release v1.6.6

add translation.

.gitattributes

@@ -1,8 +1,11 @@
+* text=auto eol=lf
+
 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/convert_pytorch_to_ggml.py linguist-vendored
 backend-python/utils/midi.py linguist-vendored
 build/** linguist-vendored
 finetune/lora/** linguist-vendored

.github/dependabot.yml

@@ -0,0 +1,9 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    commit-message:
+      prefix: "chore"
+      include: "scope"

.github/workflows/release.yml

@@ -48,29 +48,27 @@ 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/ai00_rwkv_server/releases/latest/download/webgpu_server_windows_x86_64.exe ./backend-rust/webgpu_server.exe
+          Start-BitsTransfer https://github.com/josStorer/web-rwkv-converter/releases/latest/download/web-rwkv-converter_windows_x86_64.exe ./backend-rust/web-rwkv-converter.exe
+          Start-BitsTransfer https://github.com/josStorer/LibreHardwareMonitor.Console/releases/latest/download/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 {$_}} | Set-Content ./py310/python310._pth
           ./py310/python ./backend-python/get-pip.py
-          ./py310/python -m pip install Cython==0.29.36
+          ./py310/python -m pip install Cython==3.0.4
           Copy-Item -Path "${{ 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==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 ..
+          ./py310/python -m pip install cyac==1.9
           go install github.com/wailsapp/wails/v2/cmd/wails@latest
+          del ./backend-python/rwkv_pip/cpp/librwkv.dylib
+          del ./backend-python/rwkv_pip/cpp/librwkv.so
+          (Get-Content -Path ./backend-golang/app.go) -replace "//go:custom_build windows ", "" | Set-Content -Path ./backend-golang/app.go
+          (Get-Content -Path ./backend-golang/utils.go) -replace "//go:custom_build windows ", "" | Set-Content -Path ./backend-golang/utils.go
           make
           Rename-Item -Path "build/bin/RWKV-Runner.exe" -NewName "RWKV-Runner_windows_x64.exe"
 
@@ -86,29 +84,21 @@ 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: |
+          wget https://github.com/josStorer/ai00_rwkv_server/releases/latest/download/webgpu_server_linux_x86_64 -O ./backend-rust/webgpu_server
+          wget https://github.com/josStorer/web-rwkv-converter/releases/latest/download/web-rwkv-converter_linux_x86_64 -O ./backend-rust/web-rwkv-converter
           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 ..
+          sudo apt-get install build-essential libgtk-3-dev libwebkit2gtk-4.0-dev libasound2-dev
           go install github.com/wailsapp/wails/v2/cmd/wails@latest
-          rm -rf ./backend-python/wkv_cuda_utils
+          rm ./backend-python/rwkv_pip/wkv_cuda.pyd
+          rm ./backend-python/rwkv_pip/rwkv5.pyd
+          rm ./backend-python/rwkv_pip/rwkv6.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
+          rm ./backend-python/rwkv_pip/cpp/librwkv.dylib
+          rm ./backend-python/rwkv_pip/cpp/rwkv.dll
+          rm ./backend-python/rwkv_pip/webgpu/web_rwkv_py.cp310-win_amd64.pyd
           make
           mv build/bin/RWKV-Runner build/bin/RWKV-Runner_linux_x64
 
@@ -124,23 +114,18 @@ 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 ..
+          wget https://github.com/josStorer/ai00_rwkv_server/releases/latest/download/webgpu_server_darwin_aarch64 -O ./backend-rust/webgpu_server
+          wget https://github.com/josStorer/web-rwkv-converter/releases/latest/download/web-rwkv-converter_darwin_aarch64 -O ./backend-rust/web-rwkv-converter
           go install github.com/wailsapp/wails/v2/cmd/wails@latest
-          rm -rf ./backend-python/wkv_cuda_utils
+          rm ./backend-python/rwkv_pip/wkv_cuda.pyd
+          rm ./backend-python/rwkv_pip/rwkv5.pyd
+          rm ./backend-python/rwkv_pip/rwkv6.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
+          rm ./backend-python/rwkv_pip/cpp/rwkv.dll
+          rm ./backend-python/rwkv_pip/cpp/librwkv.so
+          rm ./backend-python/rwkv_pip/webgpu/web_rwkv_py.cp310-win_amd64.pyd
           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

.gitignore

@@ -8,6 +8,7 @@ __pycache__
 *.st
 *.safetensors
 *.bin
+*.mid
 /config.json
 /cache.json
 /presets.json
@@ -18,6 +19,7 @@ __pycache__
 /cmd-helper.bat
 /install-py-dep.bat
 /backend-python/wkv_cuda
+/backend-python/rwkv*
 *.exe
 *.old
 .DS_Store
@@ -26,3 +28,4 @@ __pycache__
 train_log.txt
 finetune/json2binidx_tool/data
 /wsl.state
+/components

CURRENT_CHANGE.md

@@ -1,14 +1,17 @@
 ## Changes
 
-- webgpu support (AMD, Intel, Nvidia, Apple)
-- add rwkv-cuda-beta support (faster)
-- add misc API (`/models` and `/dashboard/billing/credit_grants`)
-- allow multiple systems
-- allow completions input to be null
+- improve refreshRemoteModels
+- reduce precompiled web_rwkv_py size
+- webgpu(Python) max_buffer_size (12B support) and turbo
+- improve role-playing effect
+- update manifest.json (a lot of new models)
+- bump webgpu(ai00_server) mode to v0.3.8
+- improve details
 
