = k[i];
+    p_ = (ww > k_) ? ww : k_;
+    e1 = expf(ww - p_);
+    e2 = expf(k_ - p_);
+    t1[i] = e1 * aa_ + e2 * v_;
+    t2[i] = e1 * bb_ + e2;
+    p[i] = p_;
+  }
+};
+
+/*
+   Equivalent Python code:
+   kx = xx * k_mix + sx * (1 - k_mix)
+   vx = xx * v_mix + sx * (1 - v_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+*/
+
+struct Mix {
+  const half *xx;
+  const half *sx;
+  const half *k_mix;
+  const half *v_mix;
+  const half *r_mix;
+  /* out */ half *kx;
+  /* out */ half *vx;
+  /* out */ half *rx;
+
+  __device__ void operator()(int i) const {
+    half xx_ = xx[i];
+    half sx_ = sx[i];
+    half k_mix_ = k_mix[i];
+    half v_mix_ = v_mix[i];
+    half r_mix_ = r_mix[i];
+    kx[i] = __hadd(__hmul(xx_, k_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+    vx[i] = __hadd(__hmul(xx_, v_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), v_mix_)));
+    rx[i] = __hadd(__hmul(xx_, r_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+  }
+};
+
+using torch::Tensor;
+
+void gemm_fp16_cublas_tensor(Tensor a, Tensor b, Tensor c);
+
+Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
+             Tensor v_mix, Tensor r_mix, Tensor kw,
+             /* imm */ Tensor kx, Tensor vw, /* imm */ Tensor vx, Tensor rw,
+             /* imm */ Tensor rx, Tensor ow, Tensor t_first,
+             /* imm */ Tensor k, Tensor pp, Tensor ww, Tensor aa, Tensor bb,
+             Tensor t_decay, /* imm */ Tensor v, /* in & out */ Tensor r,
+             /* out */ Tensor x_plus_out, /* out */ Tensor t1,
+             /* out */ Tensor t2, /* out */ Tensor p) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  element_wise(Mix{data_ptr<half>(xx), data_ptr<half>(sx),
+                   data_ptr<half>(k_mix), data_ptr<half>(v_mix),
+                   data_ptr<half>(r_mix), data_ptr<half>(kx),
+                   data_ptr<half>(vx), data_ptr<half>(rx)},
+               x.numel());
+
+  gemm_fp16_cublas_tensor(kx, kw, k);
+  gemm_fp16_cublas_tensor(vx, vw, v);
+  gemm_fp16_cublas_tensor(rx, rw, r);
+  at::sigmoid_(r);
+
+  element_wise(WkvForwardOne{data_ptr<float>(t_first), data_ptr<float>(k),
+                             data_ptr<float>(pp), data_ptr<float>(aa),
+                             data_ptr<float>(bb), data_ptr<float>(t_decay),
+                             data_ptr<float>(v), data_ptr<float>(t1),
+                             data_ptr<float>(t2), data_ptr<float>(p),
+                             data_ptr<half>(r)},
+               x.numel());
+
+  gemm_fp16_cublas_tensor(r, ow, x_plus_out);
+  x_plus_out += x;
+  return xx;
+}
--- a/backend-python/rwkv_pip/beta/cuda/att_one_v5.cu
+++ b/backend-python/rwkv_pip/beta/cuda/att_one_v5.cu
@@ -0,0 +1,109 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "element_wise.h"
+#include "util.h"
+
+// Equivalent Python code:
+// s1 = t_first * a + s
+// s2 = a + t_decay * s
+struct Fused1 {
+  const float *t_first;
+  const float *t_decay;
+  const float *a;
+  const float *s;
+  const int32_t inner_size;
+  /* out */ float *s1;
+  /* out */ float *s2;
+
+  __device__ void operator()(int i) const {
+    const int j = i / inner_size;
+    s1[i] = t_first[j] * a[i] + s[i];
+    s2[i] = a[i] + t_decay[j] * s[i];
+  }
+};
+
+/*
+   Equivalent Python code:
+   kx = xx * k_mix + sx * (1 - k_mix)
+   vx = xx * v_mix + sx * (1 - v_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+*/
+
+struct Mix {
+  const half *xx;
+  const half *sx;
+  const half *k_mix;
+  const half *v_mix;
+  const half *r_mix;
+  /* out */ half *kx;
+  /* out */ half *vx;
+  /* out */ half *rx;
+
+  __device__ void operator()(int i) const {
+    half xx_ = xx[i];
+    half sx_ = sx[i];
+    half k_mix_ = k_mix[i];
+    half v_mix_ = v_mix[i];
+    half r_mix_ = r_mix[i];
+    kx[i] = __hadd(__hmul(xx_, k_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+    vx[i] = __hadd(__hmul(xx_, v_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), v_mix_)));
+    rx[i] = __hadd(__hmul(xx_, r_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+  }
+};
+
+using torch::Tensor;
+
+void gemm_fp16_cublas_tensor(Tensor a, Tensor b, Tensor c);
+
+Tensor att_one_v5(Tensor x, Tensor sx, Tensor s, Tensor ln_w, Tensor ln_b,
+                  Tensor lx_w, Tensor lx_b, Tensor k_mix, Tensor v_mix,
+                  Tensor r_mix, Tensor kw,
+                  /* imm */ Tensor kx, Tensor vw, /* imm */ Tensor vx,
+                  Tensor rw,
+                  /* imm */ Tensor rx, Tensor ow, Tensor t_first,
+                  /* imm */ Tensor k, Tensor t_decay, /* imm */ Tensor v,
+                  /* imm */ Tensor r, /* imm */ Tensor s1,
+                  /* out */ Tensor x_plus_out, /* out */ Tensor s2) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  element_wise(Mix{data_ptr<half>(xx), data_ptr<half>(sx),
+                   data_ptr<half>(k_mix), data_ptr<half>(v_mix),
+                   data_ptr<half>(r_mix), data_ptr<half>(kx),
+                   data_ptr<half>(vx), data_ptr<half>(rx)},
+               x.numel());
+
+  int H = t_decay.size(0);
+  int S = x.size(-1) / H;
+  gemm_fp16_cublas_tensor(rx, rw, r);
+  r = at::reshape(r, {H, 1, S});
+  gemm_fp16_cublas_tensor(kx, kw, k);
+  k = at::reshape(k, {H, S, 1});
+  gemm_fp16_cublas_tensor(vx, vw, v);
+  v = at::reshape(v, {H, 1, S});
+
+  {
+    Tensor a = at::matmul(k, v);
+
+    // s1 = t_first * a + s
+    // s2 = a + t_decay * s
+    element_wise(Fused1{data_ptr<float>(t_first), data_ptr<float>(t_decay),
+                        data_ptr<float>(a), data_ptr<float>(s),
+                        static_cast<int32_t>(a.size(1) * a.size(2)),
+                        data_ptr<float>(s1), data_ptr<float>(s2)},
+                 a.numel());
+  }
+
+  Tensor out = at::matmul(r, s1);
+  out = at::flatten(out);
+  out = at::squeeze(at::group_norm(at::unsqueeze(out, 0), H, lx_w, lx_b), 0);
+  out = at::_cast_Half(out);
+
+  gemm_fp16_cublas_tensor(out, ow, x_plus_out);
+  x_plus_out += x;
+  return xx;
+}
--- a/backend-python/rwkv_pip/beta/cuda/att_seq.cu
+++ b/backend-python/rwkv_pip/beta/cuda/att_seq.cu
@@ -0,0 +1,178 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "util.h"
+#include "element_wise.h"
+
+using torch::Tensor;
+
+void gemm_fp16_cublas(const void *a, const void *b, void *c, int m,
+                      int n, int k, bool output_fp32);
+
+// based on `kernel_wkv_forward`, fusing more operations
+__global__ void kernel_wkv_forward_new(
+    const int B, const int T, const int C, const float *__restrict__ const _w,
+    const float *__restrict__ const _u, const float *__restrict__ const _k,
+    const float *__restrict__ const _v, const half *__restrict__ const r,
+    half *__restrict__ const _y, float *__restrict__ const _aa,
+    float *__restrict__ const _bb, float *__restrict__ const _pp) {
+  const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+  const int _b = idx / C;
+  const int _c = idx % C;
+  const int _offset = _b * T * C + _c;
+  const int _state_offset = _b * C + _c;
+
+  float u = _u[_c];
+  float w = _w[_c];
+  const float *__restrict__ const k = _k + _offset;
+  const float *__restrict__ const v = _v + _offset;
+  half *__restrict__ const y = _y + _offset;
+
+  float aa = _aa[_state_offset];
+  float bb = _bb[_state_offset];
+  float pp = _pp[_state_offset];
+  for (int i = 0; i < T; i++) {
+    const int ii = i * C;
+    const float kk = k[ii];
+    const float vv = v[ii];
+    float ww = u + kk;
+    float p = max(pp, ww);
+    float e1 = exp(pp - p);
+    float e2 = exp(ww - p);
+    y[ii] = __float2half((e1 * aa + e2 * vv) / (e1 * bb + e2));
+    ww = w + pp;
+    p = max(ww, kk);
+    e1 = exp(ww - p);
+    e2 = exp(kk - p);
+    aa = e1 * aa + e2 * vv;
+    bb = e1 * bb + e2;
+    pp = p;
+  }
+  _aa[_state_offset] = aa;
+  _bb[_state_offset] = bb;
+  _pp[_state_offset] = pp;
+}
+
+void cuda_wkv_forward_new(int B, int T, int C, float *w, float *u, float *k,
+                          float *v, half *r, half *y, float *aa, float *bb,
+                          float *pp) {
+  dim3 threadsPerBlock(min(C, 32));
+  assert(B * C % threadsPerBlock.x == 0);
+  dim3 numBlocks(B * C / threadsPerBlock.x);
+  kernel_wkv_forward_new<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, r,
+                                                         y, aa, bb, pp);
+}
+
+__global__ void _att_mix(const half *xx, const half *sx, const half *k_mix,
+                        const half *v_mix, const half *r_mix,
+                        const int outer_size, const int inner_size, half *kx,
+                        half *vx, half *rx) {
+  for (int idx2 = blockIdx.x * blockDim.x + threadIdx.x; idx2 < inner_size;
+       idx2 += blockDim.x * gridDim.x) {
+    half k_mix_ = k_mix[idx2];
+    half v_mix_ = v_mix[idx2];
+    half r_mix_ = r_mix[idx2];
+    for (int row = 0; row < outer_size; ++row) {
+      int idx1 = row * inner_size + idx2;
+      half xx_ = xx[idx1];
+      half sx_ = sx[idx1];
+      kx[idx1] = __hadd(__hmul(xx_, k_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+      vx[idx1] = __hadd(__hmul(xx_, v_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), v_mix_)));
+      rx[idx1] = __hadd(__hmul(xx_, r_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+    }
+  }
+}
+
+void att_mix(const half *xx, const half *sx, const half *k_mix,
+            const half *v_mix, const half *r_mix, const int outer_size,
+            const int inner_size, half *kx, half *vx, half *rx) {
+  // 256 is good enough on most GPUs
+  const int32_t BLOCK_SIZE = 256;
+  assert(inner_size % BLOCK_SIZE == 0);
+  _att_mix<<<inner_size / BLOCK_SIZE, BLOCK_SIZE>>>(
+      xx, sx, k_mix, v_mix, r_mix, outer_size, inner_size, kx, vx, rx);
+}
+
+struct InplaceSigmoid {
+  __device__ __forceinline__ half operator()(int i) const {
+    ptr[i] = __float2half(1.0 / (1.0 + exp(-__half2float(ptr[i]))));
+  }
+  half *ptr;
+};
+
+struct InplaceMul {
+  __device__ __forceinline__ half operator()(int i) const {
+    y[i] = __hmul(x[i], y[i]);
+  }
+  half *y;
+  half *x;
+};
+
+/*
+   Equivalent Python code:
+
+   xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+   sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+   kx = xx * k_mix + sx * (1 - k_mix)
+   vx = xx * v_mix + sx * (1 - v_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+
+   r = torch.sigmoid(gemm(rx, rw))
+   k = gemm(kx, kw, output_dtype=torch.float32)
+   v = gemm(vx, vw, output_dtype=torch.float32)
+
+   T = x.shape[0]
+   for t in range(T):
+       kk = k[t]
+       vv = v[t]
+       ww = t_first + kk
+       p = torch.maximum(pp, ww)
+       e1 = torch.exp(pp - p)
+       e2 = torch.exp(ww - p)
+       sx[t] = ((e1 * aa + e2 * vv) / (e1 * bb + e2)).to(dtype=x.dtype)
+       ww = t_decay + pp
+       p = torch.maximum(ww, kk)
+       e1 = torch.exp(ww - p)
+       e2 = torch.exp(kk - p)
+       aa = e1 * aa + e2 * vv
+       bb = e1 * bb + e2
+       pp = p
+   out = gemm(r * sx, ow)
+   return x + out, xx[-1,:], aa, bb, pp
+*/
+Tensor att_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor v_mix, Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               Tensor ow, Tensor t_first, Tensor pp, Tensor aa, Tensor bb,
+               Tensor t_decay, /* imm */ Tensor buf, /* out */ Tensor x_plus_out) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  sx = at::cat({sx.unsqueeze(0), xx.slice(0, 0, -1)}, 0);
+  char* buf_ptr = (char*)buf.data_ptr();
+  half* kx = (half*)buf_ptr;
+  half* vx = kx + x.numel();
+  half* rx = vx + x.numel();
+  half* wkv_y = rx + x.numel();
+  att_mix(data_ptr<half>(xx), data_ptr<half>(sx), data_ptr<half>(k_mix),
+         data_ptr<half>(v_mix), data_ptr<half>(r_mix), xx.size(0), xx.size(1),
+         kx, vx, rx);
+  float* k = reinterpret_cast<float*>(wkv_y + x.numel());
+  float* v = k + x.size(0) * kw.size(1);
+  half* r = reinterpret_cast<half*>(v + x.size(0) * vw.size(1));
+
+  gemm_fp16_cublas(kx, kw.data_ptr(), k, x.size(0), kw.size(1), kw.size(0), true);
+  gemm_fp16_cublas(vx, vw.data_ptr(), v, x.size(0), vw.size(1), vw.size(0), true);
+  gemm_fp16_cublas(rx, rw.data_ptr(), r, x.size(0), rw.size(1), rw.size(0), false);
+  element_wise(InplaceSigmoid{r}, x.size(0) * rw.size(1));
+  cuda_wkv_forward_new(1, x.size(0), x.size(1), data_ptr<float>(t_decay),
+                       data_ptr<float>(t_first), k, v, r,
+                       wkv_y, data_ptr<float>(aa),
+                       data_ptr<float>(bb), data_ptr<float>(pp));
+  element_wise(InplaceMul{wkv_y, r}, x.numel());
+  gemm_fp16_cublas(wkv_y, ow.data_ptr(), x_plus_out.data_ptr(), x.size(0), ow.size(1), ow.size(0), false);
+  x_plus_out += x;
+  return xx;
+}
--- a/backend-python/rwkv_pip/beta/cuda/element_wise.h
+++ b/backend-python/rwkv_pip/beta/cuda/element_wise.h
@@ -0,0 +1,21 @@
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+
+template <typename Func> __global__ void _element_wise(Func func, int n) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n;
+       i += blockDim.x * gridDim.x) {
+    func(i);
+  }
+}
+
+// NOTE: packed data type (e.g. float4) is a overkill for current sizes
+// (4096 in 7B model and 768 in 0.1B model),
+// and is not faster than the plain float version.
+template <typename Func>
+void element_wise(Func func, int n) {
+  // 256 is good enough on most GPUs
+  const int32_t BLOCK_SIZE = 256;
+  assert(n % BLOCK_SIZE == 0);
+  _element_wise<<<n / BLOCK_SIZE, BLOCK_SIZE>>>(func, n);
+}
--- a/backend-python/rwkv_pip/beta/cuda/ffn.cu
+++ b/backend-python/rwkv_pip/beta/cuda/ffn.cu
@@ -0,0 +1,165 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "element_wise.h"
+#include "util.h"
+
+using torch::Tensor;
+
+void gemm_fp16_cublas(const void *a, const void *b, void *c, int ori_m,
+                      int ori_n, int ori_k, bool output_fp32);
+
+__global__ void _ffn_seq_mix(const half *xx, const half *sx, const half *k_mix,
+                             const half *r_mix, const int outer_size,
+                             const int inner_size, half *kx, half *rx) {
+  for (int idx2 = blockIdx.x * blockDim.x + threadIdx.x; idx2 < inner_size;
+       idx2 += blockDim.x * gridDim.x) {
+    half k_mix_ = k_mix[idx2];
+    half r_mix_ = r_mix[idx2];
+    for (int row = 0; row < outer_size; ++row) {
+      int idx1 = row * inner_size + idx2;
+      half xx_ = xx[idx1];
+      half sx_ = sx[idx1];
+      kx[idx1] = __hadd(__hmul(xx_, k_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+      rx[idx1] = __hadd(__hmul(xx_, r_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+    }
+  }
+}
+
+void ffn_seq_mix(const half *xx, const half *sx, const half *k_mix,
+                 const half *r_mix, const int outer_size, const int inner_size,
+                 half *kx, half *rx) {
+  // 256 is good enough on most GPUs
+  const int32_t BLOCK_SIZE = 256;
+  assert(inner_size % BLOCK_SIZE == 0);
+  _ffn_seq_mix<<<inner_size / BLOCK_SIZE, BLOCK_SIZE>>>(
+      xx, sx, k_mix, r_mix, outer_size, inner_size, kx, rx);
+}
+
+struct InplaceSigmoid {
+  __device__ __forceinline__ void operator()(int i) const {
+    ptr[i] = __float2half(1.0 / (1.0 + exp(-__half2float(ptr[i]))));
+  }
+  half *ptr;
+};
+
+struct InplaceReLUAndSquare {
+  __device__ __forceinline__ void operator()(int i) const {
+    // __hmax is not defined in old cuda
+    if (__hgt(ptr[i], __float2half(0))) {
+      ptr[i] = __hmul(ptr[i], ptr[i]);
+    } else {
+      ptr[i] = __float2half(0);
+    }
+  }
+  half *ptr;
+};
+
+struct InplaceFma {
+  __device__ __forceinline__ void operator()(int i) const {
+    a[i] = __hfma(a[i], b[i], c[i]);
+  }
+  half *a;
+  const half *b;
+  const half *c;
+};
+
+/*
+   Equivalent Python code:
+
+   xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+   sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+   kx = xx * k_mix + sx * (1 - k_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+
+   r = torch.sigmoid(gemm(rx, rw))
+   vx = torch.square(torch.relu(gemm(kx, kw)))
+   out = r * gemm(vx, vw)
+   return x + out, xx[-1,:]
+*/
+Tensor ffn_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               /* imm */ Tensor buf,
+               /* out */ Tensor x_plus_out) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  sx = at::cat({sx.unsqueeze(0), xx.slice(0, 0, -1)}, 0);
+  char *buf_ptr = (char *)buf.data_ptr();
+  half *kx = (half *)buf_ptr;
+  half *rx = kx + x.numel();
+  half *vx = rx + x.numel();
+  half *r = vx + x.size(0) * kw.size(1);
+  ffn_seq_mix(data_ptr<half>(xx), data_ptr<half>(sx), data_ptr<half>(k_mix),
+              data_ptr<half>(r_mix), xx.size(0), xx.size(1), kx, rx);
+
+  gemm_fp16_cublas(rx, rw.data_ptr(), r, x.size(0), rw.size(1), x.size(1),
+                   false);
+  element_wise(InplaceSigmoid{r}, x.size(0) * rw.size(1));
+  gemm_fp16_cublas(kx, kw.data_ptr(), vx, x.size(0), kw.size(1), x.size(1),
+                   false);
+  element_wise(InplaceReLUAndSquare{vx}, x.size(0) * kw.size(1));
+  gemm_fp16_cublas(vx, vw.data_ptr(), x_plus_out.data_ptr(), x.size(0),
+                   vw.size(1), vw.size(0), false);
+  element_wise(InplaceFma{data_ptr<half>(x_plus_out), r, data_ptr<half>(x)},
+               x_plus_out.numel());
+  return xx;
+}
+
+struct FfnOneMix {
+  __device__ __forceinline__ void operator()(int idx) {
+    half k_mix_ = k_mix[idx];
+    half r_mix_ = r_mix[idx];
+    half xx_ = xx[idx];
+    half sx_ = sx[idx];
+    kx[idx] = __hadd(__hmul(xx_, k_mix_),
+                     __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+    rx[idx] = __hadd(__hmul(xx_, r_mix_),
+                     __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+  }
+  half *k_mix;
+  half *r_mix;
+  half *xx;
+  half *sx;
+  half *kx;
+  half *rx;
+};
+
+/*
+  Equivalent Python code:
+
+  xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+  kx = xx * k_mix + sx * (1 - k_mix)
+  rx = xx * r_mix + sx * (1 - r_mix)
+
+  r = torch.sigmoid(gemm(rx, rw))
+  vx = torch.square(torch.relu(gemm(kx, kw)))
+  out = r * gemm(vx, vw)
+  return x + out, xx
+*/
+Tensor ffn_one(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               /* imm */ Tensor buf,
+               /* out */ Tensor x_plus_out) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  char *buf_ptr = (char *)buf.data_ptr();
+  half *kx = (half *)buf_ptr;
+  half *rx = kx + x.numel();
+  half *vx = rx + x.numel();
+  half *r = vx + x.size(0) * kw.size(1);
+  element_wise(FfnOneMix{data_ptr<half>(k_mix), data_ptr<half>(r_mix),
+                         data_ptr<half>(xx), data_ptr<half>(sx), kx, rx},
+               x.numel());
+  // vector * matrix, so m = 1
+  gemm_fp16_cublas(rx, rw.data_ptr(), r, 1, rw.size(1), rw.size(0), false);
+  element_wise(InplaceSigmoid{r}, rw.size(1));
+  gemm_fp16_cublas(kx, kw.data_ptr(), vx, 1, kw.size(1), kw.size(0), false);
+  element_wise(InplaceReLUAndSquare{vx}, kw.size(1));
+  gemm_fp16_cublas(vx, vw.data_ptr(), x_plus_out.data_ptr(), 1, vw.size(1),
+                   vw.size(0), false);
+  element_wise(InplaceFma{data_ptr<half>(x_plus_out), r, data_ptr<half>(x)},
+               x_plus_out.numel());
+  return xx;
+}
--- a/backend-python/rwkv_pip/beta/cuda/gemm_fp16_cublas.cpp
+++ b/backend-python/rwkv_pip/beta/cuda/gemm_fp16_cublas.cpp
@@ -0,0 +1,128 @@
+#include <cublas_v2.h>
+#include <cuda.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#define CUBLAS_CHECK(condition)                                                \
+  for (cublasStatus_t _cublas_check_status = (condition);                      \
+       _cublas_check_status != CUBLAS_STATUS_SUCCESS;)                         \
+    throw std::runtime_error("cuBLAS error " +                                 \
+                             std::to_string(_cublas_check_status) + " at " +   \
+                             std::to_string(__LINE__));
+
+#define CUDA_CHECK(condition)                                                  \
+  for (cudaError_t _cuda_check_status = (condition);                           \
+       _cuda_check_status != cudaSuccess;)                                     \
+    throw std::runtime_error(                                                  \
+        "CUDA error " + std::string(cudaGetErrorString(_cuda_check_status)) +  \
+        " at " + std::to_string(__LINE__));
+
+cublasHandle_t get_cublas_handle() {
+  static cublasHandle_t cublas_handle = []() {
+    cublasHandle_t handle = nullptr;
+    CUBLAS_CHECK(cublasCreate(&handle));
+#if CUDA_VERSION < 11000
+    CUBLAS_CHECK(cublasSetMathMode(handle, CUBLAS_TENSOR_OP_MATH));
+#else
+    CUBLAS_CHECK(cublasSetMathMode(handle, CUBLAS_DEFAULT_MATH));
+#endif // CUDA_VERSION < 11000
+    return handle;
+  }();
+  return cublas_handle;
+}
+
+/*
+  NOTE: blas gemm is column-major by default, but we need row-major output.
+  The data of row-major, transposed matrix is exactly the same as the
+  column-major, non-transposed matrix, and C = A * B ---> C^T = B^T * A^T
+ */
+void gemm_fp16_cublas(const void *a, const void *b, void *c, int ori_m,
+                      int ori_n, int ori_k, bool output_fp32) {
+  const auto cuda_data_type = CUDA_R_16F;
+  const auto cuda_c_data_type = output_fp32 ? CUDA_R_32F : CUDA_R_16F;
+  const auto compute_type = CUDA_R_32F;
+  const float sp_alpha = 1.f;
+  // use CUBLAS_OP_N. see the notes above
+  const cublasOperation_t cublas_trans_a = CUBLAS_OP_N;
+  const cublasOperation_t cublas_trans_b = CUBLAS_OP_N;
+  // m = (B^T).size(0) = B.size(1) = n;
+  const int cublas_m = ori_n;
+  const int cublas_k = ori_k;
+  // comptiable with rwkv one mode, where 1-D tensor * 2-D tensor
+  // const int n = a.dense_dim() == 1 ? 1 : a.size(0);
+  const int cublas_n = ori_m;
+  const int cublas_lda = cublas_m;
+  const int cublas_ldb = cublas_k;
+  const int cublas_ldc = cublas_m;
+  cublasHandle_t cublas_handle = get_cublas_handle();
+
+#if CUDA_VERSION >= 11000
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT;
+#else
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT_TENSOR_OP;
+#endif
+  const float sp_beta = 0.f;
+  CUBLAS_CHECK(cublasGemmEx(
+      cublas_handle, cublas_trans_a, cublas_trans_b, cublas_m, cublas_n,
+      cublas_k, &sp_alpha, b, cuda_data_type, cublas_lda,
+      a, cuda_data_type, cublas_ldb, &sp_beta, c,
+      cuda_c_data_type, cublas_ldc, compute_type, algo));
+}
+
+/*
+  NOTE: blas gemm is column-major by default, but we need row-major output.
+  The data of row-major, transposed matrix is exactly the same as the
+  column-major, non-transposed matrix, and C = A * B ---> C^T = B^T * A^T
+ */
+void gemm_fp16_cublas_tensor(torch::Tensor a, torch::Tensor b, torch::Tensor c) {
+  if (a.sizes().size() == 1) {
+    assert(b.sizes().size() == 2);
+    a = at::unsqueeze(a, 0);
+  }
+  const auto cuda_data_type = CUDA_R_16F;
+  const auto cuda_c_data_type =
+      c.dtype() == torch::kFloat32 ? CUDA_R_32F : CUDA_R_16F;
+  const auto compute_type = CUDA_R_32F;
+  const float sp_alpha = 1.f;
+  // swap a and b, and use CUBLAS_OP_N. see the notes above
+  std::swap(a, b);
+  const cublasOperation_t cublas_trans_a = CUBLAS_OP_N;
+  const cublasOperation_t cublas_trans_b = CUBLAS_OP_N;
+  // m = (B^T).size(0) = B.size(1), and = A.size(1) after swap,
+  // negative axis is used because of the existence of batch matmul.
+  const int m = a.size(-1);
+  const int k = a.size(-2);
+  const int n = b.size(-2);
+  const int cublas_lda = m;
+  const int cublas_ldb = k;
+  const int cublas_ldc = m;
+  cublasHandle_t cublas_handle = get_cublas_handle();
+
+#if CUDA_VERSION >= 11000
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT;
+#else
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT_TENSOR_OP;
+#endif
+  const float sp_beta = 0.f;
+  if (a.sizes().size() == 2 && b.sizes().size() == 2) {
+    CUBLAS_CHECK(cublasGemmEx(
+        cublas_handle, cublas_trans_a, cublas_trans_b, m, n, k, &sp_alpha,
+        a.data_ptr(), cuda_data_type, cublas_lda, b.data_ptr(), cuda_data_type,
+        cublas_ldb, &sp_beta, c.data_ptr(), cuda_c_data_type, cublas_ldc,
+        compute_type, algo));
+  } else {
+    // batch matmul
+    assert(a.sizes().size() == 3 && b.sizes().size() == 3);
+
+    const long long int cublas_stride_a = m * k;
+    const long long int cublas_stride_b = k * n;
+    const long long int cublas_stride_c = m * n;
+    CUBLAS_CHECK(cublasGemmStridedBatchedEx(
+        cublas_handle, cublas_trans_a, cublas_trans_b, m,
+        n, k, &sp_alpha, a.data_ptr(), cuda_data_type, cublas_lda,
+        cublas_stride_a, b.data_ptr(), cuda_data_type, cublas_ldb, cublas_stride_b,
+        &sp_beta, c.data_ptr(), cuda_c_data_type, cublas_ldc, cublas_stride_c,
+        a.size(0), compute_type, algo));
+  }
+}
--- a/backend-python/rwkv_pip/beta/cuda/operators.cu
+++ b/backend-python/rwkv_pip/beta/cuda/operators.cu
@@ -0,0 +1,246 @@
+#include <stdio.h>
+#include <assert.h>
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#define MIN_VALUE (-1e38)
+typedef at::Half fp16;
+__half *cast(fp16 *ptr) {
+    return reinterpret_cast<__half *>(ptr);
+}
+
+template <typename F>
+__global__ void kernel_wkv_forward(const int B, const int T, const int C,
+                               const float *__restrict__ const _w, const float *__restrict__ const _u, const F *__restrict__ const _k, const F *__restrict__ const _v,
+                               F *__restrict__ const _y, float *__restrict__ const _aa, float *__restrict__ const _bb, float *__restrict__ const _pp) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int _b = idx / C;
+    const int _c = idx % C;
+    const int _offset = _b * T * C + _c;
+    const int _state_offset = _b * C + _c;
+
+    float u = _u[_c];
+    float w = _w[_c];
+    const F *__restrict__ const k = _k + _offset;
+    const F *__restrict__ const v = _v + _offset;
+    F *__restrict__ const y = _y + _offset;
+
+    float aa = _aa[_state_offset];
+    float bb = _bb[_state_offset];
+    float pp = _pp[_state_offset];
+    for (int i = 0; i < T; i++) {
+        const int ii = i * C;
+        const float kk = float(k[ii]);
+        const float vv = float(v[ii]);
+        float ww = u + kk;
+        float p = max(pp, ww);
+        float e1 = exp(pp - p);
+        float e2 = exp(ww - p);
+        y[ii] = F((e1 * aa + e2 * vv) / (e1 * bb + e2));
+        ww = w + pp;
+        p = max(ww, kk);
+        e1 = exp(ww - p);
+        e2 = exp(kk - p);
+        aa = e1 * aa + e2 * vv;
+        bb = e1 * bb + e2;
+        pp = p;
+    }
+    _aa[_state_offset] = aa;
+    _bb[_state_offset] = bb;
+    _pp[_state_offset] = pp;
+}
+
+template <typename F>
+void cuda_wkv_forward(int B, int T, int C, float *w, float *u, F *k, F *v, F *y, float *aa, float *bb, float *pp) {
+    dim3 threadsPerBlock( min(C, 32) );
+    assert(B * C % threadsPerBlock.x == 0);
+    dim3 numBlocks(B * C / threadsPerBlock.x);
+    kernel_wkv_forward<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, y, aa, bb, pp);
+}
+
+template void cuda_wkv_forward<fp16>(
+    int B, int T, int C,
+    float *w, float *u, fp16 *k, fp16 *v, fp16 *y,
+    float *aa, float *bb, float *pp);
+template void cuda_wkv_forward<float>(
+    int B, int T, int C,
+    float *w, float *u, float *k, float *v, float *y,
+    float *aa, float *bb, float *pp);
+
+__global__ void kernel_mm_seq_fp32i8(
+    const int B, const int N, const int M,
+    const float *__restrict__ const x, const int x_stride,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const float *__restrict__ const mx,
+    const float *__restrict__ const rx,
+    const float *__restrict__ const my,
+    const float *__restrict__ const ry,
+    float *__restrict__ const y, const int y_stride) {
+
+    const int i = blockIdx.x * blockDim.x + threadIdx.x;
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i < B && k < M) {
+        float y_local = 0;
+        for (int j = 0; j < N; ++j) {
+            y_local += x[i * x_stride + j] * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * rx[k] * ry[j] + mx[k] + my[j]
+            );
+        }
+        y[i * y_stride + k] = y_local;
+    }
+}
+
+template <typename F>
+void cuda_mm8_seq(int B, int N, int M,
+                  F *x, int x_stride,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  F *y, int y_stride);
+
+template <>
+void cuda_mm8_seq<float>(int B, int N, int M,
+                         float *x, int x_stride,
+                         uint8_t *w, int w_stride,
+                         float *mx, float *rx,
+                         float *my, float *ry,
+                         float *y, int y_stride) {
+    dim3 blockSize(1, 128);
+    dim3 gridSize((B + blockSize.x - 1) / blockSize.x, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_seq_fp32i8<<<gridSize, blockSize>>>(
+        B, N, M, x, x_stride, w, w_stride,
+        mx, rx, my, ry, y, y_stride);
+}
+
+__global__ void kernel_mm_seq_fp16i8(
+    const int B, const int N, const int M,
+    const __half *__restrict__ const x, const int x_stride,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const __half *__restrict__ const mx,
+    const __half *__restrict__ const rx,
+    const __half *__restrict__ const my,
+    const __half *__restrict__ const ry,
+    __half *__restrict__ const y, const int y_stride) {
+
+    const int i = blockIdx.x * blockDim.x + threadIdx.x;
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i < B && k < M) {
+        float y_local = 0;
+        for (int j = 0; j < N; ++j) {
+            y_local += __half2float(x[i * x_stride + j]) * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * __half2float(rx[k]) * __half2float(ry[j])
+                + __half2float(mx[k]) + __half2float(my[j])
+            );
+        }
+        y[i * y_stride + k] = __float2half(y_local);
+    }
+}
+
+template <>
+void cuda_mm8_seq<fp16>(int B, int N, int M,
+                        fp16 *x, int x_stride,
+                        uint8_t *w, int w_stride,
+                        fp16 *mx, fp16 *rx,
+                        fp16 *my, fp16 *ry,
+                        fp16 *y, int y_stride) {
+    dim3 blockSize(1, 128);
+    dim3 gridSize((B + blockSize.x - 1) / blockSize.x, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_seq_fp16i8<<<gridSize, blockSize>>>(
+        B, N, M, cast(x), x_stride, w, w_stride,
+        cast(mx), cast(rx), cast(my), cast(ry), cast(y), y_stride);
+}
+
+#define MM8_ONE_JSPLIT 24
+#define MM8_ONE_TILE 1024
+
+__global__ void kernel_mm_one_fp32i8(
+    const int N, const int M,
+    const float *__restrict__ const x,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const float *__restrict__ const mx,
+    const float *__restrict__ const rx,
+    const float *__restrict__ const my,
+    const float *__restrict__ const ry,
+    float *__restrict__ const y) {
+
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+    const int j0 = min(N, blockIdx.x * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+    const int j1 = min(N, (blockIdx.x + 1) * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+
+    if (k < M) {
+        float y_local = 0;
+        for (int j = j0; j < j1; ++j) {
+            y_local += x[j] * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * rx[k] * ry[j] + mx[k] + my[j]
+            );
+        }
+        atomicAdd(&y[k], y_local);
+    }
+}
+
+template <typename F>
+void cuda_mm8_one(int N, int M,
+                  F *x,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  float *y);
+
+template <>
+void cuda_mm8_one<float>(int N, int M,
+                        float *x,
+                        uint8_t *w, int w_stride,
+                        float *mx, float *rx,
+                        float *my, float *ry,
+                        float *y) {
+    dim3 blockSize(1, MM8_ONE_TILE);
+    dim3 gridSize(MM8_ONE_JSPLIT, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_one_fp32i8<<<gridSize, blockSize>>>(
+        N, M, x, w, w_stride,
+        mx, rx, my, ry, y);
+}
+
+__global__ void kernel_mm_one_fp16i8(
+    const int N, const int M,
+    const __half *__restrict__ const x,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const __half *__restrict__ const mx,
+    const __half *__restrict__ const rx,
+    const __half *__restrict__ const my,
+    const __half *__restrict__ const ry,
+    float *__restrict__ const y) {
+
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+    const int j0 = min(N, blockIdx.x * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+    const int j1 = min(N, (blockIdx.x + 1) * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+
+    if (k < M) {
+        float y_local = 0;
+        for (int j = j0; j < j1; ++j) {
+            y_local += __half2float(x[j]) * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * __half2float(rx[k]) * __half2float(ry[j])
+                + __half2float(mx[k]) + __half2float(my[j])
+            );
+        }
+        atomicAdd(&y[k], y_local);
+    }
+}
+
+template <>
+void cuda_mm8_one<fp16>(int N, int M,
+                        fp16 *x,
+                        uint8_t *w, int w_stride,
+                        fp16 *mx, fp16 *rx,
+                        fp16 *my, fp16 *ry,
+                        float *y) {
+    dim3 blockSize(1, MM8_ONE_TILE);
+    dim3 gridSize(MM8_ONE_JSPLIT, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_one_fp16i8<<<gridSize, blockSize>>>(
+        N, M, cast(x), w, w_stride,
+        cast(mx), cast(rx), cast(my), cast(ry), y);
+}
--- a/backend-python/rwkv_pip/beta/cuda/util.h
+++ b/backend-python/rwkv_pip/beta/cuda/util.h
@@ -0,0 +1,7 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+
+template <typename T> T *data_ptr(torch::Tensor x) { return x.data_ptr<T>(); }
+template <> inline half *data_ptr(torch::Tensor x) {
+  return reinterpret_cast<half *>(x.data_ptr<at::Half>());
+}
--- a/backend-python/rwkv_pip/beta/cuda/wrapper.cpp
+++ b/backend-python/rwkv_pip/beta/cuda/wrapper.cpp
@@ -0,0 +1,181 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+#include <iostream>
+#include <c10/cuda/CUDAGuard.h>
+
+typedef at::Half fp16;
+
+template <typename F>
+void cuda_wkv_forward(int B, int T, int C,
+                      float *w, float *u, F *k, F *v, F *y,
+                      float *aa, float *bb, float *pp);
+template <typename F>
+void cuda_mm8_seq(int B, int N, int M,
+                  F *x, int x_stride,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  F *y, int y_stride);
+template <typename F>
+void cuda_mm8_one(int N, int M,
+                  F *x,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  float *y);
+
+void wkv_forward(int64_t B, int64_t T, int64_t C,
+                 torch::Tensor &w, torch::Tensor &u,
+                 torch::Tensor &k, torch::Tensor &v, torch::Tensor &y,
+                 torch::Tensor &aa, torch::Tensor &bb, torch::Tensor &pp) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (k.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_wkv_forward(B, T, C,
+                         w.data_ptr<float>(), u.data_ptr<float>(),
+                         k.data_ptr<fp16>(), v.data_ptr<fp16>(), y.data_ptr<fp16>(),
+                         aa.data_ptr<float>(), bb.data_ptr<float>(), pp.data_ptr<float>());
+        break;
+    case c10::ScalarType::Float:
+        cuda_wkv_forward(B, T, C,
+                         w.data_ptr<float>(), u.data_ptr<float>(),
+                         k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(),
+                         aa.data_ptr<float>(), bb.data_ptr<float>(), pp.data_ptr<float>());
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+
+void mm8_seq(int64_t B, int64_t N, int64_t M,
+             torch::Tensor &x, torch::Tensor &w,
+             torch::Tensor &mx, torch::Tensor &rx,
+             torch::Tensor &my, torch::Tensor &ry,
+             torch::Tensor &y) {
+    assert(x.stride(1) == 1);
+    assert(w.stride(1) == 1);
+    assert(mx.stride(0) == 1 && rx.stride(0) == 1);
+    assert(my.stride(0) == 1 && ry.stride(0) == 1);
+    assert(y.stride(1) == 1);
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (x.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_mm8_seq(
+            B, N, M,
+            x.data_ptr<fp16>(), x.stride(0),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<fp16>(), rx.data_ptr<fp16>(),
+            my.data_ptr<fp16>(), ry.data_ptr<fp16>(),
+            y.data_ptr<fp16>(), y.stride(0));
+        break;
+    case c10::ScalarType::Float:
+        cuda_mm8_seq(
+            B, N, M,
+            x.data_ptr<float>(), x.stride(0),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<float>(), rx.data_ptr<float>(),
+            my.data_ptr<float>(), ry.data_ptr<float>(),
+            y.data_ptr<float>(), y.stride(0));
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+void mm8_one(int64_t N, int64_t M,
+             torch::Tensor &x, torch::Tensor &w,
+             torch::Tensor &mx, torch::Tensor &rx,
+             torch::Tensor &my, torch::Tensor &ry,
+             torch::Tensor &y) {
+    assert(x.stride(0) == 1);
+    assert(w.stride(1) == 1);
+    assert(mx.stride(0) == 1 && rx.stride(0) == 1);
+    assert(my.stride(0) == 1 && ry.stride(0) == 1);
+    assert(y.stride(0) == 1);
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (x.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_mm8_one(
+            N, M,
+            x.data_ptr<fp16>(),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<fp16>(), rx.data_ptr<fp16>(),
+            my.data_ptr<fp16>(), ry.data_ptr<fp16>(),
+            y.data_ptr<float>());
+        break;
+    case c10::ScalarType::Float:
+        cuda_mm8_one(
+            N, M,
+            x.data_ptr<float>(),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<float>(), rx.data_ptr<float>(),
+            my.data_ptr<float>(), ry.data_ptr<float>(),
+            y.data_ptr<float>());
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+
+using torch::Tensor;
+
+#ifndef DISABLE_CUBLAS_GEMM
+void gemm_fp16_cublas_tensor(Tensor a, Tensor b, Tensor c);
+#endif
+
+Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
+             Tensor v_mix, Tensor r_mix, Tensor kw,
+             /* imm */ Tensor kx, Tensor vw, /* imm */ Tensor vx, Tensor rw,
+             /* imm */ Tensor rx, Tensor ow, Tensor t_first,
+             /* imm */ Tensor k, Tensor pp, Tensor ww, Tensor aa, Tensor bb,
+             Tensor t_decay, /* imm */ Tensor v, /* in & out */ Tensor r,
+             /* out */ Tensor x_plus_out, /* out */ Tensor t1,
+             /* out */ Tensor t2, /* out */ Tensor p);
+
+Tensor att_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor v_mix, Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               Tensor ow, Tensor t_first, Tensor pp, Tensor aa, Tensor bb,
+               Tensor t_decay, /* imm */ Tensor buf, /* out */ Tensor x_plus_out);
+
+Tensor att_one_v5(Tensor x, Tensor sx, Tensor s, Tensor ln_w, Tensor ln_b,
+                  Tensor lx_w, Tensor lx_b, Tensor k_mix, Tensor v_mix,
+                  Tensor r_mix, Tensor kw,
+                  /* imm */ Tensor kx, Tensor vw, /* imm */ Tensor vx,
+                  Tensor rw,
+                  /* imm */ Tensor rx, Tensor ow, Tensor t_first,
+                  /* imm */ Tensor k, Tensor t_decay, /* imm */ Tensor v,
+                  /* imm */ Tensor r, /* imm */ Tensor s1,
+                  /* out */ Tensor x_plus_out, /* out */ Tensor s2);
+
+Tensor ffn_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               /* imm */ Tensor buf,
+               /* out */ Tensor x_plus_out);
+
+Tensor ffn_one(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               /* imm */ Tensor buf,
+               /* out */ Tensor x_plus_out);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("wkv_forward", &wkv_forward, "wkv forward");
+    m.def("mm8_seq", &mm8_seq, "mm8 seq");
+    m.def("mm8_one", &mm8_one, "mm8 one");
+    m.def("gemm_fp16_cublas", &gemm_fp16_cublas_tensor, "gemv fp16 cublas");
+    m.def("att_one", &att_one, "att one");
+    m.def("att_one_v5", &att_one_v5, "att one v5");
+    m.def("att_seq", &att_seq, "att seq");
+    m.def("ffn_seq", &ffn_seq, "ffn seq");
+    m.def("ffn_one", &ffn_one, "ffn one");
+}
+
+TORCH_LIBRARY(rwkv, m) {
+    m.def("wkv_forward", wkv_forward);
+    m.def("mm8_seq", mm8_seq);
+    m.def("mm8_one", mm8_one);
+    m.def("gemm_fp16_cublas", gemm_fp16_cublas_tensor);
+    m.def("att_one", att_one);
+    m.def("att_one_v5", &att_one_v5);
+    m.def("att_seq", att_seq);
+    m.def("ffn_seq", ffn_seq);
+    m.def("ffn_one", ffn_one);
+}
--- a/backend-python/rwkv_pip/beta/model.py
+++ b/backend-python/rwkv_pip/beta/model.py
--- a/backend-python/rwkv_pip/beta/wkv_cuda.pyd
+++ b/backend-python/rwkv_pip/beta/wkv_cuda.pyd
--- a/backend-python/rwkv_pip/cuda/att_one.cu
+++ b/backend-python/rwkv_pip/cuda/att_one.cu
@@ -0,0 +1,124 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "element_wise.h"
+#include "util.h"
+
+// Equivalent Python code:
+// ww = t_first + k
+// p = torch.maximum(pp, ww)
+// e1 = torch.exp(pp - p)
+// e2 = torch.exp(ww - p)
+// wkv = ((e1 * aa + e2 * v) / (e1 * bb + e2)).to(dtype=x.dtype)
+// ww = t_decay + pp
+// p = torch.maximum(ww, k)
+// e1 = torch.exp(ww - p)
+// e2 = torch.exp(k - p)
+// t1 = e1 * aa + e2 * v
+// t2 = e1 * bb + e2
+// r = r * wkv
+// return t1, t2, p, r
+struct WkvForwardOne {
+  const float *t_first;
+  const float *k;
+  const float *pp;
+  const float *aa;
+  const float *bb;
+  const float *t_decay;
+  const float *v;
+  /* out */ float *t1;
+  /* out */ float *t2;
+  /* out */ float *p;
+  /* in & out */ half *r;
+
+  __device__ void operator()(int i) const {
+    float ww = t_first[i] + k[i];
+    float pp_ = pp[i];
+    float p_ = (pp_ > ww) ? pp_ : ww;
+    float e1 = expf(pp_ - p_);
+    float e2 = expf(ww - p_);
+    float aa_ = aa[i];
+    float bb_ = bb[i];
+    float v_ = v[i];
+    r[i] = __hmul(r[i], __float2half(((e1 * aa_ + e2 * v_) / (e1 * bb_ + e2))));
+    ww = t_decay[i] + pp_;
+    float k_ = k[i];
+    p_ = (ww > k_) ? ww : k_;
+    e1 = expf(ww - p_);
+    e2 = expf(k_ - p_);
+    t1[i] = e1 * aa_ + e2 * v_;
+    t2[i] = e1 * bb_ + e2;
+    p[i] = p_;
+  }
+};
+
+/*
+   Equivalent Python code:
+   kx = xx * k_mix + sx * (1 - k_mix)
+   vx = xx * v_mix + sx * (1 - v_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+*/
+
+struct Mix {
+  const half *xx;
+  const half *sx;
+  const half *k_mix;
+  const half *v_mix;
+  const half *r_mix;
+  /* out */ half *kx;
+  /* out */ half *vx;
+  /* out */ half *rx;
+
+  __device__ void operator()(int i) const {
+    half xx_ = xx[i];
+    half sx_ = sx[i];
+    half k_mix_ = k_mix[i];
+    half v_mix_ = v_mix[i];
+    half r_mix_ = r_mix[i];
+    kx[i] = __hadd(__hmul(xx_, k_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+    vx[i] = __hadd(__hmul(xx_, v_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), v_mix_)));
+    rx[i] = __hadd(__hmul(xx_, r_mix_),
+                   __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+  }
+};
+
+using torch::Tensor;
+
+void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+
+Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
+             Tensor v_mix, Tensor r_mix, Tensor kw,
+             /* imm */ Tensor kx, Tensor vw, /* imm */ Tensor vx, Tensor rw,
+             /* imm */ Tensor rx, Tensor ow, Tensor t_first,
+             /* imm */ Tensor k, Tensor pp, Tensor ww, Tensor aa, Tensor bb,
+             Tensor t_decay, /* imm */ Tensor v, /* in & out */ Tensor r,
+             /* out */ Tensor x_plus_out, /* out */ Tensor t1,
+             /* out */ Tensor t2, /* out */ Tensor p) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  element_wise(Mix{data_ptr<half>(xx), data_ptr<half>(sx),
+                   data_ptr<half>(k_mix), data_ptr<half>(v_mix),
+                   data_ptr<half>(r_mix), data_ptr<half>(kx),
+                   data_ptr<half>(vx), data_ptr<half>(rx)},
+               x.numel());
+
+  gemm_fp16_cublas(kx, kw, k);
+  gemm_fp16_cublas(vx, vw, v);
+  gemm_fp16_cublas(rx, rw, r);
+  at::sigmoid_(r);
+
+  element_wise(WkvForwardOne{data_ptr<float>(t_first), data_ptr<float>(k),
+                             data_ptr<float>(pp), data_ptr<float>(aa),
+                             data_ptr<float>(bb), data_ptr<float>(t_decay),
+                             data_ptr<float>(v), data_ptr<float>(t1),
+                             data_ptr<float>(t2), data_ptr<float>(p),
+                             data_ptr<half>(r)},
+               x.numel());
+
+  gemm_fp16_cublas(r, ow, x_plus_out);
+  x_plus_out += x;
+  return xx;
+}
--- a/backend-python/rwkv_pip/cuda/att_seq.cu
+++ b/backend-python/rwkv_pip/cuda/att_seq.cu
@@ -0,0 +1,179 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "util.h"
+#include "element_wise.h"
+
+using torch::Tensor;
+
+void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+void gemm_fp16_cublas(const void *a, const void *b, void *c, int m,
+                      int n, int k, bool output_fp32);
+
+// based on `kernel_wkv_forward`, fusing more operations
+__global__ void kernel_wkv_forward_new(
+    const int B, const int T, const int C, const float *__restrict__ const _w,
+    const float *__restrict__ const _u, const float *__restrict__ const _k,
+    const float *__restrict__ const _v, const half *__restrict__ const r,
+    half *__restrict__ const _y, float *__restrict__ const _aa,
+    float *__restrict__ const _bb, float *__restrict__ const _pp) {
+  const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+  const int _b = idx / C;
+  const int _c = idx % C;
+  const int _offset = _b * T * C + _c;
+  const int _state_offset = _b * C + _c;
+
+  float u = _u[_c];
+  float w = _w[_c];
+  const float *__restrict__ const k = _k + _offset;
+  const float *__restrict__ const v = _v + _offset;
+  half *__restrict__ const y = _y + _offset;
+
+  float aa = _aa[_state_offset];
+  float bb = _bb[_state_offset];
+  float pp = _pp[_state_offset];
+  for (int i = 0; i < T; i++) {
+    const int ii = i * C;
+    const float kk = k[ii];
+    const float vv = v[ii];
+    float ww = u + kk;
+    float p = max(pp, ww);
+    float e1 = exp(pp - p);
+    float e2 = exp(ww - p);
+    y[ii] = __float2half((e1 * aa + e2 * vv) / (e1 * bb + e2));
+    ww = w + pp;
+    p = max(ww, kk);
+    e1 = exp(ww - p);
+    e2 = exp(kk - p);
+    aa = e1 * aa + e2 * vv;
+    bb = e1 * bb + e2;
+    pp = p;
+  }
+  _aa[_state_offset] = aa;
+  _bb[_state_offset] = bb;
+  _pp[_state_offset] = pp;
+}
+
+void cuda_wkv_forward_new(int B, int T, int C, float *w, float *u, float *k,
+                          float *v, half *r, half *y, float *aa, float *bb,
+                          float *pp) {
+  dim3 threadsPerBlock(min(C, 32));
+  assert(B * C % threadsPerBlock.x == 0);
+  dim3 numBlocks(B * C / threadsPerBlock.x);
+  kernel_wkv_forward_new<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, r,
+                                                         y, aa, bb, pp);
+}
+
+__global__ void _att_mix(const half *xx, const half *sx, const half *k_mix,
+                        const half *v_mix, const half *r_mix,
+                        const int outer_size, const int inner_size, half *kx,
+                        half *vx, half *rx) {
+  for (int idx2 = blockIdx.x * blockDim.x + threadIdx.x; idx2 < inner_size;
+       idx2 += blockDim.x * gridDim.x) {
+    half k_mix_ = k_mix[idx2];
+    half v_mix_ = v_mix[idx2];
+    half r_mix_ = r_mix[idx2];
+    for (int row = 0; row < outer_size; ++row) {
+      int idx1 = row * inner_size + idx2;
+      half xx_ = xx[idx1];
+      half sx_ = sx[idx1];
+      kx[idx1] = __hadd(__hmul(xx_, k_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+      vx[idx1] = __hadd(__hmul(xx_, v_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), v_mix_)));
+      rx[idx1] = __hadd(__hmul(xx_, r_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+    }
+  }
+}
+
+void att_mix(const half *xx, const half *sx, const half *k_mix,
+            const half *v_mix, const half *r_mix, const int outer_size,
+            const int inner_size, half *kx, half *vx, half *rx) {
+  // 256 is good enough on most GPUs
+  const int32_t BLOCK_SIZE = 256;
+  assert(inner_size % BLOCK_SIZE == 0);
+  _att_mix<<<inner_size / BLOCK_SIZE, BLOCK_SIZE>>>(
+      xx, sx, k_mix, v_mix, r_mix, outer_size, inner_size, kx, vx, rx);
+}
+
+struct InplaceSigmoid {
+  __device__ __forceinline__ half operator()(int i) const {
+    ptr[i] = __float2half(1.0 / (1.0 + exp(-__half2float(ptr[i]))));
+  }
+  half *ptr;
+};
+
+struct InplaceMul {
+  __device__ __forceinline__ half operator()(int i) const {
+    y[i] = __hmul(x[i], y[i]);
+  }
+  half *y;
+  half *x;
+};
+
+/*
+   Equivalent Python code:
+
+   xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+   sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+   kx = xx * k_mix + sx * (1 - k_mix)
+   vx = xx * v_mix + sx * (1 - v_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+
+   r = torch.sigmoid(gemm(rx, rw))
+   k = gemm(kx, kw, output_dtype=torch.float32)
+   v = gemm(vx, vw, output_dtype=torch.float32)
+
+   T = x.shape[0]
+   for t in range(T):
+       kk = k[t]
+       vv = v[t]
+       ww = t_first + kk
+       p = torch.maximum(pp, ww)
+       e1 = torch.exp(pp - p)
+       e2 = torch.exp(ww - p)
+       sx[t] = ((e1 * aa + e2 * vv) / (e1 * bb + e2)).to(dtype=x.dtype)
+       ww = t_decay + pp
+       p = torch.maximum(ww, kk)
+       e1 = torch.exp(ww - p)
+       e2 = torch.exp(kk - p)
+       aa = e1 * aa + e2 * vv
+       bb = e1 * bb + e2
+       pp = p
+   out = gemm(r * sx, ow)
+   return x + out, xx[-1,:], aa, bb, pp
+*/
+Tensor att_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor v_mix, Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               Tensor ow, Tensor t_first, Tensor pp, Tensor aa, Tensor bb,
+               Tensor t_decay, /* imm */ Tensor buf, /* out */ Tensor x_plus_out) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  sx = at::cat({sx.unsqueeze(0), xx.slice(0, 0, -1)}, 0);
+  char* buf_ptr = (char*)buf.data_ptr();
+  half* kx = (half*)buf_ptr;
+  half* vx = kx + x.numel();
+  half* rx = vx + x.numel();
+  half* wkv_y = rx + x.numel();
+  att_mix(data_ptr<half>(xx), data_ptr<half>(sx), data_ptr<half>(k_mix),
+         data_ptr<half>(v_mix), data_ptr<half>(r_mix), xx.size(0), xx.size(1),
+         kx, vx, rx);
+  float* k = reinterpret_cast<float*>(wkv_y + x.numel());
+  float* v = k + x.size(0) * kw.size(1);
+  half* r = reinterpret_cast<half*>(v + x.size(0) * vw.size(1));
+
+  gemm_fp16_cublas(kx, kw.data_ptr(), k, x.size(0), kw.size(1), kw.size(0), true);
+  gemm_fp16_cublas(vx, vw.data_ptr(), v, x.size(0), vw.size(1), vw.size(0), true);
+  gemm_fp16_cublas(rx, rw.data_ptr(), r, x.size(0), rw.size(1), rw.size(0), false);
+  element_wise(InplaceSigmoid{r}, x.size(0) * rw.size(1));
+  cuda_wkv_forward_new(1, x.size(0), x.size(1), data_ptr<float>(t_decay),
+                       data_ptr<float>(t_first), k, v, r,
+                       wkv_y, data_ptr<float>(aa),
+                       data_ptr<float>(bb), data_ptr<float>(pp));
+  element_wise(InplaceMul{wkv_y, r}, x.numel());
+  gemm_fp16_cublas(wkv_y, ow.data_ptr(), x_plus_out.data_ptr(), x.size(0), ow.size(1), ow.size(0), false);
+  x_plus_out += x;
+  return xx;
+}
--- a/backend-python/rwkv_pip/cuda/element_wise.h
+++ b/backend-python/rwkv_pip/cuda/element_wise.h
@@ -0,0 +1,21 @@
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+
+template <typename Func> __global__ void _element_wise(Func func, int n) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n;
+       i += blockDim.x * gridDim.x) {
+    func(i);
+  }
+}
+
+// NOTE: packed data type (e.g. float4) is a overkill for current sizes
+// (4096 in 7B model and 768 in 0.1B model),
+// and is not faster than the plain float version.
+template <typename Func>
+void element_wise(Func func, int n) {
+  // 256 is good enough on most GPUs
+  const int32_t BLOCK_SIZE = 256;
+  assert(n % BLOCK_SIZE == 0);
+  _element_wise<<<n / BLOCK_SIZE, BLOCK_SIZE>>>(func, n);
+}
--- a/backend-python/rwkv_pip/cuda/ffn.cu
+++ b/backend-python/rwkv_pip/cuda/ffn.cu
@@ -0,0 +1,165 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#include "element_wise.h"
+#include "util.h"
+
+using torch::Tensor;
+
+void gemm_fp16_cublas(const void *a, const void *b, void *c, int ori_m,
+                      int ori_n, int ori_k, bool output_fp32);
+
+__global__ void _ffn_seq_mix(const half *xx, const half *sx, const half *k_mix,
+                             const half *r_mix, const int outer_size,
+                             const int inner_size, half *kx, half *rx) {
+  for (int idx2 = blockIdx.x * blockDim.x + threadIdx.x; idx2 < inner_size;
+       idx2 += blockDim.x * gridDim.x) {
+    half k_mix_ = k_mix[idx2];
+    half r_mix_ = r_mix[idx2];
+    for (int row = 0; row < outer_size; ++row) {
+      int idx1 = row * inner_size + idx2;
+      half xx_ = xx[idx1];
+      half sx_ = sx[idx1];
+      kx[idx1] = __hadd(__hmul(xx_, k_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+      rx[idx1] = __hadd(__hmul(xx_, r_mix_),
+                        __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+    }
+  }
+}
+
+void ffn_seq_mix(const half *xx, const half *sx, const half *k_mix,
+                 const half *r_mix, const int outer_size, const int inner_size,
+                 half *kx, half *rx) {
+  // 256 is good enough on most GPUs
+  const int32_t BLOCK_SIZE = 256;
+  assert(inner_size % BLOCK_SIZE == 0);
+  _ffn_seq_mix<<<inner_size / BLOCK_SIZE, BLOCK_SIZE>>>(
+      xx, sx, k_mix, r_mix, outer_size, inner_size, kx, rx);
+}
+
+struct InplaceSigmoid {
+  __device__ __forceinline__ void operator()(int i) const {
+    ptr[i] = __float2half(1.0 / (1.0 + exp(-__half2float(ptr[i]))));
+  }
+  half *ptr;
+};
+
+struct InplaceReLUAndSquare {
+  __device__ __forceinline__ void operator()(int i) const {
+    // __hmax is not defined in old cuda
+    if (__hgt(ptr[i], __float2half(0))) {
+      ptr[i] = __hmul(ptr[i], ptr[i]);
+    } else {
+      ptr[i] = __float2half(0);
+    }
+  }
+  half *ptr;
+};
+
+struct InplaceFma {
+  __device__ __forceinline__ void operator()(int i) const {
+    a[i] = __hfma(a[i], b[i], c[i]);
+  }
+  half *a;
+  const half *b;
+  const half *c;
+};
+
+/*
+   Equivalent Python code:
+
+   xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+   sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
+   kx = xx * k_mix + sx * (1 - k_mix)
+   rx = xx * r_mix + sx * (1 - r_mix)
+
+   r = torch.sigmoid(gemm(rx, rw))
+   vx = torch.square(torch.relu(gemm(kx, kw)))
+   out = r * gemm(vx, vw)
+   return x + out, xx[-1,:]
+*/
+Tensor ffn_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               /* imm */ Tensor buf,
+               /* out */ Tensor x_plus_out) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  sx = at::cat({sx.unsqueeze(0), xx.slice(0, 0, -1)}, 0);
+  char *buf_ptr = (char *)buf.data_ptr();
+  half *kx = (half *)buf_ptr;
+  half *rx = kx + x.numel();
+  half *vx = rx + x.numel();
+  half *r = vx + x.size(0) * kw.size(1);
+  ffn_seq_mix(data_ptr<half>(xx), data_ptr<half>(sx), data_ptr<half>(k_mix),
+              data_ptr<half>(r_mix), xx.size(0), xx.size(1), kx, rx);
+
+  gemm_fp16_cublas(rx, rw.data_ptr(), r, x.size(0), rw.size(1), x.size(1),
+                   false);
+  element_wise(InplaceSigmoid{r}, x.size(0) * rw.size(1));
+  gemm_fp16_cublas(kx, kw.data_ptr(), vx, x.size(0), kw.size(1), x.size(1),
+                   false);
+  element_wise(InplaceReLUAndSquare{vx}, x.size(0) * kw.size(1));
+  gemm_fp16_cublas(vx, vw.data_ptr(), x_plus_out.data_ptr(), x.size(0),
+                   vw.size(1), vw.size(0), false);
+  element_wise(InplaceFma{data_ptr<half>(x_plus_out), r, data_ptr<half>(x)},
+               x_plus_out.numel());
+  return xx;
+}
+
+struct FfnOneMix {
+  __device__ __forceinline__ void operator()(int idx) {
+    half k_mix_ = k_mix[idx];
+    half r_mix_ = r_mix[idx];
+    half xx_ = xx[idx];
+    half sx_ = sx[idx];
+    kx[idx] = __hadd(__hmul(xx_, k_mix_),
+                     __hmul(sx_, __hsub(__float2half(1), k_mix_)));
+    rx[idx] = __hadd(__hmul(xx_, r_mix_),
+                     __hmul(sx_, __hsub(__float2half(1), r_mix_)));
+  }
+  half *k_mix;
+  half *r_mix;
+  half *xx;
+  half *sx;
+  half *kx;
+  half *rx;
+};
+
+/*
+  Equivalent Python code:
+
+  xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
+  kx = xx * k_mix + sx * (1 - k_mix)
+  rx = xx * r_mix + sx * (1 - r_mix)
+
+  r = torch.sigmoid(gemm(rx, rw))
+  vx = torch.square(torch.relu(gemm(kx, kw)))
+  out = r * gemm(vx, vw)
+  return x + out, xx
+*/
+Tensor ffn_one(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
+               Tensor r_mix, Tensor kw, Tensor vw, Tensor rw,
+               /* imm */ Tensor buf,
+               /* out */ Tensor x_plus_out) {
+  Tensor xx = at::layer_norm(x, {x.size(-1)}, ln_w, ln_b);
+  char *buf_ptr = (char *)buf.data_ptr();
+  half *kx = (half *)buf_ptr;
+  half *rx = kx + x.numel();
+  half *vx = rx + x.numel();
+  half *r = vx + x.size(0) * kw.size(1);
+  element_wise(FfnOneMix{data_ptr<half>(k_mix), data_ptr<half>(r_mix),
+                         data_ptr<half>(xx), data_ptr<half>(sx), kx, rx},
+               x.numel());
+  // vector * matrix, so m = 1
+  gemm_fp16_cublas(rx, rw.data_ptr(), r, 1, rw.size(1), rw.size(0), false);
+  element_wise(InplaceSigmoid{r}, rw.size(1));
+  gemm_fp16_cublas(kx, kw.data_ptr(), vx, 1, kw.size(1), kw.size(0), false);
+  element_wise(InplaceReLUAndSquare{vx}, kw.size(1));
+  gemm_fp16_cublas(vx, vw.data_ptr(), x_plus_out.data_ptr(), 1, vw.size(1),
+                   vw.size(0), false);
+  element_wise(InplaceFma{data_ptr<half>(x_plus_out), r, data_ptr<half>(x)},
+               x_plus_out.numel());
+  return xx;
+}
--- a/backend-python/rwkv_pip/cuda/gemm_fp16_cublas.cpp
+++ b/backend-python/rwkv_pip/cuda/gemm_fp16_cublas.cpp
@@ -0,0 +1,86 @@
+#include <cublas_v2.h>
+#include <cuda.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+
+#define CUBLAS_CHECK(condition)                                                \
+  for (cublasStatus_t _cublas_check_status = (condition);                      \
+       _cublas_check_status != CUBLAS_STATUS_SUCCESS;)                         \
+    throw std::runtime_error("cuBLAS error " +                                 \
+                             std::to_string(_cublas_check_status) + " at " +   \
+                             std::to_string(__LINE__));
+
+#define CUDA_CHECK(condition)                                                  \
+  for (cudaError_t _cuda_check_status = (condition);                           \
+       _cuda_check_status != cudaSuccess;)                                     \
+    throw std::runtime_error(                                                  \
+        "CUDA error " + std::string(cudaGetErrorString(_cuda_check_status)) +  \
+        " at " + std::to_string(__LINE__));
+
+cublasHandle_t get_cublas_handle() {
+  static cublasHandle_t cublas_handle = []() {
+    cublasHandle_t handle = nullptr;
+    CUBLAS_CHECK(cublasCreate(&handle));
+#if CUDA_VERSION < 11000
+    CUBLAS_CHECK(cublasSetMathMode(handle, CUBLAS_TENSOR_OP_MATH));
+#else
+    CUBLAS_CHECK(cublasSetMathMode(handle, CUBLAS_DEFAULT_MATH));
+#endif // CUDA_VERSION < 11000
+    return handle;
+  }();
+  return cublas_handle;
+}
+
+/*
+  NOTE: blas gemm is column-major by default, but we need row-major output.
+  The data of row-major, transposed matrix is exactly the same as the
+  column-major, non-transposed matrix, and C = A * B ---> C^T = B^T * A^T
+ */
+void gemm_fp16_cublas(torch::Tensor a, torch::Tensor b, torch::Tensor c) {
+  const auto cuda_data_type = CUDA_R_16F;
+  const auto cuda_c_data_type =
+      c.dtype() == torch::kFloat32 ? CUDA_R_32F : CUDA_R_16F;
+  const auto compute_type = CUDA_R_32F;
+  const float sp_alpha = 1.f;
+  // swap a and b, and use CUBLAS_OP_N. see the notes above
+  std::swap(a, b);
+  const cublasOperation_t cublas_trans_a = CUBLAS_OP_N;
+  const cublasOperation_t cublas_trans_b = CUBLAS_OP_N;
+  // m = (B^T).size(0) = B.size(1), and = A.size(1) after swap,
+  // negative axis is used because of the existence of batch matmul.
+  const int m = a.size(-1);
+  const int k = a.size(-2);
+  const int n = b.size(-2);
+  const int cublas_lda = m;
+  const int cublas_ldb = k;
+  const int cublas_ldc = m;
+  cublasHandle_t cublas_handle = get_cublas_handle();
+
+#if CUDA_VERSION >= 11000
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT;
+#else
+  cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT_TENSOR_OP;
+#endif
+  const float sp_beta = 0.f;
+  if (a.sizes().size() == 2 && b.sizes().size() == 2) {
+    CUBLAS_CHECK(cublasGemmEx(
+        cublas_handle, cublas_trans_a, cublas_trans_b, m, n, k, &sp_alpha,
+        a.data_ptr(), cuda_data_type, cublas_lda, b.data_ptr(), cuda_data_type,
+        cublas_ldb, &sp_beta, c.data_ptr(), cuda_c_data_type, cublas_ldc,
+        compute_type, algo));
+  } else {
+    // batch matmul
+    assert(a.sizes().size() == 3 && b.sizes().size() == 3);
+
+    const long long int cublas_stride_a = m * k;
+    const long long int cublas_stride_b = k * n;
+    const long long int cublas_stride_c = m * n;
+    CUBLAS_CHECK(cublasGemmStridedBatchedEx(
+        cublas_handle, cublas_trans_a, cublas_trans_b, m,
+        n, k, &sp_alpha, a.data_ptr(), cuda_data_type, cublas_lda,
+        cublas_stride_a, b.data_ptr(), cuda_data_type, cublas_ldb, cublas_stride_b,
+        &sp_beta, c.data_ptr(), cuda_c_data_type, cublas_ldc, cublas_stride_c,
+        a.size(0), compute_type, algo));
+  }
+}
--- a/backend-python/rwkv_pip/cuda/operators.cu
+++ b/backend-python/rwkv_pip/cuda/operators.cu
@@ -0,0 +1,246 @@
+#include <stdio.h>
+#include <assert.h>
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+#define MIN_VALUE (-1e38)
+typedef at::Half fp16;
+__half *cast(fp16 *ptr) {
+    return reinterpret_cast<__half *>(ptr);
+}
+
+template <typename F>
+__global__ void kernel_wkv_forward(const int B, const int T, const int C,
+                               const float *__restrict__ const _w, const float *__restrict__ const _u, const F *__restrict__ const _k, const F *__restrict__ const _v,
+                               F *__restrict__ const _y, float *__restrict__ const _aa, float *__restrict__ const _bb, float *__restrict__ const _pp) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int _b = idx / C;
+    const int _c = idx % C;
+    const int _offset = _b * T * C + _c;
+    const int _state_offset = _b * C + _c;
+
+    float u = _u[_c];
+    float w = _w[_c];
+    const F *__restrict__ const k = _k + _offset;
+    const F *__restrict__ const v = _v + _offset;
+    F *__restrict__ const y = _y + _offset;
+
+    float aa = _aa[_state_offset];
+    float bb = _bb[_state_offset];
+    float pp = _pp[_state_offset];
+    for (int i = 0; i < T; i++) {
+        const int ii = i * C;
+        const float kk = float(k[ii]);
+        const float vv = float(v[ii]);
+        float ww = u + kk;
+        float p = max(pp, ww);
+        float e1 = exp(pp - p);
+        float e2 = exp(ww - p);
+        y[ii] = F((e1 * aa + e2 * vv) / (e1 * bb + e2));
+        ww = w + pp;
+        p = max(ww, kk);
+        e1 = exp(ww - p);
+        e2 = exp(kk - p);
+        aa = e1 * aa + e2 * vv;
+        bb = e1 * bb + e2;
+        pp = p;
+    }
+    _aa[_state_offset] = aa;
+    _bb[_state_offset] = bb;
+    _pp[_state_offset] = pp;
+}
+
+template <typename F>
+void cuda_wkv_forward(int B, int T, int C, float *w, float *u, F *k, F *v, F *y, float *aa, float *bb, float *pp) {
+    dim3 threadsPerBlock( min(C, 32) );
+    assert(B * C % threadsPerBlock.x == 0);
+    dim3 numBlocks(B * C / threadsPerBlock.x);
+    kernel_wkv_forward<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, y, aa, bb, pp);
+}
+
+template void cuda_wkv_forward<fp16>(
+    int B, int T, int C,
+    float *w, float *u, fp16 *k, fp16 *v, fp16 *y,
+    float *aa, float *bb, float *pp);
+template void cuda_wkv_forward<float>(
+    int B, int T, int C,
+    float *w, float *u, float *k, float *v, float *y,
+    float *aa, float *bb, float *pp);
+
+__global__ void kernel_mm_seq_fp32i8(
+    const int B, const int N, const int M,
+    const float *__restrict__ const x, const int x_stride,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const float *__restrict__ const mx,
+    const float *__restrict__ const rx,
+    const float *__restrict__ const my,
+    const float *__restrict__ const ry,
+    float *__restrict__ const y, const int y_stride) {
+
+    const int i = blockIdx.x * blockDim.x + threadIdx.x;
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i < B && k < M) {
+        float y_local = 0;
+        for (int j = 0; j < N; ++j) {
+            y_local += x[i * x_stride + j] * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * rx[k] * ry[j] + mx[k] + my[j]
+            );
+        }
+        y[i * y_stride + k] = y_local;
+    }
+}
+
+template <typename F>
+void cuda_mm8_seq(int B, int N, int M,
+                  F *x, int x_stride,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  F *y, int y_stride);
+
+template <>
+void cuda_mm8_seq<float>(int B, int N, int M,
+                         float *x, int x_stride,
+                         uint8_t *w, int w_stride,
+                         float *mx, float *rx,
+                         float *my, float *ry,
+                         float *y, int y_stride) {
+    dim3 blockSize(1, 128);
+    dim3 gridSize((B + blockSize.x - 1) / blockSize.x, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_seq_fp32i8<<<gridSize, blockSize>>>(
+        B, N, M, x, x_stride, w, w_stride,
+        mx, rx, my, ry, y, y_stride);
+}
+
+__global__ void kernel_mm_seq_fp16i8(
+    const int B, const int N, const int M,
+    const __half *__restrict__ const x, const int x_stride,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const __half *__restrict__ const mx,
+    const __half *__restrict__ const rx,
+    const __half *__restrict__ const my,
+    const __half *__restrict__ const ry,
+    __half *__restrict__ const y, const int y_stride) {
+
+    const int i = blockIdx.x * blockDim.x + threadIdx.x;
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i < B && k < M) {
+        float y_local = 0;
+        for (int j = 0; j < N; ++j) {
+            y_local += __half2float(x[i * x_stride + j]) * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * __half2float(rx[k]) * __half2float(ry[j])
+                + __half2float(mx[k]) + __half2float(my[j])
+            );
+        }
+        y[i * y_stride + k] = __float2half(y_local);
+    }
+}
+
+template <>
+void cuda_mm8_seq<fp16>(int B, int N, int M,
+                        fp16 *x, int x_stride,
+                        uint8_t *w, int w_stride,
+                        fp16 *mx, fp16 *rx,
+                        fp16 *my, fp16 *ry,
+                        fp16 *y, int y_stride) {
+    dim3 blockSize(1, 128);
+    dim3 gridSize((B + blockSize.x - 1) / blockSize.x, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_seq_fp16i8<<<gridSize, blockSize>>>(
+        B, N, M, cast(x), x_stride, w, w_stride,
+        cast(mx), cast(rx), cast(my), cast(ry), cast(y), y_stride);
+}
+
+#define MM8_ONE_JSPLIT 24
+#define MM8_ONE_TILE 1024
+
+__global__ void kernel_mm_one_fp32i8(
+    const int N, const int M,
+    const float *__restrict__ const x,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const float *__restrict__ const mx,
+    const float *__restrict__ const rx,
+    const float *__restrict__ const my,
+    const float *__restrict__ const ry,
+    float *__restrict__ const y) {
+
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+    const int j0 = min(N, blockIdx.x * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+    const int j1 = min(N, (blockIdx.x + 1) * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+
+    if (k < M) {
+        float y_local = 0;
+        for (int j = j0; j < j1; ++j) {
+            y_local += x[j] * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * rx[k] * ry[j] + mx[k] + my[j]
+            );
+        }
+        atomicAdd(&y[k], y_local);
+    }
+}
+
+template <typename F>
+void cuda_mm8_one(int N, int M,
+                  F *x,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  float *y);
+
+template <>
+void cuda_mm8_one<float>(int N, int M,
+                        float *x,
+                        uint8_t *w, int w_stride,
+                        float *mx, float *rx,
+                        float *my, float *ry,
+                        float *y) {
+    dim3 blockSize(1, MM8_ONE_TILE);
+    dim3 gridSize(MM8_ONE_JSPLIT, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_one_fp32i8<<<gridSize, blockSize>>>(
+        N, M, x, w, w_stride,
+        mx, rx, my, ry, y);
+}
+
+__global__ void kernel_mm_one_fp16i8(
+    const int N, const int M,
+    const __half *__restrict__ const x,
+    const uint8_t *__restrict__ const w, const int w_stride,
+    const __half *__restrict__ const mx,
+    const __half *__restrict__ const rx,
+    const __half *__restrict__ const my,
+    const __half *__restrict__ const ry,
+    float *__restrict__ const y) {
+
+    const int k = blockIdx.y * blockDim.y + threadIdx.y;
+    const int j0 = min(N, blockIdx.x * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+    const int j1 = min(N, (blockIdx.x + 1) * ((N + MM8_ONE_JSPLIT - 1) / MM8_ONE_JSPLIT));
+
+    if (k < M) {
+        float y_local = 0;
+        for (int j = j0; j < j1; ++j) {
+            y_local += __half2float(x[j]) * (
+                (float(w[j * w_stride + k]) + 0.5f)
+                * __half2float(rx[k]) * __half2float(ry[j])
+                + __half2float(mx[k]) + __half2float(my[j])
+            );
+        }
+        atomicAdd(&y[k], y_local);
+    }
+}
+
+template <>
+void cuda_mm8_one<fp16>(int N, int M,
+                        fp16 *x,
+                        uint8_t *w, int w_stride,
+                        fp16 *mx, fp16 *rx,
+                        fp16 *my, fp16 *ry,
+                        float *y) {
+    dim3 blockSize(1, MM8_ONE_TILE);
+    dim3 gridSize(MM8_ONE_JSPLIT, (M + blockSize.y - 1) / blockSize.y);
+    kernel_mm_one_fp16i8<<<gridSize, blockSize>>>(
+        N, M, cast(x), w, w_stride,
+        cast(mx), cast(rx), cast(my), cast(ry), y);
+}
--- a/backend-python/rwkv_pip/cuda/rwkv5.cu
+++ b/backend-python/rwkv_pip/cuda/rwkv5.cu
@@ -0,0 +1,88 @@
+#include <stdio.h>
+#include <assert.h>
+#include "ATen/ATen.h"
+typedef at::BFloat16 bf16;
+typedef at::Half fp16;
+typedef float fp32;
+
+template <typename F>
+__global__ void kernel_forward(const int B, const int T, const int C, const int H, float *__restrict__ _state,
+                               const F *__restrict__ const _r, const F *__restrict__ const _k, const F *__restrict__ const _v, const float *__restrict__ _w, const F *__restrict__ _u,
+                               F *__restrict__ const _y)
+{
+    const int b = blockIdx.x / H;
+    const int h = blockIdx.x % H;
+    const int i = threadIdx.x;
+    _w += h*_N_;
+    _u += h*_N_;
+    _state += h*_N_*_N_ + i*_N_; // wrong if B > 1 !!!
+
+    __shared__ float r[_N_], k[_N_], u[_N_], w[_N_];
+    
+    float state[_N_];
+    #pragma unroll
+    for (int j = 0; j < _N_; j++)
+        state[j] = _state[j];
+    
+    __syncthreads();
+    u[i] = float(_u[i]);
+    w[i] = _w[i];
+    __syncthreads();
+
+    for (int t = b*T*C + h*_N_ + i; t < (b+1)*T*C + h*_N_ + i; t += C)
+    {
+        __syncthreads();
+        r[i] = float(_r[t]);
+        k[i] = float(_k[t]);
+        __syncthreads();
+
+        const float v = float(_v[t]);
+        float y = 0;
+
+        #pragma unroll
+        for (int j = 0; j < _N_; j+=4)
+        {
+            const float4& r_ = (float4&)(r[j]);
+            const float4& k_ = (float4&)(k[j]);
+            const float4& w_ = (float4&)(w[j]);
+            const float4& u_ = (float4&)(u[j]);
+            float4& s = (float4&)(state[j]);
+            float4 x;
+
+            x.x = k_.x * v;
+            x.y = k_.y * v;
+            x.z = k_.z * v;
+            x.w = k_.w * v;
+
+            y += r_.x * (u_.x * x.x + s.x);
+            y += r_.y * (u_.y * x.y + s.y);
+            y += r_.z * (u_.z * x.z + s.z);
+            y += r_.w * (u_.w * x.w + s.w);
+
+            s.x = s.x * w_.x + x.x;
+            s.y = s.y * w_.y + x.y;
+            s.z = s.z * w_.z + x.z;
+            s.w = s.w * w_.w + x.w;
+        }
+        _y[t] = F(y);
+    }
+    #pragma unroll
+    for (int j = 0; j < _N_; j++)
+        _state[j] = state[j];
+}
+
+void cuda_forward_bf16(int B, int T, int C, int H, float *state, bf16 *r, bf16 *k, bf16 *v, float *w, bf16 *u, bf16 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}
+void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *k, fp16 *v, float *w, fp16 *u, fp16 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}
+void cuda_forward_fp32(int B, int T, int C, int H, float *state, fp32 *r, fp32 *k, fp32 *v, float *w, fp32 *u, fp32 *y)
+{
+    assert(H*_N_ == C);
+    kernel_forward<<<dim3(B * H), dim3(_N_)>>>(B, T, C, H, state, r, k, v, w, u, y);
+}
--- a/backend-python/rwkv_pip/cuda/rwkv5_op.cpp
+++ b/backend-python/rwkv_pip/cuda/rwkv5_op.cpp
@@ -0,0 +1,30 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+typedef at::BFloat16 bf16;
+typedef at::Half fp16;
+typedef float fp32;
+
+void cuda_forward_bf16(int B, int T, int C, int H, float *state, bf16 *r, bf16 *k, bf16 *v, float *w, bf16 *u, bf16 *y);
+void cuda_forward_fp16(int B, int T, int C, int H, float *state, fp16 *r, fp16 *k, fp16 *v, float *w, fp16 *u, fp16 *y);
+void cuda_forward_fp32(int B, int T, int C, int H, float *state, fp32 *r, fp32 *k, fp32 *v, float *w, fp32 *u, fp32 *y);
+
+void forward_bf16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    cuda_forward_bf16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), w.data_ptr<float>(), u.data_ptr<bf16>(), y.data_ptr<bf16>());
+}
+void forward_fp16(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    cuda_forward_fp16(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp16>(), k.data_ptr<fp16>(), v.data_ptr<fp16>(), w.data_ptr<float>(), u.data_ptr<fp16>(), y.data_ptr<fp16>());
+}
+void forward_fp32(int64_t B, int64_t T, int64_t C, int64_t H, torch::Tensor &state, torch::Tensor &r, torch::Tensor &k, torch::Tensor &v, torch::Tensor &w, torch::Tensor &u, torch::Tensor &y) {
+    cuda_forward_fp32(B, T, C, H, state.data_ptr<float>(), r.data_ptr<fp32>(), k.data_ptr<fp32>(), v.data_ptr<fp32>(), w.data_ptr<float>(), u.data_ptr<fp32>(), y.data_ptr<fp32>());
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward_bf16", &forward_bf16, "rwkv5 forward_bf16");
+    m.def("forward_fp16", &forward_fp16, "rwkv5 forward_fp16");
+    m.def("forward_fp32", &forward_fp32, "rwkv5 forward_fp32");
+}
+TORCH_LIBRARY(rwkv5, m) {
+    m.def("forward_bf16", forward_bf16);
+    m.def("forward_fp16", forward_fp16);
+    m.def("forward_fp32", forward_fp32);
+}
--- a/backend-python/rwkv_pip/cuda/util.h
+++ b/backend-python/rwkv_pip/cuda/util.h
@@ -0,0 +1,7 @@
+#include "ATen/ATen.h"
+#include <cuda_fp16.h>
+
+template <typename T> T *data_ptr(torch::Tensor x) { return x.data_ptr<T>(); }
+template <> inline half *data_ptr(torch::Tensor x) {
+  return reinterpret_cast<half *>(x.data_ptr<at::Half>());
+}
--- a/backend-python/rwkv_pip/cuda/wrapper.cpp
+++ b/backend-python/rwkv_pip/cuda/wrapper.cpp
@@ -0,0 +1,141 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+#include <iostream>
+#include <c10/cuda/CUDAGuard.h>
+
+typedef at::Half fp16;
+
+template <typename F>
+void cuda_wkv_forward(int B, int T, int C,
+                      float *w, float *u, F *k, F *v, F *y,
+                      float *aa, float *bb, float *pp);
+template <typename F>
+void cuda_mm8_seq(int B, int N, int M,
+                  F *x, int x_stride,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  F *y, int y_stride);
+template <typename F>
+void cuda_mm8_one(int N, int M,
+                  F *x,
+                  uint8_t *w, int w_stride,
+                  F *mx, F *rx,
+                  F *my, F *ry,
+                  float *y);
+
+void wkv_forward(int64_t B, int64_t T, int64_t C,
+                 torch::Tensor &w, torch::Tensor &u,
+                 torch::Tensor &k, torch::Tensor &v, torch::Tensor &y,
+                 torch::Tensor &aa, torch::Tensor &bb, torch::Tensor &pp) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (k.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_wkv_forward(B, T, C,
+                         w.data_ptr<float>(), u.data_ptr<float>(),
+                         k.data_ptr<fp16>(), v.data_ptr<fp16>(), y.data_ptr<fp16>(),
+                         aa.data_ptr<float>(), bb.data_ptr<float>(), pp.data_ptr<float>());
+        break;
+    case c10::ScalarType::Float:
+        cuda_wkv_forward(B, T, C,
+                         w.data_ptr<float>(), u.data_ptr<float>(),
+                         k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(),
+                         aa.data_ptr<float>(), bb.data_ptr<float>(), pp.data_ptr<float>());
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+
+void mm8_seq(int64_t B, int64_t N, int64_t M,
+             torch::Tensor &x, torch::Tensor &w,
+             torch::Tensor &mx, torch::Tensor &rx,
+             torch::Tensor &my, torch::Tensor &ry,
+             torch::Tensor &y) {
+    assert(x.stride(1) == 1);
+    assert(w.stride(1) == 1);
+    assert(mx.stride(0) == 1 && rx.stride(0) == 1);
+    assert(my.stride(0) == 1 && ry.stride(0) == 1);
+    assert(y.stride(1) == 1);
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (x.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_mm8_seq(
+            B, N, M,
+            x.data_ptr<fp16>(), x.stride(0),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<fp16>(), rx.data_ptr<fp16>(),
+            my.data_ptr<fp16>(), ry.data_ptr<fp16>(),
+            y.data_ptr<fp16>(), y.stride(0));
+        break;
+    case c10::ScalarType::Float:
+        cuda_mm8_seq(
+            B, N, M,
+            x.data_ptr<float>(), x.stride(0),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<float>(), rx.data_ptr<float>(),
+            my.data_ptr<float>(), ry.data_ptr<float>(),
+            y.data_ptr<float>(), y.stride(0));
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+void mm8_one(int64_t N, int64_t M,
+             torch::Tensor &x, torch::Tensor &w,
+             torch::Tensor &mx, torch::Tensor &rx,
+             torch::Tensor &my, torch::Tensor &ry,
+             torch::Tensor &y) {
+    assert(x.stride(0) == 1);
+    assert(w.stride(1) == 1);
+    assert(mx.stride(0) == 1 && rx.stride(0) == 1);
+    assert(my.stride(0) == 1 && ry.stride(0) == 1);
+    assert(y.stride(0) == 1);
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(w));
+    switch (x.scalar_type()) {
+    case c10::ScalarType::Half:
+        cuda_mm8_one(
+            N, M,
+            x.data_ptr<fp16>(),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<fp16>(), rx.data_ptr<fp16>(),
+            my.data_ptr<fp16>(), ry.data_ptr<fp16>(),
+            y.data_ptr<float>());
+        break;
+    case c10::ScalarType::Float:
+        cuda_mm8_one(
+            N, M,
+            x.data_ptr<float>(),
+            w.data_ptr<uint8_t>(), w.stride(0),
+            mx.data_ptr<float>(), rx.data_ptr<float>(),
+            my.data_ptr<float>(), ry.data_ptr<float>(),
+            y.data_ptr<float>());
+        break;
+    default:
+        assert(false && "Only FP16 and FP32 are currently supported");
+    }
+}
+
+using torch::Tensor;
+
+#ifndef DISABLE_CUBLAS_GEMM
+void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+#endif
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("wkv_forward", &wkv_forward, "wkv forward");
+    m.def("mm8_seq", &mm8_seq, "mm8 seq");
+    m.def("mm8_one", &mm8_one, "mm8 one");
+#ifndef DISABLE_CUBLAS_GEMM
+    m.def("gemm_fp16_cublas", &gemm_fp16_cublas, "gemv fp16 cublas");
+#endif
+}
+
+TORCH_LIBRARY(rwkv, m) {
+    m.def("wkv_forward", wkv_forward);
+    m.def("mm8_seq", mm8_seq);
+    m.def("mm8_one", mm8_one);
+#ifndef DISABLE_CUBLAS_GEMM
+    m.def("gemm_fp16_cublas", gemm_fp16_cublas);
+#endif
+}
--- a/backend-python/rwkv_pip/model.py
+++ b/backend-python/rwkv_pip/model.py
--- a/backend-python/rwkv_pip/rwkv5.pyd
+++ b/backend-python/rwkv_pip/rwkv5.pyd
--- a/backend-python/rwkv_pip/tokenizer-midi.json
+++ b/backend-python/rwkv_pip/tokenizer-midi.json
--- a/backend-python/rwkv_pip/utils.py
+++ b/backend-python/rwkv_pip/utils.py
@@ -16,6 +16,7 @@ class PIPELINE_ARGS:
        top_k=0,
        alpha_frequency=0.2,
        alpha_presence=0.2,
+        alpha_decay=0.996,
        token_ban=[],
        token_stop=[],
        chunk_len=256,
@@ -25,6 +26,7 @@ class PIPELINE_ARGS:
        self.top_k = top_k
        self.alpha_frequency = alpha_frequency  # Frequency Penalty (as in GPT-3)
        self.alpha_presence = alpha_presence  # Presence Penalty (as in GPT-3)
+        self.alpha_decay = alpha_decay  # gradually decay the penalty
        self.token_ban = token_ban  # ban the generation of some tokens
        self.token_stop = token_stop  # stop generation whenever you see any token here
        self.chunk_len = (
@@ -33,7 +35,7 @@ class PIPELINE_ARGS:


 class PIPELINE:
-    def __init__(self, model, WORD_NAME):
+    def __init__(self, model, WORD_NAME: str):
        self.model = model
        if WORD_NAME == "cl100k_base":
            import tiktoken
@@ -47,9 +49,15 @@ class PIPELINE:
                os.path.dirname(os.path.abspath(__file__)) + "/rwkv_vocab_v20230424.txt"
            )
        else:
-            from tokenizers import Tokenizer
+            if WORD_NAME.endswith(".txt"):
+                sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+                from rwkv_tokenizer import TRIE_TOKENIZER

-            self.tokenizer = Tokenizer.from_file(WORD_NAME)
+                self.tokenizer = TRIE_TOKENIZER(WORD_NAME)
+            else:
+                from tokenizers import Tokenizer
+
+                self.tokenizer = Tokenizer.from_file(WORD_NAME)

    def refine_context(self, context):
        context = context.strip().split("\n")
@@ -78,7 +86,7 @@ class PIPELINE:
            sorted_ids = np.argsort(probs)
            sorted_probs = probs[sorted_ids][::-1]
            cumulative_probs = np.cumsum(sorted_probs)
-            cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
+            cutoff = float(sorted_probs[np.argmax(cumulative_probs >= top_p)])
            probs[probs < cutoff] = 0
            if top_k < len(probs) and top_k > 0:
                probs[sorted_ids[:-top_k]] = 0
@@ -92,7 +100,7 @@ class PIPELINE:
            sorted_probs = probs[sorted_ids]
            sorted_probs = torch.flip(sorted_probs, dims=(0,))
            cumulative_probs = torch.cumsum(sorted_probs, dim=-1).cpu().numpy()
-            cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
+            cutoff = float(sorted_probs[np.argmax(cumulative_probs >= top_p)])
            probs[probs < cutoff] = 0
            if top_k < len(probs) and top_k > 0:
                probs[sorted_ids[:-top_k]] = 0
@@ -127,10 +135,13 @@ class PIPELINE:
            if token in args.token_stop:
                break
            all_tokens += [token]
+            for xxx in occurrence:
+                occurrence[xxx] *= args.alpha_decay
            if token not in occurrence:
                occurrence[token] = 1
            else:
                occurrence[token] += 1
+            # print(occurrence) # debug

            # output
            tmp = self.decode(all_tokens[out_last:])
--- a/backend-python/rwkv_pip/wkv_cuda.pyd
+++ b/backend-python/rwkv_pip/wkv_cuda.pyd
--- a/backend-python/utils/log.py
+++ b/backend-python/utils/log.py
@@ -2,6 +2,8 @@ import json
 import logging
 from typing import Any
 from fastapi import Request
+from pydantic import BaseModel
+from enum import Enum


 logger = logging.getLogger()
@@ -14,12 +16,21 @@ fh.setFormatter(formatter)
 logger.addHandler(fh)