 ## Install
 
 - Windows: https://github.com/josStorer/RWKV-Runner/blob/master/build/windows/Readme_Install.txt
 - MacOS: https://github.com/josStorer/RWKV-Runner/blob/master/build/darwin/Readme_Install.txt
 - Linux: https://github.com/josStorer/RWKV-Runner/blob/master/build/linux/Readme_Install.txt
-- Server-Deploy-Examples: https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples
+- Simple Deploy Example: https://github.com/josStorer/RWKV-Runner/blob/master/README.md#simple-deploy-example
+- Server Deploy Examples: https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples

Makefile

@@ -8,16 +8,26 @@ endif
 
 build-windows:
 	@echo ---- build for windows
-	wails build -upx -ldflags "-s -w" -platform windows/amd64
+	wails build -upx -ldflags '-s -w -extldflags "-static"' -platform windows/amd64
 
 build-macos:
 	@echo ---- build for macos
-	wails build -ldflags "-s -w" -platform darwin/universal
+	wails build -ldflags '-s -w' -platform darwin/universal
 
 build-linux:
 	@echo ---- build for linux
-	wails build -upx -ldflags "-s -w" -platform linux/amd64
+	wails build -upx -ldflags '-s -w' -platform linux/amd64
+
+build-web:
+	@echo ---- build for web
+	cd frontend && npm run build
 
 dev:
 	wails dev
 
+dev-web:
+	cd frontend && npm run dev
+
+preview:
+	cd frontend && npm run preview
+

README.md

@@ -21,7 +21,7 @@ English | [简体中文](README_ZH.md) | [日本語](README_JA.md)
 [![MacOS][MacOS-image]][MacOS-url]
 [![Linux][Linux-image]][Linux-url]
 
-[FAQs](https://github.com/josStorer/RWKV-Runner/wiki/FAQs) | [Preview](#Preview) | [Download][download-url] | [Server-Deploy-Examples](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples)
+[FAQs](https://github.com/josStorer/RWKV-Runner/wiki/FAQs) | [Preview](#Preview) | [Download][download-url] | [Simple Deploy Example](#Simple-Deploy-Example) | [Server Deploy Examples](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples) | [MIDI Hardware Input](#MIDI-Input)
 
 [license-image]: http://img.shields.io/badge/license-MIT-blue.svg
 
@@ -47,28 +47,74 @@ English | [简体中文](README_ZH.md) | [日本語](README_JA.md)
 
 </div>
 
-#### 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`.
+## Tips
 
-#### 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`).
+- You can deploy [backend-python](./backend-python/) on a server and use this program as a client only. Fill in
+  your server address in the Settings `API URL`.
 
-#### 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.
+- If you are deploying and providing public services, please limit the request size through API gateway to prevent
+  excessive resource usage caused by submitting overly long prompts. Additionally, please restrict the upper limit of
+  requests' max_tokens based on your actual
+  situation: https://github.com/josStorer/RWKV-Runner/blob/master/backend-python/utils/rwkv.py#L567, the default is set
+  as le=102400, which may result in significant resource consumption for individual responses in extreme cases.
+
+- Default configs has enabled custom CUDA kernel acceleration, which is much faster and consumes much less VRAM. If you
+  encounter possible compatibility issues (output garbled), go to the Configs page and turn
+  off `Use Custom CUDA kernel to Accelerate`, or try to upgrade your gpu driver.
+
+- 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.
 
 ## Features
 
-- RWKV model management and one-click startup
-- Fully compatible with the OpenAI API, making every ChatGPT client an RWKV client. After starting the model,
+- RWKV model management and one-click startup.
+- Front-end and back-end separation, if you don't want to use the client, also allows for separately deploying the
+  front-end service, or the back-end inference service, or the back-end inference service with a WebUI.
+  [Simple Deploy Example](#Simple-Deploy-Example) | [Server Deploy Examples](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples)
+- Compatible with the OpenAI API, making every ChatGPT client an RWKV client. After starting the model,
   open http://127.0.0.1:8000/docs to view more details.
-- Automatic dependency installation, requiring only a lightweight executable program
-- Configs with 2G to 32G VRAM are included, works well on almost all computers
-- User-friendly chat and completion interaction interface included
-- 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
+- Automatic dependency installation, requiring only a lightweight executable program.
+- Pre-set multi-level VRAM configs, works well on almost all computers. In Configs page, switch Strategy to WebGPU, it
+  can also run on AMD, Intel, and other graphics cards.
+- User-friendly chat, completion, and composition interaction interface included. Also supports chat presets, attachment
+  uploads, MIDI hardware input, and track editing.
+  [Preview](#Preview) | [MIDI Hardware Input](#MIDI-Input)
+- Built-in WebUI option, one-click start of Web service, sharing your hardware resources.
+- Easy-to-understand and operate parameter configuration, along with various operation guidance prompts.
+- Built-in model conversion tool.
+- Built-in download management and remote model inspection.
+- Built-in one-click LoRA Finetune. (Windows Only)
+- Can also be used as an OpenAI ChatGPT and GPT-Playground client. (Fill in the API URL and API Key in Settings page)
+- Multilingual localization.
+- Theme switching.
+- Automatic updates.
+
+## Simple Deploy Example
+
+```bash
+git clone https://github.com/josStorer/RWKV-Runner
+
+# Then
+cd RWKV-Runner
+python ./backend-python/main.py #The backend inference service has been started, request /switch-model API to load the model, refer to the API documentation: http://127.0.0.1:8000/docs
+
+# Or
+cd RWKV-Runner/frontend
+npm ci
+npm run build #Compile the frontend
+cd ..
+python ./backend-python/webui_server.py #Start the frontend service separately
+# Or
+python ./backend-python/main.py --webui #Start the frontend and backend service at the same time
+
+# Help Info
+python ./backend-python/main.py -h
+```
 
 ## API Concurrency Stress Testing
 
@@ -131,40 +177,98 @@ for i in np.argsort(embeddings_cos_sim)[::-1]:
     print(f"{embeddings_cos_sim[i]:.10f} - {values[i]}")
 ```
 
+## MIDI Input
+
+Tip: You can download https://github.com/josStorer/sgm_plus and unzip it to the program's `assets/sound-font` directory
+to use it as an offline sound source. Please note that if you are compiling the program from source code, do not place
+it in the source code directory.
+
+If you don't have a MIDI keyboard, you can use virtual MIDI input software like `Virtual Midi Controller 3 LE`, along
+with [loopMIDI](https://www.tobias-erichsen.de/wp-content/uploads/2020/01/loopMIDISetup_1_0_16_27.zip), to use a regular
+computer keyboard as MIDI input.
+
+### USB MIDI Connection
+
+- USB MIDI devices are plug-and-play, and you can select your input device in the Composition page
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/13bb92c3-4504-482d-ab82-026ac6c31095)
+
+### Mac MIDI Bluetooth Connection
+
+- For Mac users who want to use Bluetooth input,
+  please install [Bluetooth MIDI Connect](https://apps.apple.com/us/app/bluetooth-midi-connect/id1108321791), then click
+  the tray icon to connect after launching,
+  afterwards, you can select your input device in the Composition page.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c079a109-1e3d-45c1-bbf5-eed85da1550e)
+
+### Windows MIDI Bluetooth Connection
+
+- Windows seems to have implemented Bluetooth MIDI support only for UWP (Universal Windows Platform) apps. Therefore, it
+  requires multiple steps to establish a connection. We need to create a local virtual MIDI device and then launch a UWP
+  application. Through this UWP application, we will redirect Bluetooth MIDI input to the virtual MIDI device, and then
+  this software will listen to the input from the virtual MIDI device.
+- So, first, you need to
+  download [loopMIDI](https://www.tobias-erichsen.de/wp-content/uploads/2020/01/loopMIDISetup_1_0_16_27.zip)
+  to create a virtual MIDI device. Click the plus sign in the bottom left corner to create the device.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b75998ff-115c-4ddd-b97c-deeb5c106255)
+- Next, you need to download [Bluetooth LE Explorer](https://apps.microsoft.com/detail/9N0ZTKF1QD98) to discover and
+  connect to Bluetooth MIDI devices. Click "Start" to search for devices, and then click "Pair" to bind the MIDI device.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c142c3ea-a973-4531-9807-4c385d640a2b)
+- Finally, you need to install [MIDIberry](https://apps.microsoft.com/detail/9N39720H2M05),
+  This UWP application can redirect Bluetooth MIDI input to the virtual MIDI device. After launching it, double-click
+  your actual Bluetooth MIDI device name in the input field, and in the output field, double-click the virtual MIDI
+  device name we created earlier.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/5ad6a1d9-4f68-4d95-ae17-4296107d1669)
+- Now, you can select the virtual MIDI device as the input in the Composition page. Bluetooth LE Explorer no longer
+  needs to run, and you can also close the loopMIDI window, it will run automatically in the background. Just keep
+  MIDIberry open.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/1c371821-c7b7-4c18-8e42-9e315efbe427)
+
 ## Related Repositories:
 
+- RWKV-5-World: https://huggingface.co/BlinkDL/rwkv-5-world/tree/main
 - 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
+- ai00_rwkv_server: https://github.com/cgisky1980/ai00_rwkv_server
+- rwkv.cpp: https://github.com/saharNooby/rwkv.cpp
+- web-rwkv-py: https://github.com/cryscan/web-rwkv-py
 
 ## Preview
 
 ### Homepage
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/d7f24d80-f382-428d-8b28-edf87e1549e2)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c9b9cdd0-63f9-4319-9f74-5bf5d7df5a67)
 
 ### Chat
 
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/80009872-528f-4932-aeb2-f724fa892e7c)
 