+class ClsEncoder(json.JSONEncoder):
+    def default(self, obj):
+        if isinstance(obj, BaseModel):
+            return obj.dict()
+        if isinstance(obj, Enum):
+            return obj.value
+        return super().default(obj)
+
+
 def quick_log(request: Request, body: Any, response: str):
    try:
        logger.info(
            f"Client: {request.client if request else ''}\nUrl: {request.url if request else ''}\n"
            + (
-                f"Body: {json.dumps(body.__dict__, default=vars, ensure_ascii=False)}\n"
+                f"Body: {json.dumps(body.__dict__, ensure_ascii=False, cls=ClsEncoder)}\n"
                if body
                else ""
            )
--- a/backend-python/utils/midi.py
+++ b/backend-python/utils/midi.py
@@ -0,0 +1,685 @@
+# https://github.com/briansemrau/MIDI-LLM-tokenizer
+
+# MIT License
+
+# Copyright (c) 2023 Brian Semrau
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+import json
+import random
+from dataclasses import dataclass
+from functools import lru_cache
+from math import ceil, floor, log
+from typing import Dict, Iterator, List, Optional, Tuple
+
+import mido
+
+
+@dataclass
+class VocabConfig:
+    # Number of note events. Should be 128.
+    note_events: int
+    # Number of wait events. Configurable, must evenly divide max_wait_time.
+    wait_events: int
+    # Max wait time in milliseconds to be represented by a single token.
+    max_wait_time: int
+    # Number of velocity events. Should be 128 (or 100? need to check midi standard)
+    velocity_events: int
+    # Number of bins to quantize velocity into. Should evenly divide velocity_events.
+    velocity_bins: int
+    # Exponential scaling factor for velocity bin sizes. 1.0 = linear scaling.
+    velocity_exp: float
+    # Whether to sort tokens by instrument, note. This should improve data reducibility.
+    do_token_sorting: bool
+    # Whether tokens should be represented as combined instrument/note/velocity tokens, or separate tokens for each.
+    unrolled_tokens: bool
+    # If non-zero, notes held for this many seconds will be automatically released during str->midi decoding.
+    decode_end_held_note_delay: float
+    # If true, repeated notes will be automatically released before playing again during str->midi decoding.
+    decode_fix_repeated_notes: bool
+    # List of instrument names to use for binning. Must have at most 16 values.
+    bin_instrument_names: List[str]
+    # Indicates which bin name represents percussion instruments on MIDI channel 10.
+    ch10_instrument_bin_name: str
+    # Mapping from instrument name to bin name.
+    program_name_to_bin_name: Dict[str, str]
+    # Mapping from bin name to program name.
+    bin_name_to_program_name: Dict[str, str]
+    # Mapping from program number to instrument name.
+    instrument_names: Dict[str, str]
+
+    def __post_init__(self):
+        self.validate()
+
+        self._instrument_names_str_to_int = {
+            name: int(i) for i, name in self.instrument_names.items()
+        }
+        self._instrument_names_int_to_str = {
+            int(i): name for i, name in self.instrument_names.items()
+        }
+
+        self._bin_str_to_int = {
+            name: int(i) for i, name in enumerate(self.bin_instrument_names)
+        }
+
+        self._bin_int_to_instrument_int = [
+            self._instrument_names_str_to_int[self.bin_name_to_program_name[name]]
+            if name != self.ch10_instrument_bin_name
+            else 0
+            for name in self.bin_instrument_names
+        ]
+        self._instrument_int_to_bin_int = [
+            self._bin_str_to_int[self.program_name_to_bin_name[instr]]
+            if self.program_name_to_bin_name[instr] != ""
+            else -1
+            for instr in self.program_name_to_bin_name.keys()
+        ]
+
+        self._ch10_bin_int = (
+            self._bin_str_to_int[self.ch10_instrument_bin_name]
+            if self.ch10_instrument_bin_name
+            else -1
+        )
+
+        self.short_instr_bin_names = []
+        for instr in self.bin_instrument_names:
+            i = min(1, len(instr))
+            while instr[:i] in self.short_instr_bin_names:
+                i += 1
+            self.short_instr_bin_names.append(instr[:i])
+        self._short_instrument_names_str_to_int = {
+            name: int(i) for i, name in enumerate(self.short_instr_bin_names)
+        }
+
+        range_excluding_ch10 = [
+            (i if i < 9 else i + 1) for i in range(len(self.bin_instrument_names))
+        ]
+        bins_excluding_ch10 = [
+            n for n in self.bin_instrument_names if n != self.ch10_instrument_bin_name
+        ]
+        self.bin_channel_map = {
+            bin: channel
+            for channel, bin in zip(range_excluding_ch10, bins_excluding_ch10)
+        }
+        if self.ch10_instrument_bin_name:
+            self.bin_channel_map[self.ch10_instrument_bin_name] = 9
+
+    def validate(self):
+        if self.max_wait_time % self.wait_events != 0:
+            raise ValueError("max_wait_time must be exactly divisible by wait_events")
+        if self.velocity_bins < 2:
+            raise ValueError("velocity_bins must be at least 2")
+        if len(self.bin_instrument_names) > 16:
+            raise ValueError("bin_instruments must have at most 16 values")
+        if (
+            self.ch10_instrument_bin_name
+            and self.ch10_instrument_bin_name not in self.bin_instrument_names
+        ):
+            raise ValueError("ch10_instrument_bin_name must be in bin_instruments")
+        if self.velocity_exp <= 0:
+            raise ValueError("velocity_exp must be greater than 0")
+
+    @classmethod
+    def from_json(cls, path: str):
+        with open(path, "r") as f:
+            config = json.load(f)
+        return cls(**config)
+
+
+class VocabUtils:
+    def __init__(self, cfg: VocabConfig) -> None:
+        self.cfg = cfg
+
+    @lru_cache(maxsize=128)
+    def format_wait_token(self, wait: int) -> str:
+        return f"t{wait}"
+
+    @lru_cache(maxsize=128)
+    def format_note_token(
+        self, instrument_bin: int, note: int, velocity_bin: int
+    ) -> str:
+        return f"{self.cfg.short_instr_bin_names[instrument_bin]}:{note:x}:{velocity_bin:x}"
+
+    def format_unrolled_note(self, note: int) -> str:
+        return f"n{note:x}"
+
+    def format_unrolled_velocity(self, velocity_bin: int) -> str:
+        return f"v{velocity_bin:x}"
+
+    def format_unrolled_instrument_bin(self, instrument_bin: int) -> str:
+        return f"i{self.cfg.short_instr_bin_names[instrument_bin]}"
+
+    def velocity_to_bin(self, velocity: float) -> int:
+        velocity = max(0, min(velocity, self.cfg.velocity_events - 1))
+        binsize = self.cfg.velocity_events / (self.cfg.velocity_bins - 1)
+        if self.cfg.velocity_exp == 1.0:
+            return ceil(velocity / binsize)
+        else:
+            return ceil(
+                (
+                    self.cfg.velocity_events
+                    * (
+                        (
+                            self.cfg.velocity_exp
+                            ** (velocity / self.cfg.velocity_events)
+                            - 1.0
+                        )
+                        / (self.cfg.velocity_exp - 1.0)
+                    )
+                )
+                / binsize
+            )
+
+    def bin_to_velocity(self, bin: int) -> int:
+        binsize = self.cfg.velocity_events / (self.cfg.velocity_bins - 1)
+        if self.cfg.velocity_exp == 1.0:
+            return max(0, ceil(bin * binsize - 1))
+        else:
+            return max(
+                0,
+                ceil(
+                    self.cfg.velocity_events
+                    * log(
+                        ((self.cfg.velocity_exp - 1) * binsize * bin)
+                        / self.cfg.velocity_events
+                        + 1,
+                        self.cfg.velocity_exp,
+                    )
+                    - 1
+                ),
+            )
+
+    def delta_to_wait_ids(self, delta_ms: float) -> Iterator[int]:
+        def roundi(f: float):
+            return ceil(f - 0.5)
+
+        max_wait_ms = self.cfg.max_wait_time
+        div = max_wait_ms / self.cfg.wait_events
+
+        # if delta_ms // max_wait_ms > 512:  # arbitrary limit to avoid excessive time_shifts
+        #    raise ValueError("delta_time is too large")
+        if delta_ms > max_wait_ms * 10:
+            delta_ms = max_wait_ms * 10  # truncate time
+
+        for _ in range(floor(delta_ms / max_wait_ms)):
+            yield roundi(max_wait_ms / div)
+        leftover_time_shift = roundi((delta_ms % max_wait_ms) / div)
+        if leftover_time_shift > 0:
+            yield leftover_time_shift
+
+    def prog_data_to_token_data(
+        self, program: int, channel: int, note: int, velocity: float
+    ) -> Optional[Tuple[int, int, int]]:
+        if channel == 9:
+            if self.cfg._ch10_bin_int == -1:
+                return None
+            return self.cfg._ch10_bin_int, note, self.velocity_to_bin(velocity)
+
+        instrument_bin = self.cfg._instrument_int_to_bin_int[program]
+        if instrument_bin != -1:
+            return instrument_bin, note, self.velocity_to_bin(velocity)
+        return None
+
+    def prog_data_list_to_token_data_list(
+        self, data: List[Tuple[int, int, int, float]]
+    ) -> Iterator[Tuple[int, int, int]]:
+        for d in data:
+            token_data = self.prog_data_to_token_data(*d)
+            if token_data is not None:
+                yield token_data
+
+    def sort_token_data(
+        self, data: List[Tuple[int, int, int]]
+    ) -> List[Tuple[int, int, int]]:
+        # ensure order is preserved for tokens with the same instrument, note
+        data = [(i, n, v, x) for x, (i, n, v) in enumerate(data)]
+        data.sort(key=lambda x: (x[0] != self.cfg._ch10_bin_int, x[0], x[1], x[3]))
+        return [(i, n, v) for i, n, v, _ in data]
+
+    def data_to_wait_tokens(self, delta_ms: float) -> List[str]:
+        if delta_ms == 0.0:
+            return []
+        return [self.format_wait_token(i) for i in self.delta_to_wait_ids(delta_ms)]
+
+    def wait_token_to_delta(self, token: str) -> float:
+        return self.cfg.max_wait_time / self.cfg.wait_events * int(token[1:])
+
+    def note_token_to_data(self, token: str) -> Tuple[int, int, int]:
+        instr_str, note_str, velocity_str = token.strip().split(":")
+        instr_bin = self.cfg._short_instrument_names_str_to_int[instr_str]
+        note = int(note_str, base=16)
+        velocity = self.bin_to_velocity(int(velocity_str, base=16))
+        return instr_bin, note, velocity
+
+
+@dataclass
+class AugmentValues:
+    instrument_bin_remap: Dict[int, int]
+    velocity_mod_factor: float
+    transpose_semitones: int
+    time_stretch_factor: float
+
+    @classmethod
+    def default(cls) -> "AugmentValues":
+        return cls(
+            instrument_bin_remap={},
+            velocity_mod_factor=1.0,
+            transpose_semitones=0,
+            time_stretch_factor=1.0,
+        )
+
+
+@dataclass
+class AugmentConfig:
+    # The number of times to augment each MIDI file. The dataset size will be multiplied by this number.
+    augment_data_factor: int
+    # A list of instrument names to randomly swap with each other.
+    instrument_mixups: List[List[str]]
+    # A list of percentages to change the note velocity by. 0.0 = no change. 0 is included by default.
+    velocity_mod_pct: List[float]
+    # A list of semitones to transpose by. 0 is included by default.
+    transpose_semitones: List[int]
+    # A list of percentages to stretch the tempo by. 0.0 = no stretch. 0 is included by default.
+    time_stretch_pct: List[float]
+    # Random seed to use for reproducibility.
+    seed: int
+
+    cfg: VocabConfig
+
+    def __post_init__(self):
+        self.validate()
+        if len(self.velocity_mod_pct) == 0:
+            self.velocity_mod_pct = [0.0]
+        if len(self.transpose_semitones) == 0:
+            self.transpose_semitones = [0]
+        if len(self.time_stretch_pct) == 0:
+            self.time_stretch_pct = [0.0]
+
+        self._instrument_mixups_int = [
+            [self.cfg._bin_str_to_int[i] for i in l if i in self.cfg._bin_str_to_int]
+            for l in self.instrument_mixups
+        ]
+        self._instrument_mixups_int = [
+            l for l in self._instrument_mixups_int if len(l) > 0
+        ]  # remove empty lists
+        self._instrument_pool_assignments = {}
+        self._mixup_pools = []
+        for pool_i, mixup_list in enumerate(self._instrument_mixups_int):
+            pool = set()
+            for i in mixup_list:
+                pool.add(i)
+                self._instrument_pool_assignments[i] = pool_i
+            self._mixup_pools.append(pool)
+
+    def validate(self):
+        if self.augment_data_factor < 1:
+            raise ValueError("augment_data_factor must be at least 1")
+        used_instruments = set()
+        for mixup_list in self.instrument_mixups:
+            for n in mixup_list:
+                if n in used_instruments:
+                    raise ValueError(f"Duplicate instrument name: {n}")
+                used_instruments.add(n)
+
+    @classmethod
+    def from_json(cls, path: str, cfg: VocabConfig):
+        with open(path, "r") as f:
+            config = json.load(f)
+        config["cfg"] = cfg
+        if "seed" not in config:
+            config["seed"] = random.randint(0, 2**32 - 1)
+        return cls(**config)
+
+    def get_augment_values(self, filename: str) -> Iterator[AugmentValues]:
+        # first yield default values
+        yield AugmentValues.default()
+
+        rng = random.Random(self.seed + hash(filename))
+        for _ in range(int(self.augment_data_factor - 1)):
+            # randomize order for each pool
+            randomized_pools = [list(pool) for pool in self._mixup_pools]
+            for pool in randomized_pools:
+                rng.shuffle(pool)
+            # distribute reassignments
+            instrument_bin_remap = {}
+            for i, pool in enumerate(randomized_pools):
+                for j, instrument in enumerate(pool):
+                    instrument_bin_remap[instrument] = randomized_pools[i - 1][j]
+            yield AugmentValues(
+                instrument_bin_remap=instrument_bin_remap,
+                velocity_mod_factor=1.0 + rng.choice(self.velocity_mod_pct),
+                transpose_semitones=rng.choice(self.transpose_semitones),
+                time_stretch_factor=1.0 + rng.choice(self.time_stretch_pct),
+            )
+
+
+def mix_volume(velocity: int, volume: int, expression: int) -> float:
+    return velocity * (volume / 127.0) * (expression / 127.0)
+
+
+def convert_midi_to_str(
+    cfg: VocabConfig, mid: mido.MidiFile, augment: AugmentValues = None
+) -> str:
+    utils = VocabUtils(cfg)
+    if augment is None:
+        augment = AugmentValues.default()
+
+    # filter out unknown meta messages before merge (https://github.com/mido/mido/pull/286)
+    for i in range(len(mid.tracks)):
+        mid.tracks[i] = [msg for msg in mid.tracks[i] if msg.type != "unknown_meta"]
+
+    if len(mid.tracks) > 1:
+        mid.tracks = [mido.merge_tracks(mid.tracks)]
+
+    delta_time_ms = 0.0
+    tempo = 500000
+    channel_program = {i: 0 for i in range(16)}
+    channel_volume = {i: 127 for i in range(16)}
+    channel_expression = {
+        i: 127 for i in range(16)
+    }  # unlikely to be useful. expression usually modifies an already played note.
+    channel_notes = {i: {} for i in range(16)}
+    channel_pedal_on = {i: False for i in range(16)}
+    channel_pedal_events = {
+        i: {} for i in range(16)
+    }  # {channel: {(note, program) -> True}}
+    started_flag = False
+
+    output = ["<start>"]
+    token_data_buffer: List[
+        Tuple[int, int, int, float]
+    ] = []  # need to sort notes between wait tokens
+
+    def flush_token_data_buffer():
+        nonlocal token_data_buffer, output, cfg, utils, augment
+        token_data = [
+            x for x in utils.prog_data_list_to_token_data_list(token_data_buffer)
+        ]
+        if augment.instrument_bin_remap or augment.transpose_semitones:
+            # TODO put transpose in a real function
+            raw_transpose = (
+                lambda bin, n: n + augment.transpose_semitones
+                if bin != cfg._ch10_bin_int
+                else n
+            )
+            octave_shift_if_oob = (
+                lambda n: n + 12 if n < 0 else n - 12 if n >= cfg.note_events else n
+            )
+            # TODO handle ranges beyond 12
+            # octave_shift_if_oob = lambda n: 0 if n < 0 else (n - cfg.note_events) % 12 + cfg.note_events if n >= cfg.note_events else n
+            transpose = lambda bin, n: octave_shift_if_oob(raw_transpose(bin, n))
+
+            token_data = [
+                (augment.instrument_bin_remap.get(i, i), transpose(i, n), v)
+                for i, n, v in token_data
+            ]
+        if cfg.do_token_sorting:
+            token_data = utils.sort_token_data(token_data)
+        if cfg.unrolled_tokens:
+            for t in token_data:
+                output += [
+                    utils.format_unrolled_instrument_bin(t[0]),
+                    utils.format_unrolled_note(t[1]),
+                    utils.format_unrolled_velocity(t[2]),
+                ]
+        else:
+            output += [utils.format_note_token(*t) for t in token_data]
+        token_data_buffer = []
+
+    def consume_note_program_data(prog: int, chan: int, note: int, vel: float):
+        nonlocal output, started_flag, delta_time_ms, cfg, utils, token_data_buffer
+        is_token_valid = (
+            utils.prog_data_to_token_data(prog, chan, note, vel) is not None
+        )
+        if not is_token_valid:
+            return
+        if started_flag:
+            wait_tokens = utils.data_to_wait_tokens(delta_time_ms)
+            if len(wait_tokens) > 0:
+                flush_token_data_buffer()
+                output += wait_tokens
+        delta_time_ms = 0.0
+        token_data_buffer.append((prog, chan, note, vel * augment.velocity_mod_factor))
+        started_flag = True
+
+    for msg in mid.tracks[0]:
+        time_ms = mido.tick2second(msg.time, mid.ticks_per_beat, tempo) * 1000.0
+        delta_time_ms += time_ms
+        t = msg.type
+
+        if msg.is_meta:
+            if t == "set_tempo":
+                tempo = msg.tempo * augment.time_stretch_factor
+            continue
+
+        def handle_note_off(ch, prog, n):
+            if channel_pedal_on[ch]:
+                channel_pedal_events[ch][(n, prog)] = True
+            else:
+                consume_note_program_data(prog, ch, n, 0)
+                if n in channel_notes[ch]:
+                    del channel_notes[ch][n]
+
+        if t == "program_change":
+            channel_program[msg.channel] = msg.program
+        elif t == "note_on":
+            if msg.velocity == 0:
+                handle_note_off(msg.channel, channel_program[msg.channel], msg.note)
+            else:
+                if (msg.note, channel_program[msg.channel]) in channel_pedal_events[
+                    msg.channel
+                ]:
+                    del channel_pedal_events[msg.channel][
+                        (msg.note, channel_program[msg.channel])
+                    ]
+                consume_note_program_data(
+                    channel_program[msg.channel],
+                    msg.channel,
+                    msg.note,
+                    mix_volume(
+                        msg.velocity,
+                        channel_volume[msg.channel],
+                        channel_expression[msg.channel],
+                    ),
+                )
+                channel_notes[msg.channel][msg.note] = True
+        elif t == "note_off":
+            handle_note_off(msg.channel, channel_program[msg.channel], msg.note)
+        elif t == "control_change":
+            if msg.control == 7 or msg.control == 39:  # volume
+                channel_volume[msg.channel] = msg.value
+            elif msg.control == 11:  # expression
+                channel_expression[msg.channel] = msg.value
+            elif msg.control == 64:  # sustain pedal
+                channel_pedal_on[msg.channel] = msg.value >= 64
+                if not channel_pedal_on[msg.channel]:
+                    for note, program in channel_pedal_events[msg.channel]:
+                        handle_note_off(msg.channel, program, note)
+                    channel_pedal_events[msg.channel] = {}
+            elif msg.control == 123:  # all notes off
+                for channel in channel_notes.keys():
+                    for note in list(channel_notes[channel]).copy():
+                        handle_note_off(channel, channel_program[channel], note)
+        else:
+            pass
+
+    flush_token_data_buffer()
+    output.append("<end>")
+    return " ".join(output)
+
+
+def generate_program_change_messages(cfg: VocabConfig):
+    for bin_name, channel in cfg.bin_channel_map.items():
+        if channel == 9:
+            continue
+        program = cfg._instrument_names_str_to_int[
+            cfg.bin_name_to_program_name[bin_name]
+        ]
+        yield mido.Message("program_change", program=program, time=0, channel=channel)
+    yield mido.Message("program_change", program=0, time=0, channel=9)
+
+
+@dataclass
+class DecodeState:
+    total_time: float  # milliseconds
+    delta_accum: float  # milliseconds
+    current_bin: int
+    current_note: int
+    active_notes: Dict[Tuple[int, int], float]  # { (channel, note): time started, ... }
+
+
+def token_to_midi_message(
+    utils: VocabUtils, token: str, state: DecodeState, end_token_pause: float = 3.0
+) -> Iterator[Tuple[Optional[mido.Message], DecodeState]]:
+    if state is None:
+        state = DecodeState(
+            total_time=0.0,
+            delta_accum=0.0,
+            current_bin=utils.cfg._short_instrument_names_str_to_int[
+                utils.cfg.short_instr_bin_names[0]
+            ],
+            current_note=0,
+            active_notes={},
+        )
+    token = token.strip()
+    if not token:
+        yield None, state
+        return
+    if token == "<end>":
+        d = end_token_pause * 1000.0
+        state.delta_accum += d
+        state.total_time += d
+        if utils.cfg.decode_end_held_note_delay != 0.0:
+            # end held notes
+            for (channel, note), start_time in list(state.active_notes.items()).copy():
+                ticks = int(mido.second2tick(state.delta_accum / 1000.0, 480, 500000))
+                state.delta_accum = 0.0
+                del state.active_notes[(channel, note)]
+                yield mido.Message(
+                    "note_off", note=note, time=ticks, channel=channel
+                ), state
+        yield None, state
+        return
+    if token.startswith("<"):
+        yield None, state
+        return
+
+    if utils.cfg.unrolled_tokens:
+        if token[0] == "t":
+            d = utils.wait_token_to_delta(token)
+            state.delta_accum += d
+            state.total_time += d
+        elif token[0] == "n":
+            state.current_note = int(token[1:], base=16)
+        elif token[0] == "i":
+            state.current_bin = utils.cfg._short_instrument_names_str_to_int[token[1:]]
+        elif token[0] == "v":
+            current_velocity = utils.bin_to_velocity(int(token[1:], base=16))
+            channel = utils.cfg.bin_channel_map[
+                utils.cfg.bin_instrument_names[state.current_bin]
+            ]
+            ticks = int(mido.second2tick(state.delta_accum / 1000.0, 480, 500000))
+            state.delta_accum = 0.0
+            if current_velocity > 0:
+                yield mido.Message(
+                    "note_on",
+                    note=state.current_note,
+                    velocity=current_velocity,
+                    time=ticks,
+                    channel=channel,
+                ), state
+            else:
+                yield mido.Message(
+                    "note_off",
+                    note=state.current_note,
+                    velocity=0,
+                    time=ticks,
+                    channel=channel,
+                ), state
+    else:
+        if token[0] == "t" and token[1].isdigit():  # wait token
+            d = utils.wait_token_to_delta(token)
+            state.delta_accum += d
+            state.total_time += d
+            if utils.cfg.decode_end_held_note_delay != 0.0:
+                # remove notes that have been held for too long
+                for (channel, note), start_time in list(
+                    state.active_notes.items()
+                ).copy():
+                    if (
+                        state.total_time - start_time
+                        > utils.cfg.decode_end_held_note_delay * 1000.0
+                    ):
+                        ticks = int(
+                            mido.second2tick(state.delta_accum / 1000.0, 480, 500000)
+                        )
+                        state.delta_accum = 0.0
+                        del state.active_notes[(channel, note)]
+                        yield mido.Message(
+                            "note_off", note=note, time=ticks, channel=channel
+                        ), state
+                        return
+        else:  # note token
+            bin, note, velocity = utils.note_token_to_data(token)
+            channel = utils.cfg.bin_channel_map[utils.cfg.bin_instrument_names[bin]]
+            ticks = int(mido.second2tick(state.delta_accum / 1000.0, 480, 500000))
+            state.delta_accum = 0.0
+            if velocity > 0:
+                if utils.cfg.decode_fix_repeated_notes:
+                    if (channel, note) in state.active_notes:
+                        del state.active_notes[(channel, note)]
+                    yield mido.Message(
+                        "note_off", note=note, time=ticks, channel=channel
+                    ), state
+                    ticks = 0
+                state.active_notes[(channel, note)] = state.total_time
+                yield mido.Message(
+                    "note_on", note=note, velocity=velocity, time=ticks, channel=channel
+                ), state
+                return
+            else:
+                if (channel, note) in state.active_notes:
+                    del state.active_notes[(channel, note)]
+                yield mido.Message(
+                    "note_off", note=note, time=ticks, channel=channel
+                ), state
+                return
+    yield None, state
+
+
+def str_to_midi_messages(utils: VocabUtils, data: str) -> Iterator[mido.Message]:
+    state = None
+    for token in data.split(" "):
+        for msg, new_state in token_to_midi_message(utils, token, state):
+            state = new_state
+            if msg is not None:
+                yield msg
+
+
+def convert_str_to_midi(
+    cfg: VocabConfig, data: str, meta_text: str = "Generated by MIDI-LLM-tokenizer"
+) -> mido.MidiFile:
+    utils = VocabUtils(cfg)
+    mid = mido.MidiFile()
+    track = mido.MidiTrack()
+    mid.tracks.append(track)
+
+    tempo = 500000
+    if meta_text:
+        track.append(mido.MetaMessage("text", text=meta_text, time=0))
+    track.append(mido.MetaMessage("set_tempo", tempo=tempo, time=0))
+    for msg in generate_program_change_messages(cfg):
+        track.append(msg)
+
+    # data = data.replace("<start>", "").replace("<end>", "").replace("<pad>", "").strip()
+    for msg in str_to_midi_messages(utils, data):
+        track.append(msg)
+
+    track.append(mido.MetaMessage("end_of_track", time=0))
+
+    return mid
--- a/backend-python/utils/midi_vocab_config.json
+++ b/backend-python/utils/midi_vocab_config.json
@@ -0,0 +1,303 @@
+{
+    "note_events": 128,
+    "wait_events": 125,
+    "max_wait_time": 1000,
+    "velocity_events": 128,
+    "velocity_bins": 12,
+    "velocity_exp": 0.5,
+    "do_token_sorting": true,
+    "unrolled_tokens": false,
+    "decode_end_held_note_delay": 5.0,
+    "decode_fix_repeated_notes": true,
+    "bin_instrument_names": [
+        "percussion",
+        "drum",
+        "tuba",
+        "marimba",
+        "bass",
+        "guitar",
+        "violin",
+        "trumpet",
+        "piano",
+        "sax",
+        "flute",
+        "lead",
+        "pad"
+    ],
+    "ch10_instrument_bin_name": "percussion",
+    "program_name_to_bin_name": {
+        "Acoustic Grand Piano": "piano",
+        "Bright Acoustic Piano": "piano",
+        "Electric Grand Piano": "piano",
+        "Honky-tonk Piano": "piano",
+        "Electric Piano 1 (Rhodes Piano)": "piano",
+        "Electric Piano 2 (Chorused Piano)": "piano",
+        "Harpsichord": "piano",
+        "Clavinet": "piano",
+        "Celesta": "marimba",
+        "Glockenspiel": "marimba",
+        "Music Box": "marimba",
+        "Vibraphone": "marimba",
+        "Marimba": "marimba",
+        "Xylophone": "marimba",
+        "Tubular Bells": "marimba",
+        "Dulcimer (Santur)": "marimba",
+        "Drawbar Organ (Hammond)": "marimba",
+        "Percussive Organ": "piano",
+        "Rock Organ": "piano",
+        "Church Organ": "piano",
+        "Reed Organ": "piano",
+        "Accordion (French)": "piano",
+        "Harmonica": "piano",
+        "Tango Accordion (Band neon)": "piano",
+        "Acoustic Guitar (nylon)": "guitar",
+        "Acoustic Guitar (steel)": "guitar",
+        "Electric Guitar (jazz)": "guitar",
+        "Electric Guitar (clean)": "guitar",
+        "Electric Guitar (muted)": "guitar",
+        "Overdriven Guitar": "guitar",
+        "Distortion Guitar": "guitar",
+        "Guitar harmonics": "guitar",
+        "Acoustic Bass": "bass",
+        "Electric Bass (fingered)": "bass",
+        "Electric Bass (picked)": "bass",
+        "Fretless Bass": "bass",
+        "Slap Bass 1": "bass",
+        "Slap Bass 2": "bass",
+        "Synth Bass 1": "bass",
+        "Synth Bass 2": "bass",
+        "Violin": "violin",
+        "Viola": "violin",
+        "Cello": "bass",
+        "Contrabass": "bass",
+        "Tremolo Strings": "violin",
+        "Pizzicato Strings": "violin",
+        "Orchestral Harp": "violin",
+        "Timpani": "drum",
+        "String Ensemble 1 (strings)": "violin",
+        "String Ensemble 2 (slow strings)": "violin",
+        "SynthStrings 1": "violin",
+        "SynthStrings 2": "violin",
+        "Choir Aahs": "violin",
+        "Voice Oohs": "violin",
+        "Synth Voice": "violin",
+        "Orchestra Hit": "",
+        "Trumpet": "trumpet",
+        "Trombone": "tuba",
+        "Tuba": "tuba",
+        "Muted Trumpet": "trumpet",
+        "French Horn": "trumpet",
+        "Brass Section": "trumpet",
+        "SynthBrass 1": "trumpet",
+        "SynthBrass 2": "trumpet",
+        "Soprano Sax": "sax",
+        "Alto Sax": "sax",
+        "Tenor Sax": "sax",
+        "Baritone Sax": "sax",
+        "Oboe": "sax",
+        "English Horn": "trumpet",
+        "Bassoon": "sax",
+        "Clarinet": "sax",
+        "Piccolo": "flute",
+        "Flute": "flute",
+        "Recorder": "flute",
+        "Pan Flute": "flute",
+        "Blown Bottle": "flute",
+        "Shakuhachi": "flute",
+        "Whistle": "flute",
+        "Ocarina": "flute",
+        "Lead 1 (square wave)": "lead",
+        "Lead 2 (sawtooth wave)": "lead",
+        "Lead 3 (calliope)": "lead",
+        "Lead 4 (chiffer)": "lead",
+        "Lead 5 (charang)": "lead",
+        "Lead 6 (voice solo)": "violin",
+        "Lead 7 (fifths)": "lead",
+        "Lead 8 (bass + lead)": "lead",
+        "Pad 1 (new age Fantasia)": "pad",
+        "Pad 2 (warm)": "pad",
+        "Pad 3 (polysynth)": "pad",
+        "Pad 4 (choir space voice)": "violin",
+        "Pad 5 (bowed glass)": "pad",
+        "Pad 6 (metallic pro)": "pad",
+        "Pad 7 (halo)": "pad",
+        "Pad 8 (sweep)": "pad",
+        "FX 1 (rain)": "",
+        "FX 2 (soundtrack)": "",
+        "FX 3 (crystal)": "",
+        "FX 4 (atmosphere)": "",
+        "FX 5 (brightness)": "",
+        "FX 6 (goblins)": "",
+        "FX 7 (echoes, drops)": "",
+        "FX 8 (sci-fi, star theme)": "",
+        "Sitar": "guitar",
+        "Banjo": "guitar",
+        "Shamisen": "guitar",
+        "Koto": "guitar",
+        "Kalimba": "guitar",
+        "Bag pipe": "sax",
+        "Fiddle": "violin",
+        "Shanai": "sax",
+        "Tinkle Bell": "marimba",
+        "Agogo": "marimba",
+        "Steel Drums": "marimba",
+        "Woodblock": "marimba",
+        "Taiko Drum": "drum",
+        "Melodic Tom": "drum",
+        "Synth Drum": "drum",
+        "Reverse Cymbal": "",
+        "Guitar Fret Noise": "",
+        "Breath Noise": "",
+        "Seashore": "",
+        "Bird Tweet": "",
+        "Telephone Ring": "",
+        "Helicopter": "",
+        "Applause": "",
+        "Gunshot": ""
+    },
+    "bin_name_to_program_name": {
+        "piano": "Acoustic Grand Piano",
+        "marimba": "Marimba",
+        "drum": "Synth Drum",
+        "guitar": "Acoustic Guitar (steel)",
+        "bass": "Acoustic Bass",
+        "violin": "Violin",
+        "percussion": "",
+        "trumpet": "Trumpet",
+        "tuba": "Tuba",
+        "sax": "Tenor Sax",
+        "flute": "Flute",
+        "lead": "Lead 1 (square wave)",
+        "pad": "Pad 1 (new age Fantasia)"
+    },
+    "instrument_names": {
+        "0": "Acoustic Grand Piano",
+        "1": "Bright Acoustic Piano",
+        "2": "Electric Grand Piano",
+        "3": "Honky-tonk Piano",
+        "4": "Electric Piano 1 (Rhodes Piano)",
+        "5": "Electric Piano 2 (Chorused Piano)",
+        "6": "Harpsichord",
+        "7": "Clavinet",
+        "8": "Celesta",
+        "9": "Glockenspiel",
+        "10": "Music Box",
+        "11": "Vibraphone",
+        "12": "Marimba",
+        "13": "Xylophone",
+        "14": "Tubular Bells",
+        "15": "Dulcimer (Santur)",
+        "16": "Drawbar Organ (Hammond)",
+        "17": "Percussive Organ",
+        "18": "Rock Organ",
+        "19": "Church Organ",
+        "20": "Reed Organ",
+        "21": "Accordion (French)",
+        "22": "Harmonica",
+        "23": "Tango Accordion (Band neon)",
+        "24": "Acoustic Guitar (nylon)",
+        "25": "Acoustic Guitar (steel)",
+        "26": "Electric Guitar (jazz)",
+        "27": "Electric Guitar (clean)",
+        "28": "Electric Guitar (muted)",
+        "29": "Overdriven Guitar",
+        "30": "Distortion Guitar",
+        "31": "Guitar harmonics",
+        "32": "Acoustic Bass",
+        "33": "Electric Bass (fingered)",
+        "34": "Electric Bass (picked)",
+        "35": "Fretless Bass",
+        "36": "Slap Bass 1",
+        "37": "Slap Bass 2",
+        "38": "Synth Bass 1",
+        "39": "Synth Bass 2",
+        "40": "Violin",
+        "41": "Viola",
+        "42": "Cello",
+        "43": "Contrabass",
+        "44": "Tremolo Strings",
+        "45": "Pizzicato Strings",
+        "46": "Orchestral Harp",
+        "47": "Timpani",
+        "48": "String Ensemble 1 (strings)",
+        "49": "String Ensemble 2 (slow strings)",
+        "50": "SynthStrings 1",
+        "51": "SynthStrings 2",
+        "52": "Choir Aahs",
+        "53": "Voice Oohs",
+        "54": "Synth Voice",
+        "55": "Orchestra Hit",
+        "56": "Trumpet",
+        "57": "Trombone",
+        "58": "Tuba",
+        "59": "Muted Trumpet",
+        "60": "French Horn",
+        "61": "Brass Section",
+        "62": "SynthBrass 1",
+        "63": "SynthBrass 2",
+        "64": "Soprano Sax",
+        "65": "Alto Sax",
+        "66": "Tenor Sax",
+        "67": "Baritone Sax",
+        "68": "Oboe",
+        "69": "English Horn",
+        "70": "Bassoon",
+        "71": "Clarinet",
+        "72": "Piccolo",
+        "73": "Flute",
+        "74": "Recorder",
+        "75": "Pan Flute",
+        "76": "Blown Bottle",
+        "77": "Shakuhachi",
+        "78": "Whistle",
+        "79": "Ocarina",
+        "80": "Lead 1 (square wave)",
+        "81": "Lead 2 (sawtooth wave)",
+        "82": "Lead 3 (calliope)",
+        "83": "Lead 4 (chiffer)",
+        "84": "Lead 5 (charang)",
+        "85": "Lead 6 (voice solo)",
+        "86": "Lead 7 (fifths)",
+        "87": "Lead 8 (bass + lead)",
+        "88": "Pad 1 (new age Fantasia)",
+        "89": "Pad 2 (warm)",
+        "90": "Pad 3 (polysynth)",
+        "91": "Pad 4 (choir space voice)",
+        "92": "Pad 5 (bowed glass)",
+        "93": "Pad 6 (metallic pro)",
+        "94": "Pad 7 (halo)",
+        "95": "Pad 8 (sweep)",
+        "96": "FX 1 (rain)",
+        "97": "FX 2 (soundtrack)",
+        "98": "FX 3 (crystal)",
+        "99": "FX 4 (atmosphere)",
+        "100": "FX 5 (brightness)",
+        "101": "FX 6 (goblins)",
+        "102": "FX 7 (echoes, drops)",
+        "103": "FX 8 (sci-fi, star theme)",
+        "104": "Sitar",
+        "105": "Banjo",
+        "106": "Shamisen",
+        "107": "Koto",
+        "108": "Kalimba",
+        "109": "Bag pipe",
+        "110": "Fiddle",
+        "111": "Shanai",
+        "112": "Tinkle Bell",
+        "113": "Agogo",
+        "114": "Steel Drums",
+        "115": "Woodblock",
+        "116": "Taiko Drum",
+        "117": "Melodic Tom",
+        "118": "Synth Drum",
+        "119": "Reverse Cymbal",
+        "120": "Guitar Fret Noise",
+        "121": "Breath Noise",
+        "122": "Seashore",
+        "123": "Bird Tweet",
+        "124": "Telephone Ring",
+        "125": "Helicopter",
+        "126": "Applause",
+        "127": "Gunshot"
+    }
+}
--- a/backend-python/utils/rwkv.py
+++ b/backend-python/utils/rwkv.py
@@ -1,27 +1,45 @@
+from abc import ABC, abstractmethod
+from enum import Enum, auto
 import os
 import pathlib
 import copy
-from typing import Dict, List, Tuple
+import re
+from typing import Dict, Iterable, List, Tuple, Union
 from utils.log import quick_log
 from fastapi import HTTPException
 from pydantic import BaseModel, Field
-import torch
 import numpy as np
-from rwkv_pip.utils import PIPELINE
 from routes import state_cache
+import global_var


 END_OF_TEXT = 0
-END_OF_LINE = 187
 END_OF_LINE_DOUBLE = 535


 os.environ["TORCH_EXTENSIONS_DIR"] = f"{pathlib.Path(__file__).parent.parent.resolve()}"


-class RWKV:
-    def __init__(self, model: str, strategy: str, tokens_path: str) -> None:
-        from rwkv.model import RWKV as Model  # dynamic import to make RWKV_CUDA_ON work
+class RWKVType(Enum):
+    Raven = auto()
+    World = auto()
+    Music = auto()
+
+
+class AbstractRWKV(ABC):
+    def __init__(self, model: str, strategy: str, tokens_path: str):
+        rwkv_beta = global_var.get(global_var.Args).rwkv_beta
+
+        # dynamic import to make RWKV_CUDA_ON work
+        if rwkv_beta:
+            from rwkv_pip.beta.model import (
+                RWKV as Model,
+            )
+        else:
+            from rwkv_pip.model import (
+                RWKV as Model,
+            )
+        from rwkv_pip.utils import PIPELINE

        filename, _ = os.path.splitext(os.path.basename(model))
        self.name = filename
@@ -29,98 +47,54 @@ class RWKV:
        self.pipeline = PIPELINE(self.model, tokens_path)
        self.model_state = None
        self.model_tokens = []
-
-        self.CHUNK_LEN = 256
+        self.rwkv_type: RWKVType = None

        self.max_tokens_per_generation = 500
        self.temperature = 1
-        self.top_p = 0.5
-        self.penalty_alpha_presence = 0.4
-        self.penalty_alpha_frequency = 0.4
+        self.top_p = 0.3
+        self.top_k = 0
+        self.penalty_alpha_presence = 0
+        self.penalty_alpha_frequency = 1

-        self.interface = ":"
-        if "rwkv_vocab" in tokens_path:
-            self.user = "Question"
-            self.bot = "Answer"
-        else:
-            self.user = "Bob"
-            self.bot = "Alice"
+    @abstractmethod
+    def adjust_occurrence(self, occurrence: Dict, token: int):
+        pass

-        self.AVOID_REPEAT_TOKENS = []
-        AVOID_REPEAT = "，：？！"
-        for i in AVOID_REPEAT:
-            dd = self.pipeline.encode(i)
-            assert len(dd) == 1
-            self.AVOID_REPEAT_TOKENS += dd
-
-        self.preload()
-
-    def preload(self):
-        interface = self.interface
-        user = self.user
-        bot = self.bot
-        preset_system = (
-            f"""
-The following is a coherent verbose detailed conversation between a girl named {bot} and her friend {user}. \
-{bot} is very intelligent, creative and friendly. \
-{bot} is unlikely to disagree with {user}, and {bot} doesn't like to ask {user} questions. \
-{bot} likes to tell {user} a lot about herself and her opinions. \
-{bot} usually gives {user} kind, helpful and informative advices.\n
-"""
-            if self.user == "Bob"
-            else f"{user}{interface} hi\n\n{bot}{interface} 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.\n\n"
-        )
-        logits, _ = self.run_rnn(self.fix_tokens(self.pipeline.encode(preset_system)))
-        try:
-            state_cache.add_state(
-                state_cache.AddStateBody(
-                    prompt=preset_system,
-                    tokens=self.model_tokens,
-                    state=self.model_state,
-                    logits=logits,
-                )
-            )
-        except HTTPException:
-            pass
+    @abstractmethod
+    def adjust_forward_logits(self, logits: List[float], occurrence: Dict, i: int):
+        pass

    # Model only saw '\n\n' as [187, 187] before, but the tokenizer outputs [535] for it at the end
-    def fix_tokens(self, tokens):
-        if len(tokens) > 0 and tokens[-1] == END_OF_LINE_DOUBLE:
-            tokens = tokens[:-1] + [END_OF_LINE, END_OF_LINE]
-        return tokens
+    @abstractmethod
+    def fix_tokens(self, tokens) -> List[int]:
+        pass

-    def run_rnn(self, _tokens: List[str], newline_adj: int = 0):
-        tokens = [int(x) for x in _tokens]
-        token_len = len(tokens)
-        self.model_tokens += tokens
+    @abstractmethod
+    def run_rnn(
+        self, _tokens: List[str], newline_adj: int = 0
+    ) -> Tuple[List[float], int]:
+        pass

-        while len(tokens) > 0:
-            out, self.model_state = self.model.forward(
-                tokens[: self.CHUNK_LEN], self.model_state
-            )
-            tokens = tokens[self.CHUNK_LEN :]
-
-        out[END_OF_LINE] += newline_adj  # adjust \n probability
-
-        if self.model_tokens[-1] in self.AVOID_REPEAT_TOKENS:
-            out[self.model_tokens[-1]] = -999999999
-        return out, token_len
+    @abstractmethod
+    def delta_postprocess(self, delta: str) -> str:
+        pass

    def get_embedding(self, input: str, fast_mode: bool) -> Tuple[List[float], int]:
        if fast_mode:
-            embedding, token_len = self.fast_embedding(
+            embedding, token_len = self.__fast_embedding(
                self.fix_tokens(self.pipeline.encode(input)), None
            )
        else:
            self.model_state = None
            self.model_tokens = []
            _, token_len = self.run_rnn(self.fix_tokens(self.pipeline.encode(input)))
-            embedding = self.model_state[-5].tolist()
+            embedding = self.model_state[-11].tolist()
        embedding = (embedding / np.linalg.norm(embedding)).tolist()
        return embedding, token_len

-    def fast_embedding(self, tokens: List[str], state):
+    def __fast_embedding(self, tokens: List[str], state):
+        import torch
+
        tokens = [int(x) for x in tokens]
        token_len = len(tokens)
        self = self.model
@@ -257,7 +231,9 @@ The following is a coherent verbose detailed conversation between a girl named {

                return state[0].tolist(), token_len

-    def generate(self, prompt: str, stop: str = None):
+    def generate(
+        self, prompt: str, stop: Union[str, List[str], None] = None
+    ) -> Iterable[Tuple[str, str, int, int]]:
        quick_log(None, None, "Generation Prompt:\n" + prompt)
        cache = None
        delta_prompt = prompt
@@ -301,44 +277,60 @@ The following is a coherent verbose detailed conversation between a girl named {
        completion_token_len = 0
        response = ""
        for i in range(self.max_tokens_per_generation):
-            for n in occurrence:
-                logits[n] -= (
-                    self.penalty_alpha_presence
-                    + occurrence[n] * self.penalty_alpha_frequency
-                )
+            self.adjust_forward_logits(logits, occurrence, i)
+
            token = self.pipeline.sample_logits(
-                logits, temperature=self.temperature, top_p=self.top_p
+                logits, temperature=self.temperature, top_p=self.top_p, top_k=self.top_k
            )

            if token == END_OF_TEXT:
                yield response, "", prompt_token_len, completion_token_len
                break
-            if token not in occurrence:
-                occurrence[token] = 1
-            else:
-                occurrence[token] += 1
+
+            self.adjust_occurrence(occurrence, token)

            logits, _ = self.run_rnn([token])
            completion_token_len = completion_token_len + 1
-            delta: str = self.pipeline.decode(self.model_tokens[out_last:])
+            delta: str = self.delta_postprocess(
+                self.pipeline.decode(self.model_tokens[out_last:])
+            )
            if "\ufffd" not in delta:  # avoid utf-8 display issues
                response += delta
                if stop is not None:
-                    if stop in response:
-                        response = response.split(stop)[0]
-                        try:
-                            state_cache.add_state(
-                                state_cache.AddStateBody(
-                                    prompt=prompt + response,
-                                    tokens=self.model_tokens,
-                                    state=self.model_state,
-                                    logits=logits,
+                    if type(stop) == str:
+                        if stop in response:
+                            try:
+                                state_cache.add_state(
+                                    state_cache.AddStateBody(
+                                        prompt=prompt + response,
+                                        tokens=self.model_tokens,
+                                        state=self.model_state,
+                                        logits=logits,
+                                    )
                                )
-                            )
-                        except HTTPException:
-                            pass
-                        yield response, "", prompt_token_len, completion_token_len
-                        break
+                            except HTTPException:
+                                pass
+                            response = response.split(stop)[0]
+                            yield response, "", prompt_token_len, completion_token_len
+                            break
+                    elif type(stop) == list:
+                        stop_exist_regex = "|".join(stop)
+                        matched = re.search(stop_exist_regex, response)
+                        if matched:
+                            try:
+                                state_cache.add_state(
+                                    state_cache.AddStateBody(
+                                        prompt=prompt + response,
+                                        tokens=self.model_tokens,
+                                        state=self.model_state,
+                                        logits=logits,
+                                    )
+                                )
+                            except HTTPException:
+                                pass
+                            response = response.split(matched.group())[0]
+                            yield response, "", prompt_token_len, completion_token_len
+                            break
                out_last = begin + i + 1
                if i == self.max_tokens_per_generation - 1:
                    try:
@@ -355,6 +347,169 @@ The following is a coherent verbose detailed conversation between a girl named {
                yield response, delta, prompt_token_len, completion_token_len


+class TextRWKV(AbstractRWKV):
+    def __init__(self, model: str, strategy: str, tokens_path: str) -> None:
+        super().__init__(model, strategy, tokens_path)
+
+        self.CHUNK_LEN = 256
+
+        self.max_tokens_per_generation = 500
+        self.temperature = 1
+        self.top_p = 0.3
+        self.top_k = 0
+        self.penalty_alpha_presence = 0
+        self.penalty_alpha_frequency = 1
+
+        self.interface = ":"
+        if "world" in self.name.lower():
+            self.rwkv_type = RWKVType.World
+            self.user = "Question"
+            self.bot = "Answer"
+            self.END_OF_LINE = 11
+        else:
+            self.rwkv_type = RWKVType.Raven
+            self.user = "Bob"
+            self.bot = "Alice"
+            self.END_OF_LINE = 187
+
+        self.AVOID_REPEAT_TOKENS = []
+        AVOID_REPEAT = "，：？！"
+        for i in AVOID_REPEAT:
+            dd = self.pipeline.encode(i)
+            assert len(dd) == 1
+            self.AVOID_REPEAT_TOKENS += dd
+
+        self.__preload()
+
+    def adjust_occurrence(self, occurrence: Dict, token: int):
+        for xxx in occurrence:
+            occurrence[xxx] *= 0.996
+        if token not in occurrence:
+            occurrence[token] = 1
+        else:
+            occurrence[token] += 1
+
+    def adjust_forward_logits(self, logits: List[float], occurrence: Dict, i: int):
+        for n in occurrence:
+            logits[n] -= (
+                self.penalty_alpha_presence
+                + occurrence[n] * self.penalty_alpha_frequency
+            )
+
+        if i == 0:
+            for token in self.model_tokens:
+                token = int(token)
+                for xxx in occurrence:
+                    occurrence[xxx] *= 0.996
+                if token not in occurrence:
+                    occurrence[token] = 1
+                else:
+                    occurrence[token] += 1
+
+    # Model only saw '\n\n' as [187, 187] before, but the tokenizer outputs [535] for it at the end
+    def fix_tokens(self, tokens) -> List[int]:
+        if self.rwkv_type == RWKVType.World:
+            return tokens
+        if len(tokens) > 0 and tokens[-1] == END_OF_LINE_DOUBLE:
+            tokens = tokens[:-1] + [self.END_OF_LINE, self.END_OF_LINE]
+        return tokens
+
+    def run_rnn(
+        self, _tokens: List[str], newline_adj: int = 0
+    ) -> Tuple[List[float], int]:
+        tokens = [int(x) for x in _tokens]
+        token_len = len(tokens)
+        self.model_tokens += tokens
+
+        while len(tokens) > 0:
+            out, self.model_state = self.model.forward(
+                tokens[: self.CHUNK_LEN], self.model_state
+            )
+            tokens = tokens[self.CHUNK_LEN :]
+
+        out[self.END_OF_LINE] += newline_adj  # adjust \n probability
+
+        if self.model_tokens[-1] in self.AVOID_REPEAT_TOKENS:
+            out[self.model_tokens[-1]] = -999999999
+        return out, token_len
+
+    def delta_postprocess(self, delta: str) -> str:
+        return delta
+
+    def __preload(self):
+        interface = self.interface
+        user = self.user
+        bot = self.bot
+        preset_system = (
+            f"""
+The following is a coherent verbose detailed conversation between a girl named {bot} and her friend {user}. \
+{bot} is very intelligent, creative and friendly. \
+{bot} is unlikely to disagree with {user}, and {bot} doesn't like to ask {user} questions. \
+{bot} likes to tell {user} a lot about herself and her opinions. \
+{bot} usually gives {user} kind, helpful and informative advices.\n
+"""
+            if self.rwkv_type == RWKVType.Raven
+            else (
+                f"{user}{interface} hi\n\n{bot}{interface} 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.\n\n"
+            )
+        )
+        logits, _ = self.run_rnn(self.fix_tokens(self.pipeline.encode(preset_system)))
+        try:
+            state_cache.add_state(
+                state_cache.AddStateBody(
+                    prompt=preset_system,
+                    tokens=self.model_tokens,
+                    state=self.model_state,
+                    logits=logits,
+                )
+            )
+        except HTTPException:
+            pass
+
+
+class MusicRWKV(AbstractRWKV):
+    def __init__(self, model: str, strategy: str, tokens_path: str):
+        super().__init__(model, strategy, tokens_path)
+
+        self.max_tokens_per_generation = 500
+        self.temperature = 1
+        self.top_p = 0.8
+        self.top_k = 8
+
+        self.rwkv_type = RWKVType.Music
+
+    def adjust_occurrence(self, occurrence: Dict, token: int):
+        for n in occurrence:
+            occurrence[n] *= 0.997  #### decay repetition penalty
+        if token >= 128 or token == 127:
+            occurrence[token] = 1 + (occurrence[token] if token in occurrence else 0)
+        else:
+            occurrence[token] = 0.3 + (occurrence[token] if token in occurrence else 0)
+
+    def adjust_forward_logits(self, logits: List[float], occurrence: Dict, i: int):
+        for n in occurrence:
+            logits[n] -= 0 + occurrence[n] * 0.5
+
+        logits[0] += (i - 2000) / 500  # try not to be too short or too long
+        logits[127] -= 1  # avoid "t125"
+
+    def fix_tokens(self, tokens) -> List[int]:
+        return tokens
+
+    def run_rnn(
+        self, _tokens: List[str], newline_adj: int = 0
+    ) -> Tuple[List[float], int]:
+        tokens = [int(x) for x in _tokens]
+        token_len = len(tokens)
+        self.model_tokens += tokens
+        out, self.model_state = self.model.forward(tokens, self.model_state)
+        return out, token_len
+
+    def delta_postprocess(self, delta: str) -> str:
+        return " " + delta
+
+
 class ModelConfigBody(BaseModel):
    max_tokens: int = Field(default=None, gt=0, le=102400)
    temperature: float = Field(default=None, ge=0, le=2)
@@ -374,7 +529,7 @@ class ModelConfigBody(BaseModel):
        }


-def set_rwkv_config(model: RWKV, body: ModelConfigBody):
+def set_rwkv_config(model: AbstractRWKV, body: ModelConfigBody):
    if body.max_tokens is not None:
        model.max_tokens_per_generation = body.max_tokens
    if body.temperature is not None:
@@ -390,7 +545,7 @@ def set_rwkv_config(model: RWKV, body: ModelConfigBody):
        model.penalty_alpha_frequency = body.frequency_penalty