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/e98c9038-3323-47b0-8edb-d639fafd37b2)
+
 ### Completion
 
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/bf49de8e-3b89-4543-b1ef-7cd4b19a1836)
 
 ### Composition
 
+Tip: You can download https://github.com/josStorer/sgm_plus and unzip it to the program's `assets/sound-font` directory
+to use it as an offline sound source. Please note that if you are compiling the program from source code, do not place
+it in the source code directory.
+
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/e8ad908d-3fd2-4e92-bcdb-96815cb836ee)
 
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b2ce4761-9e75-477e-a182-d0255fb8ac76)
+
 ### Configuration
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/48befdc6-e03c-4851-9bee-22f77ee2640e)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/f41060dc-5517-44af-bb3f-8ef71720016d)
 
 ### Model Management
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/367fe4f8-cc12-475f-9371-3cf62cdbf293)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b1581147-a6ce-4493-8010-e33c0ddeca0a)
 
 ### Download Management
 

README_JA.md

@@ -21,7 +21,7 @@
 [![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)
+[FAQs](https://github.com/josStorer/RWKV-Runner/wiki/FAQs) | [プレビュー](#Preview) | [ダウンロード][download-url] | [シンプルなデプロイの例](#Simple-Deploy-Example) | [サーバーデプロイ例](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples) | [MIDIハードウェア入力](#MIDI-Input)
 
 [license-image]: http://img.shields.io/badge/license-MIT-blue.svg
 
@@ -47,29 +47,71 @@
 
 </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。
+## ヒント
 
-#### 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`)。
+- サーバーに [backend-python](./backend-python/)
+  をデプロイし、このプログラムをクライアントとして使用することができます。設定された`API URL`にサーバーアドレスを入力してください。
 
-#### 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。
+- もし、あなたがデプロイし、外部に公開するサービスを提供している場合、APIゲートウェイを使用してリクエストのサイズを制限し、
+  長すぎるプロンプトの提出がリソースを占有しないようにしてください。さらに、実際の状況に応じて、リクエストの max_tokens
+  の上限を制限してください：https://github.com/josStorer/RWKV-Runner/blob/master/backend-python/utils/rwkv.py#L567
+  、デフォルトは le=102400 ですが、極端な場合には単一の応答が大量のリソースを消費する可能性があります。
+
+- 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 (文字化けを出力する)
+  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、あるいは、GPUドライバーをアップグレードしてみてください。
+
+- 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、
+- フロントエンドとバックエンドの分離は、クライアントを使用しない場合でも、フロントエンドサービス、またはバックエンド推論サービス、またはWebUIを備えたバックエンド推論サービスを個別に展開することを可能にします。
+  [シンプルなデプロイの例](#Simple-Deploy-Example) | [サーバーデプロイ例](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples)
+- OpenAI API と互換性があり、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
+- 事前設定された多段階のVRAM設定、ほとんどのコンピュータで動作します。配置ページで、ストラテジーをWebGPUに切り替えると、AMD、インテル、その他のグラフィックカードでも動作します
+- ユーザーフレンドリーなチャット、完成、および作曲インターフェイスが含まれています。また、チャットプリセット、添付ファイルのアップロード、MIDIハードウェア入力、トラック編集もサポートしています。
+  [プレビュー](#Preview) | [MIDIハードウェア入力](#MIDI-Input)
+- 内蔵WebUIオプション、Webサービスのワンクリック開始、ハードウェアリソースの共有
+- 分かりやすく操作しやすいパラメータ設定、各種操作ガイダンスプロンプトとともに
 - 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
+- 内蔵のLoRA微調整機能を搭載しています (Windowsのみ)
+- 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（設定ページで `API URL` と `API Key`
+  を入力してください）
 - 多言語ローカライズ
 - テーマ切り替え
 - 自動アップデート
 
+## Simple Deploy Example
+
+```bash
+git clone https://github.com/josStorer/RWKV-Runner
+
+# Then
+cd RWKV-Runner
+python ./backend-python/main.py #The backend inference service has been started, request /switch-model API to load the model, refer to the API documentation: http://127.0.0.1:8000/docs
+
+# Or
+cd RWKV-Runner/frontend
+npm ci
+npm run build #Compile the frontend
+cd ..
+python ./backend-python/webui_server.py #Start the frontend service separately
+# Or
+python ./backend-python/main.py --webui #Start the frontend and backend service at the same time
+
+# Help Info
+python ./backend-python/main.py -h
+```
+
 ## API 同時実行ストレステスト
 
 ```bash
@@ -91,8 +133,8 @@ body.json:
 
 ## 埋め込み API の例
 
-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.
+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
@@ -132,40 +174,98 @@ for i in np.argsort(embeddings_cos_sim)[::-1]:
     print(f"{embeddings_cos_sim[i]:.10f} - {values[i]}")
 ```
 
+## MIDI Input
+
+Tip: You can download https://github.com/josStorer/sgm_plus and unzip it to the program's `assets/sound-font` directory
+to use it as an offline sound source. Please note that if you are compiling the program from source code, do not place
+it in the source code directory.
+
+MIDIキーボードをお持ちでない場合、`Virtual Midi Controller 3 LE`
+などの仮想MIDI入力ソフトウェアを使用することができます。[loopMIDI](https://www.tobias-erichsen.de/wp-content/uploads/2020/01/loopMIDISetup_1_0_16_27.zip)
+を組み合わせて、通常のコンピュータキーボードをMIDI入力として使用できます。
+
+### USB MIDI Connection
+
+- USB MIDI devices are plug-and-play, and you can select your input device in the Composition page
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/13bb92c3-4504-482d-ab82-026ac6c31095)
+
+### Mac MIDI Bluetooth Connection
+
+- For Mac users who want to use Bluetooth input,
+  please install [Bluetooth MIDI Connect](https://apps.apple.com/us/app/bluetooth-midi-connect/id1108321791), then click
+  the tray icon to connect after launching,
+  afterwards, you can select your input device in the Composition page.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c079a109-1e3d-45c1-bbf5-eed85da1550e)
+
+### Windows MIDI Bluetooth Connection
+
+- Windows seems to have implemented Bluetooth MIDI support only for UWP (Universal Windows Platform) apps. Therefore, it
+  requires multiple steps to establish a connection. We need to create a local virtual MIDI device and then launch a UWP
+  application. Through this UWP application, we will redirect Bluetooth MIDI input to the virtual MIDI device, and then
+  this software will listen to the input from the virtual MIDI device.
+- So, first, you need to
+  download [loopMIDI](https://www.tobias-erichsen.de/wp-content/uploads/2020/01/loopMIDISetup_1_0_16_27.zip)
+  to create a virtual MIDI device. Click the plus sign in the bottom left corner to create the device.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b75998ff-115c-4ddd-b97c-deeb5c106255)
+- Next, you need to download [Bluetooth LE Explorer](https://apps.microsoft.com/detail/9N0ZTKF1QD98) to discover and
+  connect to Bluetooth MIDI devices. Click "Start" to search for devices, and then click "Pair" to bind the MIDI device.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c142c3ea-a973-4531-9807-4c385d640a2b)
+- Finally, you need to install [MIDIberry](https://apps.microsoft.com/detail/9N39720H2M05),
+  This UWP application can redirect Bluetooth MIDI input to the virtual MIDI device. After launching it, double-click
+  your actual Bluetooth MIDI device name in the input field, and in the output field, double-click the virtual MIDI
+  device name we created earlier.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/5ad6a1d9-4f68-4d95-ae17-4296107d1669)
+- Now, you can select the virtual MIDI device as the input in the Composition page. Bluetooth LE Explorer no longer
+  needs to run, and you can also close the loopMIDI window, it will run automatically in the background. Just keep
+  MIDIberry open.
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/1c371821-c7b7-4c18-8e42-9e315efbe427)
+
 ## 関連リポジトリ:
 
+- RWKV-5-World: https://huggingface.co/BlinkDL/rwkv-5-world/tree/main
 - 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
+- ai00_rwkv_server: https://github.com/cgisky1980/ai00_rwkv_server
+- rwkv.cpp: https://github.com/saharNooby/rwkv.cpp
+- web-rwkv-py: https://github.com/cryscan/web-rwkv-py
 
-## プレビュー
+## Preview
 
 ### ホームページ
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/d7f24d80-f382-428d-8b28-edf87e1549e2)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c9b9cdd0-63f9-4319-9f74-5bf5d7df5a67)
 
 ### チャット
 
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/80009872-528f-4932-aeb2-f724fa892e7c)
 
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/e98c9038-3323-47b0-8edb-d639fafd37b2)
+
 ### 補完
 
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/bf49de8e-3b89-4543-b1ef-7cd4b19a1836)
 
 ### 作曲
 
+Tip: You can download https://github.com/josStorer/sgm_plus and unzip it to the program's `assets/sound-font` directory
+to use it as an offline sound source. Please note that if you are compiling the program from source code, do not place
+it in the source code directory.
+
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/e8ad908d-3fd2-4e92-bcdb-96815cb836ee)
 
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b2ce4761-9e75-477e-a182-d0255fb8ac76)
+
 ### コンフィグ
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/48befdc6-e03c-4851-9bee-22f77ee2640e)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/f41060dc-5517-44af-bb3f-8ef71720016d)
 
 ### モデル管理
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/367fe4f8-cc12-475f-9371-3cf62cdbf293)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b1581147-a6ce-4493-8010-e33c0ddeca0a)
 
 ### ダウンロード管理
 

README_ZH.md

@@ -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/1zdzZ_a0uM3gDqi6pXIZVAA?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) | [简明服务部署示例](#Simple-Deploy-Example) | [服务器部署示例](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples) | [MIDI硬件输入](#MIDI-Input)
 
 [license-image]: http://img.shields.io/badge/license-MIT-blue.svg
 
@@ -46,28 +46,65 @@ API兼容的接口，这意味着一切ChatGPT客户端都是RWKV客户端。
 
 </div>
 
-#### 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`
+## 小贴士
 
-#### 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`)
+- 你可以在服务器部署[backend-python](./backend-python/)，然后将此程序仅用作客户端，在设置的`API URL`中填入你的服务器地址
 
-#### 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
+- 如果你正在部署并对外提供公开服务，请通过API网关限制请求大小，避免过长的prompt提交占用资源。此外，请根据你的实际情况，限制请求的
+  max_tokens 上限: https://github.com/josStorer/RWKV-Runner/blob/master/backend-python/utils/rwkv.py#L567,
+  默认le=102400, 这可能导致极端情况下单个响应消耗大量资源
+
+- 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(输出乱码)
+  问题，前往配置页面，关闭`使用自定义CUDA算子加速`，或更新你的显卡驱动
+
+- 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
 
 ## 功能
 
 - RWKV模型管理，一键启动
-- 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
+- 前后端分离，如果你不想使用客户端，也允许单独部署前端服务，或后端推理服务，或具有WebUI的后端推理服务。
+  [简明服务部署示例](#Simple-Deploy-Example) | [服务器部署示例](https://github.com/josStorer/RWKV-Runner/tree/master/deploy-examples)
+- 与OpenAI API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API文档
 - 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续写交互页面
-- 易于理解和操作的参数配置
+- 预设多级显存配置，几乎在各种电脑上工作良好。通过配置页面切换Strategy到WebGPU，还可以在AMD，Intel等显卡上运行
+- 自带用户友好的聊天，续写，作曲交互页面。支持聊天预设，附件上传，MIDI硬件输入及音轨编辑。
+  [预览](#Preview) | [MIDI硬件输入](#MIDI-Input)
+- 内置WebUI选项，一键启动Web服务，共享硬件资源
+- 易于理解和操作的参数配置，及各类操作引导提示
 - 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
+- 内置一键LoRA微调 (仅限Windows)
+- 也可用作 OpenAI ChatGPT 和 GPT Playground 客户端 (在设置内填写API URL和API Key)
 - 多语言本地化
 - 主题切换
 - 自动更新
 
+## Simple Deploy Example
+
+```bash
+git clone https://github.com/josStorer/RWKV-Runner
+
+# 然后
+cd RWKV-Runner
+python ./backend-python/main.py #后端推理服务已启动, 调用/switch-model载入模型, 参考API文档: http://127.0.0.1:8000/docs
+
+# 或者
+cd RWKV-Runner/frontend
+npm ci
+npm run build #编译前端
+cd ..
+python ./backend-python/webui_server.py #单独启动前端服务
+# 或者
+python ./backend-python/main.py --webui #同时启动前后端服务
+
+# 帮助参数
+python ./backend-python/main.py -h
+```
+
 ## API并发压力测试
 