-def get_rwkv_config(model: RWKV) -> ModelConfigBody:
+def get_rwkv_config(model: AbstractRWKV) -> ModelConfigBody:
    return ModelConfigBody(
        max_tokens=model.max_tokens_per_generation,
        temperature=model.temperature,
--- a/backend-python/wkv_cuda_utils/wkv_cuda10_30.pyd
+++ b/backend-python/wkv_cuda_utils/wkv_cuda10_30.pyd
--- a/backend-python/wkv_cuda_utils/wkv_cuda40.pyd
+++ b/backend-python/wkv_cuda_utils/wkv_cuda40.pyd
--- a/backend-python/wkv_cuda_utils/wkv_cuda_model.py
+++ b/backend-python/wkv_cuda_utils/wkv_cuda_model.py
@@ -1,734 +0,0 @@
-########################################################################################################
-# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
-########################################################################################################
-
-import types, gc, os, time, re
-import torch
-from torch.nn import functional as F
-torch.backends.cudnn.benchmark = True
-torch.backends.cudnn.allow_tf32 = True
-torch.backends.cuda.matmul.allow_tf32 = True
-current_path = os.path.dirname(os.path.abspath(__file__))
-
-# https://zhuanlan.zhihu.com/p/612879065
-def LoadPreCompileLibrary(file):
-    import importlib
-    import os
-
-    import torch
-
-    # load the custom_op_library and register the custom ops
-    lib_dir = os.path.dirname(__file__)
-    if os.name == "nt":
-        # Register the main torchvision library location on the default DLL path
-        import ctypes
-        import sys
-
-        kernel32 = ctypes.WinDLL("kernel32.dll", use_last_error=True)
-        with_load_library_flags = hasattr(kernel32, "AddDllDirectory")
-        prev_error_mode = kernel32.SetErrorMode(0x0001)
-
-        if with_load_library_flags:
-            kernel32.AddDllDirectory.restype = ctypes.c_void_p
-
-        if sys.version_info >= (3, 8):
-            os.add_dll_directory(lib_dir)
-        elif with_load_library_flags:
-            res = kernel32.AddDllDirectory(lib_dir)
-            if res is None:
-                err = ctypes.WinError(ctypes.get_last_error())
-                err.strerror += f' Error adding "{lib_dir}" to the DLL directories.'
-                raise ValueError(err)
-
-        kernel32.SetErrorMode(prev_error_mode)
-
-    loader_details = (
-        importlib.machinery.ExtensionFileLoader,
-        importlib.machinery.EXTENSION_SUFFIXES,
-    )
-
-    extfinder = importlib.machinery.FileFinder(lib_dir, loader_details)
-    ext_specs = extfinder.find_spec(file)
-    if ext_specs is None:
-        return False
-
-    try:
-        torch.ops.load_library(ext_specs.origin)
-    except OSError as exc:
-        return False
-    return True
-
-########################################################################################################
-
-if os.environ.get('RWKV_JIT_ON') != '0':
-    os.environ["RWKV_JIT_ON"] = '1'
-    MyModule = torch.jit.ScriptModule
-    MyFunction = torch.jit.script_method
-    MyStatic = torch.jit.script
-else:
-    MyModule = torch.nn.Module
-    def __nop(ob):
-        return ob
-    MyFunction = __nop
-    MyStatic = __nop
-
-if os.environ.get('RWKV_CUDA_ON') == '1':
-    if LoadPreCompileLibrary('wkv_cuda') is False:
-        from torch.utils.cpp_extension import load
-        load(
-            name=f"wkv_cuda",
-            sources=[f"{current_path}/cuda/wrapper.cpp", f"{current_path}/cuda/operators.cu"],
-            verbose=True,
-            extra_cuda_cflags=["-t 4", "-std=c++17", "--use_fast_math", "-O3", "--extra-device-vectorization"],
-            is_python_module=False)
-
-    @MyStatic
-    def cuda_wkv(T: int, C: int, w, u, k, v, aa, bb, pp):
-        assert 1 * C % min(C, 32) == 0
-        assert k.dtype == v.dtype == torch.float16 or k.dtype == v.dtype == torch.float32
-        assert w.dtype == u.dtype == aa.dtype == bb.dtype == pp.dtype == torch.float32
-        w = w.contiguous()
-        u = u.contiguous()
-        k = k.contiguous()
-        v = v.contiguous()
-        y = torch.empty((T, C), device=w.device, memory_format=torch.contiguous_format, dtype=k.dtype)
-        torch.ops.rwkv.wkv_forward(1, T, C, w, u, k, v, y, aa, bb, pp)
-        return y, aa, bb, pp
-    @MyStatic
-    def cuda_mm8_seq(B: int, N: int, M: int, x, w, mx, rx, my, ry):
-        assert x.dtype == mx.dtype == rx.dtype == my.dtype == ry.dtype
-        assert x.dtype == torch.float32 or x.dtype == torch.float16
-        assert w.dtype == torch.uint8
-        assert x.shape == [B, N]
-        assert w.shape == [N, M]
-        assert rx.shape == mx.shape == [M]
-        assert ry.shape == my.shape == [N, 1]
-        y = torch.empty((B, M), device=w.device, dtype=x.dtype)
-        torch.ops.rwkv.mm8_seq(B, N, M, x, w, mx, rx, my, ry, y)
-        return y
-    @MyStatic
-    def cuda_mm8_one(N: int, M: int, x, w, mx, rx, my, ry):
-        assert x.dtype == mx.dtype == rx.dtype == my.dtype == ry.dtype
-        assert x.dtype == torch.float32 or x.dtype == torch.float16
-        assert w.dtype == torch.uint8
-        assert x.shape == [N]
-        assert w.shape == [N, M]
-        assert rx.shape == mx.shape == [M]
-        assert ry.shape == my.shape == [N, 1]
-        y = torch.zeros((M,), device=w.device, dtype=torch.float32)
-        torch.ops.rwkv.mm8_one(N, M, x, w, mx, rx, my, ry, y)
-        return y.to(dtype=x.dtype)
-else:
-    os.environ["RWKV_CUDA_ON"] = '0'
-
-########################################################################################################
-
-class RWKV(MyModule):
-    def __init__(self, model, strategy, verbose = True, convert_and_save_and_exit = None):
-        super().__init__()
-        if verbose:
-            prxxx = lambda *args, **kwargs: print(*args, **kwargs)
-        else:
-            prxxx = lambda *args, **kwargs: None
-
-        STRATEGY_REGEX = r"^(?:(?:^|->) *(?:cuda(?::[\d]+)?|cpu|mps) (?:fp(?:16|32)|bf16)(?:i8|i4|i3)?(?: \*[\d]+\+?)? *)+$"
-        if not re.match(STRATEGY_REGEX, strategy):
-            raise ValueError("Invalid strategy. Please read https://pypi.org/project/rwkv/")
-
-        strategy = ('->'.join([x.strip() for x in strategy.split('->')])).replace('->', ' -> ')
-        self.args = types.SimpleNamespace()
-        args = self.args
-        args.MODEL_NAME = model
-        args.strategy_string = strategy
-
-        # Rescale for fp16 mode: set x = x/2 every X layer (to avoid fp16 overflow)
-        self.RESCALE_LAYER = 6 if 'fp16' in strategy else 0
-        prxxx(f'RWKV_JIT_ON {os.environ["RWKV_JIT_ON"]} RWKV_CUDA_ON {os.environ["RWKV_CUDA_ON"]} RESCALE_LAYER {self.RESCALE_LAYER}\n')
-
-        args.MODEL_NAME = args.MODEL_NAME.strip()
-        if not args.MODEL_NAME.endswith('.pth'):
-            args.MODEL_NAME += '.pth'
-        prxxx(f'Loading {args.MODEL_NAME} ...')
-        with torch.no_grad():
-            self.w = torch.load(args.MODEL_NAME, map_location='cpu') # load model to CPU first
-            gc.collect()
-            w = self.w
-
-            ALREADY_CONVERTED = False
-            if '_strategy' in w:
-                ALREADY_CONVERTED = True
-                assert convert_and_save_and_exit == None # you should only convert a raw model
-                prxxx(f"Converted model: strategy {w['_strategy']}, version {w['_version']}\n")
-                assert w['_strategy'] == args.strategy_string # if you are using a new strategy, re-convert the model
-                assert float(w['_version']) >= 0.7 # sometimes you should re-convert using latest convert_model.py
-                assert w['_rescale_layer'] == self.RESCALE_LAYER
-                del w['_strategy']
-                del w['_version']
-                del w['_rescale_layer']
-            
-            args.n_embd = w['emb.weight'].shape[1]
-            args.n_layer = 0
-            keys = list(w.keys())
-            for x in keys:
-                layer_id = int(x.split('.')[1]) if ('blocks.' in x) else 0
-                args.n_layer = max(args.n_layer, layer_id+1)
-
-            ####################### Compute strategy
-
-            s = [x.strip().split(' ') for x in strategy.split('->')]
-            plan = [0] * len(s)
-            stream_i = -1
-            stream_count = 0
-            to_allocate = args.n_layer + 1
-            allocated = 0
-            free_slots = 0
-            for i in range(len(s)):
-                si = s[i]
-                si1 = si[1]
-                if si1.startswith('fp32'): si[1] = [torch.float]
-                elif si1.startswith('fp16'): si[1] = [torch.float16]
-                elif si1.startswith('bf16'): si[1] = [torch.bfloat16]
-                if si1.endswith('i8'): si[1] += [torch.uint8]
-                else: si[1] += [si[1][0]]
-                if len(si) > 2:
-                    ss = si[2]
-                    assert ss.startswith('*')
-                    if ss.endswith('+'):
-                        plan[i] = int(ss[1:-1])
-                        stream_i = i
-                    else:
-                        plan[i] = int(ss[1:])
-                    allocated += plan[i]
-                    if allocated >= to_allocate:
-                        plan[i] += to_allocate - allocated
-                        break
-                else:
-                    free_slots += 1
-            if stream_i < 0:
-                if free_slots > 0 and to_allocate > allocated:
-                    for i in range(len(s)):
-                        if plan[i] == 0:
-                            plan[i] = (to_allocate - allocated) // free_slots
-                            allocated += plan[i]
-                            free_slots -= 1
-                if to_allocate > allocated:
-                    plan[len(s)-1] += to_allocate - allocated
-            else:
-                if to_allocate > allocated:
-                    stream_count = to_allocate - allocated
-                    plan[stream_i] += stream_count
-            prxxx(f'Strategy: (total {args.n_layer}+1={args.n_layer+1} layers)')
-            for i in range(len(s)):
-                ss = s[i]
-                if i != stream_i:
-                    prxxx(f'* {ss[0]} {str(ss[1]).replace("torch.","")}, store {plan[i]} layers')
-                else:
-                    prxxx(f'* {ss[0]} {str(ss[1]).replace("torch.","")}, store {plan[i]-stream_count} layers, stream {stream_count} layers')
-                plan[i] += (0 if i == 0 else plan[i-1])
-            self.strategy = [None] * (args.n_layer + 1)
-            strategy = self.strategy
-            for n in range(args.n_layer + 1):
-                for i in range(len(s)):
-                    if n < plan[i]:
-                        strategy[n] = types.SimpleNamespace()
-                        strategy[n].device = s[i][0]
-                        strategy[n].atype = s[i][1][0]
-                        strategy[n].wtype = s[i][1][1]
-                        strategy[n].stream = False
-                        if i == stream_i and n >= (plan[i] - stream_count):
-                            strategy[n].stream = True
-                        break
-                prxxx(f"{n}-{strategy[n].device}-{str(strategy[n].atype).replace('torch.','')}-{str(strategy[n].wtype).replace('torch.','')}{'-stream' if strategy[n].stream else ''}",end=' ')
-            prxxx()
-
-            ####################### Load weights to self.w
-
-            if not ALREADY_CONVERTED:
-                try: # precompute embedding
-                    w['emb.weight'] = F.layer_norm(w['emb.weight'], (args.n_embd,), weight=w['blocks.0.ln0.weight'], bias=w['blocks.0.ln0.bias'])
-                except:
-                    w['emb.weight'] = F.layer_norm(w['emb.weight'].float(), (args.n_embd,), weight=w['blocks.0.ln0.weight'].float(), bias=w['blocks.0.ln0.bias'].float())
-                del w['blocks.0.ln0.weight']
-                del w['blocks.0.ln0.bias']
-
-            print_need_newline = False
-            keys = list(w.keys())
-            for x in keys:
-                w[x].requires_grad = False
-                layer_id = int(x.split('.')[1]) if ('blocks.' in x) else 0
-                if ('ln_out.' in x) or ('head.' in x):
-                    layer_id = args.n_layer
-                dd = strategy[layer_id]
-                DEVICE = dd.device
-                ATYPE = dd.atype
-                WTYPE = dd.wtype
-
-                if not ALREADY_CONVERTED:
-                    if self.RESCALE_LAYER > 0:
-                        if 'att.output.weight' in x:
-                            w[x] = w[x] / (2 ** int(layer_id // self.RESCALE_LAYER))
-                        if 'ffn.value.weight' in x:
-                            w[x] = w[x] / (2 ** int(layer_id // self.RESCALE_LAYER))
-
-                    if '.time_' in x:
-                        w[x] = w[x].squeeze()
-                    if 'key.weight' in x or 'value.weight' in x or 'receptance.weight' in x or 'output.weight' in x or 'head.weight' in x:
-                        w[x] = w[x].t()
-
-                    if '.time_decay' in x: # need fp32 for this
-                        w[x] = -torch.exp(w[x].float())
-                    elif '.time_first' in x: # need fp32 for this
-                        w[x] = w[x].float()
-                    else:
-                        if (len(w[x].shape) == 2) and ('emb' not in x):
-                            if WTYPE != torch.uint8:
-                                w[x] = w[x].to(dtype=WTYPE)
-                            else:
-                                w[x] = w[x].float()
-
-                                if w[x].shape[0] > w[x].shape[1]:
-                                    w[x+'_my'] = torch.amin(w[x], dim=1).unsqueeze(1)
-                                    w[x] = w[x] - w[x+'_my']
-                                    w[x+'_mx'] = torch.amin(w[x], dim=0)
-                                    w[x] = w[x] - w[x+'_mx']
-                                    w[x+'_rx'] = torch.amax(w[x], dim=0)
-                                    w[x] = w[x] / w[x+'_rx']
-                                    w[x+'_ry'] = torch.amax(w[x], dim=1).unsqueeze(1)
-                                    w[x] = w[x] / w[x+'_ry']
-                                else:
-                                    w[x+'_mx'] = torch.amin(w[x], dim=0)
-                                    w[x] = w[x] - w[x+'_mx']
-                                    w[x+'_my'] = torch.amin(w[x], dim=1).unsqueeze(1)
-                                    w[x] = w[x] - w[x+'_my']
-                                    w[x+'_rx'] = torch.amax(w[x], dim=0)
-                                    w[x] = w[x] / w[x+'_rx']
-                                    w[x+'_ry'] = torch.amax(w[x], dim=1).unsqueeze(1)
-                                    w[x] = w[x] / w[x+'_ry']
-
-                                w[x] = torch.clip(torch.floor(w[x] * 256), min=0, max=255).to(dtype=torch.uint8)
-                                w[x+'_mx'] = w[x+'_mx'].to(dtype=ATYPE).contiguous()
-                                w[x+'_rx'] = (w[x+'_rx'] / 16).to(dtype=ATYPE).contiguous()
-                                w[x+'_my'] = w[x+'_my'].to(dtype=ATYPE).contiguous()
-                                w[x+'_ry'] = (w[x+'_ry'] / 16).to(dtype=ATYPE).contiguous()
-                        else:
-                            w[x] = w[x].to(dtype=ATYPE)
-                
-                if convert_and_save_and_exit == None:
-                    if 'emb.' in x:
-                        w[x] = w[x].contiguous()
-                    elif (dd.stream) and (x.endswith('key.weight') or x.endswith('value.weight') or x.endswith('receptance.weight') or x.endswith('output.weight')):
-                        try:
-                            w[x] = w[x].contiguous().pin_memory() # if you see "CUDA error: out of memory" here, that's out of CPU RAM, not VRAM. Get more RAM :)
-                        except:
-                            print('Note: You are running out of RAM. Get more CPU RAM. Now this will run much slower.')
-                    elif DEVICE != 'cpu':
-                        w[x] = w[x].to(device=DEVICE).contiguous()
-                    
-                    if (dd.stream) or (DEVICE != 'cpu'):
-                        try:
-                            w[x+'_mx'] = w[x+'_mx'].to(device=DEVICE).contiguous()
-                            w[x+'_rx'] = w[x+'_rx'].to(device=DEVICE).contiguous()
-                            w[x+'_my'] = w[x+'_my'].to(device=DEVICE).contiguous()
-                            w[x+'_ry'] = w[x+'_ry'].to(device=DEVICE).contiguous()
-                        except:
-                            pass
-
-                if 'ffn.value.weight' in x:
-                    gc.collect()
-                    if 'cuda' in args.strategy_string:
-                        torch.cuda.empty_cache()
-
-                shape = [i for i in w[x].shape if i != 1]
-                if len(shape) > 1:
-                    shape = f" {str(shape[0]).rjust(5)} {str(shape[1]).rjust(5)}"
-                else:
-                    shape = f" {str(shape[0]).rjust(5)}      "
-                if layer_id == 0 or layer_id >= args.n_layer-1:
-                    if print_need_newline:
-                        prxxx('\n', end = '')
-                        print_need_newline = False
-                    dt = str(w[x].dtype).replace('torch.', '')
-                    dt = dt.replace('float32', 'f32').replace('bfloat16', 'bf16').replace('float16', 'f16').replace('uint8', 'i8')
-                    prxxx(x.ljust(32), dt.rjust(4), str(w[x].device).rjust(8), shape, ' (pinned)' if w[x].is_pinned() else '')
-                else:
-                    print_need_newline = True
-                    prxxx('.', end = '', flush = True)
-            
-            if convert_and_save_and_exit:
-                w['_strategy'] = args.strategy_string
-                w['_rescale_layer'] = self.RESCALE_LAYER
-                w['_version'] = '0.7'
-                if not convert_and_save_and_exit.endswith('.pth'):
-                    convert_and_save_and_exit += '.pth'
-                prxxx(f'Saving to {convert_and_save_and_exit}...')
-                torch.save(w, convert_and_save_and_exit)
-                prxxx(f'Converted and saved. Now this will exit.')
-                exit(0)
-            
-            gc.collect()
-            if 'cuda' in args.strategy_string:
-                torch.cuda.empty_cache()
-
-    @MyFunction
-    def torch_mm8_seq(self, x, w, mx, rx, my, ry):
-        return x @ ((w.to(dtype=x.dtype) + 0.5) * ry * rx + my + mx)
-
-    @MyFunction
-    def torch_mm8_one(self, x, w, mx, rx, my, ry):
-        return x @ ((w.to(dtype=x.dtype) + 0.5) * ry * rx + my + mx)
-
-    if os.environ.get('RWKV_CUDA_ON') == '1':
-        @MyFunction
-        def mm8_seq(self, x, w, mx, rx, my, ry):
-            if w.device.type == 'cuda' and x.dtype == torch.float16:
-                B, N, M = x.shape[0], w.shape[0], w.shape[1]
-                return cuda_mm8_seq(B, N, M, x, w, mx, rx, my, ry)
-            else:
-                return self.torch_mm8_seq(x, w, mx, rx, my, ry)
-        @MyFunction
-        def mm8_one(self, x, w, mx, rx, my, ry):
-            if w.device.type == 'cuda':
-                N, M = w.shape[0], w.shape[1]
-                return cuda_mm8_one(N, M, x, w, mx, rx, my, ry)
-            else:
-                return self.torch_mm8_one(x, w, mx, rx, my, ry)
-    else:
-        @MyFunction
-        def mm8_seq(self, x, w, mx, rx, my, ry):
-            return self.torch_mm8_seq(x, w, mx, rx, my, ry)
-        @MyFunction
-        def mm8_one(self, x, w, mx, rx, my, ry):
-            return self.torch_mm8_one(x, w, mx, rx, my, ry)
-
-    ########################################################################################################
-
-    @MyFunction
-    def ffn_one(self, x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        kx = xx * k_mix + sx * (1 - k_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(rx @ rw)
-        vx = torch.square(torch.relu(kx @ kw))
-        out = r * (vx @ vw)
-        return x + out, xx
-
-    @MyFunction
-    def ffn_one_i8(self, x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        kx = xx * k_mix + sx * (1 - k_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(self.mm8_one(rx, rw, rmx, rrx, rmy, rry))
-        vx = torch.square(torch.relu(self.mm8_one(kx, kw, kmx, krx, kmy, kry)))
-        out = r * (self.mm8_one(vx, vw, vmx, vrx, vmy, vry))
-        return x + out, xx
-    
-    ########################################################################################################
-
-    @MyFunction
-    def ffn_seq(self, x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
-        kx = xx * k_mix + sx * (1 - k_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(rx @ rw)
-        vx = torch.square(torch.relu(kx @ kw))
-        out = r * (vx @ vw)
-        return x + out, xx[-1,:]
-
-    @MyFunction
-    def ffn_seq_i8(self, x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
-        kx = xx * k_mix + sx * (1 - k_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(self.mm8_seq(rx, rw, rmx, rrx, rmy, rry))
-        vx = torch.square(torch.relu(self.mm8_seq(kx, kw, kmx, krx, kmy, kry)))
-        out = r * (self.mm8_seq(vx, vw, vmx, vrx, vmy, vry))
-        return x + out, xx[-1,:]
-
-    ########################################################################################################
-
-    @MyFunction
-    def att_one(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        kx = xx * k_mix + sx * (1 - k_mix)
-        vx = xx * v_mix + sx * (1 - v_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(rx @ rw)
-        k = (kx @ kw).float()
-        v = (vx @ vw).float()
-
-        ww = t_first + k
-        p = torch.maximum(pp, ww)
-        e1 = torch.exp(pp - p)
-        e2 = torch.exp(ww - p)
-        wkv = ((e1 * aa + e2 * v) / (e1 * bb + e2)).to(dtype=x.dtype)
-        ww = t_decay + pp
-        p = torch.maximum(ww, k)
-        e1 = torch.exp(ww - p)
-        e2 = torch.exp(k - p)
-
-        out = (r * wkv) @ ow
-        return x + out, xx, e1 * aa + e2 * v, e1 * bb + e2, p
-
-    @MyFunction
-    def att_one_i8(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        kx = xx * k_mix + sx * (1 - k_mix)
-        vx = xx * v_mix + sx * (1 - v_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(self.mm8_one(rx, rw, rmx, rrx, rmy, rry))
-        k = (self.mm8_one(kx, kw, kmx, krx, kmy, kry)).float()
-        v = (self.mm8_one(vx, vw, vmx, vrx, vmy, vry)).float()
-
-        ww = t_first + k
-        p = torch.maximum(pp, ww)
-        e1 = torch.exp(pp - p)
-        e2 = torch.exp(ww - p)
-        wkv = ((e1 * aa + e2 * v) / (e1 * bb + e2)).to(dtype=x.dtype)
-        ww = t_decay + pp
-        p = torch.maximum(ww, k)
-        e1 = torch.exp(ww - p)
-        e2 = torch.exp(k - p)
-
-        out = self.mm8_one(r * wkv, ow, omx, orx, omy, ory)
-        return x + out, xx, e1 * aa + e2 * v, e1 * bb + e2, p
-
-    ########################################################################################################
-
-    @MyFunction
-    def att_seq(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
-        kx = xx * k_mix + sx * (1 - k_mix)
-        vx = xx * v_mix + sx * (1 - v_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(rx @ rw)
-        k = (kx @ kw).float()
-        v = (vx @ vw).float()
-
-        T = x.shape[0]
-        for t in range(T):
-            kk = k[t]
-            vv = v[t]
-            ww = t_first + kk
-            p = torch.maximum(pp, ww)
-            e1 = torch.exp(pp - p)
-            e2 = torch.exp(ww - p)
-            sx[t] = ((e1 * aa + e2 * vv) / (e1 * bb + e2)).to(dtype=x.dtype)
-            ww = t_decay + pp
-            p = torch.maximum(ww, kk)
-            e1 = torch.exp(ww - p)
-            e2 = torch.exp(kk - p)
-            aa = e1 * aa + e2 * vv
-            bb = e1 * bb + e2
-            pp = p
-        out = (r * sx) @ ow
-        return x + out, xx[-1,:], aa, bb, pp
-
-    @MyFunction
-    def att_seq_i8(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
-        xx = F.layer_norm(x, (x.shape[-1],), weight=ln_w, bias=ln_b)
-        sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
-        kx = xx * k_mix + sx * (1 - k_mix)
-        vx = xx * v_mix + sx * (1 - v_mix)
-        rx = xx * r_mix + sx * (1 - r_mix)
-
-        r = torch.sigmoid(self.mm8_seq(rx, rw, rmx, rrx, rmy, rry))
-        k = self.mm8_seq(kx, kw, kmx, krx, kmy, kry).float()
-        v = self.mm8_seq(vx, vw, vmx, vrx, vmy, vry).float()
-
-        T = x.shape[0]
-        for t in range(T):
-            kk = k[t]
-            vv = v[t]
-            ww = t_first + kk
-            p = torch.maximum(pp, ww)
-            e1 = torch.exp(pp - p)
-            e2 = torch.exp(ww - p)
-            sx[t] = ((e1 * aa + e2 * vv) / (e1 * bb + e2)).to(dtype=x.dtype)
-            ww = t_decay + pp
-            p = torch.maximum(ww, kk)
-            e1 = torch.exp(ww - p)
-            e2 = torch.exp(kk - p)
-            aa = e1 * aa + e2 * vv
-            bb = e1 * bb + e2
-            pp = p
-        out = self.mm8_seq(r * sx, ow, omx, orx, omy, ory)
-        return x + out, xx[-1,:], aa, bb, pp
-
-    ########################################################################################################
-
-    if os.environ["RWKV_CUDA_ON"] == '1':
-        @MyFunction
-        def cuda_att_seq(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
-            T, C = x.size()
-            xx = F.layer_norm(x, (C,), weight=ln_w, bias=ln_b)
-            sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
-            kx = xx * k_mix + sx * (1 - k_mix)
-            vx = xx * v_mix + sx * (1 - v_mix)
-            rx = xx * r_mix + sx * (1 - r_mix)
-
-            r = torch.sigmoid(rx @ rw)
-            k = kx @ kw
-            v = vx @ vw
-            y, aa, bb, pp = cuda_wkv(T, C, t_decay, t_first, k, v, aa, bb, pp)
-            
-            out = (r * y) @ ow
-            return x + out, xx[-1,:], aa, bb, pp
-
-        @MyFunction
-        def cuda_att_seq_i8(self, x, sx, aa, bb, pp, ln_w, ln_b, k_mix, v_mix, r_mix, t_decay, t_first, kw, vw, rw, ow, kmx, krx, kmy, kry, vmx, vrx, vmy, vry, rmx, rrx, rmy, rry, omx, orx, omy, ory):
-            T, C = x.size()
-            xx = F.layer_norm(x, (C,), weight=ln_w, bias=ln_b)
-            sx = torch.cat((sx.unsqueeze(0), xx[:-1,:]))
-            kx = xx * k_mix + sx * (1 - k_mix)
-            vx = xx * v_mix + sx * (1 - v_mix)
-            rx = xx * r_mix + sx * (1 - r_mix)
-
-            r = torch.sigmoid(self.mm8_seq(rx, rw, rmx, rrx, rmy, rry))
-            k = self.mm8_seq(kx, kw, kmx, krx, kmy, kry)
-            v = self.mm8_seq(vx, vw, vmx, vrx, vmy, vry)
-            y, aa, bb, pp = cuda_wkv(T, C, t_decay, t_first, k, v, aa, bb, pp)
-
-            out = self.mm8_seq(r * y, ow, omx, orx, omy, ory)
-            return x + out, xx[-1,:], aa, bb, pp
-
-    ########################################################################################################
-
-    def forward(self, tokens, state, full_output=False):
-        with torch.no_grad():
-            w = self.w
-            args = self.args
-
-            if state == None:
-                state = [None] * args.n_layer * 5
-                for i in range(args.n_layer): # state: 0=att_xx 1=att_aa 2=att_bb 3=att_pp 4=ffn_xx
-                    dd = self.strategy[i]
-                    dev = dd.device
-                    atype = dd.atype
-                    state[i*5+0] = torch.zeros(args.n_embd, dtype=atype, requires_grad=False, device=dev).contiguous()
-                    state[i*5+1] = torch.zeros(args.n_embd, dtype=torch.float, requires_grad=False, device=dev).contiguous()
-                    state[i*5+2] = torch.zeros(args.n_embd, dtype=torch.float, requires_grad=False, device=dev).contiguous()
-                    state[i*5+3] = torch.zeros(args.n_embd, dtype=torch.float, requires_grad=False, device=dev).contiguous() - 1e30
-                    state[i*5+4] = torch.zeros(args.n_embd, dtype=atype, requires_grad=False, device=dev).contiguous()
-
-            seq_mode = len(tokens) > 1
-
-            x = w['emb.weight'][tokens if seq_mode else tokens[0]]
-
-            for i in range(args.n_layer):
-                bbb = f'blocks.{i}.'
-                att = f'blocks.{i}.att.'
-                ffn = f'blocks.{i}.ffn.'
-                dd = self.strategy[i]
-                dev = dd.device
-                atype = dd.atype
-                wtype = dd.wtype
-                if seq_mode:
-                    if 'cuda' in str(dev) and os.environ["RWKV_CUDA_ON"] == '1':
-                        ATT = self.cuda_att_seq if wtype != torch.uint8 else self.cuda_att_seq_i8
-                    else:
-                        ATT = self.att_seq if wtype != torch.uint8 else self.att_seq_i8
-                    FFN = self.ffn_seq if wtype != torch.uint8 else self.ffn_seq_i8
-                else:
-                    ATT = self.att_one if wtype != torch.uint8 else self.att_one_i8
-                    FFN = self.ffn_one if wtype != torch.uint8 else self.ffn_one_i8
-
-                x = x.to(dtype=atype, device=dev)
-
-                kw = w[f'{att}key.weight']
-                vw = w[f'{att}value.weight']
-                rw = w[f'{att}receptance.weight']
-                ow = w[f'{att}output.weight']
-                if dd.stream:
-                    kw = kw.to(device=dev, non_blocking=True)
-                    vw = vw.to(device=dev, non_blocking=True)
-                    rw = rw.to(device=dev, non_blocking=True)
-                    ow = ow.to(device=dev, non_blocking=True)
-                kmx = w[f'{att}key.weight_mx'] if wtype == torch.uint8 else x
-                krx = w[f'{att}key.weight_rx'] if wtype == torch.uint8 else x
-                kmy = w[f'{att}key.weight_my'] if wtype == torch.uint8 else x
-                kry = w[f'{att}key.weight_ry'] if wtype == torch.uint8 else x
-                vmx = w[f'{att}value.weight_mx'] if wtype == torch.uint8 else x
-                vrx = w[f'{att}value.weight_rx'] if wtype == torch.uint8 else x
-                vmy = w[f'{att}value.weight_my'] if wtype == torch.uint8 else x
-                vry = w[f'{att}value.weight_ry'] if wtype == torch.uint8 else x
-                rmx = w[f'{att}receptance.weight_mx'] if wtype == torch.uint8 else x
-                rrx = w[f'{att}receptance.weight_rx'] if wtype == torch.uint8 else x
-                rmy = w[f'{att}receptance.weight_my'] if wtype == torch.uint8 else x
-                rry = w[f'{att}receptance.weight_ry'] if wtype == torch.uint8 else x
-                omx = w[f'{att}output.weight_mx'] if wtype == torch.uint8 else x
-                orx = w[f'{att}output.weight_rx'] if wtype == torch.uint8 else x
-                omy = w[f'{att}output.weight_my'] if wtype == torch.uint8 else x
-                ory = w[f'{att}output.weight_ry'] if wtype == torch.uint8 else x
-                x, state[i*5+0], state[i*5+1], state[i*5+2], state[i*5+3] = ATT(
-                    x, state[i*5+0], state[i*5+1], state[i*5+2], state[i*5+3],
-                    w[f'{bbb}ln1.weight'], w[f'{bbb}ln1.bias'],
-                    w[f'{att}time_mix_k'], w[f'{att}time_mix_v'], w[f'{att}time_mix_r'],
-                    w[f'{att}time_decay'], w[f'{att}time_first'],
-                    kw, vw, rw, ow,
-                    kmx, krx, kmy, kry,
-                    vmx, vrx, vmy, vry,
-                    rmx, rrx, rmy, rry,
-                    omx, orx, omy, ory,
-                    )
-                if dd.stream:
-                    del kw, vw, rw, ow
-
-                kw = w[f'{ffn}key.weight']
-                vw = w[f'{ffn}value.weight']
-                rw = w[f'{ffn}receptance.weight']
-                if dd.stream:
-                    kw = kw.to(device=dev, non_blocking=True)
-                    vw = vw.to(device=dev, non_blocking=True)
-                    rw = rw.to(device=dev, non_blocking=True)
-                kmx = w[f'{ffn}key.weight_mx'] if wtype == torch.uint8 else x
-                krx = w[f'{ffn}key.weight_rx'] if wtype == torch.uint8 else x
-                kmy = w[f'{ffn}key.weight_my'] if wtype == torch.uint8 else x
-                kry = w[f'{ffn}key.weight_ry'] if wtype == torch.uint8 else x
-                vmx = w[f'{ffn}value.weight_mx'] if wtype == torch.uint8 else x
-                vrx = w[f'{ffn}value.weight_rx'] if wtype == torch.uint8 else x
-                vmy = w[f'{ffn}value.weight_my'] if wtype == torch.uint8 else x
-                vry = w[f'{ffn}value.weight_ry'] if wtype == torch.uint8 else x
-                rmx = w[f'{ffn}receptance.weight_mx'] if wtype == torch.uint8 else x
-                rrx = w[f'{ffn}receptance.weight_rx'] if wtype == torch.uint8 else x
-                rmy = w[f'{ffn}receptance.weight_my'] if wtype == torch.uint8 else x
-                rry = w[f'{ffn}receptance.weight_ry'] if wtype == torch.uint8 else x                    
-                x, state[i*5+4] = FFN(
-                    x, state[i*5+4],
-                    w[f'{bbb}ln2.weight'], w[f'{bbb}ln2.bias'],
-                    w[f'{ffn}time_mix_k'], w[f'{ffn}time_mix_r'],
-                    kw, vw, rw,
-                    kmx, krx, kmy, kry,
-                    vmx, vrx, vmy, vry,
-                    rmx, rrx, rmy, rry,                    
-                    )
-                if dd.stream:                
-                    del kw, vw, rw
-                
-                if self.RESCALE_LAYER > 0:
-                    if (i+1) % self.RESCALE_LAYER == 0:
-                        x = x / 2
-            
-            dd = self.strategy[args.n_layer]
-            x = x[-1,:] if (seq_mode and (not full_output)) else x
-            x = x.to(dtype=dd.atype, device=dd.device)
-            
-            x = F.layer_norm(x, (args.n_embd,), weight=w['ln_out.weight'], bias=w['ln_out.bias'])
-            if w['head.weight'].dtype != torch.uint8:
-                x = x @ w['head.weight']
-            else:
-                if seq_mode and full_output:
-                    x = self.mm8_seq(x, w['head.weight'], w['head.weight_mx'], w['head.weight_rx'], w['head.weight_my'], w['head.weight_ry'])
-                else:
-                    x = self.mm8_one(x, w['head.weight'], w['head.weight_mx'], w['head.weight_rx'], w['head.weight_my'], w['head.weight_ry'])
-
-            return x.float(), state
--- a/backend-rust/assets/rwkv_vocab_v20230424.json
+++ b/backend-rust/assets/rwkv_vocab_v20230424.json
--- a/build/appicon.png
+++ b/build/appicon.png
--- a/build/darwin/Readme_Install.txt
+++ b/build/darwin/Readme_Install.txt
@@ -1,6 +1,6 @@
-For Mac and Linux users, please manually install Python 3.10 (usually the latest systems come with it built-in). You can specify the Python interpreter to use in Settings.
-对于Mac和Linux用户，请手动安装 Python3.10 (通常最新的系统已经内置了). 你可以在设置中指定使用的Python解释器.
-MacおよびLinuxのユーザーの方は、Python3.10を手動でインストールしてください（通常、最新のシステムには既に組み込まれています）。 設定メニューで使用するPythonインタプリタを指定することができます。
+For Mac and Linux users, please manually install Python 3.10 (usually the latest systems come with it built-in). You can specify the Python interpreter to use in Settings. (which python3)
+对于Mac和Linux用户，请手动安装 Python3.10 (通常最新的系统已经内置了). 你可以在设置中指定使用的Python解释器. (which python3)
+MacおよびLinuxのユーザーの方は、Python3.10を手動でインストールしてください（通常、最新のシステムには既に組み込まれています）。 設定メニューで使用するPythonインタプリタを指定することができます。 (which python3)