 ```bash
@@ -128,40 +165,88 @@ for i in np.argsort(embeddings_cos_sim)[::-1]:
     print(f"{embeddings_cos_sim[i]:.10f} - {values[i]}")
 ```
 
+## MIDI Input
+
+小贴士: 你可以下载 https://github.com/josStorer/sgm_plus, 并解压到程序的`assets/sound-font`目录, 以使用离线音源. 注意,
+如果你正在从源码编译程序, 请不要将其放置在源码目录中
+
+如果你没有MIDI键盘, 你可以使用像 `Virtual Midi Controller 3 LE` 这样的虚拟MIDI输入软件,
+配合[loopMIDI](https://www.tobias-erichsen.de/wp-content/uploads/2020/01/loopMIDISetup_1_0_16_27.zip), 使用普通电脑键盘作为MIDI输入
+
+### USB MIDI 连接
+
+- USB MIDI设备是即插即用的, 你能够在作曲页面选择你的输入设备
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/a448c34a-56d8-46eb-8dc2-dd11e8e0c4ce)
+
+### Mac MIDI 蓝牙连接
+
+- 对于想要使用蓝牙输入的Mac用户,
+  请安装[Bluetooth MIDI Connect](https://apps.apple.com/us/app/bluetooth-midi-connect/id1108321791), 启动后点击托盘连接,
+  之后你可以在作曲页面选择你的输入设备
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c079a109-1e3d-45c1-bbf5-eed85da1550e)
+
+### Windows MIDI 蓝牙连接
+
+- Windows似乎只为UWP实现了蓝牙MIDI支持, 因此需要多个步骤进行连接, 我们需要创建一个本地的虚拟MIDI设备, 然后启动一个UWP应用,
+  通过此UWP应用将蓝牙MIDI输入重定向到虚拟MIDI设备, 然后本软件监听虚拟MIDI设备的输入
+- 因此, 首先你需要下载[loopMIDI](https://www.tobias-erichsen.de/wp-content/uploads/2020/01/loopMIDISetup_1_0_16_27.zip),
+  用于创建虚拟MIDI设备, 点击左下角的加号创建设备
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/b75998ff-115c-4ddd-b97c-deeb5c106255)
+- 然后, 你需要下载[Bluetooth LE Explorer](https://apps.microsoft.com/detail/9N0ZTKF1QD98), 以发现并连接蓝牙MIDI设备,
+  点击Start搜索设备, 然后点击Pair绑定MIDI设备
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/c142c3ea-a973-4531-9807-4c385d640a2b)
+- 最后, 你需要安装[MIDIberry](https://apps.microsoft.com/detail/9N39720H2M05), 这个UWP应用能将MIDI蓝牙输入重定向到虚拟MIDI设备,
+  启动后, 在输入栏, 双击你实际的蓝牙MIDI设备名称, 在输出栏, 双击我们先前创建的虚拟MIDI设备名称
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/5ad6a1d9-4f68-4d95-ae17-4296107d1669)
+- 现在, 你可以在作曲页面选择虚拟MIDI设备作为输入. Bluetooth LE Explorer不再需要运行, loopMIDI窗口也可以退出, 它会自动在后台运行,
+  仅保持MIDIberry打开即可
+- ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/6460c355-884e-4b28-a2eb-8ab7a2e3a01a)
+
 ## 相关仓库:
 
+- RWKV-5-World: https://huggingface.co/BlinkDL/rwkv-5-world/tree/main
 - 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
+- ai00_rwkv_server: https://github.com/cgisky1980/ai00_rwkv_server
+- rwkv.cpp: https://github.com/saharNooby/rwkv.cpp
+- web-rwkv-py: https://github.com/cryscan/web-rwkv-py
 
 ## Preview
 
 ### 主页
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/ff2b1eef-dd3b-4cbf-98fb-b5a1ecee43e1)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/cd82674e-3ee3-4175-bd9c-a11d45437327)
 
 ### 聊天
 
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/9570e73b-dca2-4316-9e92-09961f3c48c4)
 
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/54bb0e2b-cdc4-4ea0-8d16-9beaf57c232c)
+
 ### 续写
 
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/69f9ba7a-2fe8-4a5e-94cb-aa655aa409e2)
 
 ### 作曲
 
+小贴士: 你可以下载 https://github.com/josStorer/sgm_plus, 并解压到程序的`assets/sound-font`目录, 以使用离线音源. 注意,
+如果你正在从源码编译程序, 请不要将其放置在源码目录中
+
 ![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/95b34893-80c2-4706-87f9-bc141032ed4b)
 
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/3cb31ca8-d708-42f1-8768-1605fb0b2174)
+
 ### 配置
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/59460f69-b172-4c7a-86cb-573262543076)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/0f4d4f21-8abe-4f4d-8c4f-cd7d5607f20e)
 
 ### 模型管理
 
-![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/551121ee-1bfe-421b-a9d1-24125126ab4b)
+![image](https://github.com/josStorer/RWKV-Runner/assets/13366013/871f2d2a-7e41-4be7-9b32-be1b3e00dc3e)
 
 ### 下载管理
 

backend-golang/app.go

@@ -1,13 +1,17 @@
 package backend_golang
 
 import (
+	"bufio"
 	"context"
 	"errors"
+	"io"
 	"net/http"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"syscall"
+	"time"
 
 	"github.com/fsnotify/fsnotify"
 	"github.com/minio/selfupdate"
@@ -41,20 +45,37 @@ func (a *App) OnStartup(ctx context.Context) {
 		a.cmdPrefix = "cd " + a.exDir + " && "
 	}
 
+	os.Chmod(a.exDir+"backend-rust/webgpu_server", 0777)
+	os.Chmod(a.exDir+"backend-rust/web-rwkv-converter", 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()
+	trainLogPath := a.exDir + "lora-models/train_log.txt"
+	if !a.FileExists(trainLogPath) {
+		f, err := os.Create(trainLogPath)
+		if err == nil {
+			f.Close()
+		}
 	}
 
 	a.downloadLoop()
+	a.midiLoop()
+	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")
+		watcher.Add(a.exDir + "./lora-models")
+		watcher.Add(a.exDir + "./models")
 		go func() {
 			for {
 				select {
@@ -62,7 +83,7 @@ func (a *App) OnStartup(ctx context.Context) {
 					if !ok {
 						return
 					}
-					wruntime.EventsEmit(ctx, "fsnotify", event.Name)