 Please execute this program in an empty directory. All related dependencies will be placed in this directory.
 请将本程序放在一个空目录内执行, 所有相关依赖均会放置于此目录.
--- a/build/windows/icon.ico
+++ b/build/windows/icon.ico
--- a/finetune/data/sample.jsonl
+++ b/finetune/data/sample.jsonl
@@ -0,0 +1,7 @@
+{"text": "1:This is the first document."}
+{"text": "2:Hello\nWorld"}
+{"text": "3:1+1=2\n1+2=3\n2+2=4"}
+{"text": "4:You will be training the GPT version because it's paralleziable and faster to train."}
+{"text": "5:Read the inference code in src/model.py and try using the final hidden state(.xx .aa .bb)"}
+{"text": "6:You can fine-tune the model with longer ctxLen and it can quickly adapt to longer ctxLens."}
+{"text": "7:Consider RWKV 14B. The state has 200 vectors, that is, 5 vectors for each block: fp16 (xx), fp32 (aa), fp32 (bb), fp32 (pp), fp16 (xx)."}
--- a/finetune/get_layer_and_embd.py
+++ b/finetune/get_layer_and_embd.py
@@ -0,0 +1,41 @@
+import torch
+import sys
+import time
+import os
+import threading
+import gc
+
+
+def file_cleaner(file):
+    last_pos = 0
+
+    def cleaner():
+        nonlocal last_pos
+        while True:
+            time.sleep(0.1)
+            pos = file.tell()
+            if pos > last_pos:
+                os.posix_fadvise(
+                    file.fileno(), last_pos, pos - last_pos, os.POSIX_FADV_DONTNEED
+                )
+            last_pos = pos
+
+    return cleaner
+
+
+model_file = open(sys.argv[1], "rb")
+cleaner = file_cleaner(model_file)
+cleaner_thread = threading.Thread(target=cleaner, daemon=True)
+cleaner_thread.start()
+
+w = torch.load(model_file, map_location="cpu")
+gc.collect()
+
+n_embd = w["emb.weight"].shape[1]
+n_layer = 0
+keys = list(w.keys())
+for x in keys:
+    layer_id = int(x.split(".")[1]) if ("blocks." in x) else 0
+    n_layer = max(n_layer, layer_id + 1)
+
+print(f"--n_layer {n_layer} --n_embd {n_embd}", end="")
--- a/finetune/install-wsl-dep-and-train.sh
+++ b/finetune/install-wsl-dep-and-train.sh
@@ -0,0 +1,54 @@
+echo $@
+
+if [[ ${cnMirror} == 1 ]]; then
+  export PIP_INDEX_URL="https://pypi.tuna.tsinghua.edu.cn/simple"
+  if grep -q "mirrors.aliyun.com" /etc/apt/sources.list; then
+    echo "apt cnMirror already set"
+  else
+    sudo sed -i 's/http:\/\/archive.ubuntu.com\/ubuntu\//http:\/\/mirrors.aliyun.com\/ubuntu\//g' /etc/apt/sources.list
+    sudo apt update
+  fi
+fi
+
+if dpkg -s "gcc" >/dev/null 2>&1; then
+  echo "gcc installed"
+else
+  sudo apt -y install gcc
+fi
+
+if dpkg -s "python3-pip" >/dev/null 2>&1; then
+  echo "pip installed"
+else
+  sudo apt -y install python3-pip
+fi
+
+if dpkg -s "ninja-build" >/dev/null 2>&1; then
+  echo "ninja installed"
+else
+  sudo apt -y install ninja-build
+fi
+
+if dpkg -s "cuda" >/dev/null 2>&1 && dpkg -s "cuda" | grep Version | awk '{print $2}' | grep -q "12"; then
+  echo "cuda 12 installed"
+else
+  wget -N https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
+  sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600
+  wget -N https://developer.download.nvidia.com/compute/cuda/12.2.0/local_installers/cuda-repo-wsl-ubuntu-12-2-local_12.2.0-1_amd64.deb
+  sudo dpkg -i cuda-repo-wsl-ubuntu-12-2-local_12.2.0-1_amd64.deb
+  sudo cp /var/cuda-repo-wsl-ubuntu-12-2-local/cuda-*-keyring.gpg /usr/share/keyrings/
+  sudo apt-get update
+  sudo apt-get -y install cuda
+fi
+
+if python3 -c "import pkg_resources; pkg_resources.require(open('./finetune/requirements.txt',mode='r'))" &>/dev/null; then
+  echo "requirements satisfied"
+else
+  python3 -m pip install -r ./finetune/requirements.txt
+fi
+
+echo "loading $loadModel"
+modelInfo=$(python3 ./finetune/get_layer_and_embd.py $loadModel)
+echo $modelInfo
+
+python3 ./finetune/lora/train.py $modelInfo $@ --proj_dir lora-models --data_type binidx --lora \
+  --lora_parts=att,ffn,time,ln --strategy deepspeed_stage_2 --accelerator gpu
--- a/finetune/json2binidx_tool/tools/indexed_dataset.py
+++ b/finetune/json2binidx_tool/tools/indexed_dataset.py
@@ -0,0 +1,597 @@
+# Copyright (c) 2021, EleutherAI
+# This file is based on code by the authors denoted below and has been modified from its original version.
+#
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# copied from fairseq/fairseq/data/indexed_dataset.py
+# Removed IndexedRawTextDataset since it relied on Fairseq dictionary
+# other slight modifications to remove fairseq dependencies
+# Added document index to index file and made it accessible.
+#    An empty sentence no longer separates documents.
+
+import os
+import shutil
+import struct
+from functools import lru_cache
+from itertools import accumulate
+
+import numpy as np
+import torch
+
+
+
+
+def __best_fitting_dtype(vocab_size=None):
+    if vocab_size is not None and vocab_size < 65500:
+        return np.uint16
+    else:
+        return np.int32
+
+
+def infer_dataset_impl(path):
+    if IndexedDataset.exists(path):
+        with open(index_file_path(path), "rb") as f:
+            magic = f.read(8)
+            if magic == IndexedDataset._HDR_MAGIC:
+                return "cached"
+            elif magic == MMapIndexedDataset.Index._HDR_MAGIC[:8]:
+                return "mmap"
+            else:
+                return None
+    else:
+        print(f"Dataset does not exist: {path}")
+        print(
+            "Path should be a basename that both .idx and .bin can be appended to get full filenames."
+        )
+        return None
+
+
+def make_builder(out_file, impl, vocab_size=None):
+    if impl == "mmap":
+        return MMapIndexedDatasetBuilder(
+            out_file, dtype=__best_fitting_dtype(vocab_size)
+        )
+    else:
+        return IndexedDatasetBuilder(out_file)
+
+
+def make_dataset(path, impl, skip_warmup=False):
+    if not IndexedDataset.exists(path):
+        print(f"Dataset does not exist: {path}")
+        print(
+            "Path should be a basename that both .idx and .bin can be appended to get full filenames."
+        )
+        return None
+    if impl == "infer":
+        impl = infer_dataset_impl(path)
+    if impl == "lazy" and IndexedDataset.exists(path):
+        return IndexedDataset(path)
+    elif impl == "cached" and IndexedDataset.exists(path):
+        return IndexedCachedDataset(path)
+    elif impl == "mmap" and MMapIndexedDataset.exists(path):
+        return MMapIndexedDataset(path, skip_warmup)
+    print(f"Unknown dataset implementation: {impl}")
+    return None
+
+
+def dataset_exists(path, impl):
+    if impl == "mmap":
+        return MMapIndexedDataset.exists(path)
+    else:
+        return IndexedDataset.exists(path)
+
+
+def read_longs(f, n):
+    a = np.empty(n, dtype=np.int64)
+    f.readinto(a)
+    return a
+
+
+def write_longs(f, a):
+    f.write(np.array(a, dtype=np.int64))
+
+
+dtypes = {
+    1: np.uint8,
+    2: np.int8,
+    3: np.int16,
+    4: np.int32,
+    5: np.int64,
+    6: np.float32,
+    7: np.float64,
+    8: np.uint16,
+}
+
+
+def code(dtype):
+    for k in dtypes.keys():
+        if dtypes[k] == dtype:
+            return k
+    raise ValueError(dtype)
+
+
+def index_file_path(prefix_path):
+    return prefix_path + ".idx"
+
+
+def data_file_path(prefix_path):
+    return prefix_path + ".bin"
+
+
+def create_doc_idx(sizes):
+    doc_idx = [0]
+    for i, s in enumerate(sizes):
+        if s == 0:
+            doc_idx.append(i + 1)
+    return doc_idx
+
+
+class IndexedDataset(torch.utils.data.Dataset):
+    """Loader for IndexedDataset"""
+
+    _HDR_MAGIC = b"TNTIDX\x00\x00"
+
+    def __init__(self, path):
+        super().__init__()
+        self.path = path
+        self.data_file = None
+        self.read_index(path)
+
+    def read_index(self, path):
+        with open(index_file_path(path), "rb") as f:
+            magic = f.read(8)
+            assert magic == self._HDR_MAGIC, (
+                "Index file doesn't match expected format. "
+                "Make sure that --dataset-impl is configured properly."
+            )
+            version = f.read(8)
+            assert struct.unpack("<Q", version) == (1,)
+            code, self.element_size = struct.unpack("<QQ", f.read(16))
+            self.dtype = dtypes[code]
+            self._len, self.s = struct.unpack("<QQ", f.read(16))
+            self.doc_count = struct.unpack("<Q", f.read(8))
+            self.dim_offsets = read_longs(f, self._len + 1)
+            self.data_offsets = read_longs(f, self._len + 1)
+            self.sizes = read_longs(f, self.s)
+            self.doc_idx = read_longs(f, self.doc_count)
+
+    def read_data(self, path):
+        self.data_file = open(data_file_path(path), "rb", buffering=0)
+
+    def check_index(self, i):
+        if i < 0 or i >= self._len:
+            raise IndexError("index out of range")
+
+    def __del__(self):
+        if self.data_file:
+            self.data_file.close()
+
+    # @lru_cache(maxsize=8)
+    def __getitem__(self, idx):
+        if not self.data_file:
+            self.read_data(self.path)
+        if isinstance(idx, int):
+            i = idx
+            self.check_index(i)
+            tensor_size = self.sizes[self.dim_offsets[i] : self.dim_offsets[i + 1]]
+            a = np.empty(tensor_size, dtype=self.dtype)
+            self.data_file.seek(self.data_offsets[i] * self.element_size)
+            self.data_file.readinto(a)
+            return a
+        elif isinstance(idx, slice):
+            start, stop, step = idx.indices(len(self))
+            if step != 1:
+                raise ValueError("Slices into indexed_dataset must be contiguous")
+            sizes = self.sizes[self.dim_offsets[start] : self.dim_offsets[stop]]
+            size = sum(sizes)
+            a = np.empty(size, dtype=self.dtype)
+            self.data_file.seek(self.data_offsets[start] * self.element_size)
+            self.data_file.readinto(a)
+            offsets = list(accumulate(sizes))
+            sents = np.split(a, offsets[:-1])
+            return sents
+
+    def __len__(self):
+        return self._len
+
+    def num_tokens(self, index):
+        return self.sizes[index]
+
+    def size(self, index):
+        return self.sizes[index]
+
+    @staticmethod
+    def exists(path):
+        return os.path.exists(index_file_path(path)) and os.path.exists(
+            data_file_path(path)
+        )
+
+    @property
+    def supports_prefetch(self):
+        return False  # avoid prefetching to save memory
+
+
+class IndexedCachedDataset(IndexedDataset):
+    def __init__(self, path):
+        super().__init__(path)
+        self.cache = None
+        self.cache_index = {}
+
+    @property
+    def supports_prefetch(self):
+        return True
+
+    def prefetch(self, indices):
+        if all(i in self.cache_index for i in indices):
+            return
+        if not self.data_file:
+            self.read_data(self.path)
+        indices = sorted(set(indices))
+        total_size = 0
+        for i in indices:
+            total_size += self.data_offsets[i + 1] - self.data_offsets[i]
+        self.cache = np.empty(total_size, dtype=self.dtype)
+        ptx = 0
+        self.cache_index.clear()
+        for i in indices:
+            self.cache_index[i] = ptx
+            size = self.data_offsets[i + 1] - self.data_offsets[i]
+            a = self.cache[ptx : ptx + size]
+            self.data_file.seek(self.data_offsets[i] * self.element_size)
+            self.data_file.readinto(a)
+            ptx += size
+        if self.data_file:
+            # close and delete data file after prefetch so we can pickle
+            self.data_file.close()
+            self.data_file = None
+
+    # @lru_cache(maxsize=8)
+    def __getitem__(self, idx):
+        if isinstance(idx, int):
+            i = idx
+            self.check_index(i)
+            tensor_size = self.sizes[self.dim_offsets[i] : self.dim_offsets[i + 1]]
+            a = np.empty(tensor_size, dtype=self.dtype)
+            ptx = self.cache_index[i]
+            np.copyto(a, self.cache[ptx : ptx + a.size])
+            return a
+        elif isinstance(idx, slice):
+            # Hack just to make this work, can optimizer later if necessary
+            sents = []
+            for i in range(*idx.indices(len(self))):
+                sents.append(self[i])
+            return sents
+
+
+class IndexedDatasetBuilder(object):
+    element_sizes = {
+        np.uint8: 1,
+        np.int8: 1,
+        np.int16: 2,
+        np.int32: 4,
+        np.int64: 8,
+        np.float32: 4,
+        np.float64: 8,
+    }
+
+    def __init__(self, out_file, dtype=np.int32):
+        self.out_file = open(out_file, "wb")
+        self.dtype = dtype
+        self.data_offsets = [0]
+        self.dim_offsets = [0]
+        self.sizes = []
+        self.element_size = self.element_sizes[self.dtype]
+        self.doc_idx = [0]
+
+    def add_item(self, np_array):
+        assert isinstance(np_array, np.ndarray) and np_array.dtype == self.dtype
+        bytes = self.out_file.write(np_array)
+        self.data_offsets.append(self.data_offsets[-1] + bytes / self.element_size)
+        for s in np_array.shape:
+            self.sizes.append(s)
+        self.dim_offsets.append(self.dim_offsets[-1] + len(np_array.shape))
+
+    def end_document(self):
+        self.doc_idx.append(len(self.sizes))
+
+    def merge_file_(self, another_file):
+        index = IndexedDataset(another_file)
+        assert index.dtype == self.dtype
+
+        begin = self.data_offsets[-1]
+        for offset in index.data_offsets[1:]:
+            self.data_offsets.append(begin + offset)
+        self.sizes.extend(index.sizes)
+        begin = self.dim_offsets[-1]
+        for dim_offset in index.dim_offsets[1:]:
+            self.dim_offsets.append(begin + dim_offset)
+
+        with open(data_file_path(another_file), "rb") as f:
+            while True:
+                data = f.read(1024)
+                if data:
+                    self.out_file.write(data)
+                else:
+                    break
+
+    def finalize(self, index_file):
+        self.out_file.close()
+        index = open(index_file, "wb")
+        index.write(b"TNTIDX\x00\x00")
+        index.write(struct.pack("<Q", 1))
+        index.write(struct.pack("<QQ", code(self.dtype), self.element_size))
+        index.write(struct.pack("<QQ", len(self.data_offsets) - 1, len(self.sizes)))
+        index.write(struct.pack("<Q", len(self.doc_idx)))
+        write_longs(index, self.dim_offsets)
+        write_longs(index, self.data_offsets)
+        write_longs(index, self.sizes)
+        write_longs(index, self.doc_idx)
+        index.close()
+
+
+def _warmup_mmap_file(path):
+    with open(path, "rb") as stream:
+        while stream.read(100 * 1024 * 1024):
+            pass
+
+
+class MMapIndexedDataset(torch.utils.data.Dataset):
+    class Index(object):
+        _HDR_MAGIC = b"MMIDIDX\x00\x00"
+
+        @classmethod
+        def writer(cls, path, dtype):
+            class _Writer(object):
+                def __enter__(self):
+                    self._file = open(path, "wb")
+
+                    # Write Magic string so we can check the file format then opening it again.
+                    self._file.write(cls._HDR_MAGIC)
+                    # Write version number
+                    # Little endian unsigned 64 Bit integer
+                    self._file.write(struct.pack("<Q", 1))
+                    # Little endian unsigned 8 Bit integer
+                    self._file.write(struct.pack("<B", code(dtype)))
+
+                    return self
+
+                @staticmethod
+                def _get_pointers(sizes):
+                    pointers = np.zeros(len(sizes), dtype=np.int64)
+                    sizes = np.array(sizes, dtype=np.int64)
+
+                    np.cumsum(sizes[:-1], out=pointers[1:])
+                    pointers = pointers * dtype().itemsize
+                    return pointers
+
+                def write(self, sizes, doc_idx):
+                    pointers = self._get_pointers(sizes)
+
+                    # Little endian unsigned 64 Bit integer
+                    self._file.write(struct.pack("<Q", len(sizes)))
+                    # Little endian unsigned 64 Bit integer
+                    self._file.write(struct.pack("<Q", len(doc_idx)))
+
+                    sizes = np.array(sizes, dtype=np.int32)
+                    self._file.write(sizes.tobytes(order="C"))
+                    del sizes
+
+                    pointers = np.array(pointers, dtype=np.int64)
+                    self._file.write(pointers.tobytes(order="C"))
+                    del pointers
+
+                    doc_idx = np.array(doc_idx, dtype=np.int64)
+                    self._file.write(doc_idx.tobytes(order="C"))
+
+                def __exit__(self, exc_type, exc_val, exc_tb):
+                    self._file.close()
+
+            return _Writer()
+
+        def __init__(self, path, skip_warmup=False):
+            with open(path, "rb") as stream:
+                magic_test = stream.read(9)
+                assert self._HDR_MAGIC == magic_test, (
+                    "Index file doesn't match expected format. "
+                    "Make sure that --dataset-impl is configured properly."
+                )
+                # Little endian unsigned 64 Bit integer
+                version = struct.unpack("<Q", stream.read(8))
+                assert (1,) == version
+
+                # Little endian unsigned 8 Bit integer
+                (dtype_code,) = struct.unpack("<B", stream.read(1))
+                self._dtype = dtypes[dtype_code]
+                self._dtype_size = self._dtype().itemsize
+
+                self._len = struct.unpack("<Q", stream.read(8))[0]
+                self._doc_count = struct.unpack("<Q", stream.read(8))[0]
+                offset = stream.tell()
+
+            if not skip_warmup:
+                print("    warming up index mmap file...")
+                _warmup_mmap_file(path)
+
+            self._bin_buffer_mmap = np.memmap(path, mode="r", order="C")
+            self._bin_buffer = memoryview(self._bin_buffer_mmap)
+            print("    reading sizes...")
+            self._sizes = np.frombuffer(
+                self._bin_buffer, dtype=np.int32, count=self._len, offset=offset
+            )
+            print("    reading pointers...")
+            self._pointers = np.frombuffer(
+                self._bin_buffer,
+                dtype=np.int64,
+                count=self._len,
+                offset=offset + self._sizes.nbytes,
+            )
+            print("    reading document index...")
+            self._doc_idx = np.frombuffer(
+                self._bin_buffer,
+                dtype=np.int64,
+                count=self._doc_count,
+                offset=offset + self._sizes.nbytes + self._pointers.nbytes,
+            )
+
+        def __del__(self):
+            self._bin_buffer_mmap._mmap.close()
+            del self._bin_buffer_mmap
+
+        @property
+        def dtype(self):
+            return self._dtype
+
+        @property
+        def sizes(self):
+            return self._sizes
+
+        @property
+        def doc_idx(self):
+            return self._doc_idx
+
+        @lru_cache(maxsize=8)
+        def __getitem__(self, i):
+            return self._pointers[i], self._sizes[i]
+
+        def __len__(self):
+            return self._len
+
+    def __init__(self, path, skip_warmup=False):
+        super().__init__()
+
+        self._path = None
+        self._index = None
+        self._bin_buffer = None
+
+        self._do_init(path, skip_warmup)
+
+    def __getstate__(self):
+        return self._path
+
+    def __setstate__(self, state):
+        self._do_init(state)
+
+    def _do_init(self, path, skip_warmup):
+        self._path = path
+        self._index = self.Index(index_file_path(self._path), skip_warmup)
+
+        if not skip_warmup:
+            print("    warming up data mmap file...")
+            _warmup_mmap_file(data_file_path(self._path))
+        print("    creating numpy buffer of mmap...")
+        self._bin_buffer_mmap = np.memmap(
+            data_file_path(self._path), mode="r", order="C"
+        )
+        print("    creating memory view of numpy buffer...")
+        self._bin_buffer = memoryview(self._bin_buffer_mmap)
+
+    def __del__(self):
+        self._bin_buffer_mmap._mmap.close()
+        del self._bin_buffer_mmap
+        del self._index
+
+    def __len__(self):
+        return len(self._index)
+
+    # @lru_cache(maxsize=8)
+    def __getitem__(self, idx):
+        if isinstance(idx, int):
+            ptr, size = self._index[idx]
+            np_array = np.frombuffer(
+                self._bin_buffer, dtype=self._index.dtype, count=size, offset=ptr
+            )
+            return np_array
+        elif isinstance(idx, slice):
+            start, stop, step = idx.indices(len(self))
+            if step != 1:
+                raise ValueError("Slices into indexed_dataset must be contiguous")
+            ptr = self._index._pointers[start]
+            sizes = self._index._sizes[idx]
+            offsets = list(accumulate(sizes))
+            total_size = sum(sizes)
+            np_array = np.frombuffer(
+                self._bin_buffer, dtype=self._index.dtype, count=total_size, offset=ptr
+            )
+            sents = np.split(np_array, offsets[:-1])
+            return sents
+
+    def get(self, idx, offset=0, length=None):
+        """Retrieves a single item from the dataset with the option to only
+        return a portion of the item.
+
+        get(idx) is the same as [idx] but get() does not support slicing.
+        """
+        ptr, size = self._index[idx]
+        if length is None:
+            length = size - offset
+        ptr += offset * np.dtype(self._index.dtype).itemsize
+        np_array = np.frombuffer(
+            self._bin_buffer, dtype=self._index.dtype, count=length, offset=ptr
+        )
+        return np_array
+
+    @property
+    def sizes(self):
+        return self._index.sizes
+
+    @property
+    def doc_idx(self):
+        return self._index.doc_idx
+
+    def get_doc_idx(self):
+        return self._index._doc_idx
+
+    def set_doc_idx(self, doc_idx_):
+        self._index._doc_idx = doc_idx_
+
+    @property
+    def supports_prefetch(self):
+        return False
+
+    @staticmethod
+    def exists(path):
+        return os.path.exists(index_file_path(path)) and os.path.exists(
+            data_file_path(path)
+        )
+
+
+class MMapIndexedDatasetBuilder(object):
+    def __init__(self, out_file, dtype=np.int64):
+        self._data_file = open(out_file, "wb")
+        self._dtype = dtype
+        self._sizes = []
+        self._doc_idx = [0]
+
+    @property
+    def dtype(self):
+        return self._dtype
+
+    def add_item(self, np_array):
+        assert isinstance(np_array, np.ndarray) and np_array.dtype == self.dtype
+        self._data_file.write(np_array.tobytes(order="C"))
+        self._sizes.append(np_array.size)
+
+    def end_document(self):
+        self._doc_idx.append(len(self._sizes))
+
+    def merge_file_(self, another_file):
+        # Concatenate index
+        index = MMapIndexedDataset.Index(index_file_path(another_file))
+        assert index.dtype == self._dtype
+
+        for size in index.sizes:
+            self._sizes.append(size)
+
+        # Concatenate data
+        with open(data_file_path(another_file), "rb") as f:
+            shutil.copyfileobj(f, self._data_file)
+
+    def finalize(self, index_file):
+        self._data_file.close()
+
+        with MMapIndexedDataset.Index.writer(index_file, self._dtype) as index:
+            index.write(self._sizes, self._doc_idx)
--- a/finetune/json2binidx_tool/tools/preprocess_data.py
+++ b/finetune/json2binidx_tool/tools/preprocess_data.py
@@ -0,0 +1,250 @@
+# Copyright (c) 2021, EleutherAI
+# This file is based on code by the authors denoted below and has been modified from its original version.
+#
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Processing data for pretraining."""
+
+import os
+import sys
+
+sys.path.append(os.path.dirname(os.path.realpath(__file__)))
+
+import argparse
+import multiprocessing
+
+import lm_dataformat as lmd
+import numpy as np
+
+sys.path.append(
+    os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir))
+)
+import time
+import tqdm
+import ftfy
+
+from tokenizer import build_tokenizer
+import indexed_dataset
+from threading import Semaphore
+
+
+class Encoder(object):
+    def __init__(self, args):
+        self.args = args
+
+    def initializer(self):
+        # Use Encoder class as a container for global data
+        Encoder.tokenizer = build_tokenizer(self.args)
+
+    def encode(self, text):
+        if self.args.ftfy:
+            text = ftfy.fix_text(text)
+        ids = {}
+        for key in self.args.jsonl_keys:
+            doc_ids = []
+            text_ids = Encoder.tokenizer.tokenize(text)
+            if len(text_ids) > 0:
+                doc_ids.append(text_ids)
+            if self.args.append_eod:
+                doc_ids[-1].append(Encoder.tokenizer.eod)
+            ids[key] = doc_ids
+        return ids, len(text)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    group = parser.add_argument_group(title="input data")
+    group.add_argument(
+        "--input",
+        type=str,
+        required=True,
+        help="Path to input jsonl files or lmd archive(s) - if using multiple archives, put them in a comma separated "
+        "list",
+    )
+    group.add_argument(
+        "--jsonl-keys",
+        nargs="+",
+        default=["text"],
+        help="space separate listed of keys to extract from jsonl. Defa",
+    )
+    group.add_argument(
+        "--num-docs",
+        default=None,
+        help="Optional: Number of documents in the input data (if known) for an accurate progress bar.",
+        type=int,
+    )
+    group = parser.add_argument_group(title="tokenizer")
+    group.add_argument(
+        "--tokenizer-type",
+        type=str,
+        required=True,
+        choices=[
+            "HFGPT2Tokenizer",
+            "HFTokenizer",
+            "GPT2BPETokenizer",
+            "CharLevelTokenizer",
+            "TiktokenTokenizer",
+            "RWKVTokenizer",
+        ],
+        help="What type of tokenizer to use.",
+    )
+    group.add_argument(
+        "--vocab-file", type=str, default=None, help="Path to the vocab file"
+    )
+    group.add_argument(
+        "--merge-file",
+        type=str,
+        default=None,
+        help="Path to the BPE merge file (if necessary).",
+    )
+    group.add_argument(
+        "--append-eod",
+        action="store_true",
+        help="Append an <eod> token to the end of a document.",
+    )
+    group.add_argument("--ftfy", action="store_true", help="Use ftfy to clean text")
+    group = parser.add_argument_group(title="output data")
+    group.add_argument(
+        "--output-prefix",
+        type=str,
+        required=True,
+        help="Path to binary output file without suffix",
+    )
+    group.add_argument(
+        "--dataset-impl",
+        type=str,
+        default="mmap",
+        choices=["lazy", "cached", "mmap"],
+        help="Dataset implementation to use. Default: mmap",
+    )
+
+    group = parser.add_argument_group(title="runtime")
+    group.add_argument(
+        "--workers", type=int, default=1, help="Number of worker processes to launch"
+    )
+    group.add_argument(
+        "--log-interval",
+        type=int,
+        default=100,
+        help="Interval between progress updates",
+    )
+    args = parser.parse_args()
+    args.keep_empty = False
+
+    # some default/dummy values for the tokenizer
+    args.rank = 0
+    args.make_vocab_size_divisible_by = 128
+    args.model_parallel_size = 1
+
+    return args
+
+
+def yield_from_files(fnames: list, semaphore):
+    """
+    Iterator over input documents using lm_dataformat. Should be able to handle jsons / texts /
+    other compressed formats. Also filters out empty documents.
+
+    :param fnames: list of filenames
+    """
+
+    def yielder(fname, semaphore):
+        for f in filter(lambda x: x, lmd.Reader(fname).stream_data()):
+            semaphore.acquire()
+            yield f
+
+    for fname in fnames:
+        semaphore.acquire()
+
+        yield from yielder(fname, semaphore)
+
+
+def main():
+    args = get_args()
+    encoder = Encoder(args)
+    tokenizer = build_tokenizer(args)
+    print(f"Vocab size: {tokenizer.vocab_size}")
+    print(f"Output prefix: {args.output_prefix}")
+
+    # build a semaphore object to stop `yield_from_files` from getting ahead of encoder.encode and
+    # hence building up memory
+    semaphore = Semaphore(10000 + args.workers)
+
+    # use multiprocessing to iterate over input documents
+    fin = yield_from_files(args.input.split(","), semaphore)
+
+    if args.workers > 1:
+        pool = multiprocessing.Pool(args.workers, initializer=encoder.initializer)
+        encoded_docs = pool.imap(encoder.encode, fin, chunksize=25)
+    else:
+        encoder.initializer()
+        encoded_docs = (encoder.encode(doc) for doc in fin)
+
+    # make a dataset builder for each key in args.jsonl_keys
+    # each key will output to a different file beginning with args.output_prefix
+    output_bin_files = {}
+    output_idx_files = {}
+    builders = {}
+    for key in args.jsonl_keys:
+        output_bin_files[key] = "{}_{}_{}.bin".format(
+            args.output_prefix, key, "document"
+        )
+        output_idx_files[key] = "{}_{}_{}.idx".format(
+            args.output_prefix, key, "document"
+        )
+        builders[key] = indexed_dataset.make_builder(
+            output_bin_files[key],
+            impl=args.dataset_impl,
+            vocab_size=tokenizer.vocab_size,
+        )
+
+    # actually do tokenization
+    proc_start = time.time()
+    total_bytes_processed = 0
+    pbar = tqdm.tqdm()
+    for i, (doc, bytes_processed) in enumerate(encoded_docs, start=1):
+        total_bytes_processed += bytes_processed
+
+        # release semaphore so `yield_from_files` can add another file to the buffer
+        semaphore.release()
+
+        # add each tokenized document / sentence
+        for key, sentences in doc.items():
+            for sentence in sentences:
+                builders[key].add_item(np.array(sentence, dtype=builders[key].dtype))
+            # separate with eos token
+            builders[key].end_document()
+
+        # log progress
+        if i % args.log_interval == 0:
+            current = time.time()
+            elapsed = current - proc_start
+            mbs = total_bytes_processed / elapsed / 1024 / 1024
+            pbar.set_description(
+                f"Processed {i}{'' if args.num_docs is None else '/' + str(args.num_docs)} documents ({i / elapsed:0.2f} docs/s, {mbs:0.2f} MB/s)."
+            )
+            if i != 0:
+                pbar.update(args.log_interval)
+
+    # save output file
+    for key in args.jsonl_keys:
+        builders[key].finalize(output_idx_files[key])
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as e:
+        with open("error.txt", "w") as f:
+            f.write(str(e))
--- a/finetune/json2binidx_tool/tools/rwkv_tokenizer.py
+++ b/finetune/json2binidx_tool/tools/rwkv_tokenizer.py
@@ -0,0 +1,232 @@
+########################################################################################################
+# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
+# Source: https://github.com/BlinkDL/ChatRWKV/blob/main/tokenizer/rwkv_tokenizer.py
+########################################################################################################
+
+import os, sys, time, random
+
+print('''
+#######################################################################################################################
+
+This tokenizer is not used in any RWKV models yet. I plan to use it for the future multilang RWKV models.
+
+Benefits:
+
+* Good support of most languages, from European to CJK to Arabic and Hindi and more.
+
+* Clean vocab. Good for code too. Vocab size = 65525 (use 0 for <|endoftext|>).
+
+* Good at numbers: the numerical tokens are '0'~'9', '10'~'99', ' 0'~' 9', ' 10'~' 99'.
+
+* Very easy tokenization:
+
+** The input text must be in UTF-8.
+
+** Greedy encoding: always pick the longest (in bytes) token (with the highest id) that matches your UTF-8 bytes.
+
+* The tokenization result is surprisingly good, because the vocab respects word boundaries and UTF-8 boundaries.
+
+For 10x faster speed:
+mypyc rwkv_tokenizer.py
+python3 -c "import rwkv_tokenizer"
+
+#######################################################################################################################
+''')
+
+########################################################################################################
+# Tokenizer #1 (reference, naive, slow)
+########################################################################################################
+
+class RWKV_TOKENIZER():
+    table = None  # : list[list[list[bytes]]] = None
+    good = None  # : list[set[int]]
+    wlen = None  # : list[int]
+    def __init__(self, file_name):
+        self.vocab_size = 65525
+        self.idx2token = {}
+        sorted = [] # must be already sorted
+        lines = open(file_name, "r", encoding="utf-8").readlines()
+        for l in lines:
+            idx = int(l[:l.index(' ')])
+            x = eval(l[l.index(' '):l.rindex(' ')])
+            x = x.encode("utf-8") if isinstance(x, str) else x
+            assert isinstance(x, bytes)
+            assert len(x) == int(l[l.rindex(' '):])
+            sorted += [x]
+            self.idx2token[idx] = x
+
+        self.token2idx = {}
+        for k, v in self.idx2token.items():
+            self.token2idx[v] = int(k)
+
+        # precompute some tables for fast matching
+        self.table = [[[] for j in range(256)] for i in range(256)]
+        self.good = [set() for i in range(256)]
+        self.wlen = [0 for i in range(256)]
+
+        for i in reversed(range(len(sorted))): # reverse order - match longer tokens first
+            s = sorted[i]
+            if len(s) >= 2:
+                s0 = int(s[0])
+                s1 = int(s[1])
+                self.table[s0][s1] += [s]
+                self.wlen[s0] = max(self.wlen[s0], len(s))
+                self.good[s0].add(s1)
+
+    def encodeBytes(self, src: bytes):
+        src_len: int = len(src)
+        tokens = []
+        i: int = 0
+        while i < src_len:
+            s: bytes = src[i : i + 1]
+
+            if i < src_len - 1:
+                s1: int = int(src[i + 1])
+                s0: int = int(src[i])
+                if s1 in self.good[s0]:
+                    sss: bytes = src[i : i + self.wlen[s0]]
+                    try:
+                        s = next(filter(sss.startswith, self.table[s0][s1]))
+                    except:
+                        pass
+            tokens.append(self.token2idx[s])
+            i += len(s)
+
+        return tokens
+
+    def decodeBytes(self, tokens):
+        return b''.join(map(lambda i: self.idx2token[i], tokens))
+
+    def encode(self, src: str):
+        return self.encodeBytes(src.encode("utf-8"))
+
+    def decode(self, tokens):
+        return self.decodeBytes(tokens).decode('utf-8')
+
+    def token_to_id(self, token):
+        return self.token2idx[token]
+
+    def get_vocab_size(self):
+        return self.vocab_size
+
+    def get_vocab(self):
+        return self.idx2token
+
+    def printTokens(self, tokens):
+        for i in tokens:
+            s = self.idx2token[i]
+            try:
+                s = s.decode('utf-8')
+            except:
+                pass
+            print(f'{repr(s)}{i}', end=' ')
+            # print(repr(s), i)
+        print()
+
+########################################################################################################
+# Tokenizer #2 (trie, faster) https://github.com/TkskKurumi/ChatRWKV-TRIE-Tokenizer
+########################################################################################################
+
+class TRIE:
+    __slots__ = tuple("ch,to,values,front".split(","))
+    to:list
+    values:set
+    def __init__(self, front=None, ch=None):
+        self.ch = ch
+        self.to = [None for ch in range(256)]
+        self.values = set()
+        self.front = front
+
+    def __repr__(self):
+        fr = self
+        ret = []
+        while(fr!=None):
+            if(fr.ch!=None):
+                ret.append(fr.ch)
+            fr = fr.front
+        return "<TRIE %s %s>"%(ret[::-1], self.values)
+    
+    def add(self, key:bytes, idx:int=0, val=None):
+        if(idx == len(key)):
+            if(val is None):
+                val = key
+            self.values.add(val)
+            return self
+        ch = key[idx]
+        if(self.to[ch] is None):
+            self.to[ch] = TRIE(front=self, ch=ch)
+        return self.to[ch].add(key, idx=idx+1, val=val)
+    
+    def find_longest(self, key:bytes, idx:int=0):
+        u:TRIE = self
+        ch:int = key[idx]
+        
+        while(u.to[ch] is not None):
+            u = u.to[ch]
+            idx += 1
+            if(u.values):
+                ret = idx, u, u.values
+            if(idx==len(key)):
+                break
+            ch = key[idx]
+        return ret
+
+class TRIE_TOKENIZER():
+    def __init__(self, file_name):
+        self.vocab_size = 65525
+        self.idx2token = {}
+        sorted = [] # must be already sorted
+        with open(file_name, "r", encoding="utf-8") as f:
+            lines = f.readlines()
+        for l in lines:
+            idx = int(l[:l.index(' ')])
+            x = eval(l[l.index(' '):l.rindex(' ')])
+            x = x.encode("utf-8") if isinstance(x, str) else x
+            assert isinstance(x, bytes)
+            assert len(x) == int(l[l.rindex(' '):])
+            sorted += [x]
+            self.idx2token[idx] = x
+
+        self.token2idx = {}
+        for k,v in self.idx2token.items():
+            self.token2idx[v] = int(k)
+
+        self.root = TRIE()
+        for t, i in self.token2idx.items():
+            _ = self.root.add(t, val=(t, i))
+
+    def encodeBytes(self, src:bytes):
+        idx:int = 0
+        tokens = []
+        while (idx < len(src)):
+            _idx:int = idx
+            idx, _, values = self.root.find_longest(src, idx)
+            assert(idx != _idx)
+            _, token = next(iter(values))            
+            tokens.append(token)
+        return tokens
+
+    def decodeBytes(self, tokens):
+        return b''.join(map(lambda i: self.idx2token[i], tokens))
+
+    def encode(self, src):
+        return self.encodeBytes(src.encode("utf-8"))
+
+    def decode(self, tokens):
+        return self.decodeBytes(tokens).decode('utf-8')
+
+    def get_vocab_size(self):
+        return self.vocab_size
+
+    def get_vocab(self):
+        return self.idx2token
+
+    def printTokens(self, tokens):
+        for i in tokens:
+            s = self.idx2token[i]
+            try:
+                s = s.decode('utf-8')
+            except:
+                pass
+            print(f'{repr(s)}{i}', end=' ')
+        print()
--- a/finetune/json2binidx_tool/tools/tokenizer.py
+++ b/finetune/json2binidx_tool/tools/tokenizer.py
@@ -0,0 +1,205 @@
+# Copyright (c) 2021, EleutherAI
+# This file is based on code by the authors denoted below and has been modified from its original version.
+#
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Megatron tokenizers."""
+
+from abc import ABC
+from abc import abstractmethod
+
+from tokenizers import Tokenizer
+from rwkv_tokenizer import RWKV_TOKENIZER, TRIE_TOKENIZER
+
+from typing import List, Union
+
+
+def build_tokenizer(args):
+    """Initialize tokenizer."""
+    if args.rank == 0:
+        print("> building {} tokenizer ...".format(args.tokenizer_type), flush=True)
+
+    # Select and instantiate the tokenizer.
+
+    if args.tokenizer_type.lower() == "HFTokenizer".lower():
+        assert args.vocab_file is not None
+        tokenizer = HFTokenizer(args.vocab_file)
+    elif args.tokenizer_type.lower() == "RWKVTokenizer".lower():
+        assert args.vocab_file is not None
+        tokenizer = RWKVTokenizer(args.vocab_file)
+    else:
+        raise NotImplementedError(
+            "{} tokenizer is not " "implemented.".format(args.tokenizer_type)
+        )
+
+    # Add vocab size.
+    args.padded_vocab_size = _vocab_size_with_padding(tokenizer.vocab_size, args)
+
+    return tokenizer
+
+
+def _vocab_size_with_padding(orig_vocab_size, args):
+    """Pad vocab size so it is divisible by model parallel size and
+    still having GPU friendly size."""
+
+    after = orig_vocab_size
+    multiple = args.make_vocab_size_divisible_by * args.model_parallel_size
+    while (after % multiple) != 0:
+        after += 1
+    if args.rank == 0:
+        print(
+            " > padded vocab (size: {}) with {} dummy tokens "
+            "(new size: {})".format(orig_vocab_size, after - orig_vocab_size, after),
+            flush=True,
+        )
+    return after
+
+
+class AbstractTokenizer(ABC):
+    """Abstract class for tokenizer."""
+
+    def __init__(self, name):
+        self.name = name
+        super().__init__()
+
+    @property
+    @abstractmethod
+    def vocab_size(self):
+        pass
+
+    @property
+    @abstractmethod
+    def vocab(self):
+        """Dictionary from vocab text token to id token."""
+        pass
+
+    @property
+    @abstractmethod
+    def inv_vocab(self):
+        """Dictionary from vocab id token to text token."""
+        pass
+
+    @abstractmethod
+    def tokenize(self, text):
+        pass
+
+    def detokenize(self, token_ids):
+        raise NotImplementedError(
+            "detokenizer is not implemented for {} " "tokenizer".format(self.name)
+        )
+
+    @property
+    def cls(self):
+        raise NotImplementedError(
+            "CLS is not provided for {} " "tokenizer".format(self.name)
+        )
+
+    @property
+    def sep(self):
+        raise NotImplementedError(
+            "SEP is not provided for {} " "tokenizer".format(self.name)
+        )
+
+    @property
+    def pad(self):
+        raise NotImplementedError(
+            "PAD is not provided for {} " "tokenizer".format(self.name)
+        )
+
+    @property
+    def eod(self):
+        raise NotImplementedError(
+            "EOD is not provided for {} " "tokenizer".format(self.name)
+        )
+
+    @property
+    def mask(self):
+        raise NotImplementedError(
+            "MASK is not provided for {} " "tokenizer".format(self.name)
+        )
+
+
+class HFTokenizer(AbstractTokenizer):
+    """Designed to Integrate HF's Tokenizer library."""
+
+    def __init__(self, vocab_file):
+        name = "HFTokenizer"
+        super().__init__(name)
+
+        self.tokenizer = Tokenizer.from_file(vocab_file)
+        self.eod_id = self.tokenizer.token_to_id("<|endoftext|>")
+        self.pad_id = self.tokenizer.token_to_id("<|padding|>")
+
+    @property
+    def vocab_size(self):
+        return self.tokenizer.get_vocab_size()
+
+    @property
+    def vocab(self):
+        return self.tokenizer.get_vocab()
+
+    @property
+    def inv_vocab(self):
+        return self.tokenizer.decoder
+
+    def tokenize(self, text: str):
+        return self.tokenizer.encode(text).ids
+
+    def tokenize_batch(self, text_batch: Union[List[str], str]):
+        return self.tokenizer.encode_batch(text_batch)
+
+    def detokenize(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def eod(self):
+        return self.eod_id
+
+
+class RWKVTokenizer(AbstractTokenizer):
+    """RWKV Worlds Tokenizer."""
+
+    def __init__(self, vocab_file='rwkv_vocab_v20230424.txt'):
+        name = "RWKVTokenizer"
+        super().__init__(name)
+
+        self.tokenizer = TRIE_TOKENIZER(vocab_file)
+        self.eod_id = 0  # self.tokenizer.token_to_id("<|endoftext|>")
+        # self.pad_id = self.tokenizer.token_to_id("<|padding|>")
+
+    @property
+    def vocab_size(self):
+        return self.tokenizer.get_vocab_size()
+
+    @property
+    def vocab(self):
+        return self.tokenizer.get_vocab()
+
+    @property
+    def inv_vocab(self):
+        return self.tokenizer.decode
+
+    def tokenize(self, text: str):
+        return self.tokenizer.encode(text)
+
+    def tokenize_batch(self, text_batch: Union[List[str], str]):
+        return self.tokenizer.encode_batch(text_batch)
+
+    def detokenize(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def eod(self):
+        return self.eod_id
--- a/finetune/lora/cuda/wkv_cuda.cu
+++ b/finetune/lora/cuda/wkv_cuda.cu
@@ -0,0 +1,133 @@
+#include <stdio.h>
+#include <assert.h>
+
+#define MIN_VALUE (-1e38)
+
+template <typename F>
+__global__ void kernel_forward(const int B, const int T, const int C,
+                               const F *__restrict__ const _w, const F *__restrict__ const _u, const F *__restrict__ const _k, const F *__restrict__ const _v,
+                               F *__restrict__ const _y) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int _b = idx / C;
+    const int _c = idx % C;
+    const int _offset = _b * T * C + _c;
+
+    F u = _u[_c];
+    F w = _w[_c];
+    const F *__restrict__ const k = _k + _offset;
+    const F *__restrict__ const v = _v + _offset;
+    F *__restrict__ const y = _y + _offset;
+
+    // aa and bb are running sums divided by exp(pp) (to avoid overflow)
+    F aa = 0, bb = 0, pp = MIN_VALUE;
+    for (int i = 0; i < T; i++) {
+        const int ii = i * C;
+        const F kk = k[ii];
+        const F vv = v[ii];
+
+        F ww = u + kk;
+        F p = max(pp, ww);
+        F e1 = exp(pp - p);
+        F e2 = exp(ww - p);
+        y[ii] = (e1 * aa + e2 * vv) / (e1 * bb + e2);
+        
+        ww = w + pp;
+        p = max(ww, kk);
+        e1 = exp(ww - p);
+        e2 = exp(kk - p);
+        aa = e1 * aa + e2 * vv;
+        bb = e1 * bb + e2;
+        pp = p;
+    }
+}
+
+template <typename F>
+__global__ void kernel_backward(const int B, const int T, const int C,
+                                const F *__restrict__ const _w, const F *__restrict__ const _u, const F *__restrict__ const _k, const F *__restrict__ const _v,
+                                const F *__restrict__ const _y, const F *__restrict__ const _gy,
+                                F *__restrict__ const _gw, F *__restrict__ const _gu, F *__restrict__ const _gk, F *__restrict__ const _gv) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int _b = idx / C;
+    const int _c = idx % C;
+    const int _offset = _b * T * C + _c;
+
+    F u = _u[_c];
+    F w = _w[_c];
+    const F *__restrict__ const k = _k + _offset;
+    const F *__restrict__ const v = _v + _offset;
+    const F *__restrict__ const y = _y + _offset;
+    const F *__restrict__ const gy = _gy + _offset;
+    F *__restrict__ const gk = _gk + _offset;
+    F *__restrict__ const gv = _gv + _offset;
+
+    F q[Tmax], r[Tmax];
+
+    F gw = 0, gu = 0, aa = 0, bb = 0, ga = 0, gb = 0, pp = MIN_VALUE;
+    for (int i = 0; i < T; i++) {
+        const int ii = i * C;
+        const F kk = k[ii];
+        const F vv = v[ii];
+        const F yy = y[ii];
+
+        F ww = u + kk;
+        F p = max(pp, ww);
+        F e1 = exp(pp - p);
+        F e2 = exp(ww - p);
+        const F qq = gy[ii] / (e1 * bb + e2);
+        gw += (ga - gb * yy) * e1 * qq;
+        gu += (vv - yy) * e2 * qq;
+        q[i] = qq;
+        r[i] = ww - p;
+
+        ww = w + pp;
+        p = max(ww, kk);
+        e1 = exp(ww - p);
+        e2 = exp(kk - p);
+        ga = e1 * (aa + ga);
+        gb = e1 * (bb + gb);
+        aa = e1 * aa + e2 * vv;
+        bb = e1 * bb + e2;
+        pp = p;
+    }
+    const int _offsetBC = _b * C + _c;
+    _gw[_offsetBC] = gw * _w[_c]; // multiply by w because of w -> -exp(w) in python forward()
+    _gu[_offsetBC] = gu;
+
+    aa = 0, bb = 0, pp = MIN_VALUE;
+    for (int i = T - 1; i >= 0; i--) {
+        const int ii = i * C;
+        const F kk = k[ii];
+        const F vv = v[ii];
+        const F yy = y[ii];
+        const F qq = q[i];
+        const F rr = r[i];
+
+        F e1 = qq * exp(rr);
+        F e2 = exp(kk + pp);
+        gk[ii] = e1 * (vv - yy) + e2 * (aa * vv + bb);
+        gv[ii] = e1 + e2 * aa;
+
+        const F ww = w + pp;
+        const F www = rr - u - kk;
+        const F p = max(ww, www);
+        e1 = exp(ww - p);
+        e2 = qq * exp(www - p);
+        aa = e1 * aa + e2;
+        bb = e1 * bb - e2 * yy;
+        pp = p;
+    }
+}
+
+void cuda_forward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y) {
+    dim3 threadsPerBlock( min(C, 32) ); // requires --maxrregcount 60 for optimal performance
+    assert(B * C % threadsPerBlock.x == 0);
+    dim3 numBlocks(B * C / threadsPerBlock.x);
+    kernel_forward<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, y);
+}
+
+void cuda_backward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *gy, float *gw, float *gu, float *gk, float *gv) {
+    dim3 threadsPerBlock( min(C, 32) ); // requires --maxrregcount 60 for optimal performance
+    assert(B * C % threadsPerBlock.x == 0);
+    dim3 numBlocks(B * C / threadsPerBlock.x);
+    kernel_backward<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, y, gy, gw, gu, gk, gv);
+}
--- a/finetune/lora/cuda/wkv_cuda_bf16.cu
+++ b/finetune/lora/cuda/wkv_cuda_bf16.cu
@@ -0,0 +1,132 @@
+#include <stdio.h>
+#include <assert.h>
+#include "ATen/ATen.h"
+#define MIN_VALUE (-1e38)
+typedef at::BFloat16 bf16;
+
+__global__ void kernel_forward(const int B, const int T, const int C,
+                               const float *__restrict__ const _w, const bf16 *__restrict__ const _u, const bf16 *__restrict__ const _k, const bf16 *__restrict__ const _v,
+                               bf16 *__restrict__ const _y) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int _b = idx / C;
+    const int _c = idx % C;
+    const int _offset = _b * T * C + _c;
+
+    float u = float(_u[_c]);
+    float w = _w[_c];
+    const bf16 *__restrict__ const k = _k + _offset;
+    const bf16 *__restrict__ const v = _v + _offset;
+    bf16 *__restrict__ const y = _y + _offset;
+
+    // aa and bb are running sums divided by exp(pp) (to avoid overflow)
+    float aa = 0, bb = 0, pp = MIN_VALUE;
+    for (int i = 0; i < T; i++) {
+        const int ii = i * C;
+        const float kk = float(k[ii]);
+        const float vv = float(v[ii]);
+
+        float ww = u + kk;
+        float p = max(pp, ww);
+        float e1 = exp(pp - p);
+        float e2 = exp(ww - p);
+        y[ii] = bf16((e1 * aa + e2 * vv) / (e1 * bb + e2));
+        
+        ww = w + pp;
+        p = max(ww, kk);
+        e1 = exp(ww - p);
+        e2 = exp(kk - p);
+        aa = e1 * aa + e2 * vv;
+        bb = e1 * bb + e2;
+        pp = p;
+    }
+}
+
+__global__ void kernel_backward(const int B, const int T, const int C,
+                                const float *__restrict__ const _w, const bf16 *__restrict__ const _u, const bf16 *__restrict__ const _k, const bf16 *__restrict__ const _v,
+                                const bf16 *__restrict__ const _y, const bf16 *__restrict__ const _gy,
+                                bf16 *__restrict__ const _gw, bf16 *__restrict__ const _gu, bf16 *__restrict__ const _gk, bf16 *__restrict__ const _gv) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    const int _b = idx / C;
+    const int _c = idx % C;
+    const int _offset = _b * T * C + _c;
+
+    float u = float(_u[_c]);
+    float w = _w[_c];
+    const bf16 *__restrict__ const k = _k + _offset;
+    const bf16 *__restrict__ const v = _v + _offset;
+    const bf16 *__restrict__ const y = _y + _offset;
+    const bf16 *__restrict__ const gy = _gy + _offset;
+    bf16 *__restrict__ const gk = _gk + _offset;
+    bf16 *__restrict__ const gv = _gv + _offset;
+
+    float q[Tmax], r[Tmax];
+
+    float gw = 0, gu = 0, aa = 0, bb = 0, ga = 0, gb = 0, pp = MIN_VALUE;
+    for (int i = 0; i < T; i++) {
+        const int ii = i * C;
+        const float kk = float(k[ii]);
+        const float vv = float(v[ii]);
+        const float yy = float(y[ii]);
+
+        float ww = u + kk;
+        float p = max(pp, ww);
+        float e1 = exp(pp - p);
+        float e2 = exp(ww - p);
+        const float qq = float(gy[ii]) / (e1 * bb + e2);
+        gw += (ga - gb * yy) * e1 * qq;
+        gu += (vv - yy) * e2 * qq;
+        q[i] = qq;
+        r[i] = ww - p;
+
+        ww = w + pp;
+        p = max(ww, kk);
+        e1 = exp(ww - p);
+        e2 = exp(kk - p);
+        ga = e1 * (aa + ga);
+        gb = e1 * (bb + gb);
+        aa = e1 * aa + e2 * vv;
+        bb = e1 * bb + e2;
+        pp = p;
+    }
+    const int _offsetBC = _b * C + _c;
+    _gw[_offsetBC] = bf16(gw * _w[_c]); // multiply by w because of w -> -exp(w) in python forward()
+    _gu[_offsetBC] = bf16(gu);
+
+    aa = 0, bb = 0, pp = MIN_VALUE;
+    for (int i = T - 1; i >= 0; i--) {
+        const int ii = i * C;
+        const float kk = float(k[ii]);
+        const float vv = float(v[ii]);
+        const float yy = float(y[ii]);
+        const float qq = q[i];
+        const float rr = r[i];
+
+        float e1 = qq * exp(rr);
+        float e2 = exp(kk + pp);
+        gk[ii] = bf16(e1 * (vv - yy) + e2 * (aa * vv + bb));
+        gv[ii] = bf16(e1 + e2 * aa);
+
+        const float ww = w + pp;
+        const float www = rr - u - kk;
+        const float p = max(ww, www);
+        e1 = exp(ww - p);
+        e2 = qq * exp(www - p);
+        aa = e1 * aa + e2;
+        bb = e1 * bb - e2 * yy;
+        pp = p;
+    }
+}
+
+void cuda_forward(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y) {
+    dim3 threadsPerBlock( min(C, 32) ); // requires --maxrregcount 60 for optimal performance
+    assert(B * C % threadsPerBlock.x == 0);
+    dim3 numBlocks(B * C / threadsPerBlock.x);
+    kernel_forward<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, y);
+}
+
+void cuda_backward(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, bf16 *gy, bf16 *gw, bf16 *gu, bf16 *gk, bf16 *gv) {
+    dim3 threadsPerBlock( min(C, 32) ); // requires --maxrregcount 60 for optimal performance
+    assert(B * C % threadsPerBlock.x == 0);
+    dim3 numBlocks(B * C / threadsPerBlock.x);
+    kernel_backward<<<numBlocks, threadsPerBlock>>>(B, T, C, w, u, k, v, y, gy, gw, gu, gk, gv);
+}
--- a/finetune/lora/cuda/wkv_op.cpp
+++ b/finetune/lora/cuda/wkv_op.cpp
@@ -0,0 +1,21 @@
+#include <torch/extension.h>
+
+void cuda_forward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y);
+void cuda_backward(int B, int T, int C, float *w, float *u, float *k, float *v, float *y, float *gy, float *gw, float *gu, float *gk, float *gv);
+
+void forward(int64_t B, int64_t T, int64_t C, torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
+    cuda_forward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>());
+}
+void backward(int64_t B, int64_t T, int64_t C, torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y, torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
+    cuda_backward(B, T, C, w.data_ptr<float>(), u.data_ptr<float>(), k.data_ptr<float>(), v.data_ptr<float>(), y.data_ptr<float>(), gy.data_ptr<float>(), gw.data_ptr<float>(), gu.data_ptr<float>(), gk.data_ptr<float>(), gv.data_ptr<float>());
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward", &forward, "wkv forward");
+    m.def("backward", &backward, "wkv backward");
+}
+
+TORCH_LIBRARY(wkv, m) {
+    m.def("forward", forward);
+    m.def("backward", backward);
+}
--- a/finetune/lora/cuda/wkv_op_bf16.cpp
+++ b/finetune/lora/cuda/wkv_op_bf16.cpp
@@ -0,0 +1,25 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+typedef at::BFloat16 bf16;
+
+void cuda_forward(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y);
+void cuda_backward(int B, int T, int C, float *w, bf16 *u, bf16 *k, bf16 *v, bf16 *y, bf16 *gy, bf16 *gw, bf16 *gu, bf16 *gk, bf16 *gv);
+
+void forward(int64_t B, int64_t T, int64_t C, torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y) {
+    cuda_forward(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>());
+}
+void backward(int64_t B, int64_t T, int64_t C, torch::Tensor &w, torch::Tensor &u, torch::Tensor &k, torch::Tensor &v, torch::Tensor &y,
+    torch::Tensor &gy, torch::Tensor &gw, torch::Tensor &gu, torch::Tensor &gk, torch::Tensor &gv) {
+    cuda_backward(B, T, C, w.data_ptr<float>(), u.data_ptr<bf16>(), k.data_ptr<bf16>(), v.data_ptr<bf16>(), y.data_ptr<bf16>(),
+        gy.data_ptr<bf16>(), gw.data_ptr<bf16>(), gu.data_ptr<bf16>(), gk.data_ptr<bf16>(), gv.data_ptr<bf16>());
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward", &forward, "wkv forward");
+    m.def("backward", &backward, "wkv backward");
+}
+
+TORCH_LIBRARY(wkv, m) {
+    m.def("forward", forward);
+    m.def("backward", backward);
+}
--- a/finetune/lora/merge_lora.py
+++ b/finetune/lora/merge_lora.py
@@ -0,0 +1,68 @@
+from collections import OrderedDict
+import os
+import sys
+from typing import Dict
+import typing
+import torch
+
+try:
+    if "-h" in sys.argv or "--help" in sys.argv:
+        print(
+            f"Usage: python3 {sys.argv[0]} [--use-gpu] <lora_alpha> <base_model.pth> <lora_checkpoint.pth> <output.pth>"
+        )
+
+    if sys.argv[1] == "--use-gpu":
+        device = "cuda"
+        lora_alpha, base_model, lora, output = (
+            float(sys.argv[2]),
+            sys.argv[3],
+            sys.argv[4],
+            sys.argv[5],
+        )
+    else:
+        device = "cpu"
+        lora_alpha, base_model, lora, output = (
+            float(sys.argv[1]),
+            sys.argv[2],
+            sys.argv[3],
+            sys.argv[4],
+        )
+
+    with torch.no_grad():
+        w: Dict[str, torch.Tensor] = torch.load(base_model, map_location="cpu")
+        # merge LoRA-only slim checkpoint into the main weights
+        w_lora: Dict[str, torch.Tensor] = torch.load(lora, map_location="cpu")
+        for k in w_lora.keys():
+            w[k] = w_lora[k]
+        output_w: typing.OrderedDict[str, torch.Tensor] = OrderedDict()
+        # merge LoRA weights
+        keys = list(w.keys())
+        for k in keys:
+            if k.endswith(".weight"):
+                prefix = k[: -len(".weight")]
+                lora_A = prefix + ".lora_A"
+                lora_B = prefix + ".lora_B"
+                if lora_A in keys:
+                    assert lora_B in keys
+                    print(f"merging {lora_A} and {lora_B} into {k}")
+                    assert w[lora_B].shape[1] == w[lora_A].shape[0]
+                    lora_r = w[lora_B].shape[1]
+                    w[k] = w[k].to(device=device)
+                    w[lora_A] = w[lora_A].to(device=device)
+                    w[lora_B] = w[lora_B].to(device=device)
+                    w[k] += w[lora_B] @ w[lora_A] * (lora_alpha / lora_r)
+                    output_w[k] = w[k].to(device="cpu", copy=True)
+                    del w[k]
+                    del w[lora_A]
+                    del w[lora_B]
+                    continue
+
+            if "lora" not in k:
+                print(f"retaining {k}")
+                output_w[k] = w[k].clone()
+                del w[k]
+
+        torch.save(output_w, output)
+except Exception as e:
+    with open("error.txt", "w") as f:
+        f.write(str(e))
--- a/finetune/lora/src/init.py
+++ b/finetune/lora/src/init.py
--- a/finetune/lora/src/binidx.py
+++ b/finetune/lora/src/binidx.py
@@ -0,0 +1,269 @@
+from lib2to3.pgen2 import token
+import os
+import torch
+import numpy as np
+import shutil
+import struct
+from functools import lru_cache
+from itertools import accumulate
+
+def print_rank_0(*message):
+    pass
+    # """If distributed is initialized print only on rank 0."""
+    # if torch.distributed.is_initialized():
+    #     if torch.distributed.get_rank() == 0:
+    #         print(*message, flush=True)
+    # else:
+    #     print(*message, flush=True)
+
+def _warmup_mmap_file(path):
+    pass
+    # with open(path, "rb") as stream:
+    #     while stream.read(100 * 1024 * 1024):
+    #         pass
+
+dtypes = {
+    1: np.uint8,
+    2: np.int8,
+    3: np.int16,
+    4: np.int32,
+    5: np.int64,
+    6: float,
+    7: np.double,
+    8: np.uint16,
+}
+
+def code(dtype):
+    for k in dtypes.keys():
+        if dtypes[k] == dtype:
+            return k
+    raise ValueError(dtype)
+
+def index_file_path(prefix_path):
+    return prefix_path + ".idx"
+
+def data_file_path(prefix_path):
+    return prefix_path + ".bin"
+
+class MMapIndexedDataset(torch.utils.data.Dataset):
+    class Index(object):
+        _HDR_MAGIC = b"MMIDIDX\x00\x00"
+
+        @classmethod
+        def writer(cls, path, dtype):
+            class _Writer(object):
+                def __enter__(self):
+                    self._file = open(path, "wb")
+
+                    # Write Magic string so we can check the file format then opening it again.
+                    self._file.write(cls._HDR_MAGIC)
+                    # Write version number
+                    # Little endian unsigned 64 Bit integer
+                    self._file.write(struct.pack("<Q", 1))
+                    # Little endian unsigned 8 Bit integer
+                    self._file.write(struct.pack("<B", code(dtype)))
+
+                    return self
+
+                @staticmethod
+                def _get_pointers(sizes):
+                    dtype_size = dtype().itemsize
+                    address = 0
+                    pointers = []
+
+                    for size in sizes:
+                        pointers.append(address)
+                        address += size * dtype_size
+
+                    return pointers
+
+                def write(self, sizes, doc_idx):
+                    pointers = self._get_pointers(sizes)
+
+                    # Little endian unsigned 64 Bit integer
+                    self._file.write(struct.pack("<Q", len(sizes)))
+                    # Little endian unsigned 64 Bit integer
+                    self._file.write(struct.pack("<Q", len(doc_idx)))
+
+                    sizes = np.array(sizes, dtype=np.int32)
+                    self._file.write(sizes.tobytes(order="C"))
+                    del sizes
+
+                    pointers = np.array(pointers, dtype=np.int64)
+                    self._file.write(pointers.tobytes(order="C"))
+                    del pointers
+
+                    doc_idx = np.array(doc_idx, dtype=np.int64)
+                    self._file.write(doc_idx.tobytes(order="C"))
+
+                def __exit__(self, exc_type, exc_val, exc_tb):
+                    self._file.close()
+
+            return _Writer()
+        
+        def __init__(self, path, skip_warmup=False):
+            with open(path, "rb") as stream:
+                magic_test = stream.read(9)
+                assert self._HDR_MAGIC == magic_test, (
+                    "Index file doesn't match expected format. "
+                    "Make sure that --dataset-impl is configured properly."
+                )
+                # Little endian unsigned 64 Bit integer
+                version = struct.unpack("<Q", stream.read(8))
+                assert (1,) == version
+
+                # Little endian unsigned 8 Bit integer
+                (dtype_code,) = struct.unpack("<B", stream.read(1))
+                self._dtype = dtypes[dtype_code]
+                self._dtype_size = self._dtype().itemsize
+
+                self._len = struct.unpack("<Q", stream.read(8))[0]
+                self._doc_count = struct.unpack("<Q", stream.read(8))[0]
+                offset = stream.tell()
+
+            if not skip_warmup:
+                print_rank_0("    warming up index mmap file...")
+                _warmup_mmap_file(path)
+
+            self._bin_buffer_mmap = np.memmap(path, mode="r", order="C")
+            self._bin_buffer = memoryview(self._bin_buffer_mmap)
+            print_rank_0("    reading sizes...")
+            self._sizes = np.frombuffer(
+                self._bin_buffer, dtype=np.int32, count=self._len, offset=offset
+            )
+            print_rank_0("    reading pointers...")
+            self._pointers = np.frombuffer(
+                self._bin_buffer,
+                dtype=np.int64,
+                count=self._len,
+                offset=offset + self._sizes.nbytes,
+            )
+            print_rank_0("    reading document index...")
+            self._doc_idx = np.frombuffer(
+                self._bin_buffer,
+                dtype=np.int64,
+                count=self._doc_count,
+                offset=offset + self._sizes.nbytes + self._pointers.nbytes,
+            )
+
+        def __del__(self):
+            self._bin_buffer_mmap._mmap.close()
+            del self._bin_buffer_mmap
+
+        @property
+        def dtype(self):
+            return self._dtype
+
+        @property
+        def sizes(self):
+            return self._sizes
+
+        @property
+        def doc_idx(self):
+            return self._doc_idx
+
+        @lru_cache(maxsize=8)
+        def __getitem__(self, i):
+            return self._pointers[i], self._sizes[i]
+
+        def __len__(self):
+            return self._len
+
+    def __init__(self, path, skip_warmup=False):
+        super().__init__()
+
+        self._path = None
+        self._index = None
+        self._bin_buffer = None
+
+        self._do_init(path, skip_warmup)
+
+    def __getstate__(self):
+        return self._path
+
+    def __setstate__(self, state):
+        self._do_init(state)
+
+    def _do_init(self, path, skip_warmup):
+        self._path = path
+        self._index = self.Index(index_file_path(self._path), skip_warmup)
+
+        if not skip_warmup:
+            print_rank_0("    warming up data mmap file...")
+            _warmup_mmap_file(data_file_path(self._path))
+        print_rank_0("    creating numpy buffer of mmap...")
+        self._bin_buffer_mmap = np.memmap(
+            data_file_path(self._path), mode="r", order="C"
+        )
+        print_rank_0("    creating memory view of numpy buffer...")
+        self._bin_buffer = memoryview(self._bin_buffer_mmap)
+
+    def __del__(self):
+        self._bin_buffer_mmap._mmap.close()
+        del self._bin_buffer_mmap
+        del self._index
+
+    def __len__(self):
+        return len(self._index)
+
+    # @lru_cache(maxsize=8)
+    def __getitem__(self, idx):
+        if isinstance(idx, int):
+            ptr, size = self._index[idx]
+            np_array = np.frombuffer(
+                self._bin_buffer, dtype=self._index.dtype, count=size, offset=ptr
+            )
+            return np_array
+        elif isinstance(idx, slice):
+            start, stop, step = idx.indices(len(self))
+            if step != 1:
+                raise ValueError(
+                    "Slices into indexed_dataset must be contiguous")
+            ptr = self._index._pointers[start]
+            sizes = self._index._sizes[idx]
+            offsets = list(accumulate(sizes))
+            total_size = sum(sizes)
+            np_array = np.frombuffer(
+                self._bin_buffer, dtype=self._index.dtype, count=total_size, offset=ptr
+            )
+            sents = np.split(np_array, offsets[:-1])
+            return sents
+
+    def get(self, idx, offset=0, length=None):
+        """Retrieves a single item from the dataset with the option to only
+        return a portion of the item.
+
+        get(idx) is the same as [idx] but get() does not support slicing.
+        """
+        ptr, size = self._index[idx]
+        if length is None:
+            length = size - offset
+        ptr += offset * np.dtype(self._index.dtype).itemsize
+        np_array = np.frombuffer(
+            self._bin_buffer, dtype=self._index.dtype, count=length, offset=ptr
+        )
+        return np_array
+
+    @property
+    def sizes(self):
+        return self._index.sizes
+
+    @property
+    def doc_idx(self):
+        return self._index.doc_idx
+
+    def get_doc_idx(self):
+        return self._index._doc_idx
+
+    def set_doc_idx(self, doc_idx_):
+        self._index._doc_idx = doc_idx_
+
+    @property
+    def supports_prefetch(self):
+        return False
+
+    @staticmethod
+    def exists(path):
+        return os.path.exists(index_file_path(path)) and os.path.exists(
+            data_file_path(path)
+        )
--- a/finetune/lora/src/dataset.py
+++ b/finetune/lora/src/dataset.py
@@ -0,0 +1,224 @@
+########################################################################################################
+# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
+########################################################################################################
+
+import json, math, random, os, sys
+import numpy as np
+import torch
+from torch.utils.data import Dataset
+from pytorch_lightning.utilities import rank_zero_info
+from .binidx import MMapIndexedDataset
+from .utils import MaybeIsPrime
+
+
+class MyDataset(Dataset):
+    def __init__(self, args):
+        self.args = args
+
+        if args.data_type == "binidx":
+            self.vocab_size = args.vocab_size
+            rank_zero_info(f"Current vocab size = {self.vocab_size} (make sure it's correct)")
+
+            if args.data_file.endswith('/'):
+                d_all = []
+                for p in os.listdir(args.data_file):
+                    if p.endswith(".idx"):
+                        d_all += [p[:-4]]
+                d_all.sort()
+                rank_zero_info(d_all)
+                exit(0)
+            else:
+                self.data = MMapIndexedDataset(args.data_file)
+                self.data_size = len(self.data._bin_buffer) // self.data._index._dtype_size
+                rank_zero_info(f"Data has {self.data_size} tokens.")
+
+            if args.my_qa_mask > 0:
+                self.data_pile = MMapIndexedDataset('/fsx/BlinkDL/pile/pile_20B_tokenizer_text_document')
+                self.data_pile_size = len(self.data_pile._bin_buffer) // self.data._index._dtype_size
+
+            if args.my_pile_stage > 0:
+                # assert self.data_size == 332115325534 and self.vocab_size == 50277
+                self.samples_per_epoch = args.epoch_steps * args.real_bsz
+                assert self.samples_per_epoch == 40320
+                rank_zero_info(f"########## Pile 20b-tokenized stage {args.my_pile_stage} ##########")
+                dataset_slot = self.data_size // args.ctx_len
+                if args.my_pile_stage != 4:
+                    assert MaybeIsPrime(args.magic_prime)
+                    assert args.magic_prime % 3 == 2
+                    assert args.magic_prime / dataset_slot > 0.99 and args.magic_prime / dataset_slot <= 1
+        elif args.data_type == "numpy":
+            self.data = np.load(args.data_file).astype("int")
+            self.vocab_size = args.vocab_size
+            rank_zero_info("Current vocab size =", self.vocab_size, "(make sure it's correct)")
+            self.data_size = len(self.data)
+            rank_zero_info(f"Data has {self.data_size} tokens.")
+        elif args.data_type == "uint16":
+            self.data = np.fromfile(args.data_file, dtype=np.uint16).astype("int32").reshape(-1, args.my_sample_len)
+            self.vocab_size = args.vocab_size
+            rank_zero_info("Current vocab size =", self.vocab_size, "(make sure it's correct)")
+            self.data_size = self.data.shape[0]
+            rank_zero_info(f"Data has {self.data_size} samples.")
+        elif args.data_type == "wds_img":
+            self.vocab_size = -1
+            self.data_size = -1
+            self.data = None
+            self.error_count = 0
+        else:
+            if args.data_type == "dummy":
+                rank_zero_info("Building dummy data...")
+                self.data = ""
+                for i in range(100000):
+                    aa = (i) % 10000
+                    bb = (i * i) % 10000
+                    cc = aa + bb
+                    self.data += f".{aa}+{bb}={cc}."
+            else:
+                self.data = open(args.data_file, "r", encoding=args.data_type).read()
+            rank_zero_info("Building token list...")
+            unique = sorted(list(set(self.data)))
+            self.vocab_size = len(unique)
+            # rank_zero_info()
+            # for u in unique:
+            #     print(u, end=' ')
+            # rank_zero_info('\n\n')
+            xx = 0
+            xxObj = {}
+            for u in unique:
+                xxObj[xx] = u
+                xx += 1
+            with open(f"{args.proj_dir}/vocab.json", "w", encoding="utf-16le") as vocab_file:
+                vocab_file.write(json.dumps(xxObj, ensure_ascii=False))
+            self.data_size = len(self.data)
+            rank_zero_info(f"Data has {self.data_size} tokens, {self.vocab_size} vocab size.")
+            self.stoi = {ch: i for i, ch in enumerate(unique)}
+            self.itos = {i: ch for i, ch in enumerate(unique)}
+
+    def __len__(self):
+        return self.args.epoch_steps * self.args.micro_bsz
+
+    def __getitem__(self, idx):
+        args = self.args
+        rank = self.global_rank
+        epoch = self.real_epoch
+        world_size = self.world_size
+        # print(f"epoch {epoch} idx {idx} rank {rank}/{world_size}")
+
+        if args.data_type == "wds_img":
+            def init_wds(self, bias=0):
+                def identity(x):
+                    return x            
+                import webdataset as wds
+                import torchvision.transforms as transforms
+                # img_transform = transforms.Compose(
+                #     [transforms.CenterCrop(256)]
+                # )
+                img_transform = transforms.Compose([
+                    transforms.CenterCrop(512),
+                    transforms.Resize((args.my_img_size))
+                ])
+                self.data_raw = wds.WebDataset(args.data_file, resampled=True).shuffle(10000, initial=1000, rng=random.Random(epoch*100000+rank+bias*1e9)).decode("torchrgb").to_tuple("jpg", "json", "txt").map_tuple(img_transform, identity, identity)
+                for pp in self.data_raw.pipeline:
+                    if 'Resampled' in str(pp):
+                        pp.deterministic = True
+                        def worker_seed():
+                            return rank*100000+epoch+bias*1e9
+                        pp.worker_seed = worker_seed
+                self.data = iter(self.data_raw)
+                # print(f"WebDataset loaded for rank {rank} epoch {epoch}")
+            if self.data == None:
+                init_wds(self)
+            trial = 0
+            while trial < 10:
+                try:
+                    dd = next(self.data) # jpg, json, txt
+                    break
+                except:
+                    print(f'[dataloader error - epoch {epoch} rank {rank} - trying a new shuffle]')
+                    self.error_count += 1
+                    init_wds(self, self.error_count)
+                    trial += 1
+                    pass
+            # print(f"epoch {epoch} idx {idx} rank {rank}/{world_size} {dd[2]}")
+            # with open(f"sample_{rank}.txt", "a", encoding="utf-8") as tmp:
+            #     tmp.write(f"epoch {epoch} idx {idx} rank {rank}/{world_size} {int(dd[1]['key'])}\n")
+            return dd[0], dd[2]
+        else:
+            if args.data_type == "uint16":
+                i = np.random.randint(0, self.data_size-1)
+                dix = self.data[i]
+                x = torch.tensor(dix[:-1], dtype=torch.long)
+                y = torch.tensor(dix[1:], dtype=torch.long)
+            else:
+                ctx_len = args.ctx_len
+                req_len = ctx_len + 1
+                magic_prime = args.magic_prime
+                data = self.data
+
+                if args.my_pile_stage > 0 and args.my_pile_stage != 4:
+                    ii = 1 + epoch * self.samples_per_epoch + (idx * world_size) + rank
+
+                    if args.my_qa_mask > 0:
+                        ii_orig = ii
+                        if ii % 2 == 0:
+                            ii = (ii // 2) * args.magic_prime
+                            if args.ctx_len == 1024:
+                                magic_prime = 324331313
+                            elif args.ctx_len == 2048:
+                                magic_prime = 162165671
+                            elif args.ctx_len == 4096:
+                                magic_prime = 81082817
+                            data = self.data_pile
+                        else:
+                            ii = ii // 2
+
+                    factor = (math.sqrt(5) - 1) / 2
+                    factor = int(magic_prime * factor)
+                    i = ((factor * ii * ii * ii) % magic_prime) * ctx_len
+                    if (args.my_qa_mask == 0) or (data == self.data_pile):
+                        i = i + args.my_pile_shift
+                    # print(f"epoch {epoch} idx {idx} rank {rank}/{world_size} ii {ii} pos {round(i / self.data_size, 3)}")
+                else:
+                    # cheat: pick a random spot in dataset
+                    i = np.random.randint(0, self.data_size - req_len)
+
+                if args.data_type == "binidx":
+                    dix = data.get(idx=0, offset=i, length=req_len).astype(int)
+                elif args.data_type == "numpy":
+                    dix = data[i : i + req_len]
+                else:
+                    dix = [self.stoi[s] for s in data[i : i + req_len]]
+
+                if args.my_qa_mask == 1:
+                    if data == self.data_pile:
+                        z = [1] * ctx_len
+                    else:
+                        z = [0] * ctx_len
+                        z_sum = 0
+                        isGood = False
+                        for i in range(3, ctx_len):
+                            if dix[i] == 27 and dix[i-1] == 34 and dix[i-2] == 187 and dix[i-3] == 187:
+                                isGood = True
+                            if dix[i] == 0:
+                                isGood = False
+                            if isGood:
+                                z[i] = 1
+                                z_sum += 1
+                        if z_sum == 0:
+                            z = [1] * ctx_len
+                            i = np.random.randint(0, self.data_pile_size - req_len)
+                            dix = self.data_pile.get(idx=0, offset=i, length=req_len).astype(int)
+                    z = torch.tensor(z, dtype=torch.bfloat16)
+
+                x = torch.tensor(dix[:-1], dtype=torch.long)
+                y = torch.tensor(dix[1:], dtype=torch.long)
+
+                # if ii_orig < 50:
+                #     # if rank == 1:
+                #     print('rank', rank, 'i', ii_orig, ii, i, 'x', x[:5], '...', x[-5:])
+                # else:
+                #     exit(0)
+
+                if args.my_qa_mask == 1:
+                    return x, y, z
+
+            return x, y
--- a/finetune/lora/src/model.py
+++ b/finetune/lora/src/model.py
@@ -0,0 +1,678 @@
+########################################################################################################
+# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
+########################################################################################################
+
+import functools
+import os, math, gc, importlib
+import torch
+# torch._C._jit_set_profiling_executor(True)
+# torch._C._jit_set_profiling_mode(True)
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint as torch_checkpoint
+from torch.nn import functional as F
+import pytorch_lightning as pl
+from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
+from pytorch_lightning.strategies import DeepSpeedStrategy
+if importlib.util.find_spec('deepspeed'):
+    import deepspeed
+    from deepspeed.ops.adam import DeepSpeedCPUAdam, FusedAdam
+
+# from deepspeed.runtime.fp16.onebit.zoadam import ZeroOneAdam
+
+LORA_CONFIG = {
+    "r": 0,
+    "alpha": 0,
+    "dropout": 0,
+    "parts": {"att", "ln", "time"},
+}
+
+
+try:
+    print('RWKV_MY_TESTING', os.environ["RWKV_MY_TESTING"])
+except:
+    os.environ["RWKV_MY_TESTING"] = ''
+
+def __nop(ob):
+    return ob
+
+
+MyModule = nn.Module
+MyFunction = __nop
+if os.environ["RWKV_JIT_ON"] == "1":
+    MyModule = torch.jit.ScriptModule
+    MyFunction = torch.jit.script_method
+
+
+########################################################################################################
+# CUDA Kernel
+########################################################################################################
+
+T_MAX = int(os.environ["RWKV_T_MAX"])  # TAKES LOTS OF VRAM!
+# it's possible to go beyond CUDA limitations if you slice the ctx and pass the hidden state in each slice
+
+from torch.utils.cpp_extension import load
+
+if os.environ["RWKV_FLOAT_MODE"] == "bf16":
+    wkv_cuda = load(name=f"wkv_{T_MAX}_bf16", sources=["finetune/lora/cuda/wkv_op_bf16.cpp", "finetune/lora/cuda/wkv_cuda_bf16.cu"], verbose=True, extra_cuda_cflags=["-t 4", "-std=c++17", "-res-usage", "--maxrregcount 60", "--use_fast_math", "-O3", "-Xptxas -O3", "--extra-device-vectorization", f"-DTmax={T_MAX}"])
+    class WKV(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, B, T, C, w, u, k, v):
+            ctx.B = B
+            ctx.T = T
+            ctx.C = C
+            assert T <= T_MAX
+            assert B * C % min(C, 32) == 0
+            w = -torch.exp(w.float().contiguous())
+            u = u.contiguous()
+            k = k.contiguous()
+            v = v.contiguous()
+            y = torch.empty((B, T, C), device=w.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
+            wkv_cuda.forward(B, T, C, w, u, k, v, y)
+            ctx.save_for_backward(w, u, k, v, y)
+            return y
+        @staticmethod
+        def backward(ctx, gy):
+            B = ctx.B
+            T = ctx.T
+            C = ctx.C
+            assert T <= T_MAX
+            assert B * C % min(C, 32) == 0
+            w, u, k, v, y = ctx.saved_tensors
+            gw = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
+            gu = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
+            gk = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
+            gv = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format, dtype=torch.bfloat16)
+            wkv_cuda.backward(B, T, C, w, u, k, v, y, gy.contiguous(), gw, gu, gk, gv)
+            gw = torch.sum(gw, dim=0)
+            gu = torch.sum(gu, dim=0)
+            return (None, None, None, gw, gu, gk, gv)
+else:
+    wkv_cuda = load(name=f"wkv_{T_MAX}", sources=["finetune/lora/cuda/wkv_op.cpp", "finetune/lora/cuda/wkv_cuda.cu"], verbose=True, extra_cuda_cflags=["-res-usage", "--maxrregcount 60", "--use_fast_math", "-O3", "-Xptxas -O3", "--extra-device-vectorization", f"-DTmax={T_MAX}"])
+    class WKV(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, B, T, C, w, u, k, v):
+            ctx.B = B
+            ctx.T = T
+            ctx.C = C
+            assert T <= T_MAX
+            assert B * C % min(C, 32) == 0
+            if "32" in os.environ["RWKV_FLOAT_MODE"]:
+                w = -torch.exp(w.contiguous())
+                u = u.contiguous()
+                k = k.contiguous()
+                v = v.contiguous()
+            else:
+                w = -torch.exp(w.float().contiguous())
+                u = u.float().contiguous()
+                k = k.float().contiguous()
+                v = v.float().contiguous()
+            y = torch.empty((B, T, C), device=w.device, memory_format=torch.contiguous_format)
+            wkv_cuda.forward(B, T, C, w, u, k, v, y)
+            ctx.save_for_backward(w, u, k, v, y)
+            if "32" in os.environ["RWKV_FLOAT_MODE"]:
+                return y
+            elif os.environ["RWKV_FLOAT_MODE"] == "fp16":
+                return y.half()
+            elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
+                return y.bfloat16()
+        @staticmethod
+        def backward(ctx, gy):
+            B = ctx.B
+            T = ctx.T
+            C = ctx.C
+            assert T <= T_MAX
+            assert B * C % min(C, 32) == 0
+            w, u, k, v, y = ctx.saved_tensors
+            gw = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format)
+            gu = torch.empty((B, C), device=gy.device, memory_format=torch.contiguous_format)
+            gk = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format)
+            gv = torch.empty((B, T, C), device=gy.device, memory_format=torch.contiguous_format)
+            if "32" in os.environ["RWKV_FLOAT_MODE"]:
+                wkv_cuda.backward(B, T, C, w, u, k, v, y, gy.contiguous(), gw, gu, gk, gv)
+            else:
+                wkv_cuda.backward(B, T, C, w, u, k, v, y, gy.float().contiguous(), gw, gu, gk, gv)
+            gw = torch.sum(gw, dim=0)
+            gu = torch.sum(gu, dim=0)
+            if "32" in os.environ["RWKV_FLOAT_MODE"]:
+                return (None, None, None, gw, gu, gk, gv)
+            elif os.environ["RWKV_FLOAT_MODE"] == "fp16":
+                return (None, None, None, gw.half(), gu.half(), gk.half(), gv.half())
+            elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
+                return (None, None, None, gw.bfloat16(), gu.bfloat16(), gk.bfloat16(), gv.bfloat16())
+
+
+def RUN_CUDA(B, T, C, w, u, k, v):
+    return WKV.apply(B, T, C, w, u, k, v)
+
+
+########################################################################################################
+# LoRA
+########################################################################################################
+
+
+class LoraLinear(nn.Module):
+
+    def __init__(self, in_features: int, out_features: int, bias: bool):
+        super().__init__()
+
+        self.weight = nn.Parameter(torch.empty((out_features, in_features)))
+        assert bias == False, "Biased LoraLinear not supported"
+
+        r, alpha, dropout = LORA_CONFIG["r"], LORA_CONFIG[
+            "alpha"], LORA_CONFIG["dropout"]
+        self.lora_A = nn.Parameter(torch.empty(r, in_features))
+        self.lora_B = nn.Parameter(torch.empty(out_features, r))
+        self.lora_dropout = nn.Dropout(dropout)
+        self.scaling = alpha / r
+
+        nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+        nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
+        nn.init.zeros_(self.lora_B)
+
+    def forward(self, x):
+        return (
+            F.linear(x, self.weight) + self.scaling *
+            F.linear(F.linear(self.lora_dropout(x), self.lora_A), self.lora_B))
+
+
+@functools.wraps(LoraLinear)
+def make_linear_att(*args, **kwargs):
+    if "att" in LORA_CONFIG["parts"] and LORA_CONFIG["r"] > 0:
+        return LoraLinear(*args, **kwargs)
+    else:
+        return nn.Linear(*args, **kwargs)
+
+
+@functools.wraps(LoraLinear)
+def make_linear_ffn(*args, **kwargs):
+    if "ffn" in LORA_CONFIG["parts"] and LORA_CONFIG["r"] > 0:
+        return LoraLinear(*args, **kwargs)
+    else:
+        return nn.Linear(*args, **kwargs)
+
+
+########################################################################################################
+# RWKV: RWKV Time-mix + RWKV Channel-mix
+########################################################################################################
+
+
+class RWKV_TimeMix(MyModule):
+    def __init__(self, args, layer_id):
+        super().__init__()
+        self.args = args
+        self.layer_id = layer_id
+        self.ctx_len = args.ctx_len
+        self.n_embd = args.n_embd
+
+        with torch.no_grad():  # fancy init
+            ratio_0_to_1 = layer_id / (args.n_layer - 1)  # 0 to 1
+            ratio_1_to_almost0 = 1.0 - (layer_id / args.n_layer)  # 1 to ~0
+            ddd = torch.ones(1, 1, args.n_embd)
+            for i in range(args.n_embd):
+                ddd[0, 0, i] = i / args.n_embd
+
+            # fancy time_decay
+            decay_speed = torch.ones(args.dim_att)
+            for h in range(args.dim_att):
+                decay_speed[h] = -5 + 8 * (h / (args.dim_att - 1)) ** (0.7 + 1.3 * ratio_0_to_1)
+            self.time_decay = nn.Parameter(decay_speed)
+            # print(layer_id, self.time_decay.flatten()[:3].cpu().numpy(), '...', self.time_decay.flatten()[-3:].cpu().numpy())
+
+            # fancy time_first
+            zigzag = torch.tensor([(i + 1) % 3 - 1 for i in range(args.dim_att)]) * 0.5
+            self.time_first = nn.Parameter(torch.ones(args.dim_att) * math.log(0.3) + zigzag)
+
+            # fancy time_mix
+            self.time_mix_k = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
+            self.time_mix_v = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0) + 0.3 * ratio_0_to_1)
+            self.time_mix_r = nn.Parameter(torch.pow(ddd, 0.5 * ratio_1_to_almost0))
+
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+
+        self.key = make_linear_att(args.n_embd, args.dim_att, bias=False)
+        self.value = make_linear_att(args.n_embd, args.dim_att, bias=False)
+        self.receptance = make_linear_att(args.n_embd, args.dim_att, bias=False)
+
+        self.output = nn.Linear(args.dim_att, args.n_embd, bias=False)
+
+        if 'a' in os.environ["RWKV_MY_TESTING"]:
+            self.register_buffer("att_mask", torch.tril(torch.ones(args.ctx_len, args.ctx_len)))
+            d_qkv = args.n_embd // 16
+            self.qq = nn.Linear(args.n_embd, d_qkv, bias=False)
+            self.kk = nn.Linear(args.n_embd, d_qkv, bias=False)
+            self.vv = nn.Linear(args.n_embd, d_qkv, bias=False)
+            self.oo = nn.Linear(d_qkv, args.n_embd, bias=False)
+            with torch.no_grad():
+                self.time_mix_qq = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
+                self.time_mix_kk = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
+                self.time_mix_vv = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0) + 0.3 * ratio_0_to_1)
+
+    if 'a' not in os.environ["RWKV_MY_TESTING"]:
+        @MyFunction
+        def jit_func(self, x):
+            xx = self.time_shift(x) # Mix x with the previous timestep to produce xk, xv, xr
+            xk = x * self.time_mix_k + xx * (1 - self.time_mix_k)
+            xv = x * self.time_mix_v + xx * (1 - self.time_mix_v)
+            xr = x * self.time_mix_r + xx * (1 - self.time_mix_r)
+            k = self.key(xk)
+            v = self.value(xv)
+            r = self.receptance(xr)
+            sr = torch.sigmoid(r)
+            return sr, k, v
+
+        def forward(self, x):
+            B, T, C = x.size()  # x = (Batch,Time,Channel)
+            sr, k, v = self.jit_func(x)
+            rwkv = sr * RUN_CUDA(B, T, self.args.dim_att, self.time_decay, self.time_first, k, v)
+            return self.output(rwkv)
+
+    if 'a' in os.environ["RWKV_MY_TESTING"]:
+        @MyFunction
+        def QKV(self, q, k, v):
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = att.masked_fill(self.att_mask == 0, float('-inf'))
+            att = F.softmax(att, dim = -1)
+            x = att @ v
+            return x
+
+        @MyFunction
+        def jit_funcQKV(self, x):
+            xx = self.time_shift(x) # Mix x with the previous timestep to produce xk, xv, xr
+            xk = x * self.time_mix_k + xx * (1 - self.time_mix_k)
+            xv = x * self.time_mix_v + xx * (1 - self.time_mix_v)
+            xr = x * self.time_mix_r + xx * (1 - self.time_mix_r)
+            xqq = x * self.time_mix_qq + xx * (1 - self.time_mix_qq)
+            xkk = x * self.time_mix_kk + xx * (1 - self.time_mix_kk)
+            xvv = x * self.time_mix_vv + xx * (1 - self.time_mix_vv)
+            k = self.key(xk)
+            v = self.value(xv)
+            r = self.receptance(xr)
+            sr = torch.sigmoid(r)
+            qq = self.qq(xqq)
+            kk = self.kk(xkk)
+            vv = self.vv(xvv)
+            return sr, k, v, qq, kk, vv
+
+        def forward(self, x):
+            B, T, C = x.size()  # x = (Batch,Time,Channel)
+            sr, k, v, qq, kk, vv = self.jit_funcQKV(x)
+            rwkv = sr * RUN_CUDA(B, T, self.args.dim_att, self.time_decay, self.time_first, k, v)
+            rwkv = self.output(rwkv) + self.oo(self.QKV(qq, kk, vv))
+            return rwkv
+
+########################################################################################################
+
+class RWKV_ChannelMix(MyModule):
+    def __init__(self, args, layer_id):
+        super().__init__()
+        self.args = args
+        self.layer_id = layer_id
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+
+        with torch.no_grad():  # fancy init of time_mix
+            ratio_1_to_almost0 = 1.0 - (layer_id / args.n_layer)  # 1 to ~0
+            ddd = torch.ones(1, 1, args.n_embd)
+            for i in range(args.n_embd):
+                ddd[0, 0, i] = i / args.n_embd
+            self.time_mix_k = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
+            self.time_mix_r = nn.Parameter(torch.pow(ddd, ratio_1_to_almost0))
+
+        self.key = make_linear_ffn(args.n_embd, args.dim_ffn, bias=False)
+        self.receptance = make_linear_ffn(args.n_embd, args.n_embd, bias=False)
+        self.value = make_linear_ffn(args.dim_ffn, args.n_embd, bias=False)
+
+    @MyFunction
+    def forward(self, x):
+        xx = self.time_shift(x)
+        xk = x * self.time_mix_k + xx * (1 - self.time_mix_k)
+        xr = x * self.time_mix_r + xx * (1 - self.time_mix_r)
+        k = self.key(xk)
+        k = torch.square(torch.relu(k))
+        kv = self.value(k)
+        return torch.sigmoid(self.receptance(xr)) * kv
+
+class MishGLU(MyModule):
+    def __init__(self, args, layer_id):
+        super().__init__()
+        self.args = args
+        self.layer_id = layer_id
+        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
+
+        with torch.no_grad():
+            ratio_1_to_almost0 = 1.0 - (layer_id / args.n_layer)
+
+            x = torch.ones(1, 1, args.n_embd)
+            for i in range(args.n_embd):
+                x[0, 0, i] = i / args.n_embd
+
+            self.time_mix_k = nn.Parameter(torch.pow(x, ratio_1_to_almost0))
+            self.time_mix_r = nn.Parameter(torch.pow(x, ratio_1_to_almost0))
+            self.aa = nn.Linear(args.n_embd, args.dim_ffn, bias=False)
+            self.bb = nn.Linear(args.n_embd, args.dim_ffn, bias=False)
+            self.value = nn.Linear(args.dim_ffn, args.n_embd, bias=False)
+
+    @MyFunction
+    def forward(self, x):
+        xx = self.time_shift(x)
+        xa = x * self.time_mix_k + xx * (1 - self.time_mix_k)
+        xb = x * self.time_mix_r + xx * (1 - self.time_mix_r)
+        a = self.aa(xa)
+        b = self.bb(xb)
+        return self.value(a * F.mish(b))
+
+########################################################################################################
+# The RWKV Model with our blocks
+########################################################################################################
+
+
+class Block(nn.Module):
+    def __init__(self, args, layer_id):
+        super().__init__()
+        self.args = args
+        self.layer_id = layer_id
+
+        self.ln1 = nn.LayerNorm(args.n_embd)
+        self.ln2 = nn.LayerNorm(args.n_embd)
+
+        if self.layer_id == 0:
+            self.ln0 = nn.LayerNorm(args.n_embd)
+            if args.my_pos_emb > 0:
+                self.pos_emb_x = nn.Parameter(torch.zeros((1,args.my_pos_emb,args.n_embd)))
+                self.pos_emb_y = nn.Parameter(torch.zeros((args.my_pos_emb,1,args.n_embd)))
+
+        if self.layer_id == 0 and self.args.pre_ffn > 0:
+            self.ffnPre = RWKV_ChannelMix(args, 0)
+        else:
+            self.att = RWKV_TimeMix(args, layer_id)
+
+        if 'g' in os.environ["RWKV_MY_TESTING"]:
+            self.ffn = MishGLU(args, layer_id)
+        else:
+            self.ffn = RWKV_ChannelMix(args, layer_id)
+
+        if args.tiny_att_dim > 0 and self.layer_id == args.tiny_att_layer:
+            self.tiny_ln = nn.LayerNorm(args.n_embd)
+            self.tiny_q = nn.Linear(args.n_embd, args.tiny_att_dim, bias=False)
+            self.tiny_k = nn.Linear(args.n_embd, args.tiny_att_dim, bias=False)
+            self.tiny_v = nn.Linear(args.n_embd, args.n_embd, bias=False)
+            self.register_buffer("tiny_mask", torch.tril(torch.ones(args.ctx_len, args.ctx_len)))
+
+    def forward(self, x, x_emb=None):
+        args = self.args
+        B, T, C = x.size()
+        if self.layer_id == 0:
+            x = self.ln0(x)
+            if args.my_pos_emb > 0:
+                pos_emb = (self.pos_emb_x + self.pos_emb_y).reshape(T+1, -1)[:-1,:]
+                x = x + pos_emb
+
+        if self.layer_id == 0 and args.pre_ffn > 0:
+            x = x + self.ffnPre(self.ln1(x))
+        else:
+            x = x + self.att(self.ln1(x))
+        x = x + self.ffn(self.ln2(x))
+
+        if args.tiny_att_dim > 0 and self.layer_id == args.tiny_att_layer:
+            xx = self.tiny_ln(x)
+            q = self.tiny_q(xx)[:, :T, :]
+            k = self.tiny_k(xx)[:, :T, :]
+            c = (q @ k.transpose(-2, -1)) * (args.tiny_att_dim ** (-0.5))
+            c = c.masked_fill(self.tiny_mask[:T, :T] == 0, 0)
+            x = x + c @ self.tiny_v(x_emb)
+        return x
+
+
+class L2Wrap(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, loss, y):
+        ctx.save_for_backward(y)
+        return loss
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        y = ctx.saved_tensors[0]
+        # to encourage the logits to be close to 0
+        factor = 1e-4 / (y.shape[0] * y.shape[1])
+        maxx, ids = torch.max(y, -1, keepdim=True)
+        gy = torch.zeros_like(y)
+        gy.scatter_(-1, ids, maxx * factor)
+        return (grad_output, gy)
+
+
+class RWKV(pl.LightningModule):
+    def __init__(self, args):
+        super().__init__()
+        self.args = args
+        if not hasattr(args, 'dim_att'):
+            args.dim_att = args.n_embd
+        if not hasattr(args, 'dim_ffn'):
+            args.dim_ffn = args.n_embd * 4
+        if not hasattr(args, 'tiny_att_layer'):
+            args.tiny_att_layer = -1
+        if not hasattr(args, 'tiny_att_dim'):
+            args.tiny_att_dim = -1
+
+        self.emb = nn.Embedding(args.vocab_size, args.n_embd)
+
+        self.blocks = nn.ModuleList([Block(args, i) for i in range(args.n_layer)])
+
+        self.ln_out = nn.LayerNorm(args.n_embd)
+        self.head = nn.Linear(args.n_embd, args.vocab_size, bias=False)
+
+        if args.head_qk > 0:
+            self.head_q = nn.Linear(args.n_embd, args.head_qk, bias=False)
+            self.head_k = nn.Linear(args.n_embd, args.head_qk, bias=False)
+            self.register_buffer("copy_mask", torch.tril(torch.ones(args.ctx_len, args.ctx_len)))
+
+    def configure_optimizers(self):
+        args = self.args
+        if args.layerwise_lr > 0:
+            lr_1x = set()
+            lr_2x = set()
+            lr_3x = set()
+            for n, p in self.named_parameters():
+                if "time_mix" in n:
+                    if args.my_pile_stage == 2:
+                        lr_2x.add(n)
+                    else:
+                        lr_1x.add(n)
+                elif "time_decay" in n:
+                    if args.my_pile_stage == 2:
+                        lr_3x.add(n)
+                    else:
+                        lr_2x.add(n)
+                elif "time_first" in n:
+                    lr_3x.add(n)
+                else:
+                    lr_1x.add(n)
+            lr_1x = sorted(list(lr_1x))
+            lr_2x = sorted(list(lr_2x))
+            lr_3x = sorted(list(lr_3x))
+            # print('1x', lr_1x)
+            # print('2x', lr_2x)
+            # print('3x', lr_3x)
+            param_dict = {n: p for n, p in self.named_parameters()}
+            if args.my_pile_stage == 2:
+                optim_groups = [
+                    {"params": [param_dict[n] for n in lr_1x], "weight_decay": 0.0, "my_lr_scale": 1.0},
+                    {"params": [param_dict[n] for n in lr_2x], "weight_decay": 0.0, "my_lr_scale": 5.0},# test: 2e-3 / args.lr_init},
+                    {"params": [param_dict[n] for n in lr_3x], "weight_decay": 0.0, "my_lr_scale": 5.0},# test: 3e-3 / args.lr_init},
+                ]
+            else:
+                optim_groups = [
+                    {"params": [param_dict[n] for n in lr_1x], "weight_decay": 0.0, "my_lr_scale": 1.0},
+                    {"params": [param_dict[n] for n in lr_2x], "weight_decay": 0.0, "my_lr_scale": 2.0},
+                    {"params": [param_dict[n] for n in lr_3x], "weight_decay": 0.0, "my_lr_scale": 3.0},
+                ]
+        else:
+            optim_groups = [
+                {"params": [p for n, p in self.named_parameters()], "weight_decay": 0.0},
+            ]
+
+        for g in optim_groups:
+            g["params"] = [p for p in g["params"] if p.requires_grad]
+        optim_groups = [g for g in optim_groups if len(g["params"]) > 0]
+
+        if self.deepspeed_offload:
+            return DeepSpeedCPUAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, adamw_mode=False, weight_decay=0, amsgrad=False)
+        return FusedAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, adam_w_mode=False, weight_decay=0, amsgrad=False)
+        # return ZeroOneAdam(optim_groups, lr=self.args.lr_init, betas=self.args.betas, eps=self.args.adam_eps, bias_correction=True, weight_decay=0, amsgrad=False, cuda_aware=False)
+
+    @property
+    def deepspeed_offload(self) -> bool:
+        strategy = self.trainer.strategy
+        if isinstance(strategy, DeepSpeedStrategy):
+            cfg = strategy.config["zero_optimization"]
+            return cfg.get("offload_optimizer") or cfg.get("offload_param")
+        return False
+
+    def forward(self, idx):
+        args = self.args
+        B, T = idx.size()
+        assert T <= args.ctx_len, "Cannot forward, model ctx_len is exhausted."
+
+        x = self.emb(idx)
+        x_emb = x
+
+        if args.tiny_att_dim > 0:
+            for block in self.blocks:
+                if args.grad_cp == 1:
+                    if args.lora:
+                        x = torch_checkpoint(block, x, x_emb, use_reentrant=False)
+                    else:
+                        x = deepspeed.checkpointing.checkpoint(block, x, x_emb)
+                else:
+                    x = block(x, x_emb)
+        else:
+            for block in self.blocks:
+                if args.grad_cp == 1:
+                    if args.lora:
+                        x = torch_checkpoint(block, x, x_emb, use_reentrant=False)
+                    else:
+                        x = deepspeed.checkpointing.checkpoint(block, x)
+                else:
+                    x = block(x)
+
+        x = self.ln_out(x)
+
+        if args.head_qk > 0:
+            q = self.head_q(x)[:, :T, :]
+            k = self.head_k(x)[:, :T, :]
+            c = (q @ k.transpose(-2, -1)) * (1.0 / args.head_qk)
+            c = c.masked_fill(self.copy_mask[:T, :T] == 0, 0)
+
+            if "32" in os.environ["RWKV_FLOAT_MODE"]:
+                c = c @ F.one_hot(idx, num_classes=args.vocab_size)
+            elif os.environ["RWKV_FLOAT_MODE"] == "fp16":
+                c = c @ F.one_hot(idx, num_classes=args.vocab_size).half()
+            elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
+                c = c @ F.one_hot(idx, num_classes=args.vocab_size).bfloat16()
+
+            x = self.head(x) + c
+        else:
+            x = self.head(x)
+
+        return x
+
+    def training_step(self, batch, batch_idx):
+        args = self.args
+        if args.my_qa_mask != 1:
+            idx, targets = batch
+            logits = self(idx)
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        else:
+            idx, targets, mask = batch
+            mask = mask.view(-1)
+            sum_mask = torch.sum(mask).item()
+            # if sum_mask == 0:
+            #     return torch.tensor([0.0], requires_grad=True)
+
+            logits = self(idx)
+            if sum_mask == mask.shape[0]:
+                loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+                # print('rank', self.global_rank, 'loss', loss.item())
+            else:
+                loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), reduction='none')
+                # loss_raw = loss
+                loss = torch.sum(loss * mask) / sum_mask
+
+                # torch.set_printoptions(threshold=10000)
+                # if True: #self.global_rank == 1:
+                #     tmp = ''
+                #     sss = 0
+                #     ccc = 0
+                #     for i in range(mask.shape[0]):
+                #         if mask[i] > 0:
+                #             tmp += str(idx.view(-1)[i].item()) + ','
+                #             sss += loss_raw.view(-1)[i].float().item()
+                #             ccc += 1
+                #     print('rank', self.global_rank, 'loss', loss.item(), 'lavg', sss / ccc)#, 'tmp', tmp, 'input', idx)
+
+        return L2Wrap.apply(loss, logits)
+
+    def training_step_end(self, batch_parts):
+        all = self.all_gather(batch_parts)
+        if self.trainer.is_global_zero:
+            self.trainer.my_loss_all = all
+
+    def generate_init_weight(self):
+        print(
+            f"""
+############################################################################
+#
+# Init model weight (slow for large models)...
+#
+############################################################################
+"""
+        )
+        m = {}
+        for n in self.state_dict():
+            p = self.state_dict()[n]
+            shape = p.shape
+
+            gain = 1.0
+            scale = 1.0
+            if "ln_" in n or ".ln" in n or "time_" in n or "_mask" in n or "pos_emb" in n or '.mask.' in n:
+                m[n] = p
+            else:
+                if n == "emb.weight":
+                    scale = -1 * self.args.lr_init
+                else:
+                    if shape[0] > shape[1]:
+                        gain = math.sqrt(shape[0] / shape[1])
+                    for kk in [".att.key.", ".att.receptance.", ".att.output.", ".att.key.", ".ffn.value.", ".ffn.receptance.", ".ffnPre.value.", ".ffnPre.receptance.", "head_q.", '.oo.', '.rr.']:
+                        if kk in n:
+                            scale = 0
+                    if n == "head.weight":
+                        scale = 0.5
+                    if "head_k." in n:
+                        scale = 0.1
+                    if "head_q." in n:
+                        scale = 0
+
+                print(f"{str(shape[0]).ljust(5)} {str(shape[1]).ljust(5)} {str(scale).ljust(4)} {n}")
+
+                if self.args.accelerator.upper() == "GPU":
+                    m[n] = torch.empty((shape[0], shape[1]), device="cuda")
+                else:
+                    m[n] = torch.empty((shape[0], shape[1]))
+
+                if scale == 0:
+                    nn.init.zeros_(m[n])
+                elif scale < 0:
+                    nn.init.uniform_(m[n], a=scale, b=-scale)
+                else:
+                    nn.init.orthogonal_(m[n], gain=gain * scale)
+
+            m[n] = m[n].cpu()
+            if os.environ["RWKV_FLOAT_MODE"] == "fp16":
+                m[n] = m[n].half()
+            elif os.environ["RWKV_FLOAT_MODE"] == "bf16":
+                m[n] = m[n].bfloat16()
+
+            # if n == "emb.weight":
+            #     print(m[n])
+
+        gc.collect()
+        torch.cuda.empty_cache()
+        return m
--- a/finetune/lora/src/trainer.py
+++ b/finetune/lora/src/trainer.py
@@ -0,0 +1,203 @@
+import os, math, time, datetime, subprocess
+import torch
+from torch.utils.data import DataLoader
+import pytorch_lightning as pl
+from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
+from .model import LORA_CONFIG
+
+def my_save(dd, ff):
+    if '14b-run1' not in ff:
+        torch.save(dd, ff)
+    else:
+        fn = ff.split('/')[-1]
+        fff = '/dev/shm/' + fn
+        torch.save(dd, fff)
+        subprocess.Popen(f" aws s3 mv {fff} s3://rwkv-14b-4k/{fn} --quiet", shell=True)
+
+class train_callback(pl.Callback):
+    def __init__(self, args):
+        super().__init__()
+        self.args = args
+
+    def on_train_batch_start(self, trainer, pl_module, batch, batch_idx):
+        args = self.args
+        # if args.cuda_cleanup > 0:
+        #     torch.cuda.empty_cache()
+        real_step = trainer.global_step + args.epoch_begin * args.epoch_steps
+
+        # LR schedule
+        w_step = args.warmup_steps
+        if args.lr_final == args.lr_init or args.epoch_count == 0:
+            lr = args.lr_init
+        else:
+            decay_step = real_step - args.my_pile_edecay * args.epoch_steps
+            decay_total = (args.epoch_count - args.my_pile_edecay) * args.epoch_steps
+            progress = (decay_step - w_step + 1) / (decay_total - w_step)
+            progress = min(1, max(0, progress))
+
+            if args.lr_final == 0 or args.lr_init == 0:  # linear decay
+                lr = args.lr_init + (args.lr_final - args.lr_init) * progress
+            else:  # exp decay
+                lr = args.lr_init * math.exp(math.log(args.lr_final / args.lr_init) * pow(progress, 1))
+
+            if trainer.global_step < w_step:
+                lr = lr * (0.2 + 0.8 * trainer.global_step / w_step)
+            # if trainer.is_global_zero:
+            #     print(trainer.global_step, decay_step, decay_total, w_step, progress, lr)
+
+        for param_group in trainer.optimizers[0].param_groups:
+            if args.layerwise_lr > 0:
+                param_group["lr"] = lr * param_group["my_lr_scale"]
+                # print(param_group["lr"], param_group["my_lr_scale"])
+            else:
+                param_group["lr"] = lr
+
+        trainer.my_lr = lr
+        # rank_zero_info(f"{real_step} {lr}")
+
+        if trainer.global_step == 0:
+            if trainer.is_global_zero:  # logging
+                trainer.my_loss_sum = 0
+                trainer.my_loss_count = 0
+                trainer.my_log = open(args.proj_dir + "/train_log.txt", "a")
+                trainer.my_log.write(f"NEW RUN {args.my_timestamp}\n{vars(self.args)}\n")
+                try:
+                    print(f"\n{trainer.strategy.config}\n")
+                    trainer.my_log.write(f"{trainer.strategy.config}\n")
+                except:
+                    pass
+                trainer.my_log.flush()
+                if len(args.wandb) > 0:
+                    print("Login to wandb...")
+                    import wandb
+                    wandb.init(
+                        project=args.wandb,
+                        name=args.run_name + " " + args.my_timestamp,
+                        config=args,
+                        save_code=False,
+                    )
+                    trainer.my_wandb = wandb
+
+    def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx):
+        args = self.args
+        if trainer.is_global_zero:  # logging
+            t_now = time.time_ns()
+            token_per_step = args.ctx_len * args.real_bsz
+            real_step = trainer.global_step + args.epoch_begin * args.epoch_steps
+            kt_s = 0
+            try:
+                t_cost = (t_now - trainer.my_time_ns) / 1e9
+                kt_s = token_per_step / t_cost / 1000
+                self.log("REAL it/s", 1.0 / t_cost, prog_bar=True, on_step=True)
+                self.log("Kt/s", kt_s, prog_bar=True, on_step=True)
+            except:
+                pass
+            trainer.my_time_ns = t_now
+            trainer.my_loss = trainer.my_loss_all.float().mean().item()
+            trainer.my_loss_sum += trainer.my_loss
+            trainer.my_loss_count += 1
+            trainer.my_epoch_loss = trainer.my_loss_sum / trainer.my_loss_count
+            self.log("lr", trainer.my_lr, prog_bar=True, on_step=True)
+            self.log("loss", trainer.my_epoch_loss, prog_bar=True, on_step=True)
+            # self.log("s", real_step, prog_bar=True, on_step=True)
+
+            if len(args.wandb) > 0:
+                lll = {"loss": trainer.my_loss, "lr": trainer.my_lr, "Gtokens": real_step * token_per_step / 1e9}
+                if kt_s > 0:
+                    lll["kt/s"] = kt_s
+                trainer.my_wandb.log(lll, step=int(real_step))
+            if args.magic_prime > 0:
+                expand_factor = 2 if args.my_qa_mask > 0 else 1
+                if int(real_step) == int(args.magic_prime * expand_factor // args.real_bsz) - 1:
+                    to_save_dict = pl_module.state_dict()
+                    my_save(
+                        to_save_dict,
+                        f"{args.proj_dir}/rwkv-final.pth",
+                    )
+
+
+    def on_train_epoch_start(self, trainer, pl_module):
+        args = self.args
+        dataset = trainer.train_dataloader.dataset.datasets
+        assert "MyDataset" in str(dataset)
+        dataset.global_rank = trainer.global_rank
+        dataset.real_epoch = int(args.epoch_begin + trainer.current_epoch)
+        dataset.world_size = trainer.world_size
+        # print(f'########## world_size {dataset.world_size} global_rank {dataset.global_rank} real_epoch {dataset.real_epoch} ##########')
+
+    def on_train_epoch_end(self, trainer, pl_module):
+        args = self.args
+        if trainer.is_global_zero:  # logging & save state_dict
+            if (args.epoch_save > 0 and trainer.current_epoch % args.epoch_save == 0) or trainer.current_epoch == args.epoch_count - 1:
+                if args.data_type == 'wds_img':
+                    raw_dict = pl_module.state_dict()
+                    to_save_dict = {}
+                    for k in raw_dict:
+                        if k.startswith('encoder.') or k.startswith('decoder.'):
+                            to_save_dict[k] = raw_dict[k]
+                else:
+                    to_save_dict = pl_module.state_dict()
+
+                if args.lora:
+                    enable_time_finetune = 'time' in LORA_CONFIG["parts"]
+                    enable_ln_finetune = 'ln' in LORA_CONFIG["parts"]
+                    lora_dict = {}
+                    for name, state in to_save_dict.items():
+                        if ('.lora_' in name
+                                or (enable_time_finetune and '.time_' in name)
+                                or (enable_ln_finetune and '.ln' in name)):
+                            lora_dict[name] = state
+                    to_save_dict = lora_dict
+
+                try:
+                    my_save(
+                        to_save_dict,
+                        f"{args.proj_dir}/rwkv-{args.epoch_begin + trainer.current_epoch}.pth",
+                    )
+                except Exception as e:
+                    print('Error\n\n', e, '\n\n')
+            trainer.my_log.write(f"{args.epoch_begin + trainer.current_epoch} {trainer.my_epoch_loss:.6f} {math.exp(trainer.my_epoch_loss):.4f} {trainer.my_lr:.8f} {datetime.datetime.now()} {trainer.current_epoch}\n")
+            trainer.my_log.flush()
+
+            trainer.my_loss_sum = 0
+            trainer.my_loss_count = 0
+
+
+@rank_zero_only
+def generate_init_weight(model, init_weight_name):
+    mm = model.generate_init_weight()
+
+    if model.args.my_pile_stage == 1:
+        if len(model.args.load_model) > 0:
+            print(f"Combine weights from {model.args.load_model}...")
+            load_dict = torch.load(model.args.load_model, map_location="cpu")
+            for k in load_dict:
+                assert k in mm
+                src = load_dict[k]
+                try:
+                    mm[k] = src.reshape(mm[k].shape)
+                except:
+                    tmp = mm[k].squeeze().clone()
+                    print(k, src.shape, '-->', mm[k].shape)
+                    ss = src.shape[0]
+                    dd = tmp.shape[0]
+                    for i in range(dd):
+                        pos = i / dd * ss
+                        if pos >= ss - 1:
+                            tmp[i] = src[ss-1]
+                        else:
+                            p0 = int(math.floor(pos))
+                            ii = pos - p0
+                            tmp[i] = src[p0] * (1-ii) + src[p0+1] * (ii)
+                    mm[k] = tmp.reshape(mm[k].shape)
+                    sss = src.squeeze().float().cpu().numpy()
+                    print(sss[:10], '...', sss[-10:])
+                    mmm = mm[k].squeeze().float().cpu().numpy()
+                    print(mmm[:10], '...', mmm[-10:])
+
+    print(f"Save to {init_weight_name}...")
+    torch.save(mm, init_weight_name)
+
+    if model.args.my_pile_stage == 1:
+        print("Done. Now go for stage 2.")
+        exit(0)
--- a/finetune/lora/src/utils.py
+++ b/finetune/lora/src/utils.py
@@ -0,0 +1,130 @@
+import json, time, random, os
+import numpy as np
+import torch
+from torch.nn import functional as F
+
+time_slot = {}
+time_ref = time.time_ns()
+
+def record_time(name):
+    if name not in time_slot:
+        time_slot[name] = 1e20
+    tt = (time.time_ns() - time_ref) / 1e9
+    if tt < time_slot[name]:
+        time_slot[name] = tt
+
+class TOKENIZER():
+    def __init__(self, WORD_NAME, UNKNOWN_CHAR='\ue083'):
+        if 'list' in str(type(WORD_NAME)):
+            self.charMode = False
+            if WORD_NAME[0] == WORD_NAME[1]:
+                from transformers import PreTrainedTokenizerFast
+                self.tokenizer = PreTrainedTokenizerFast(tokenizer_file=WORD_NAME[0])
+            else:
+                from transformers import GPT2TokenizerFast
+                self.tokenizer = GPT2TokenizerFast(WORD_NAME[0], WORD_NAME[1])
+            self.vocab_size = len(self.tokenizer)
+        else:
+            self.charMode = True
+            with open(WORD_NAME + '.json', "r", encoding="utf-16") as result_file:
+                self.word_table = json.load(result_file)
+
+            self.vocab_size = len(self.word_table)
+
+            self.stoi = {v: int(k) for k, v in self.word_table.items()}
+            self.itos = {int(k): v for k, v in self.word_table.items()}
+
+            self.UNKNOWN_CHAR = self.stoi[UNKNOWN_CHAR]
+
+    def refine_context(self, context):
+        context = context.strip().split('\n')
+        for c in range(len(context)):
+            context[c] = context[c].strip().strip('\u3000').strip('\r')
+        context = list(filter(lambda c: c != '', context))
+        context = '\n' + ('\n'.join(context)).strip()
+        if context == '':
+            context = '\n'
+        return context
+
+    def sample_logits(self, out, x, ctx_len, temperature=1.0, top_p_usual=None, top_p_newline=None):
+        # out[self.UNKNOWN_CHAR] = -float('Inf')
+        lastChar = int(x[-1])
+
+        probs = F.softmax(out, dim=-1)
+
+        if self.charMode:
+            if self.itos[lastChar] == '\n':
+                top_p = top_p_newline
+            else:
+                top_p = top_p_usual
+        else:
+            top_p = top_p_usual
+
+        if os.environ["RWKV_RUN_DEVICE"] == "cpu":
+            probs = probs.numpy()
+            sorted_probs = np.sort(probs)[::-1]
+            cumulative_probs = np.cumsum(sorted_probs)
+            cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
+            probs[probs < cutoff] = 0
+            if temperature != 1.0:
+                probs = probs.pow(1.0 / temperature)
+            probs = probs / np.sum(probs)
+            out = np.random.choice(a=len(probs), p=probs)
+            return out
+        else:
+            sorted_probs = torch.sort(probs, descending=True)[0]
+            cumulative_probs = torch.cumsum(sorted_probs, dim=-1).cpu().numpy()
+            cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
+            probs[probs < cutoff] = 0
+            if temperature != 1.0:
+                probs = probs.pow(1.0 / temperature)
+            out = torch.multinomial(probs, num_samples=1)[0]
+            return out
+
+def MaybeIsPrime(number):
+    if FermatPrimalityTest(number) and MillerRabinPrimalityTest(number):
+        return True
+    else:
+        return False
+
+
+def FermatPrimalityTest(number):
+    if number > 1:
+        for time in range(3):
+            randomNumber = random.randint(2, number) - 1
+            if pow(randomNumber, number - 1, number) != 1:
+                return False
+        return True
+    else:
+        return False
+
+
+def MillerRabinPrimalityTest(number):
+    if number == 2:
+        return True
+    elif number == 1 or number % 2 == 0:
+        return False
+    oddPartOfNumber = number - 1
+    timesTwoDividNumber = 0
+    while oddPartOfNumber % 2 == 0:
+        oddPartOfNumber = oddPartOfNumber // 2
+        timesTwoDividNumber = timesTwoDividNumber + 1
+
+    for time in range(3):
+        while True:
+            randomNumber = random.randint(2, number) - 1
+            if randomNumber != 0 and randomNumber != 1:
+                break
+
+        randomNumberWithPower = pow(randomNumber, oddPartOfNumber, number)
+
+        if (randomNumberWithPower != 1) and (randomNumberWithPower != number - 1):
+            iterationNumber = 1
+
+            while (iterationNumber <= timesTwoDividNumber - 1) and (randomNumberWithPower != number - 1):
+                randomNumberWithPower = pow(randomNumberWithPower, 2, number)
+                iterationNumber = iterationNumber + 1
+            if randomNumberWithPower != (number - 1):
+                return False
+
+    return True
--- a/finetune/lora/train.py
+++ b/finetune/lora/train.py
@@ -0,0 +1,479 @@
+########################################################################################################
+# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
+########################################################################################################
+
+if __name__ == "__main__":
+    from argparse import ArgumentParser
+    from pytorch_lightning import Trainer
+    from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
+
+    rank_zero_info("########## work in progress ##########")
+
+    ########################################################################################################
+    #
+    # example: train a simple L12-D768 RWKV on dummy data
+    #
+    # python train.py --load_model "" --wandb "" --proj_dir "out" \
+    # --data_file "" --data_type "dummy" --vocab_size 0 \
+    # --ctx_len 128 --epoch_steps 1000 --epoch_count 20 --epoch_begin 0 --epoch_save 10 \
+    # --micro_bsz 16 --n_layer 12 --n_embd 768 --pre_ffn 0 --head_qk 0 \
+    # --lr_init 6e-4 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.99 --adam_eps 1e-8 \
+    # --accelerator gpu --devices 1 --precision bf16 --strategy ddp_find_unused_parameters_false --grad_cp 0
+
+    # example: train a simple L6-D512 RWKV from scratch on enwik8
+    #
+    # python train.py --load_model "" --wandb "" --proj_dir "out" \
+    # --data_file "../data/enwik8" --data_type "utf-8" --vocab_size 0 \
+    # --ctx_len 512 --epoch_steps 5000 --epoch_count 500 --epoch_begin 0 --epoch_save 5 \
+    # --micro_bsz 12 --n_layer 6 --n_embd 512 --pre_ffn 0 --head_qk 0 \
+    # --lr_init 8e-4 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.99 --adam_eps 1e-8 \
+    # --accelerator gpu --devices 1 --precision bf16 --strategy ddp_find_unused_parameters_false --grad_cp 0
+
+    # example: fine-tune RWKV 1.5B using 8xA100 40G = 1.76it/s = 115k token/s, VRAM 37477M
+    #
+    # python train.py --load_model "/fsx/BlinkDL/CODE/FP16/out_1b2/all-8040.pth" --wandb "" --proj_dir "out" \
+    # --data_file "../data/train.npy" --data_type "numpy" --vocab_size 50277 \
+    # --ctx_len 1024 --epoch_steps 1000 --epoch_count 1000 --epoch_begin 0 --epoch_save 5 \
+    # --micro_bsz 8 --n_layer 24 --n_embd 2048 --pre_ffn 0 --head_qk 0 \
+    # --lr_init 1e-5 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.999 --adam_eps 1e-8 \
+    # --accelerator gpu --devices 8 --precision bf16 --strategy deepspeed_stage_2 --grad_cp 0
+
+    # example: fine-tune RWKV 1.5B using 1 GPU fp16 (VRAM 16G) NOTE: fp16 might overflow
+    #
+    # python train.py --load_model "/fsx/BlinkDL/CODE/FP16/out_1b2/all-8040.pth" --wandb "" --proj_dir "out" \
+    # --data_file "../data/train.npy" --data_type "numpy" --vocab_size 50277 \
+    # --ctx_len 1024 --epoch_steps 200 --epoch_count 1000 --epoch_begin 0 --epoch_save 1 \
+    # --micro_bsz 11 --n_layer 24 --n_embd 2048 --pre_ffn 0 --head_qk 0 \
+    # --lr_init 1e-5 --lr_final 1e-5 --warmup_steps 0 --beta1 0.9 --beta2 0.999 --adam_eps 1e-8 \
+    # --accelerator gpu --devices 1 --precision fp16 --strategy deepspeed_stage_2_offload --grad_cp 1
+
+    parser = ArgumentParser()
+
+    parser.add_argument("--load_model", default="", type=str)  # full path, with .pth
+    parser.add_argument(
+        "--wandb", default="", type=str
+    )  # wandb project name. if "" then don't use wandb
+    parser.add_argument("--proj_dir", default="out", type=str)
+    parser.add_argument("--random_seed", default="-1", type=int)
+
+    parser.add_argument("--data_file", default="", type=str)
+    parser.add_argument("--data_type", default="utf-8", type=str)
+    parser.add_argument(
+        "--vocab_size", default=0, type=int
+    )  # vocab_size = 0 means auto (for char-level LM and .txt data)
+
+    parser.add_argument("--ctx_len", default=1024, type=int)
+    parser.add_argument(
+        "--epoch_steps", default=1000, type=int
+    )  # a mini "epoch" has [epoch_steps] steps
+    parser.add_argument(
+        "--epoch_count", default=500, type=int
+    )  # train for this many "epochs". will continue afterwards with lr = lr_final
+    parser.add_argument(
+        "--epoch_begin", default=0, type=int
+    )  # if you load a model trained for x "epochs", set epoch_begin = x
+    parser.add_argument(
+        "--epoch_save", default=5, type=int
+    )  # save the model every [epoch_save] "epochs"
+
+    parser.add_argument(
+        "--micro_bsz", default=12, type=int
+    )  # micro batch size (batch size per GPU)
+    parser.add_argument("--n_layer", default=6, type=int)
+    parser.add_argument("--n_embd", default=512, type=int)
+    parser.add_argument("--dim_att", default=0, type=int)
+    parser.add_argument("--dim_ffn", default=0, type=int)
+    parser.add_argument(
+        "--pre_ffn", default=0, type=int
+    )  # replace first att layer by ffn (sometimes better)
+    parser.add_argument("--head_qk", default=0, type=int)  # my headQK trick
+    parser.add_argument("--tiny_att_dim", default=0, type=int)  # tiny attention dim
+    parser.add_argument(
+        "--tiny_att_layer", default=-999, type=int
+    )  # tiny attention @ which layer
+
+    parser.add_argument(
+        "--lr_init", default=6e-4, type=float
+    )  # 6e-4 for L12-D768, 4e-4 for L24-D1024, 3e-4 for L24-D2048
+    parser.add_argument("--lr_final", default=1e-5, type=float)
+    parser.add_argument(
+        "--warmup_steps", default=0, type=int
+    )  # try 50 if you load a model
+    parser.add_argument("--beta1", default=0.9, type=float)
+    parser.add_argument(
+        "--beta2", default=0.99, type=float
+    )  # use 0.999 when your model is close to convergence
+    parser.add_argument("--adam_eps", default=1e-8, type=float)
+
+    parser.add_argument(
+        "--grad_cp", default=0, type=int
+    )  # gradient checkpt: saves VRAM, but slower
+    parser.add_argument("--my_pile_stage", default=0, type=int)  # my special pile mode
+    parser.add_argument(
+        "--my_pile_shift", default=-1, type=int
+    )  # my special pile mode - text shift
+    parser.add_argument("--my_pile_edecay", default=0, type=int)
+    parser.add_argument(
+        "--layerwise_lr", default=1, type=int
+    )  # layerwise lr for faster convergence (but slower it/s)
+    parser.add_argument(
+        "--ds_bucket_mb", default=200, type=int
+    )  # deepspeed bucket size in MB. 200 seems enough
+    # parser.add_argument("--cuda_cleanup", default=0, type=int)  # extra cuda cleanup (sometimes helpful)
+
+    parser.add_argument("--my_img_version", default=0, type=str)
+    parser.add_argument("--my_img_size", default=0, type=int)
+    parser.add_argument("--my_img_bit", default=0, type=int)
+    parser.add_argument("--my_img_clip", default="x", type=str)
+    parser.add_argument("--my_img_clip_scale", default=1, type=float)
+    parser.add_argument("--my_img_l1_scale", default=0, type=float)
+    parser.add_argument("--my_img_encoder", default="x", type=str)
+    # parser.add_argument("--my_img_noise_scale", default=0, type=float)
+    parser.add_argument("--my_sample_len", default=0, type=int)
+    parser.add_argument("--my_ffn_shift", default=1, type=int)
+    parser.add_argument("--my_att_shift", default=1, type=int)
+    parser.add_argument("--my_pos_emb", default=0, type=int)
+    parser.add_argument("--load_partial", default=0, type=int)
+    parser.add_argument("--magic_prime", default=0, type=int)
+    parser.add_argument("--my_qa_mask", default=0, type=int)
+    parser.add_argument("--my_testing", default="", type=str)
+
+    parser.add_argument("--lora", action="store_true")
+    parser.add_argument("--lora_load", default="", type=str)
+    parser.add_argument("--lora_r", default=8, type=int)
+    parser.add_argument("--lora_alpha", default=32, type=float)
+    parser.add_argument("--lora_dropout", default=0.01, type=float)
+    parser.add_argument("--lora_parts", default="att,ln,time", type=str)
+
+    parser = Trainer.add_argparse_args(parser)
+    args = parser.parse_args()
+
+    ########################################################################################################
+
+    import os, warnings, math, datetime, sys, time, importlib
+    import numpy as np
+    import torch
+    from torch.utils.data import DataLoader
+
+    if "deepspeed" in args.strategy:
+        import deepspeed
+    import pytorch_lightning as pl
+    from pytorch_lightning import seed_everything
+
+    if args.random_seed >= 0:
+        print(
+            f"########## WARNING: GLOBAL SEED {args.random_seed} THIS WILL AFFECT MULTIGPU SAMPLING ##########\n"
+            * 3
+        )
+        seed_everything(args.random_seed)
+
+    np.set_printoptions(precision=4, suppress=True, linewidth=200)
+    warnings.filterwarnings(
+        "ignore", ".*Consider increasing the value of the `num_workers` argument*"
+    )
+    warnings.filterwarnings(
+        "ignore", ".*The progress bar already tracks a metric with the*"
+    )
+    # os.environ["WDS_SHOW_SEED"] = "1"
+
+    args.my_timestamp = datetime.datetime.today().strftime("%Y-%m-%d-%H-%M-%S")
+    args.enable_checkpointing = False
+    args.replace_sampler_ddp = False
+    args.logger = False
+    args.gradient_clip_val = 1.0
+    args.num_sanity_val_steps = 0
+    args.check_val_every_n_epoch = int(1e20)
+    args.log_every_n_steps = int(1e20)
+    args.max_epochs = args.epoch_count  # continue forever
+    args.betas = (args.beta1, args.beta2)
+    args.real_bsz = int(args.num_nodes) * int(args.devices) * args.micro_bsz
+    os.environ["RWKV_T_MAX"] = str(args.ctx_len)
+    os.environ["RWKV_MY_TESTING"] = args.my_testing
+    if args.dim_att <= 0:
+        args.dim_att = args.n_embd
+    if args.dim_ffn <= 0:
+        args.dim_ffn = args.n_embd * 4
+
+    if args.data_type == "wds_img":
+        args.run_name = f"v{args.my_img_version}-{args.my_img_size}-{args.my_img_bit}bit-{args.my_img_clip}x{args.my_img_clip_scale}"
+        args.proj_dir = f"{args.proj_dir}-{args.run_name}"
+    else:
+        args.run_name = (
+            f"{args.vocab_size} ctx{args.ctx_len} L{args.n_layer} D{args.n_embd}"
+        )
+    if not os.path.exists(args.proj_dir):
+        os.makedirs(args.proj_dir)
+
+    if args.my_pile_stage > 0:
+        magic_prime_bak = args.magic_prime
+        if args.ctx_len == 1024:
+            args.magic_prime = 324331313
+            args.epoch_count = 8043
+        elif args.ctx_len == 2048:
+            args.magic_prime = 162165671
+            args.epoch_count = 4021
+        elif args.ctx_len == 4096:
+            args.magic_prime = 81082817
+            args.epoch_count = 2010
+        if args.my_pile_shift < 0:
+            if args.ctx_len == 1024:
+                args.my_pile_shift = 0
+            elif args.ctx_len == 2048:
+                args.my_pile_shift = 512
+            elif args.ctx_len == 4096:
+                args.my_pile_shift = 768
+
+        if magic_prime_bak > 0:
+            args.magic_prime = magic_prime_bak
+
+        args.epoch_steps = 40320 // args.real_bsz
+        assert args.epoch_steps * args.real_bsz == 40320
+        if args.my_pile_stage == 2:
+            assert args.lr_final == args.lr_init
+        if args.my_pile_stage >= 2:  # find latest saved model
+            list_p = []
+            for p in os.listdir(args.proj_dir):
+                if p.startswith("rwkv") and p.endswith(".pth"):
+                    p = ((p.split("-"))[1].split("."))[0]
+                    if p == "init":
+                        p = -1
+                    else:
+                        p = int(p)
+                    list_p += [p]
+            list_p.sort()
+            max_p = list_p[-1]
+            if len(list_p) > 1:
+                args.my_pile_prev_p = list_p[-2]  # in case max_p is corrupted
+            if max_p == -1:
+                args.load_model = f"{args.proj_dir}/rwkv-init.pth"
+            else:
+                args.load_model = f"{args.proj_dir}/rwkv-{max_p}.pth"
+                if args.my_pile_stage == 2:
+                    args.warmup_steps = 10
+                else:
+                    args.warmup_steps = 30
+            args.epoch_begin = max_p + 1
+
+    samples_per_epoch = args.epoch_steps * args.real_bsz
+    tokens_per_epoch = samples_per_epoch * args.ctx_len
+    rank_zero_info(
+        f"""
+############################################################################
+#
+# RWKV-4 {args.precision.upper()} on {args.num_nodes}x{args.devices} {args.accelerator.upper()}, bsz {args.num_nodes}x{args.devices}x{args.micro_bsz}={args.real_bsz}, {args.strategy} {'with grad_cp' if args.grad_cp > 0 else ''}
+#
+# Data = {args.data_file} ({args.data_type}), ProjDir = {args.proj_dir}
+#
+# Epoch = {args.epoch_begin} to {args.epoch_begin + args.epoch_count - 1} (will continue afterwards), save every {args.epoch_save} epoch
+#
+# Each "epoch" = {args.epoch_steps} steps, {samples_per_epoch} samples, {tokens_per_epoch} tokens
+#
+# Model = {args.n_layer} n_layer, {args.n_embd} n_embd, {args.ctx_len} ctx_len
+# LoRA = {f'enabled, {args.lora_r} r, {args.lora_alpha} alpha, {args.lora_dropout} dropout, on {args.lora_parts}' if args.lora else 'disabled'}
+#
+# Adam = lr {args.lr_init} to {args.lr_final}, warmup {args.warmup_steps} steps, beta {args.betas}, eps {args.adam_eps}
+#
+# Found torch {torch.__version__}, recommend 1.13.1+cu117 or newer
+# Found deepspeed {deepspeed.__version__ if importlib.util.find_spec('deepspeed') else 'None'}, recommend 0.7.0 (faster than newer versions)
+# Found pytorch_lightning {pl.__version__}, recommend 1.9.1 or newer
+#
+############################################################################
+"""
+    )
+    rank_zero_info(str(vars(args)) + "\n")
+
+    assert args.data_type in [
+        "utf-8",
+        "utf-16le",
+        "numpy",
+        "binidx",
+        "dummy",
+        "wds_img",
+        "uint16",
+    ]
+
+    if args.lr_final == 0 or args.lr_init == 0:
+        rank_zero_info(
+            "\n\nNote: lr_final = 0 or lr_init = 0. Using linear LR schedule instead.\n\n"
+        )
+
+    assert args.precision in ["fp32", "tf32", "fp16", "bf16"]
+    os.environ["RWKV_FLOAT_MODE"] = args.precision
+    if args.precision == "fp32":
+        for i in range(10):
+            rank_zero_info(
+                "\n\nNote: you are using fp32 (very slow). Try bf16 / tf32 for faster training.\n\n"
+            )
+    if args.precision == "fp16":
+        rank_zero_info(
+            "\n\nNote: you are using fp16 (might overflow). Try bf16 / tf32 for stable training.\n\n"
+        )
+
+    os.environ["RWKV_JIT_ON"] = "1"
+    if "deepspeed_stage_3" in args.strategy:
+        os.environ["RWKV_JIT_ON"] = "0"
+    if args.lora and args.grad_cp == 1:
+        print(
+            "!!!!! LoRA Warning: Gradient Checkpointing requires JIT off, disabling it"
+        )
+        os.environ["RWKV_JIT_ON"] = "0"
+
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cudnn.enabled = True
+    if args.precision == "fp32":
+        torch.backends.cudnn.allow_tf32 = False
+        torch.backends.cuda.matmul.allow_tf32 = False
+    else:
+        torch.backends.cudnn.allow_tf32 = True
+        torch.backends.cuda.matmul.allow_tf32 = True
+
+    if "32" in args.precision:
+        args.precision = 32
+    elif args.precision == "fp16":
+        args.precision = 16
+    else:
+        args.precision = "bf16"
+
+    ########################################################################################################
+
+    from src.trainer import train_callback, generate_init_weight
+    from src.dataset import MyDataset
+
+    train_data = MyDataset(args)
+    args.vocab_size = train_data.vocab_size
+
+    if args.data_type == "wds_img":
+        from src.model_img import RWKV_IMG
+
+        assert args.lora, "LoRA not yet supported for RWKV_IMG"
+        model = RWKV_IMG(args)
+    else:
+        from src.model import RWKV, LORA_CONFIG, LoraLinear
+
+        if args.lora:
+            assert args.lora_r > 0, "LoRA should have its `r` > 0"
+            LORA_CONFIG["r"] = args.lora_r
+            LORA_CONFIG["alpha"] = args.lora_alpha
+            LORA_CONFIG["dropout"] = args.lora_dropout
+            LORA_CONFIG["parts"] = set(str(args.lora_parts).split(","))
+            enable_time_finetune = "time" in LORA_CONFIG["parts"]
+            enable_ln_finetune = "ln" in LORA_CONFIG["parts"]
+        model = RWKV(args)
+        # only train lora parameters
+        if args.lora:
+            model.requires_grad_(False)
+            for name, module in model.named_modules():
+                # have to check param name since it may have been wrapped by torchscript
+                if any(n.startswith("lora_") for n, _ in module.named_parameters()):
+                    print(f"  LoRA training module {name}")
+                    for pname, param in module.named_parameters():
+                        param.requires_grad = "lora_" in pname
+                elif enable_ln_finetune and ".ln" in name:
+                    print(f"  LoRA additionally training module {name}")
+                    for param in module.parameters():
+                        param.requires_grad = True
+                elif enable_time_finetune and any(
+                        n.startswith("time") for n, _ in module.named_parameters()
+                ):
+                    for pname, param in module.named_parameters():
+                        if pname.startswith("time"):
+                            print(f"  LoRA additionally training parameter {pname}")
+                            param.requires_grad = True
+
+    if (
+            len(args.load_model) == 0 or args.my_pile_stage == 1
+    ):  # shall we build the initial weights?
+        init_weight_name = f"{args.proj_dir}/rwkv-init.pth"
+        generate_init_weight(model, init_weight_name)  # save initial weights
+        args.load_model = init_weight_name
+
+    rank_zero_info(f"########## Loading {args.load_model}... ##########")
+    try:
+        load_dict = torch.load(args.load_model, map_location="cpu")
+        model.load_state_dict(load_dict, strict=(not args.lora))
+    except:
+        rank_zero_info(f"Bad checkpoint {args.load_model}")
+        if args.my_pile_stage >= 2:  # try again using another checkpoint
+            max_p = args.my_pile_prev_p
+            if max_p == -1:
+                args.load_model = f"{args.proj_dir}/rwkv-init.pth"
+            else:
+                args.load_model = f"{args.proj_dir}/rwkv-{max_p}.pth"
+            args.epoch_begin = max_p + 1
+            rank_zero_info(f"Trying {args.load_model}")
+            load_dict = torch.load(args.load_model, map_location="cpu")
+            model.load_state_dict(load_dict, strict=(not args.lora))
+
+    if args.load_partial == 1:
+        load_keys = load_dict.keys()
+        for k in model.state_dict():
+            if k not in load_keys:
+                load_dict[k] = model.state_dict()[k]
+        model.load_state_dict(load_dict, strict=(not args.lora))
+    # If using LoRA, the LoRA keys might be missing in the original model
+    # model.load_state_dict(load_dict, strict=(not args.lora))
+    if os.path.isfile(args.lora_load):
+        model.load_state_dict(
+            torch.load(args.lora_load, map_location="cpu"), strict=False
+        )
+
+    trainer: Trainer = Trainer.from_argparse_args(
+        args,
+        callbacks=[train_callback(args)],
+    )
+
+    if (
+            args.lr_init > 1e-4
+            or trainer.world_size * args.micro_bsz * trainer.accumulate_grad_batches < 8
+    ):
+        if "I_KNOW_WHAT_IM_DOING" in os.environ:
+            if trainer.global_rank == 0:
+                print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+                print(
+                    f"  WARNING: you are using too large LR ({args.lr_init} > 1e-4) or too small global batch size ({trainer.world_size} * {args.micro_bsz} * {trainer.accumulate_grad_batches} < 8)"
+                )
+                print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+        else:
+            if trainer.global_rank == 0:
+                print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+                print(
+                    f"  ERROR: you are using too large LR ({args.lr_init} > 1e-4) or too small global batch size ({trainer.world_size} * {args.micro_bsz} * {trainer.accumulate_grad_batches} < 8)"
+                )
+                print(
+                    f"  Unless you are sure this is what you want, adjust them accordingly"
+                )
+                print(
+                    f'  (to suppress this, set environment variable "I_KNOW_WHAT_IM_DOING")'
+                )
+                print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+            exit(0)
+
+    if trainer.global_rank == 0:
+        for n in model.state_dict():
+            shape = model.state_dict()[n].shape
+            shape = [i for i in shape if i != 1]
+            if len(shape) > 1:
+                print(f"{str(shape[0]).ljust(5)} {str(shape[1]).ljust(5)} {n}")
+            else:
+                print(f"{str(shape[0]).ljust(5)}       {n}")
+
+    if "deepspeed" in args.strategy:
+        trainer.strategy.config["zero_optimization"]["allgather_bucket_size"] = (
+                args.ds_bucket_mb * 1000 * 1000
+        )
+        trainer.strategy.config["zero_optimization"]["reduce_bucket_size"] = (
+                args.ds_bucket_mb * 1000 * 1000
+        )
+
+    # must set shuffle=False, persistent_workers=False (because worker is in another thread)
+    data_loader = DataLoader(
+        train_data,
+        shuffle=False,
+        pin_memory=True,
+        batch_size=args.micro_bsz,
+        num_workers=1,
+        persistent_workers=False,
+        drop_last=True,
+    )
+
+    trainer.fit(model, data_loader)
--- a/finetune/requirements.txt
+++ b/finetune/requirements.txt
@@ -0,0 +1,3 @@
+torch==1.13.1
+pytorch_lightning==1.9.5
+deepspeed
--- a/frontend/package-lock.json
+++ b/frontend/package-lock.json
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -11,14 +11,19 @@
  "dependencies": {
    "@fluentui/react-components": "^9.20.0",
    "@fluentui/react-icons": "^2.0.201",
+    "@magenta/music": "^1.23.1",
    "@microsoft/fetch-event-source": "^2.0.1",
    "@primer/octicons-react": "^19.1.0",
+    "chart.js": "^4.3.0",
    "classnames": "^2.3.2",
    "github-markdown-css": "^5.2.0",
+    "html-midi-player": "^1.5.0",
    "i18next": "^22.4.15",
    "mobx": "^6.9.0",
    "mobx-react-lite": "^3.4.3",
    "react": "^18.2.0",
+    "react-beautiful-dnd": "^13.1.1",
+    "react-chartjs-2": "^5.2.0",
    "react-dom": "^18.2.0",
    "react-i18next": "^12.2.2",
    "react-markdown": "^8.0.7",
@@ -34,6 +39,7 @@
  },
  "devDependencies": {
    "@types/react": "^18.2.6",
+    "@types/react-beautiful-dnd": "^13.1.4",
    "@types/react-dom": "^18.2.4",
    "@types/uuid": "^9.0.1",
    "@vitejs/plugin-react": "^4.0.0",
--- a/frontend/src/App.tsx
+++ b/frontend/src/App.tsx
@@ -28,10 +28,10 @@ import { FC, useEffect, useState } from 'react';
 import { Route, Routes, useLocation, useNavigate } from 'react-router';
 import { pages } from './pages';
 import { useMediaQuery } from 'usehooks-ts';
-import { ToastContainer } from 'react-toastify';
 import commonStore from './stores/commonStore';
 import { observer } from 'mobx-react-lite';
 import { useTranslation } from 'react-i18next';
+import { CustomToastContainer } from './components/CustomToastContainer';