+					wruntime.EventsEmit(a.ctx, "fsnotify", event.Name)
 				case _, ok := <-watcher.Errors:
 					if !ok {
 						return
@@ -73,13 +94,81 @@ func (a *App) OnStartup(ctx context.Context) {
 	}
 }
 
+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()
+}
+
+type ProgressReader struct {
+	reader io.Reader
+	total  int64
+	err    error
+}
+
+func (pr *ProgressReader) Read(p []byte) (n int, err error) {
+	n, err = pr.reader.Read(p)
+	pr.err = err
+	pr.total += int64(n)
+	return
+}
+
 func (a *App) UpdateApp(url string) (broken bool, err error) {
 	resp, err := http.Get(url)
 	if err != nil {
 		return false, err
 	}
 	defer resp.Body.Close()
-	err = selfupdate.Apply(resp.Body, selfupdate.Options{})
+	pr := &ProgressReader{reader: resp.Body}
+
+	ticker := time.NewTicker(250 * time.Millisecond)
+	defer ticker.Stop()
+
+	go func() {
+		for {
+			<-ticker.C
+			wruntime.EventsEmit(a.ctx, "updateApp", &DownloadStatus{
+				Name:        filepath.Base(url),
+				Path:        "",
+				Url:         url,
+				Transferred: pr.total,
+				Size:        resp.ContentLength,
+				Speed:       0,
+				Progress:    100 * (float64(pr.total) / float64(resp.ContentLength)),
+				Downloading: pr.err == nil && pr.total < resp.ContentLength,
+				Done:        pr.total == resp.ContentLength,
+			})
+			if pr.err != nil || pr.total == resp.ContentLength {
+				break
+			}
+		}
+	}()
+	err = selfupdate.Apply(pr, selfupdate.Options{})
 	if err != nil {
 		if rerr := selfupdate.RollbackError(err); rerr != nil {
 			return true, rerr

backend-golang/download.go

@@ -33,9 +33,9 @@ type DownloadStatus struct {
 
 var downloadList []*DownloadStatus
 
-func existsInDownloadList(url string) bool {
+func existsInDownloadList(path string, url string) bool {
 	for _, ds := range downloadList {
-		if ds.Url == url {
+		if ds.Path == path || ds.Url == url {
 			return true
 		}
 	}
@@ -88,7 +88,7 @@ func (a *App) ContinueDownload(url string) {
 }
 
 func (a *App) AddToDownloadList(path string, url string) {
-	if !existsInDownloadList(url) {
+	if !existsInDownloadList(a.exDir+path, url) {
 		downloadList = append(downloadList, &DownloadStatus{
 			resp:        nil,
 			Name:        filepath.Base(path),

backend-golang/file.go

@@ -14,6 +14,13 @@ import (
 	wruntime "github.com/wailsapp/wails/v2/pkg/runtime"
 )
 
+func (a *App) SaveFile(path string, savedContent []byte) error {
+	if err := os.WriteFile(a.exDir+path, savedContent, 0644); err != nil {
+		return err
+	}
+	return nil
+}
+
 func (a *App) SaveJson(fileName string, jsonData any) error {
 	text, err := json.MarshalIndent(jsonData, "", "  ")
 	if err != nil {
@@ -53,12 +60,12 @@ type FileInfo struct {
 	ModTime string `json:"modTime"`
 }
 
-func (a *App) ReadFileInfo(fileName string) (FileInfo, error) {
+func (a *App) ReadFileInfo(fileName string) (*FileInfo, error) {
 	info, err := os.Stat(a.exDir + fileName)
 	if err != nil {
-		return FileInfo{}, err
+		return nil, err
 	}
-	return FileInfo{
+	return &FileInfo{
 		Name:    info.Name(),
 		Size:    info.Size(),
 		IsDir:   info.IsDir(),
@@ -145,6 +152,20 @@ func (a *App) OpenSaveFileDialogBytes(filterPattern string, defaultFileName stri
 	return path, nil
 }
 
+// Only return the path of the selected file, because communication between frontend and backend is slow. Use AssetServer Handler to read the file.
+func (a *App) OpenOpenFileDialog(filterPattern string) (string, error) {
+	path, err := wruntime.OpenFileDialog(a.ctx, wruntime.OpenDialogOptions{
+		Filters: []wruntime.FileFilter{{Pattern: filterPattern}},
+	})
+	if err != nil {
+		return "", err
+	}
+	if path == "" {
+		return "", nil
+	}
+	return path, nil
+}
+
 func (a *App) OpenFileFolder(path string, relative bool) error {
 	var absPath string
 	var err error
@@ -181,3 +202,12 @@ func (a *App) OpenFileFolder(path string, relative bool) error {
 	}
 	return errors.New("unsupported OS")
 }
+
+func (a *App) StartFile(path string) error {
+	cmd, err := CmdHelper(true, path)
+	if err != nil {
+		return err
+	}
+	err = cmd.Start()
+	return err
+}

backend-golang/midi.go

@@ -0,0 +1,170 @@
+package backend_golang
+
+import (
+	"errors"
+	"fmt"
+	"time"
+
+	"github.com/mattrtaylor/go-rtmidi"
+	"github.com/wailsapp/wails/v2/pkg/runtime"
+)
+
+type Port struct {
+	Name string `json:"name"`
+}
+type MIDIMessage struct {
+	MessageType string `json:"messageType"`
+	Channel     int    `json:"channel"`
+	Note        int    `json:"note"`
+	Velocity    int    `json:"velocity"`
+	Control     int    `json:"control"`
+	Value       int    `json:"value"`
+}
+
+var ports []Port