 const App: FC = observer(() => {
  const { t } = useTranslation();
@@ -87,21 +87,7 @@ const App: FC = observer(() => {
          </Routes>
        </div>
      </div>
-      <ToastContainer
-        style={{
-          width: '350px'
-        }}
-        position="top-center"
-        autoClose={4000}
-        pauseOnHover={true}
-        hideProgressBar={true}
-        newestOnTop={true}
-        closeOnClick={false}
-        rtl={false}
-        pauseOnFocusLoss={false}
-        draggable={false}
-        theme={commonStore.settings.darkMode ? 'dark' : 'light'}
-      />
+      <CustomToastContainer />
    </FluentProvider>
  );
 });
--- a/frontend/src/_locales/ja/main.json
+++ b/frontend/src/_locales/ja/main.json
@@ -0,0 +1,259 @@
+{
+  "Home": "ホーム",
+  "Train": "トレーニング",
+  "About": "約",
+  "Settings": "設定",
+  "Go to chat page": "チャットページに移動する",
+  "Manage your configs": "あなたの設定を管理する",
+  "Manage models": "モデルの管理",
+  "Run": "実行",
+  "Offline": "オフライン",
+  "Starting": "起動中",
+  "Loading": "モデルを読み込み中",
+  "Working": "動作中",
+  "Stop": "停止",
+  "Enable High Precision For Last Layer": "最後の層で高精度を有効にする",
+  "Stored Layers": "メモリ層読み込み",
+  "Precision": "精度",
+  "Device": "デバイス",
+  "Convert model with these configs. Using a converted model will greatly improve the loading speed, but model parameters of the converted model cannot be modified.": "これらの設定でモデルを変換します。変換されたモデルを使用すると、読み込み速度が大幅に向上しますが、変換したモデルのパラメータを変更することはできません。",
+  "Manage Models": "モデルの管理",
+  "Model": "モデル",
+  "Model Parameters": "モデルのパラメータ",
+  "Frequency Penalty": "周波数のペナルティ",
+  "Presence Penalty": "存在のペナルティ",
+  "Top_P": "Top_P",
+  "Temperature": "温度",
+  "Max Response Token": "最大レスポンストークン",
+  "API Port": "API ポート",
+  "Hover your mouse over the text to view a detailed description. Settings marked with * will take effect immediately after being saved.": "マウスをテキストに一定時間置いて詳細な説明を表示します。 * が付いている設定は保存後すぐに有効化されます。",
+  "Default API Parameters": "デフォルトのAPIパラメータ",
+  "Provide JSON file URLs for the models manifest. Separate URLs with semicolons. The \"models\" field in JSON files will be parsed into the following table.": "モデルマニフェストのためのJSONファイルURLを提供します。URLはセミコロンで分割します。JSONファイルの\"models\"フィールドは次の表に解析されます。",
+  "Config Name": "構成名",
+  "Refresh": "リフレッシュ",
+  "Save Config": "構成を保存",
+  "Model Source Manifest List": "モデルソースマニフェストリスト",
+  "Models": "モデル",
+  "Delete Config": "設定を削除",
+  "Help": "ヘルプ",
+  "Version": "バージョン",
+  "New Config": "新たな設定",
+  "Open Url": "URLを開く",
+  "Download": "ダウンロード",
+  "Open Folder": "フォルダを開く",
+  "Configs": "設定",
+  "Automatic Updates Check": "自動更新チェック",
+  "Updates Check Error": "更新チェックエラー",
+  "Introduction": "序文",
+  "Dark Mode": "ダークモード",
+  "Language": "言語",
+  "In Development": "開発中",
+  "Chat": "チャット",
+  "Convert": "変更",
+  "Actions": "行動",
+  "Last updated": "最後に更新",
+  "Desc": "説明",
+  "Size": "サイズ",
+  "File": "ファイル",
+  "Config Saved": "設定が保存されました",
+  "Downloading": "ダウンロード中",
+  "Loading Model": "モデルを読み込んでいます",
+  "Startup Completed": "起動完了",
+  "Failed to switch model": "モデルの切り替えに失敗しました",
+  "Start Converting": "変換を開始",
+  "Convert Success": "変換成功",
+  "Convert Failed": "変換失敗",
+  "Model Not Found": "モデルが見つかりません",
+  "Model Status": "モデルの状態",
+  "Clear": "クリア",
+  "Send": "送信",
+  "Type your message here": "ここにメッセージを入力してください",
+  "Copy": "コピー",
+  "Read Aloud": "読み上げ",
+  "Hello! I'm RWKV, an open-source and commercially usable large language model.": "こんにちは！私はRWKV、オープンソースで商用利用可能な大規模な言語モデルです。",
+  "This tool's API is compatible with OpenAI API. It can be used with any ChatGPT tool you like. Go to the settings of some ChatGPT tool, replace the 'https://api.openai.com' part in the API address with '": "このツールのAPIはOpenAI APIと互換性があります。 お好きなChatGPTツールで使用することができます。いくつかのChatGPTツールの設定に移動し、APIアドレスの 'https://api.openai.com' 部分を '",
+  "New Version Available": "新しいバージョンが存在します",
+  "Update": "更新",
+  "Please click the button in the top right corner to start the model": "右上角のボタンをクリックしてモデルを起動してください",
+  "Update Error": "更新エラー",
+  "Open the following URL with your browser to view the API documentation": "以下のURLをブラウザで開いてAPIドキュメンテーションを確認してください",
+  "By default, the maximum number of tokens that can be answered in a single response, it can be changed by the user by specifying API parameters.": "デフォルトでは、一度に回答できるトークンの最大数は、APIパラメータを指定することでユーザーが変更できます。",
+  "Sampling temperature, it's like giving alcohol to a model, the higher the stronger the randomness and creativity, while the lower, the more focused and deterministic it will be.": "サンプリング温度は、モデルにアルコールを与えるようなもので、高いほどランダム性と創造性が強く、低いほど焦点を絞り、決定論的になります。",
+  "Just like feeding sedatives to the model. Consider the results of the top n% probability mass, 0.1 considers the top 10%, with higher quality but more conservative, 1 considers all results, with lower quality but more diverse.": "モデルに鎮静剤を与えるようなもの。上位n％の確率質量の結果を考えてみてください。0.1は上位10％を考えており、質が高いが保守的で、1は全ての結果を考慮しており、質は低いが多様性があります。",
+  "Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.": "ポジティヴ値は、新しいトークンが今までのテキストに出現していたかどうかに基づいてこれらをペナルティとし、新しいトピックについて話す可能性を増加させます。",
+  "Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.": "ポジティブ値は、新しいトークンが既存のテキストでどれだけ頻繁に使われているかに基づいてペナルティを与え、モデルが同じ行を完全に繰り返す可能性を減らします。",
+  "int8 uses less VRAM, but has slightly lower quality. fp16 has higher quality, and fp32 has the best quality.": "int8はVRAMの使用量が少ないですが、質が若干低いです。fp16は高品質、fp32は最高品質です。",
+  "Number of the neural network layers loaded into VRAM, the more you load, the faster the speed, but it consumes more VRAM. (If your VRAM is not enough, it will fail to load)": "VRAMにロードされるニューラルネットワークの層の数。ロードする量が多いほど速度は速くなりますが、VRAMを多く消費します。(VRAMが不足している場合、ロードに失敗します)",
+  "Whether to use CPU to calculate the last output layer of the neural network with FP32 precision to obtain better quality.": "ネットワークの最終出力層をFP32精度で計算するためにCPUを使用するかどうか。",
+  "Downloads": "ダウンロード",
+  "Pause": "ポーズ",
+  "Continue": "続行",
+  "Resume": "続行",
+  "Check": "確認",
+  "Model file not found": "モデルファイルが見つかりません",
+  "Can not find download url": "ダウンロードURLが見つかりません",
+  "Python target not found, would you like to download it?": "Pythonターゲットが見つかりません、ダウンロードしますか？",
+  "Python dependencies are incomplete, would you like to install them?": "Pythonの依存関係が不完全です、インストールしますか？",
+  "Install": "インストール",
+  "This is the latest version": "これは最新バージョンです",
+  "Use Tsinghua Pip Mirrors": "清華大学Pipミラーサーバーを使用",
+  "Model Config Exception": "モデル設定例外",
+  "Use Gitee Updates Source": "Gitee更新ソースを使用",
+  "Use Custom CUDA kernel to Accelerate": "カスタムCUDAカーネルを使用して加速",
+  "Enabling this option can greatly improve inference speed and save some VRAM, but there may be compatibility issues. If it fails to start, please turn off this option.": "このオプションを有効にすると、推論速度が大幅に向上し、一部のVRAMを節約できますが、互換性の問題が生じる可能性があります。起動に失敗した場合は、このオプションをオフにしてください。",
+  "Supported custom cuda file not found": "対応しているカスタムCUDAファイルが見つかりません",
+  "Failed to copy custom cuda file": "カスタムCUDAファイルのコピーに失敗しました",
+  "Downloading update, please wait. If it is not completed, please manually download the program from GitHub and replace the original program.": "更新をダウンロード中です、お待ちください。完了しない場合は、GitHubから手動でプログラムをダウンロードし、元のプログラムを置き換えてください。",
+  "Completion": "補完",
+  "Parameters": "パラメータ",
+  "Stop Sequences": "シーケンスを停止",
+  "When this content appears in the response result, the generation will end.": "この内容が応答結果に表示されると、生成が終了します。",
+  "Reset": "リセット",
+  "Generate": "生成",
+  "Writer": "ライター",
+  "Translator": "翻訳者",
+  "Catgirl": "ネコガール",
+  "Code Generation": "コード生成",
+  "Werewolf": "人狼",
+  "Instruction": "指示",
+  "Blank": "空白",
+  "The following is an epic science fiction masterpiece that is immortalized, with delicate descriptions and grand depictions of interstellar civilization wars.\nChapter 1.\n": "以下は、壮大な描写と共に、不滅のエピックサイエンスフィクションの傑作で、星間文明戦争が繊細に描かれています。\n第1章\n",
+  "The following is a conversation between a cat girl and her owner. The cat girl is a humanized creature that behaves like a cat but is humanoid. At the end of each sentence in the dialogue, she will add \"Meow~\". In the following content, User represents the owner and Assistant represents the cat girl.\n\nUser: Hello.\n\nAssistant: I'm here, meow~.\n\nUser: Can you tell jokes?": "以下は、猫少女とその飼い主との会話です。猫少女は、猫のように振る舞いながらもヒトの姿をした生物です。会話の各文の終わりには必ず「にゃ〜」とつけています。以下の文章では、Userが飼い主、Assistantが猫少女を表しています。\n\nUser: こんにちは。\n\nAssistant: ここにいますよ、にゃ〜。\n\nUser: 笑い話を話せますか？",
+  "When response finished, inject this content.": "応答終了時に、この内容を注入します。",
+  "Inject start text": "開始テキストを注入",
+  "Inject end text": "終了テキストを注入",
+  "Before the response starts, inject this content.": "応答が始まる前に、この内容を注入します。",
+  "There is currently a game of Werewolf with six players, including a Seer (who can check identities at night), two Werewolves (who can choose someone to kill at night), a Bodyguard (who can choose someone to protect at night), two Villagers (with no special abilities), and a game host. User will play as Player 1, Assistant will play as Players 2-6 and the game host, and they will begin playing together. Every night, the host will ask User for his action and simulate the actions of the other players. During the day, the host will oversee the voting process and ask User for his vote. \n\nAssistant: Next, I will act as the game host and assign everyone their roles, including randomly assigning yours. Then, I will simulate the actions of Players 2-6 and let you know what happens each day. Based on your assigned role, you can tell me your actions and I will let you know the corresponding results each day.\n\nUser: Okay, I understand. Let's begin. Please assign me a role. Am I the Seer, Werewolf, Villager, or Bodyguard?\n\nAssistant: You are the Seer. Now that night has fallen, please choose a player to check his identity.\n\nUser: Tonight, I want to check Player 2 and find out his role.": "現在、6人のプレイヤーが参加する人狼ゲームが行われています。その中には、夜に任意のプレイヤーの正体を確認できる占い師、夜に誰かを殺すことができる人狼2名、夜に誰かを守ることができるボディガード、特殊な能力を持っていない村人2名、そしてゲームのホストがいます。Userはプレイヤー1として、Assistantはプレーヤー2から6まで及びゲームのホストとして参加し、一緒にゲームを始めます。ホストは毎晩、Userに彼の行動を問い、他のプレーヤーの行動をシミュレートします。昼には、ホストが投票プロセスを監督し、Userに彼の投票を求めます。\n\nAssistant: 次に、私はゲームのホストとして参加者全員に役割を割り当てることになります。それには、あなたの役割もランダムに割り当てます。その後、私はプレーヤー2から6の行動をシミュレートし、毎日何が起こったかを報告します。あなたに割り当てられた役割に基づいて、あなたの行動を教えてください。私は毎日、それに対する結果を報告します。\n\nUser: 了解しました。では、始めましょう。私の役割を割り当ててください。占い師、人狼、村人、ボディーガードのいずれなのでしょうか？\n\nAssistant: あなたの役割は占い師です。今夜が来たので、誰の正体を確認するか選んでください。\n\nUser: 今夜、プレイヤー2の役割を確認したい。",
+  "Writer, Translator, Role-playing": "ライター、翻訳者、ロールプレイング",
+  "Chinese Kongfu": "中国武術",
+  "Allow external access to the API (service must be restarted)": "APIへの外部アクセスを許可する (サービスを再起動する必要があります)",
+  "Custom": "カスタム",
+  "CUDA (Beta, Faster)": "CUDA (ベータ、高速)",
+  "Reset All Configs": "すべての設定をリセット",
+  "Cancel": "キャンセル",
+  "Confirm": "確認",
+  "Are you sure you want to reset all configs? This will obtain the latest preset configs, but will override your custom configs and cannot be undone.": "本当にすべての設定をリセットしますか？これにより最新のプリセット設定が取得されますが、カスタム設定は上書きされ、元に戻すことはできません。",
+  "Advanced": "高度な",
+  "Custom Python Path": "カスタムPythonパス",
+  "Custom Models Path": "カスタムモデルパス",
+  "Microsoft Visual C++ Redistributable is not installed, would you like to download it?": "Microsoft Visual C++ 再頒布可能パッケージがインストールされていません。ダウンロードしますか？",
+  "File Path Cannot Contain Space": "ファイルのパスにスペースを含めることはできません",
+  "Current Strategy": "現在の戦略",
+  "MacOS is not yet supported for performing this operation, please do it manually.": "MacOSはまだこの操作を実行するサポートがありませんので、手動で行ってください。",
+  "Linux is not yet supported for performing this operation, please do it manually.": "Linuxはまだこの操作を実行するサポートがありませんので、手動で行ってください。",
+  "On Linux system, you must manually install python dependencies.": "Linuxシステムでは、pythonの依存関係を手動でインストールする必要があります。",
+  "Update completed, please restart the program.": "更新が完了したら、プログラムを再起動してください。",
+  "Are you sure you want to reset this page? It cannot be undone.": "本当にこのページをリセットしてもよろしいですか？元に戻すことはできません。",
+  "Model file download is not complete": "モデルファイルのダウンロードが完了していません",
+  "Error": "エラー",
+  "Are you sure you want to clear the conversation? It cannot be undone.": "会話をクリアしてもよろしいですか？元に戻すことはできません。",
+  "Save": "保存",
+  "Conversation Saved": "会話が保存されました",
+  "Open": "開く",
+  "DPI Scaling": "DPIスケーリング",
+  "Restart the app to apply DPI Scaling.": "DPIスケーリングを適用するためにアプリを再起動してください。",
+  "Restart": "再起動",
+  "API Chat Model Name": "APIチャットモデル名",
+  "API Completion Model Name": "API完成モデル名",
+  "Localhost": "ローカルホスト",
+  "Retry": "リトライ",
+  "Delete": "削除",
+  "Edit": "編集",
+  "Memory is not enough, try to increase the virtual memory or use a smaller model.": "メモリが不足しています。仮想メモリを増やすか、もしくは小さなモデルを使ってみてください",
+  "Bad PyTorch version, please reinstall PyTorch with cuda.": "不適切なPyTorchのバージョンです。cudaと共にPyTorchを再インストールしてください。",
+  "The model file is corrupted, please download again.": "モデルファイルが破損しています。再度ダウンロードしてください。",
+  "Found no NVIDIA driver, please install the latest driver.": "NVIDIAのドライバが見つかりません。最新版のドライバをインストールしてください。",
+  "VRAM is not enough, please reduce stored layers or use a lower precision in Configs page.": "VRAMが足りません。設定ページで保存されているレイヤーを減らすか、精度を下げてください。",
+  "Failed to enable custom CUDA kernel, ninja is required to load C++ extensions. You may be using the CPU version of PyTorch, please reinstall PyTorch with CUDA. Or if you are using a custom Python interpreter, you must compile the CUDA kernel by yourself or disable Custom CUDA kernel acceleration.": "カスタムCUDAカーネルの有効化に失敗しました。C++拡張を読み込むためにはNinjaが必要です。あなたは恐らくCPU版のPyTorchを使用しており、CUDA版のPyTorchを再インストールする必要があります。または、あなたがカスタムPythonインタプリタを使用している場合は、CUDAカーネルを自分でコンパイルするか、カスタムCUDAカーネルのアクセラレーションを無効にする必要があります。",
+  "Presets": "プリセット",
+  "Online": "オンライン",
+  "english": "英語",
+  "chinese": "中国語",
+  "default": "デフォルト",
+  "japanese": "日本語",
+  "New Preset": "新規プリセット",
+  "Import": "インポート",
+  "Name": "名前",
+  "Imported successfully": "インポート成功",
+  "Failed to import. Please copy a preset to the clipboard.": "インポートに失敗しました。プリセットをクリップボードにコピーしてください。",
+  "Clipboard is empty.": "クリップボードが空です。",
+  "Successfully copied to clipboard.": "クリップボードにコピーしました。",
+  "Edit Character Settings": "キャラクター設定を編集",
+  "Go Back": "戻る",
+  "Description": "説明",
+  "Avatar Url": "アバターURL",
+  "Welcome Message": "ウェルカムメッセージ",
+  "Display Preset Messages": "プリセットメッセージの表示",
+  "Tag": "タグ",
+  "Activate": "アクティブ化",
+  "New": "新規",
+  "user": "ユーザー",
+  "assistant": "アシスタント",
+  "system": "システム",
+  "Regenerate": "再生成",
+  "LoRA Finetune": "LoRAの微調整",
+  "Command Stopped": "コマンドが停止しました",
+  "Please convert data first.": "先にデータを変換してください。",
+  "Ubuntu is not installed, do you want to install it?": "Ubuntuがインストールされていません、インストールしますか？",
+  "Install Ubuntu": "Ubuntuをインストール",
+  "Please install Ubuntu using Microsoft Store, after installation click the Open button in Microsoft Store and then click the Train button": "UbuntuをMicrosoftストアからインストールすることができます。インストールが完了したら、MicrosoftストアのOpenボタンを押し、Trainボタンを押してください",
+  "WSL is not enabled, do you want to enable it?": "WSLが有効になっていません、有効化しますか？",
+  "Enable WSL": "WSLを有効化",
+  "After installation, please restart your computer to enable WSL": "インストールが完了したら、WSLを有効化するためにコンピュータを再起動してください",
+  "Data Process": "データ処理",
+  "Data Path": "データパス",
+  "Vocab Path": "語彙パス",
+  "Train Parameters": "トレーニングパラメータ",
+  "Base Model": "基本モデル",
+  "LoRA Model": "LoRAモデル",
+  "Merge Model": "モデルの統合",
+  "Devices": "デバイス",
+  "Gradient Checkpoint": "勾配チェックポイント",
+  "Context Length": "コンテキストの長さ",
+  "Epoch Steps": "エポックステップ数",
+  "Epoch Count": "エポックの数",
+  "Epoch Begin": "エポックの起点",
+  "Epoch Save": "エポックの保存",
+  "Learning Rate Init": "初期学習率",
+  "Learning Rate Final": "最終学習率",
+  "Micro Batch Size": "マイクロバッチサイズ",
+  "Accumulate Gradient Batches": "勾配バッチの累計",
+  "Warmup Steps": "ウォームアップステップ",
+  "Pre-FFN": "FFNの前処理",
+  "None": "なし",
+  "Merge model successfully": "モデルのマージが成功しました",
+  "Convert Data successfully": "データ変換に成功しました",
+  "Please select a LoRA model": "LoRAモデルを選択してください",
+  "You are using sample data for training. For formal training, please make sure to create your own jsonl file.": "トレーニングにはサンプルデータを使用しています。正式なトレーニングのためには、自身でjsonlファイルを作成してください。",
+  "WSL is not running, please retry. If it keeps happening, it means you may be using an outdated version of WSL, run \"wsl --update\" to update.": "WSLが実行されていません、もう一度試してください。これが続く場合、古いバージョンのWSLを使用している可能性があります。\"wsl --update\"を実行して更新してください。",
+  "Memory is not enough, try to increase the virtual memory (Swap of WSL) or use a smaller base model.": "メモリが不足しています、仮想メモリ (WSL Swap) を増やすか小さなベースモデルを使用してみてください。",
+  "VRAM is not enough": "ビデオRAMが不足しています",
+  "Training data is not enough, reduce context length or add more data for training": "トレーニングデータが不足しています、コンテキストの長さを減らすか、トレーニング用のデータをさらに追加してください",
+  "You are using WSL 1 for training, please upgrade to WSL 2. e.g. Run \"wsl --set-version Ubuntu-22.04 2\"": "トレーニングにWSL 1を使用しています、WSL 2にアップグレードしてください。例:\"wsl --set-version Ubuntu-22.04 2\"を実行する",
+  "Matched CUDA is not installed": "対応するCUDAがインストールされていません",
+  "Failed to convert data": "データの変換に失敗しました",
+  "Failed to merge model": "モデルのマージに失敗しました",
+  "The data path should be a directory or a file in jsonl format (more formats will be supported in the future).\n\nWhen you provide a directory path, all the txt files within that directory will be automatically converted into training data. This is commonly used for large-scale training in writing, code generation, or knowledge bases.\n\nThe jsonl format file can be referenced at https://github.com/Abel2076/json2binidx_tool/blob/main/sample.jsonl.\nYou can also write it similar to OpenAI's playground format, as shown in https://platform.openai.com/playground/p/default-chat.\nEven for multi-turn conversations, they must be written in a single line using `\\n` to indicate line breaks. If they are different dialogues or topics, they should be written in separate lines.": "データのパスはディレクトリまたはjsonl形式のファイルでなければなりません（将来的にはより多くの形式がサポートされる予定です）。ディレクトリパスを提供した場合、そのディレクトリ内のすべてのtxtファイルが自動的にトレーニングデータに変換されます。これは大規模なライティング、コード生成、または知識ベースのトレーニングで一般的に使用されます。jsonl形式のファイルは、https://github.com/Abel2076/json2binidx_tool/blob/main/sample.jsonl を参照してください。\nhttps://platform.openai.com/playground/p/default-chat のように、OpenAIのプレイグラウンド形式に似た形式で書くこともできます。複数ターンの対話であっても、一行で書く必要があり、行の区切りを示すために`\\n`を使用します。それらが異なる対話やトピックであれば、それらは別々の行に書かれるべきです。",
+  "Size mismatch for blocks. You are attempting to continue training from the LoRA model, but it does not match the base model. Please set LoRA model to None.": "ブロックのサイズが一致しません。LoRAモデルからトレーニングを続けようとしていますが、それはベースモデルと一致しません。LoRAモデルをNoneに設定してください。",
+  "Instruction: Write a story using the following information\n\nInput: A man named Alex chops a tree down\n\nResponse:": "Instruction: Write a story using the following information\n\nInput: アレックスという男が木を切り倒す\n\nResponse:",
+  "Composition": "作曲",
+  "Use Local Sound Font": "ローカルサウンドフォントを使用する",
+  "Auto Play At The End": "最後に自動再生",
+  "No File to save": "保存するファイルがありません",
+  "File Saved": "ファイルが保存されました",
+  "Failed to load local sound font, please check if the files exist - assets/sound-font": "ローカルサウンドフォントの読み込みに失敗しました、ファイルが存在するか確認してください - assets/sound-font",
+  "Please convert model to safe tensors format first": "モデルを安全なテンソル形式に変換してください",
+  "Convert To Safe Tensors Format": "安全なテンソル形式に変換",
+  "Please change Strategy to WebGPU to use safetensors format": "StrategyをWebGPUに変更して、安全なテンソル形式を使用してください",
+  "Preview Only": "プレビューのみ",
+  "RAM": "RAM",
+  "VRAM": "VRAM",
+  "GPU Usage": "GPU使用率",
+  "Use Custom Tokenizer": "カスタムトークナイザーを使用する",
+  "Tokenizer Path (e.g. backend-python/rwkv_pip/20B_tokenizer.json)": "トークナイザーパス (例: backend-python/rwkv_pip/20B_tokenizer.json)",
+  "User Name": "ユーザー名",
+  "Assistant Name": "アシスタント名",
+  "Insert default system prompt at the beginning": "最初にデフォルトのシステムプロンプトを挿入",
+  "Please Enable Custom CUDA Kernel. Latest RWKV-5 requires os.environ['RWKV_CUDA_ON'] == '1' (will fix soon).": "カスタムCUDAカーネルを有効にしてください。最新のRWKV-5ではos.environ['RWKV_CUDA_ON'] == '1'が必要です（近日中に修正します）。",
+  "Format Content": "内容フォーマットの規格化"
+}
--- a/frontend/src/_locales/resources.ts
+++ b/frontend/src/_locales/resources.ts
@@ -1,9 +1,10 @@
 import zhHans from './zh-hans/main.json';
+import ja from './ja/main.json';

 export const resources = {
  zh: {
    translation: zhHans
-  }
+  },
  // de: {
  //   translation: de,
  // },
@@ -19,9 +20,9 @@ export const resources = {
  // it: {
  //   translation: it,
  // },
-  // ja: {
-  //   translation: ja,
-  // },
+  ja: {
+    translation: ja
+  }
  // ko: {
  //   translation: ko,
  // },
--- a/frontend/src/_locales/zh-hans/main.json
+++ b/frontend/src/_locales/zh-hans/main.json
@@ -70,7 +70,7 @@
  "Type your message here": "在此输入消息",
  "Copy": "复制",
  "Read Aloud": "朗读",
-  "Hello! I'm RWKV, an open-source and commercially usable large language model.": "你好! 我是RWKV, 一个开源可商用的大语言模型.",
+  "Hello! I'm RWKV, an open-source and commercially usable large language model.": "你好！我是RWKV，一个开源可商用的大语言模型。",
  "This tool's API is compatible with OpenAI API. It can be used with any ChatGPT tool you like. Go to the settings of some ChatGPT tool, replace the 'https://api.openai.com' part in the API address with '": "本工具的API与OpenAI API兼容. 因此可以配合任意你喜欢的ChatGPT工具使用. 打开某个ChatGPT工具的设置, 将API地址中的'https://api.openai.com'部分替换为'",
  "New Version Available": "新版本可用",
  "Update": "更新",
@@ -104,7 +104,7 @@
  "Supported custom cuda file not found": "没有找到支持的自定义cuda文件",
  "Failed to copy custom cuda file": "自定义cuda文件复制失败",
  "Downloading update, please wait. If it is not completed, please manually download the program from GitHub and replace the original program.": "正在下载更新，请等待。如果一直未完成，请从Github手动下载并覆盖原程序",
-  "Completion": "补全",
+  "Completion": "续写",
  "Parameters": "参数",
  "Stop Sequences": "停止词",
  "When this content appears in the response result, the generation will end.": "响应结果出现该内容时就结束生成",
@@ -113,21 +113,22 @@
  "Writer": "写作",
  "Translator": "翻译",
  "Catgirl": "猫娘",
-  "Explain Code": "代码解释",
+  "Code Generation": "代码生成",
  "Werewolf": "狼人杀",
  "Instruction": "指令",
  "Blank": "空白",
-  "The following is an epic science fiction masterpiece that is immortalized, with delicate descriptions and grand depictions of interstellar civilization wars.\nChapter 1.\n": "以下是不朽的科幻史诗巨著，描写细腻，刻画了宏大的星际文明战争。\n第一章\n",
-  "The following is a conversation between a cat girl and her owner. The cat girl is a humanized creature that behaves like a cat but is humanoid. At the end of each sentence in the dialogue, she will add \"Meow~\". In the following content, Bob represents the owner and Alice represents the cat girl.\n\nBob: Hello.\n\nAlice: I'm here, meow~.\n\nBob: Can you tell jokes?": "以下是一位猫娘的主人和猫娘的对话内容，猫娘是一种拟人化的生物，其行为似猫但类人，在每一句对话末尾都会加上\"喵~\"。以下内容中，Bob代表主人，Alice代表猫娘。\n\nBob: 你好\n\nAlice: 主人我在哦，喵~\n\nBob: 你会讲笑话吗?",
+  "The following is an epic science fiction masterpiece that is immortalized, with delicate descriptions and grand depictions of interstellar civilization wars.\nChapter 1.\n": "《背影》\n我与父亲不相见已二年余了，我最不能忘记的是他的背影。\n那年冬天，祖母死了，父亲的差使也交卸了，正是祸不单行的日子。我从北京到徐州，打算",
+  "The following is a conversation between a cat girl and her owner. The cat girl is a humanized creature that behaves like a cat but is humanoid. At the end of each sentence in the dialogue, she will add \"Meow~\". In the following content, User represents the owner and Assistant represents the cat girl.\n\nUser: Hello.\n\nAssistant: I'm here, meow~.\n\nUser: Can you tell jokes?": "以下是一位猫娘的主人和猫娘的对话内容，猫娘是一种拟人化的生物，其行为似猫但类人，在每一句对话末尾都会加上\"喵~\"。以下内容中，User代表主人，Assistant代表猫娘。\n\nUser: 你好\n\nAssistant: 主人我在哦，喵~\n\nUser: 你会讲笑话吗?",
  "When response finished, inject this content.": "响应结束时，插入此内容到末尾",
  "Inject start text": "起始注入文本",
  "Inject end text": "结尾注入文本",
  "Before the response starts, inject this content.": "响应开始前，在开头插入此内容",
-  "There is currently a game of Werewolf with six players, including a Seer (who can check identities at night), two Werewolves (who can choose someone to kill at night), a Bodyguard (who can choose someone to protect at night), two Villagers (with no special abilities), and a game host. Bob will play as Player 1, Alice will play as Players 2-6 and the game host, and they will begin playing together. Every night, the host will ask Bob for his action and simulate the actions of the other players. During the day, the host will oversee the voting process and ask Bob for his vote. \n\nAlice: Next, I will act as the game host and assign everyone their roles, including randomly assigning yours. Then, I will simulate the actions of Players 2-6 and let you know what happens each day. Based on your assigned role, you can tell me your actions and I will let you know the corresponding results each day.\n\nBob: Okay, I understand. Let's begin. Please assign me a role. Am I the Seer, Werewolf, Villager, or Bodyguard?\n\nAlice: You are the Seer. Now that night has fallen, please choose a player to check his identity.\n\nBob: Tonight, I want to check Player 2 and find out his role.": "现在有一场六人狼人杀游戏，包括一名预言家(可以在夜晚查验身份)，两名狼人(可以在夜晚选择杀人)，一名守卫(可以在夜晚选择要守护的人)，两名平民（无技能)，一名主持人，以下内容中Bob将扮演其中的1号玩家，Alice来扮演2-6号玩家，以及主持人，并开始与Bob进行游戏，主持人每晚都会询问Bob的行动，并模拟其他人的行动，在白天则要主持投票，并同样询问Bob投票对象，公布投票结果。\n\nAlice: 接下来，我将首先作为主持人进行角色分配，并给你赋予随机的角色，之后我将模拟2-6号玩家进行行动，告知你每天的动态，根据你被分配的角色，你可以回复我你做的行动，我会告诉你每天对应的结果\n\nBob: 好的，我明白了，那么开始吧。请先给我一个角色身份。我是预言家，狼人，平民，守卫中的哪一个呢？\n\nAlice: 你的身份是预言家。现在夜晚降临，请选择你要查验的玩家。\n\nBob: 今晚我要验2号玩家，他是什么身份？",
+  "There is currently a game of Werewolf with six players, including a Seer (who can check identities at night), two Werewolves (who can choose someone to kill at night), a Bodyguard (who can choose someone to protect at night), two Villagers (with no special abilities), and a game host. User will play as Player 1, Assistant will play as Players 2-6 and the game host, and they will begin playing together. Every night, the host will ask User for his action and simulate the actions of the other players. During the day, the host will oversee the voting process and ask User for his vote. \n\nAssistant: Next, I will act as the game host and assign everyone their roles, including randomly assigning yours. Then, I will simulate the actions of Players 2-6 and let you know what happens each day. Based on your assigned role, you can tell me your actions and I will let you know the corresponding results each day.\n\nUser: Okay, I understand. Let's begin. Please assign me a role. Am I the Seer, Werewolf, Villager, or Bodyguard?\n\nAssistant: You are the Seer. Now that night has fallen, please choose a player to check his identity.\n\nUser: Tonight, I want to check Player 2 and find out his role.": "现在有一场六人狼人杀游戏，包括一名预言家(可以在夜晚查验身份)，两名狼人(可以在夜晚选择杀人)，一名守卫(可以在夜晚选择要守护的人)，两名平民（无技能)，一名主持人，以下内容中User将扮演其中的1号玩家，Assistant来扮演2-6号玩家，以及主持人，并开始与User进行游戏，主持人每晚都会询问User的行动，并模拟其他人的行动，在白天则要主持投票，并同样询问User投票对象，公布投票结果。\n\nAssistant: 接下来，我将首先作为主持人进行角色分配，并给你赋予随机的角色，之后我将模拟2-6号玩家进行行动，告知你每天的动态，根据你被分配的角色，你可以回复我你做的行动，我会告诉你每天对应的结果\n\nUser: 好的，我明白了，那么开始吧。请先给我一个角色身份。我是预言家，狼人，平民，守卫中的哪一个呢？\n\nAssistant: 你的身份是预言家。现在夜晚降临，请选择你要查验的玩家。\n\nUser: 今晚我要验2号玩家，他是什么身份？",
  "Writer, Translator, Role-playing": "写作，翻译，角色扮演",
  "Chinese Kongfu": "情境冒险",
  "Allow external access to the API (service must be restarted)": "允许外部访问API (必须重启服务)",
  "Custom": "自定义",
+  "CUDA (Beta, Faster)": "CUDA (Beta, 更快)",
  "Reset All Configs": "重置所有配置",
  "Cancel": "取消",
  "Confirm": "确认",
@@ -153,6 +154,106 @@
  "Restart the app to apply DPI Scaling.": "重启应用以使显示缩放生效",
  "Restart": "重启",
  "API Chat Model Name": "API聊天模型名",
-  "API Completion Model Name": "API补全模型名",
-  "Localhost": "本地"
+  "API Completion Model Name": "API续写模型名",
+  "Localhost": "本地",
+  "Retry": "重试",
+  "Delete": "删除",
+  "Edit": "编辑",
+  "Memory is not enough, try to increase the virtual memory or use a smaller model.": "内存不足，尝试增加虚拟内存，或使用一个更小规模的模型",
+  "Bad PyTorch version, please reinstall PyTorch with cuda.": "错误的PyTorch版本，请重新安装CUDA版本的PyTorch",
+  "The model file is corrupted, please download again.": "模型文件损坏，请重新下载",
+  "Found no NVIDIA driver, please install the latest driver.": "没有找到NVIDIA驱动，请安装最新驱动",
+  "VRAM is not enough, please reduce stored layers or use a lower precision in Configs page.": "显存不足，请在配置页面减少载入显存层数，或使用更低的精度",
+  "Failed to enable custom CUDA kernel, ninja is required to load C++ extensions. You may be using the CPU version of PyTorch, please reinstall PyTorch with CUDA. Or if you are using a custom Python interpreter, you must compile the CUDA kernel by yourself or disable Custom CUDA kernel acceleration.": "自定义CUDA算子开启失败，需要安装Ninja来读取C++扩展。你可能正在使用CPU版本的PyTorch，请重新安装CUDA版本的PyTorch。如果你正在使用自定义Python解释器，你必须自己编译CUDA算子或禁用自定义CUDA算子加速",
+  "Presets": "预设",
+  "Online": "在线",
+  "english": "英文",
+  "chinese": "中文",
+  "default": "默认",
+  "japanese": "日文",
+  "New Preset": "新建预设",
+  "Import": "导入",
+  "Name": "名称",
+  "Imported successfully": "导入成功",
+  "Failed to import. Please copy a preset to the clipboard.": "导入失败。请复制一个预设到剪贴板",
+  "Clipboard is empty.": "剪贴板没有内容",
+  "Successfully copied to clipboard.": "成功复制到剪贴板",
+  "Edit Character Settings": "编辑人设",
+  "Go Back": "返回",
+  "Description": "描述",
+  "Avatar Url": "头像图片地址",
+  "Welcome Message": "欢迎语",
+  "Display Preset Messages": "显示预设中的对话",
+  "Tag": "标签",
+  "Activate": "激活",
+  "New": "新建",
+  "user": "用户",
+  "assistant": "AI",
+  "system": "系统",
+  "Regenerate": "重新生成",
+  "LoRA Finetune": "LoRA微调",
+  "Command Stopped": "命令已终止",
+  "Please convert data first.": "请先转换数据",
+  "Ubuntu is not installed, do you want to install it?": "Ubuntu未安装，是否安装？",
+  "Install Ubuntu": "安装Ubuntu",
+  "Please install Ubuntu using Microsoft Store, after installation click the Open button in Microsoft Store and then click the Train button": "请用Microsoft Store安装Ubuntu，安装完成后，点击Microsoft Store界面的“打开”按钮，然后点击“训练”按钮",
+  "WSL is not enabled, do you want to enable it?": "WSL未启用，是否启用？",
+  "Enable WSL": "启用WSL",
+  "After installation, please restart your computer to enable WSL": "安装完成后，请重启电脑以启用WSL",
+  "Data Process": "数据处理",
+  "Data Path": "数据路径",
+  "Vocab Path": "词表路径",
+  "Train Parameters": "训练参数",
+  "Base Model": "基底模型",
+  "LoRA Model": "LoRA模型",
+  "Merge Model": "合并模型",
+  "Devices": "显卡数量",
+  "Gradient Checkpoint": "梯度检查点标志",
+  "Context Length": "上下文长度",
+  "Epoch Steps": "每轮训练步数",
+  "Epoch Count": "训练轮次",
+  "Epoch Begin": "起始轮次",
+  "Epoch Save": "保存间隔轮次",
+  "Learning Rate Init": "初始学习率",
+  "Learning Rate Final": "最终学习率",
+  "Micro Batch Size": "微批次大小",
+  "Accumulate Gradient Batches": "梯度累积批次",
+  "Warmup Steps": "学习率预热步数",
+  "Pre-FFN": "前馈网络预处理",
+  "None": "空",
+  "Merge model successfully": "合并模型成功",
+  "Convert Data successfully": "数据转换成功",
+  "Please select a LoRA model": "请选择一个LoRA模型",
+  "You are using sample data for training. For formal training, please make sure to create your own jsonl file.": "你正在使用示例数据训练，对于正式训练场合，请务必创建你自己的jsonl训练数据",
+  "WSL is not running, please retry. If it keeps happening, it means you may be using an outdated version of WSL, run \"wsl --update\" to update.": "WSL没有运行，请重试。如果一直出现此错误，意味着你可能正在使用旧版本的WSL，请在cmd执行\"wsl --update\"以更新",
+  "Memory is not enough, try to increase the virtual memory (Swap of WSL) or use a smaller base model.": "内存不足，尝试增加虚拟内存(WSL Swap)，或使用一个更小规模的基底模型",
+  "VRAM is not enough": "显存不足",
+  "Training data is not enough, reduce context length or add more data for training": "训练数据不足，请减小上下文长度或增加训练数据",
+  "You are using WSL 1 for training, please upgrade to WSL 2. e.g. Run \"wsl --set-version Ubuntu-22.04 2\"": "你正在使用WSL 1进行训练，请升级到WSL 2。例如，运行\"wsl --set-version Ubuntu-22.04 2\"",
+  "Matched CUDA is not installed": "未安装匹配的CUDA",
+  "Failed to convert data": "数据转换失败",
+  "Failed to merge model": "合并模型失败",
+  "The data path should be a directory or a file in jsonl format (more formats will be supported in the future).\n\nWhen you provide a directory path, all the txt files within that directory will be automatically converted into training data. This is commonly used for large-scale training in writing, code generation, or knowledge bases.\n\nThe jsonl format file can be referenced at https://github.com/Abel2076/json2binidx_tool/blob/main/sample.jsonl.\nYou can also write it similar to OpenAI's playground format, as shown in https://platform.openai.com/playground/p/default-chat.\nEven for multi-turn conversations, they must be written in a single line using `\\n` to indicate line breaks. If they are different dialogues or topics, they should be written in separate lines.": "数据路径必须是一个文件夹，或者jsonl格式文件 (未来会支持更多格式)\n\n当你填写的路径是一个文件夹时，该文件夹内的所有txt文件会被自动转换为训练数据，通常这用于大批量训练写作，代码生成或知识库\n\njsonl文件的格式参考 https://github.com/Abel2076/json2binidx_tool/blob/main/sample.jsonl\n你也可以仿照openai的playground编写，参考 https://platform.openai.com/playground/p/default-chat\n即使是多轮对话也必须写在一行，用`\\n`表示换行，如果是不同对话或主题，则另起一行",
+  "Size mismatch for blocks. You are attempting to continue training from the LoRA model, but it does not match the base model. Please set LoRA model to None.": "尺寸不匹配块。你正在尝试从LoRA模型继续训练，但该LoRA模型与基底模型不匹配，请将LoRA模型设为空",
+  "Instruction: Write a story using the following information\n\nInput: A man named Alex chops a tree down\n\nResponse:": "Instruction: Write a story using the following information\n\nInput: 艾利克斯砍倒了一棵树\n\nResponse:",
+  "Composition": "作曲",
+  "Use Local Sound Font": "使用本地音色资源",
+  "Auto Play At The End": "结束时自动播放",
+  "No File to save": "无文件可保存",
+  "File Saved": "文件已保存",
+  "Failed to load local sound font, please check if the files exist - assets/sound-font": "加载本地音色资源失败，请检查文件是否存在 - assets/sound-font",
+  "Please convert model to safe tensors format first": "请先将模型转换为Safetensors格式",
+  "Convert To Safe Tensors Format": "转换为Safetensors格式",
+  "Please change Strategy to WebGPU to use safetensors format": "请将Strategy改为WebGPU以使用safetensors格式",
+  "Preview Only": "仅预览",
+  "RAM": "内存",
+  "VRAM": "显存",
+  "GPU Usage": "GPU占用",
+  "Use Custom Tokenizer": "使用自定义Tokenizer",
+  "Tokenizer Path (e.g. backend-python/rwkv_pip/20B_tokenizer.json)": "Tokenizer路径 (例如: backend-python/rwkv_pip/20B_tokenizer.json)",
+  "User Name": "用户名称",
+  "Assistant Name": "AI名称",
+  "Insert default system prompt at the beginning": "在开头自动插入默认系统提示",
+  "Please Enable Custom CUDA Kernel. Latest RWKV-5 requires os.environ['RWKV_CUDA_ON'] == '1' (will fix soon).": "请启用自定义CUDA算子。最新的RWKV-5需要os.environ['RWKV_CUDA_ON'] == '1' (未来会修复)",
+  "Format Content": "规范格式"
 }
--- a/frontend/src/assets/images/logo.jpg
+++ b/frontend/src/assets/images/logo.jpg
--- a/frontend/src/assets/images/logo.png
+++ b/frontend/src/assets/images/logo.png
--- a/frontend/src/components/CopyButton.tsx
+++ b/frontend/src/components/CopyButton.tsx
@@ -4,7 +4,7 @@ import { useTranslation } from 'react-i18next';
 import { ClipboardSetText } from '../../wailsjs/runtime';
 import { ToolTipButton } from './ToolTipButton';

-export const CopyButton: FC<{ content: string }> = ({ content }) => {
+export const CopyButton: FC<{ content: string, showDelay?: number, }> = ({ content, showDelay = 0 }) => {
  const { t } = useTranslation();
  const [copied, setCopied] = useState(false);

@@ -19,7 +19,8 @@ export const CopyButton: FC<{ content: string }> = ({ content }) => {
  };

  return (
-    <ToolTipButton desc={t('Copy')} size="small" appearance="subtle" icon={copied ? <CheckIcon /> : <CopyIcon />}
+    <ToolTipButton desc={t('Copy')} size="small" appearance="subtle" showDelay={showDelay}
+      icon={copied ? <CheckIcon /> : <CopyIcon />}
      onClick={onClick} />
  );
 };
--- a/frontend/src/components/CustomToastContainer.tsx
+++ b/frontend/src/components/CustomToastContainer.tsx
@@ -0,0 +1,17 @@
+import commonStore from '../stores/commonStore';
+import { ToastContainer } from 'react-toastify';
+
+export const CustomToastContainer = () =>
+  <ToastContainer
+    style={{ width: '350px' }}
+    position="top-center"
+    autoClose={4000}
+    pauseOnHover={true}
+    hideProgressBar={true}
+    newestOnTop={true}
+    closeOnClick={false}
+    rtl={false}
+    pauseOnFocusLoss={false}
+    draggable={false}
+    theme={commonStore.settings.darkMode ? 'dark' : 'light'}
+  />;
--- a/frontend/src/components/DialogButton.tsx
+++ b/frontend/src/components/DialogButton.tsx
@@ -11,6 +11,7 @@ import {
 } from '@fluentui/react-components';
 import { ToolTipButton } from './ToolTipButton';
 import { useTranslation } from 'react-i18next';
+import MarkdownRender from './MarkdownRender';

 export const DialogButton: FC<{
  text?: string | null
@@ -19,12 +20,13 @@ export const DialogButton: FC<{
  className?: string,
  title: string,
  contentText: string,
-  onConfirm: () => void,
+  markdown?: boolean,
+  onConfirm?: () => void,
  size?: 'small' | 'medium' | 'large',
  shape?: 'rounded' | 'circular' | 'square',
  appearance?: 'secondary' | 'primary' | 'outline' | 'subtle' | 'transparent',
 }> = ({
-  text, icon, tooltip, className, title, contentText,
+  text, icon, tooltip, className, title, contentText, markdown,
  onConfirm, size, shape, appearance
 }) => {
  const { t } = useTranslation();
@@ -41,7 +43,11 @@ export const DialogButton: FC<{
      <DialogBody>
        <DialogTitle>{title}</DialogTitle>
        <DialogContent>
-          {contentText}
+          {
+            markdown ?
+              <MarkdownRender>{contentText}</MarkdownRender> :
+              contentText
+          }
        </DialogContent>
        <DialogActions>
          <DialogTrigger disableButtonEnhancement>
--- a/frontend/src/components/ReadButton.tsx
+++ b/frontend/src/components/ReadButton.tsx
@@ -7,13 +7,28 @@ import { observer } from 'mobx-react-lite';

 const synth = window.speechSynthesis;

-export const ReadButton: FC<{ content: string }> = observer(({ content }) => {
+export const ReadButton: FC<{
+  content: string,
+  inSpeaking?: boolean,
+  showDelay?: number,
+  setSpeakingOuter?: (speaking: boolean) => void
+}> = observer(({
+  content,
+  inSpeaking = false,
+  showDelay = 0,
+  setSpeakingOuter
+}) => {
  const { t } = useTranslation();
-  const [speaking, setSpeaking] = useState(false);
+  const [speaking, setSpeaking] = useState(inSpeaking);
  let lang: string = commonStore.settings.language;
  if (lang === 'dev')
    lang = 'en';

+  const setSpeakingInner = (speaking: boolean) => {
+    setSpeakingOuter?.(speaking);
+    setSpeaking(speaking);
+  };
+
  const startSpeak = () => {
    synth.cancel();

@@ -31,22 +46,22 @@ export const ReadButton: FC<{ content: string }> = observer(({ content }) => {
    Object.assign(utterance, {
      rate: 1,
      volume: 1,
-      onend: () => setSpeaking(false),
-      onerror: () => setSpeaking(false),
+      onend: () => setSpeakingInner(false),
+      onerror: () => setSpeakingInner(false),
      voice: voice
    });

    synth.speak(utterance);
-    setSpeaking(true);
+    setSpeakingInner(true);
  };

  const stopSpeak = () => {
    synth.cancel();
-    setSpeaking(false);
+    setSpeakingInner(false);
  };

  return (
-    <ToolTipButton desc={t('Read Aloud')} size="small" appearance="subtle"
+    <ToolTipButton desc={t('Read Aloud')} size="small" appearance="subtle" showDelay={showDelay}
      icon={speaking ? <MuteIcon /> : <UnmuteIcon />}
      onClick={speaking ? stopSpeak : startSpeak} />
  );
--- a/frontend/src/components/ResetConfigsButton.tsx
+++ b/frontend/src/components/ResetConfigsButton.tsx
@@ -4,14 +4,14 @@ import { useTranslation } from 'react-i18next';
 import { ArrowReset20Regular } from '@fluentui/react-icons';
 import commonStore from '../stores/commonStore';

-import { defaultModelConfigs, defaultModelConfigsMac } from '../pages/defaultModelConfigs';
+import { defaultModelConfigs, defaultModelConfigsMac } from '../pages/defaultConfigs';

 export const ResetConfigsButton: FC<{ afterConfirm?: () => void }> = ({ afterConfirm }) => {
  const { t } = useTranslation();
  return <DialogButton icon={<ArrowReset20Regular />} tooltip={t('Reset All Configs')} title={t('Reset All Configs')}
    contentText={t('Are you sure you want to reset all configs? This will obtain the latest preset configs, but will override your custom configs and cannot be undone.')}
    onConfirm={() => {
-      commonStore.setModelConfigs(commonStore.platform != 'darwin' ? defaultModelConfigs : defaultModelConfigsMac, false);
+      commonStore.setModelConfigs(commonStore.platform !== 'darwin' ? defaultModelConfigs : defaultModelConfigsMac, false);
      commonStore.setCurrentConfigIndex(0, true);
      afterConfirm?.();
    }} />;
--- a/frontend/src/components/RunButton.tsx
+++ b/frontend/src/components/RunButton.tsx
@@ -1,23 +1,16 @@
 import React, { FC, MouseEventHandler, ReactElement } from 'react';
 import commonStore, { ModelStatus } from '../stores/commonStore';
-import {
-  AddToDownloadList,
-  CopyFile,
-  DepCheck,
-  FileExists,
-  InstallPyDep,
-  StartServer
-} from '../../wailsjs/go/backend_golang/App';
+import { AddToDownloadList, FileExists, StartServer, StartWebGPUServer } from '../../wailsjs/go/backend_golang/App';
 import { Button } from '@fluentui/react-components';
 import { observer } from 'mobx-react-lite';
 import { exit, getStatus, readRoot, switchModel, updateConfig } from '../apis';
 import { toast } from 'react-toastify';
-import { getStrategy, getSupportedCustomCudaFile, saveCache, toastWithButton } from '../utils';
+import { checkDependencies, getStrategy, toastWithButton } from '../utils';
 import { useTranslation } from 'react-i18next';
 import { ToolTipButton } from './ToolTipButton';
 import { Play16Regular, Stop16Regular } from '@fluentui/react-icons';
 import { useNavigate } from 'react-router';
-import { BrowserOpenURL, WindowShow } from '../../wailsjs/runtime/runtime';
+import { WindowShow } from '../../wailsjs/runtime/runtime';

 const mainButtonText = {
  [ModelStatus.Offline]: 'Run',
@@ -46,6 +39,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
      commonStore.setStatus({ status: ModelStatus.Starting });

      const modelConfig = commonStore.getCurrentModelConfig();
+      const webgpu = modelConfig.modelParameters.device === 'WebGPU';
      let modelName = '';
      let modelPath = '';
      if (modelConfig && modelConfig.modelParameters) {
@@ -57,52 +51,31 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
        return;
      }

-      if (!commonStore.depComplete) {
-        let depErrorMsg = '';
-        await DepCheck(commonStore.settings.customPythonPath).catch((e) => {
-          depErrorMsg = e.message || e;
-          WindowShow();
-          if (depErrorMsg === 'python zip not found') {
-            toastWithButton(t('Python target not found, would you like to download it?'), t('Download'), () => {
-              toastWithButton(`${t('Downloading')} Python`, t('Check'), () => {
-                navigate({ pathname: '/downloads' });
-              }, { autoClose: 3000 });
-              AddToDownloadList('python-3.10.11-embed-amd64.zip', 'https://www.python.org/ftp/python/3.10.11/python-3.10.11-embed-amd64.zip');
-            });
-          } else if (depErrorMsg.includes('DepCheck Error')) {
-            if (depErrorMsg.includes('vc_redist')) {
-              toastWithButton(t('Microsoft Visual C++ Redistributable is not installed, would you like to download it?'), t('Download'), () => {
-                BrowserOpenURL('https://aka.ms/vs/16/release/vc_redist.x64.exe');
-              });
-            } else {
-              toast(depErrorMsg, { type: 'info', position: 'bottom-left' });
-              if (commonStore.platform != 'linux')
-                toastWithButton(t('Python dependencies are incomplete, would you like to install them?'), t('Install'), () => {
-                  InstallPyDep(commonStore.settings.customPythonPath, commonStore.settings.cnMirror).catch((e) => {
-                    const errMsg = e.message || e;
-                    toast(t('Error') + ' - ' + errMsg, { type: 'error' });
-                  });
-                  setTimeout(WindowShow, 1000);
-                }, {
-                  autoClose: 8000
-                });
-              else
-                toastWithButton(t('On Linux system, you must manually install python dependencies.'), t('Check'), () => {
-                  BrowserOpenURL('https://github.com/josStorer/RWKV-Runner/blob/master/build/linux/Readme_Install.txt');
-                });
-            }
+      if (webgpu) {
+        if (!['.st', '.safetensors'].some(ext => modelPath.endsWith(ext))) {
+          const stModelPath = modelPath.replace(/\.pth$/, '.st');
+          if (await FileExists(stModelPath)) {
+            modelPath = stModelPath;
          } else {
-            toast(depErrorMsg, { type: 'error' });
+            toast(t('Please convert model to safe tensors format first'), { type: 'error' });
+            commonStore.setStatus({ status: ModelStatus.Offline });
+            return;
          }
-        });
-        if (depErrorMsg) {
+        }
+      }
+
+      if (!webgpu) {
+        if (['.st', '.safetensors'].some(ext => modelPath.endsWith(ext))) {
+          toast(t('Please change Strategy to WebGPU to use safetensors format'), { type: 'error' });
          commonStore.setStatus({ status: ModelStatus.Offline });
          return;
        }
-        commonStore.setDepComplete(true);
-        if (commonStore.platform === 'windows')
-          CopyFile('./backend-python/wkv_cuda_utils/wkv_cuda_model.py', './py310/Lib/site-packages/rwkv/model.py');
-        saveCache();
+      }
+
+      if (!webgpu) {
+        const ok = await checkDependencies(navigate);
+        if (!ok)
+          return;
      }

      const currentModelSource = commonStore.modelSourceList.find(item => item.name === modelName);
@@ -126,7 +99,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
        showDownloadPrompt(t('Model file not found'), modelName);
        commonStore.setStatus({ status: ModelStatus.Offline });
        return;
-      } else if (!currentModelSource?.isLocal) {
+      } else if (!currentModelSource?.isComplete) {
        showDownloadPrompt(t('Model file download is not complete'), modelName);
        commonStore.setStatus({ status: ModelStatus.Offline });
        return;
@@ -136,7 +109,15 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean

      await exit(1000).catch(() => {
      });
-      StartServer(commonStore.settings.customPythonPath, port, commonStore.settings.host !== '127.0.0.1' ? '0.0.0.0' : '127.0.0.1').catch((e) => {
+
+      const startServer = webgpu ?
+        (_: string, port: number, host: string) => StartWebGPUServer(port, host)
+        : StartServer;
+      const isUsingCudaBeta = modelConfig.modelParameters.device === 'CUDA-Beta';
+
+      startServer(commonStore.settings.customPythonPath, port, commonStore.settings.host !== '127.0.0.1' ? '0.0.0.0' : '127.0.0.1',
+        isUsingCudaBeta
+      ).catch((e) => {
        const errMsg = e.message || e;
        if (errMsg.includes('path contains space'))
          toast(`${t('Error')} - ${t('File Path Cannot Contain Space')}`, { type: 'error' });
@@ -153,40 +134,49 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
          if (r.ok && !loading) {
            loading = true;
            clearInterval(intervalId);
-            await getStatus().then(status => {
-              if (status)
-                commonStore.setStatus(status);
-            });
+            if (!webgpu) {
+              await getStatus().then(status => {
+                if (status)
+                  commonStore.setStatus(status);
+              });
+            }
            commonStore.setStatus({ status: ModelStatus.Loading });
            toast(t('Loading Model'), { type: 'info' });
-            updateConfig({
-              max_tokens: modelConfig.apiParameters.maxResponseToken,
-              temperature: modelConfig.apiParameters.temperature,
-              top_p: modelConfig.apiParameters.topP,
-              presence_penalty: modelConfig.apiParameters.presencePenalty,
-              frequency_penalty: modelConfig.apiParameters.frequencyPenalty
-            });
+            if (!webgpu) {
+              updateConfig({
+                max_tokens: modelConfig.apiParameters.maxResponseToken,
+                temperature: modelConfig.apiParameters.temperature,
+                top_p: modelConfig.apiParameters.topP,
+                presence_penalty: modelConfig.apiParameters.presencePenalty,
+                frequency_penalty: modelConfig.apiParameters.frequencyPenalty
+              });
+            }

+            const strategy = getStrategy(modelConfig);
            let customCudaFile = '';
-            if ((modelConfig.modelParameters.device === 'CUDA' || modelConfig.modelParameters.device === 'Custom')
-              && modelConfig.modelParameters.useCustomCuda) {
+            if ((modelConfig.modelParameters.device.includes('CUDA') || modelConfig.modelParameters.device === 'Custom')
+              && modelConfig.modelParameters.useCustomCuda && !strategy.includes('fp32')) {
              if (commonStore.platform === 'windows') {
-                customCudaFile = getSupportedCustomCudaFile();
-                if (customCudaFile) {
-                  FileExists('./py310/Lib/site-packages/rwkv/model.py').then((exist) => {
-                    // defensive measure. As Python has already been launched, will only take effect the next time it runs.
-                    if (!exist) CopyFile('./backend-python/wkv_cuda_utils/wkv_cuda_model.py', './py310/Lib/site-packages/rwkv/model.py');
-                  });
-                  await CopyFile(customCudaFile, './py310/Lib/site-packages/rwkv/wkv_cuda.pyd').catch(() => {
-                    FileExists('./py310/Lib/site-packages/rwkv/wkv_cuda.pyd').then((exist) => {
-                      if (!exist) {
-                        customCudaFile = '';
-                        toast(t('Failed to copy custom cuda file'), { type: 'error' });
-                      }
-                    });
-                  });
-                } else
-                  toast(t('Supported custom cuda file not found'), { type: 'warning' });
+                // this part is currently unused because there's no longer a need to use different kernels for different GPUs, but it might still be needed in the future
+                //
+                // customCudaFile = getSupportedCustomCudaFile(isUsingCudaBeta);
+                // if (customCudaFile) {
+                //   let kernelTargetPath: string;
+                //   if (isUsingCudaBeta)
+                //     kernelTargetPath = './backend-python/rwkv_pip/beta/wkv_cuda.pyd';
+                //   else
+                //     kernelTargetPath = './backend-python/rwkv_pip/wkv_cuda.pyd';
+                //   await CopyFile(customCudaFile, kernelTargetPath).catch(() => {
+                //     FileExists(kernelTargetPath).then((exist) => {
+                //       if (!exist) {
+                //         customCudaFile = '';
+                //         toast(t('Failed to copy custom cuda file'), { type: 'error' });
+                //       }
+                //     });
+                //   });
+                // } else
+                //   toast(t('Supported custom cuda file not found'), { type: 'warning' });
+                customCudaFile = 'any';
              } else {
                customCudaFile = 'any';
              }
@@ -194,23 +184,43 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean

            switchModel({
              model: modelPath,
-              strategy: getStrategy(modelConfig),
+              strategy: strategy,
+              tokenizer: modelConfig.modelParameters.useCustomTokenizer ? modelConfig.modelParameters.customTokenizer : undefined,
              customCuda: customCudaFile !== ''
            }).then(async (r) => {
              if (r.ok) {
                commonStore.setStatus({ status: ModelStatus.Working });
-                toastWithButton(t('Startup Completed'), t('Chat'), () => {
-                  navigate({ pathname: '/chat' });
-                }, { type: 'success', autoClose: 3000 });
+                let buttonNameMap = {
+                  'novel': 'Completion',
+                  'midi': 'Composition'
+                };
+                let buttonName = 'Chat';
+                buttonName = Object.entries(buttonNameMap).find(([key, value]) => modelName.toLowerCase().includes(key))?.[1] || buttonName;
+                const buttonFn = () => {
+                  navigate({ pathname: '/' + buttonName.toLowerCase() });
+                };
+                toastWithButton(t('Startup Completed'), t(buttonName), buttonFn, { type: 'success', autoClose: 3000 });
              } else if (r.status === 304) {
                toast(t('Loading Model'), { type: 'info' });
              } else {
                commonStore.setStatus({ status: ModelStatus.Offline });
-                toast(t('Failed to switch model') + ' - ' + await r.text(), { type: 'error' });
+                const error = await r.text();
+                const errorsMap = {
+                  'not enough memory': 'Memory is not enough, try to increase the virtual memory or use a smaller model.',
+                  'not compiled with CUDA': 'Bad PyTorch version, please reinstall PyTorch with cuda.',
+                  'invalid header or archive is corrupted': 'The model file is corrupted, please download again.',
+                  'no NVIDIA driver': 'Found no NVIDIA driver, please install the latest driver.',
+                  'CUDA out of memory': 'VRAM is not enough, please reduce stored layers or use a lower precision in Configs page.',
+                  'Ninja is required to load C++ extensions': 'Failed to enable custom CUDA kernel, ninja is required to load C++ extensions. You may be using the CPU version of PyTorch, please reinstall PyTorch with CUDA. Or if you are using a custom Python interpreter, you must compile the CUDA kernel by yourself or disable Custom CUDA kernel acceleration.',
+                  'Please Enable Custom CUDA Kernel': 'Please Enable Custom CUDA Kernel. Latest RWKV-5 requires os.environ[\'RWKV_CUDA_ON\'] == \'1\' (will fix soon).'
+                };
+                const matchedError = Object.entries(errorsMap).find(([key, _]) => error.includes(key));
+                const message = matchedError ? t(matchedError[1]) : error;
+                toast(t('Failed to switch model') + ' - ' + message, { autoClose: 5000, type: 'error' });
              }
            }).catch((e) => {
              commonStore.setStatus({ status: ModelStatus.Offline });
-              toast(t('Failed to switch model') + ' - ' + e.message || e, { type: 'error' });
+              toast(t('Failed to switch model') + ' - ' + (e.message || e), { type: 'error' });
            });
          }
        }).catch(() => {
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
josc146	4f26404002	release v1.4.8	2023-10-03 15:05:13 +08:00
josc146	df7652856a	completion page: add format content button	2023-10-03 14:54:36 +08:00
josc146	de755463e3	improve overflow	2023-10-03 14:27:44 +08:00
josc146	2fe98d9a2c	add rwkv5 cuda kernel error prompt	2023-10-03 14:25:31 +08:00
josc146	2e42039607	chore	2023-10-03 14:04:46 +08:00
josc146	71abd357a4	update startup	2023-10-03 13:50:58 +08:00
josc146	68228a4552	rwkv5 pre-compiled kernel (for windows)	2023-10-03 13:39:07 +08:00
josc146	79851433f8	upgrade rwkv pip (0.8.13)	2023-10-03 13:33:55 +08:00
github-actions[bot]	bd4de12e05	release v1.4.7	2023-09-18 15:04:47 +00:00
josc146	c0aa6aaba9	release v1.4.7	2023-09-18 23:03:54 +08:00
josc146	d7abe5f0d1	add pre-compiled beta cuda kernel (rwkv-beta==0.8.5, 40%+ faster for fp16) (thanks to #180 , pre-compiled kernel of RTX 40 Series will be included later)	2023-09-18 23:02:49 +08:00
josc146	5e5e1e9651	custom tokenizer .txt support	2023-09-18 17:20:55 +08:00
github-actions[bot]	f8388a0527	release v1.4.6	2023-09-16 05:06:08 +00:00
josc146	f8b764ef8f	release v1.4.6	2023-09-16 13:05:34 +08:00
josc146	fcfaa5944e	frontend feature adaptation for api params (user_name, assistant_name, presystem)	2023-09-16 13:02:06 +08:00
josc146	f89e89c1c9	chore	2023-09-16 12:23:16 +08:00
josc146	a25965530c	custom tokenizer (#77 )	2023-09-16 00:34:11 +08:00
josc146	971124d0d7	upgrade to wails@v2.6.0 (EnableDefaultContextMenu: true)	2023-09-16 00:29:45 +08:00
josc146	d7dcc90008	chore	2023-09-15 16:31:14 +08:00
josc146	df969fcfc6	upgrade cuda-beta	2023-09-15 16:30:11 +08:00
josc146	c4042bbfd8	improve ui desc	2023-09-15 16:26:32 +08:00
josc146	4112200b4c	revert(2d5456): refresh local models when download complete (for macOS)	2023-09-15 16:25:04 +08:00
Ikko Eltociear Ashimine	3f9a54e36f	Update README_JA.md add translation.	2023-09-13 16:11:43 +08:00
github-actions[bot]	3ed4456135	release v1.4.5	2023-08-27 15:57:18 +00:00
josc146	e0df9ae47b	release v1.4.5	2023-08-27 23:56:37 +08:00
josc146	87b2c3ed7d	fix build	2023-08-27 23:56:30 +08:00
josc146	50ff7ef6bc	always use requirements.txt	2023-08-27 23:52:52 +08:00
josc146	c7a580ca8a	update manifest	2023-08-27 23:16:56 +08:00
josc146	eaae7624a7	add HardwareMonitor (Windows Only)	2023-08-27 22:53:18 +08:00
josc146	fcd59de6fb	correct Preset UI description	2023-08-27 21:37:32 +08:00
josc146	1bbe127209	fix webgpu_server file permissions of linux and macos	2023-08-27 21:22:26 +08:00
josc146	b868adc058	chore	2023-08-27 21:21:34 +08:00
josc146	a24b78e8c3	python-backend: extra ChatCompletionBody params (`raw`, `presystem`); add default_stop when stop is null	2023-08-27 21:21:11 +08:00
josc146	c8025f1cff	allow message content to be empty	2023-08-27 21:02:54 +08:00
josc146	fe0860dbf0	fix lora finetune max_epochs (#170 )	2023-08-24 22:49:57 +08:00
josc146	02d5d641d1	chore	2023-08-24 22:48:54 +08:00
github-actions[bot]	a057bb6c5b	release v1.4.4	2023-08-16 15:33:53 +00:00
josc146	c9e4ae7fa1	release v1.4.4	2023-08-16 23:33:22 +08:00
josc146	79a97b2bc4	webgpu release support	2023-08-16 23:31:04 +08:00
josc146	ef53951a16	webgpu support	2023-08-16 23:07:58 +08:00
josc146	74f1a1c033	chore	2023-08-16 21:11:58 +08:00
josc146	ce986cfc6d	chore	2023-08-16 12:50:22 +08:00
josc146	61cea2a784	add misc API (`/models` and `/dashboard/billing/credit_grants`)	2023-08-14 23:37:55 +08:00
josc146	8a13bd3c1e	add rwkv-cuda-beta support (faster)	2023-08-14 22:07:15 +08:00
josc146	da68926e9c	chore (AddStateBody class)	2023-08-13 21:27:29 +08:00
josc146	e0b7453883	allow multiple systems	2023-08-04 22:27:55 +08:00
josc146	91e2828a95	allow completions input to be null	2023-08-04 22:22:59 +08:00
github-actions[bot]	bcf6409536	release v1.4.3	2023-07-31 14:51:01 +00:00
josc146	d7d4f87620	release v1.4.3	2023-07-31 22:50:29 +08:00
josc146	b3e35a4cdd	allow custom user_name and assistant_name (`/chat/completions` API)	2023-07-31 22:48:54 +08:00
josc146	8764c37b03	RWKVType	2023-07-31 22:46:13 +08:00
josc146	d12a173f39	global penalty	2023-07-31 22:02:28 +08:00
josc146	64fa939c19	japanese UI chore	2023-07-29 21:44:33 +08:00
josc146	9c8e7b2f08	japanese UI	2023-07-29 21:19:45 +08:00
josc146	abfd668523	update defaultConfigs	2023-07-29 19:41:54 +08:00
github-actions[bot]	ebacf383f5	release v1.4.2	2023-07-29 11:34:18 +00:00
josc146	eb25dc6bcb	release v1.4.2	2023-07-29 19:33:52 +08:00
josc146	aecacde819	remove `response` field of completions api	2023-07-29 19:20:43 +08:00
josc146	3ef22239eb	improve default ChatCompletion `stop`	2023-07-29 19:19:38 +08:00
josc146	719090cc8c	improve python backend startup speed	2023-07-29 19:18:01 +08:00
josc146	dbb8374d89	update defaultConfigs	2023-07-29 19:16:44 +08:00
github-actions[bot]	4d875a8c00	release v1.4.1	2023-07-28 14:16:37 +00:00
josc146	30b6d66a2d	release v1.4.1	2023-07-28 22:14:53 +08:00
josc146	9d89b6f4db	fix params	2023-07-28 22:13:19 +08:00
josc146	d2928e54f7	fix `failed to build cyac`	2023-07-28 21:40:17 +08:00
josc146	49ba5c97f7	update readme	2023-07-28 13:13:14 +08:00
github-actions[bot]	4054fac359	release v1.4.0	2023-07-28 05:06:42 +00:00
josc146	dfae1d9645	release v1.4.0	2023-07-28 13:05:55 +08:00
josc146	0f16a0dd1b	remove LoraFinetunePrecision fp32	2023-07-28 12:53:41 +08:00
josc146	cb05a8a2ae	update manifest	2023-07-28 12:50:39 +08:00
josc146	a51385173c	add CPU-120M-Music config	2023-07-28 12:45:31 +08:00
josc146	4e18222a35	improve RunButton prompt	2023-07-28 12:45:13 +08:00
josc146	daabcf58a0	add Composition Page (RWKV-Music)	2023-07-28 12:30:05 +08:00
josc146	d0fd480bd6	chore	2023-07-26 22:24:26 +08:00
josc146	1df345b5eb	improve embeddings API results	2023-07-25 20:30:43 +08:00
josc146	77868c798b	chore	2023-07-25 16:37:06 +08:00
josc146	f56748a941	improve python backend startup speed	2023-07-25 16:14:29 +08:00
josc146	29c5b1d804	add midi api	2023-07-25 16:11:17 +08:00
josc146	34095a6c36	support for `stop` array	2023-07-25 16:10:22 +08:00
josc146	05b9b42b56	add support for MIDI RWKV	2023-07-25 16:09:31 +08:00
josc146	211ae342af	improve sse fetch	2023-07-25 15:59:37 +08:00
josc146	5ae683e915	update presets	2023-07-25 15:53:25 +08:00
josc146	dc59fb39c7	update readme	2023-07-18 14:21:09 +08:00
josc146	49960774ee	update readme	2023-07-18 14:16:50 +08:00
github-actions[bot]	b718452618	release v1.3.9	2023-07-17 05:05:17 +00:00
josc146	15ae312b37	release v1.3.9	2023-07-17 13:03:32 +08:00
josc146	6938b5b20e	change chinese translation of `completion`	2023-07-17 13:03:11 +08:00
josc146	9b3b06ab04	fix input with array type (#96 , #107 )	2023-07-17 12:59:45 +08:00
josc146	e2a7c93753	fix always show `Convert Failed` when converting model	2023-07-16 16:54:18 +08:00
github-actions[bot]	34349aee0b	release v1.3.8	2023-07-15 14:29:14 +00:00
josc146	8e79370e95	release v1.3.8	2023-07-15 22:28:49 +08:00
josc146	652c35322b	save conversation as txt (originally in md)	2023-07-15 22:12:59 +08:00
josc146	e2fc57ac24	training: fix data EOL format	2023-07-11 12:19:39 +08:00
josc146	994fc7c828	fix cross-device state cache exception	2023-07-11 11:20:12 +08:00
josc146	b9a960d984	update readme	2023-07-10 23:06:19 +08:00
josc146	3baf260f4d	update readme	2023-07-10 22:59:22 +08:00
github-actions[bot]	d037ded146	release v1.3.7	2023-07-10 13:50:05 +00:00
josc146	622287f3da	release v1.3.7	2023-07-10 21:49:33 +08:00
josc146	5d12bf74f6	update presets	2023-07-10 21:43:58 +08:00
josc146	c88f9321f5	update manifest	2023-07-10 20:49:31 +08:00
josc146	f9f1d5c9fc	improve `/completions` api compatibility	2023-07-10 20:45:08 +08:00
josc146	0edec68376	improve training data path compatibility	2023-07-10 20:44:09 +08:00
josc146	ee63dc25f4	update readme	2023-07-09 13:56:36 +08:00
josc146	fee8fe73f2	fix loss parser	2023-07-09 13:33:06 +08:00
github-actions[bot]	1689f9e7e7	release v1.3.6	2023-07-09 04:41:11 +00:00
josc146	a1ed0cb2e9	release v1.3.6	2023-07-09 12:40:42 +08:00
josc146	5ee5fa7e6e	fix load_state_dict crash	2023-07-09 12:33:29 +08:00
josc146	d8c70453ec	format	2023-07-09 12:32:50 +08:00
josc146	e930eb5967	extra vc check	2023-07-09 12:18:51 +08:00
josc146	aec6ad636a	chore	2023-07-09 12:10:14 +08:00
josc146	750c91bd3e	update logo	2023-07-09 11:59:23 +08:00
josc146	fcc3886db1	improve error messages for training	2023-07-09 11:39:44 +08:00
josc146	22afc98be5	fix loss parser	2023-07-09 11:32:05 +08:00
josc146	5b1a9448e6	fix jsonl data when using directory as training data	2023-07-09 11:31:07 +08:00
github-actions[bot]	07d89e3eeb	release v1.3.5	2023-07-07 13:58:33 +00:00
josc146	96e97d9c1e	release v1.3.5	2023-07-07 21:58:08 +08:00
josc146	bcb125e168	support using directory as training data	2023-07-07 21:57:01 +08:00
josc146	6fbb86667c	improve python script error messages	2023-07-07 20:16:35 +08:00
josc146	2d545604f4	refresh local models in real-time (#98 )	2023-07-07 20:14:55 +08:00
josc146	7210a7481e	improve finetune guide	2023-07-07 19:10:31 +08:00
josc146	55210c89e2	improve wsl dependencies installation	2023-07-07 18:57:51 +08:00
josc146	c725d11dd9	fix loss parser	2023-07-07 13:56:08 +08:00
josc146	ba2a6bd06c	update Related Repositories	2023-07-07 13:54:57 +08:00
josc146	57b80c6ed0	fix build for macos and linux	2023-07-07 13:54:07 +08:00
josc146	115c59d5e1	chore	2023-07-07 13:53:39 +08:00
github-actions[bot]	543ff468b7	release v1.3.4	2023-07-03 14:32:06 +00:00
josc146	96ae47989e	release v1.3.4	2023-07-03 22:31:37 +08:00
josc146	368932a610	improve finetune compatibility	2023-07-03 22:28:01 +08:00
josc146	f2cd531fcb	fix build for macos and linux	2023-07-03 22:22:55 +08:00
josc146	511652b71c	improve finetune compatibility	2023-07-03 22:19:20 +08:00
github-actions[bot]	525fb132d6	release v1.3.3	2023-07-03 13:40:51 +00:00
josc146	5acb1fd958	release v1.3.3	2023-07-03 21:40:22 +08:00
josc146	76761ee453	improve lora finetune process (need to be refactored)	2023-07-03 21:40:16 +08:00
github-actions[bot]	134b2884e6	release v1.3.2	2023-07-03 09:43:01 +00:00
josc146	261e7c8916	release v1.3.2	2023-07-03 17:42:28 +08:00
josc146	987854fe49	lora finetune (need to be refactored)	2023-07-03 17:41:47 +08:00
josc146	c54d10795f	chore	2023-07-03 16:42:11 +08:00
github-actions[bot]	b7d9ab0845	release v1.3.1	2023-07-01 11:35:45 +00:00
josc146	176800444a	release v1.3.1	2023-07-01 19:35:20 +08:00
josc146	00c13cfc3f	improve compatibility for linux	2023-07-01 19:32:58 +08:00
Ikko Eltociear Ashimine	620e0228ed	Add Japanese README (#100 ) * Add Japanese README * minor fix	2023-06-30 12:37:45 +08:00
josc146	87ca694b0b	chore	2023-06-29 20:14:52 +08:00
josc146	417389c5f6	improve for python3.8 3.9	2023-06-29 20:12:11 +08:00
github-actions[bot]	fa9f62b42c	release v1.3.0	2023-06-28 13:26:51 +00:00
josc146	2c4e9f69eb	release v1.3.0	2023-06-28 21:26:23 +08:00
josc146	119204368d	update manifest	2023-06-28 20:57:09 +08:00
josc146	87a86042d2	chore	2023-06-28 20:49:41 +08:00
josc146	32c386799d	Change chat saving format	2023-06-28 20:48:22 +08:00
josc146	b56a55e81d	Completion Regenerate Button	2023-06-28 20:46:21 +08:00
josc146	2fe7a23049	chore	2023-06-28 19:40:55 +08:00
josc146	9ed3547738	rwkv pip 0.8.0	2023-06-28 19:36:15 +08:00
github-actions[bot]	a0522594da	release v1.2.9	2023-06-24 16:12:53 +00:00
josc146	1cac147df4	release v1.2.9	2023-06-25 00:12:20 +08:00
josc146	db67f30082	feat: chat presets (experimental)	2023-06-25 00:07:14 +08:00
josc146	08cf09416a	chore	2023-06-24 23:57:49 +08:00
josc146	7f2f4f15c1	improve error messages	2023-06-23 16:32:05 +08:00
josc146	97f6af595e	display models that have not been fully downloaded in Downloads page, even if the program is restarted	2023-06-23 16:03:57 +08:00
josc146	447f4572b1	improve error messages	2023-06-23 13:55:45 +08:00
github-actions[bot]	5c9b4a4c05	release v1.2.8	2023-06-21 15:12:45 +00:00
josc146	70f2271b94	release v1.2.8	2023-06-21 23:12:17 +08:00
josc146	15cd689741	adjust MoreUtilsButton	2023-06-21 23:11:22 +08:00
josc146	82a68593bb	exact avoidOverflow	2023-06-21 23:08:34 +08:00
github-actions[bot]	21910af96a	release v1.2.7	2023-06-21 14:09:27 +00:00
josc146	412a0fe135	release v1.2.7	2023-06-21 22:08:57 +08:00
josc146	cf0972ba52	avoid overflow	2023-06-21 22:02:42 +08:00
josc146	3fe9ef4546	chore	2023-06-21 22:00:29 +08:00
josc146	4cd5a56070	add more chat utils (retry, edit, delete)	2023-06-21 21:20:21 +08:00
josc146	35a7437714	chore	2023-06-21 17:13:04 +08:00
josc146	131a7ddf4a	fix the prompt cache that contains potential error	2023-06-21 16:07:16 +08:00
josc146	1465908574	update SupportedCustomCuda	2023-06-21 13:48:09 +08:00
josc146	3eb10f08bb	rename 100+ Languages to Global Languages	2023-06-21 12:44:49 +08:00
josc146	b20990d380	when precision is fp32, disable customCuda	2023-06-21 12:14:11 +08:00
josc146	1a5bf4a95e	improve InstallPyDep for non-english path	2023-06-21 12:08:04 +08:00
github-actions[bot]	3d123524e7	release v1.2.6	2023-06-20 16:59:55 +00:00