+var input rtmidi.MIDIIn
+var out rtmidi.MIDIOut
+var activeIndex int = -1
+var lastNoteTime time.Time
+
+func (a *App) midiLoop() {
+	var err error
+	input, err = rtmidi.NewMIDIInDefault()
+	if err != nil {
+		runtime.EventsEmit(a.ctx, "midiError", err.Error())
+		return
+	}
+	out, err = rtmidi.NewMIDIOutDefault()
+	if err != nil {
+		runtime.EventsEmit(a.ctx, "midiError", err.Error())
+	}
+	err = out.OpenPort(0, "")
+	if err != nil {
+		runtime.EventsEmit(a.ctx, "midiError", err.Error())
+	}
+	ticker := time.NewTicker(500 * time.Millisecond)
+	go func() {
+		for {
+			<-ticker.C
+			count, err := input.PortCount()
+			if err != nil {
+				continue
+			}
+			ports = make([]Port, count)
+			for i := 0; i < count; i++ {
+				name, err := input.PortName(i)
+				if err == nil {
+					ports[i].Name = name
+				}
+			}
+			runtime.EventsEmit(a.ctx, "midiPorts", &ports)
+		}
+	}()
+}
+
+func (a *App) OpenMidiPort(index int) error {
+	if input == nil {
+		return errors.New("failed to initialize MIDI input")
+	}
+	if activeIndex == index {
+		return nil
+	}
+	input.Destroy()
+	var err error
+	input, err = rtmidi.NewMIDIInDefault()
+	if err != nil {
+		return err
+	}
+	err = input.SetCallback(func(msg rtmidi.MIDIIn, bytes []byte, t float64) {
+		// https://www.midi.org/specifications-old/item/table-1-summary-of-midi-message
+		// https://www.rfc-editor.org/rfc/rfc6295.html
+		//
+		// msgType channel
+		//  1001     0000
+		//
+		msgType := bytes[0] >> 4
+		channel := bytes[0] & 0x0f
+		switch msgType {
+		case 0x8:
+			elapsed := time.Since(lastNoteTime)
+			lastNoteTime = time.Now()
+			runtime.EventsEmit(a.ctx, "midiMessage", &MIDIMessage{
+				MessageType: "ElapsedTime",
+				Value:       int(elapsed.Milliseconds()),
+			})
+			note := bytes[1]
+			runtime.EventsEmit(a.ctx, "midiMessage", &MIDIMessage{
+				MessageType: "NoteOff",
+				Channel:     int(channel),
+				Note:        int(note),
+			})
+		case 0x9:
+			elapsed := time.Since(lastNoteTime)
+			lastNoteTime = time.Now()
+			runtime.EventsEmit(a.ctx, "midiMessage", &MIDIMessage{
+				MessageType: "ElapsedTime",
+				Value:       int(elapsed.Milliseconds()),
+			})
+			note := bytes[1]
+			velocity := bytes[2]
+			runtime.EventsEmit(a.ctx, "midiMessage", &MIDIMessage{
+				MessageType: "NoteOn",
+				Channel:     int(channel),
+				Note:        int(note),
+				Velocity:    int(velocity),
+			})
+		case 0xb:
+			// control 12 => K1 knob, control 13 => K2 knob
+			control := bytes[1]
+			value := bytes[2]
+			runtime.EventsEmit(a.ctx, "midiMessage", &MIDIMessage{
+				MessageType: "ControlChange",
+				Channel:     int(channel),
+				Control:     int(control),
+				Value:       int(value),
+			})
+		default:
+			fmt.Printf("Unknown midi message: %v\n", bytes)
+		}
+	})
+	if err != nil {
+		return err
+	}
+	err = input.OpenPort(index, "")
+	if err != nil {
+		return err
+	}
+	activeIndex = index
+	lastNoteTime = time.Now()
+	return nil
+}
+
+func (a *App) CloseMidiPort() error {
+	if input == nil {
+		return errors.New("failed to initialize MIDI input")
+	}
+	if activeIndex == -1 {
+		return nil
+	}
+	activeIndex = -1
+	input.Destroy()
+	var err error
+	input, err = rtmidi.NewMIDIInDefault()
+	if err != nil {
+		return err
+	}
+	return nil
+}
+
+func (a *App) PlayNote(msg MIDIMessage) error {
+	if out == nil {
+		return errors.New("failed to initialize MIDI output")
+	}
+	channelByte := byte(msg.Channel)
+	if msg.MessageType == "NoteOn" {
+		out.SendMessage([]byte{0x90 | channelByte, byte(msg.Note), byte(msg.Velocity)})
+	} else if msg.MessageType == "NoteOff" {
+		out.SendMessage([]byte{0x80 | channelByte, byte(msg.Note), byte(msg.Velocity)})
+	}
+	return nil
+}

backend-golang/rwkv.go

@@ -10,7 +10,7 @@ import (
 	"strings"
 )
 
-func (a *App) StartServer(python string, port int, host string, rwkvBeta bool) (string, error) {
+func (a *App) StartServer(python string, port int, host string, webui bool, rwkvBeta bool, rwkvcpp bool, webgpu bool) (string, error) {
 	var err error
 	if python == "" {
 		python, err = GetPython()
@@ -19,17 +19,25 @@ func (a *App) StartServer(python string, port int, host string, rwkvBeta bool) (
 		return "", err
 	}
 	args := []string{python, "./backend-python/main.py"}
+	if webui {
+		args = append(args, "--webui")
+	}
 	if rwkvBeta {
 		args = append(args, "--rwkv-beta")
 	}
+	if rwkvcpp {
+		args = append(args, "--rwkv.cpp")
+	}
+	if webgpu {
+		args = append(args, "--webgpu")
+	}
 	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)
+	args = append(args, "--port", strconv.Itoa(port), "--ip", host)
 	return Cmd(args...)
 }
 
@@ -44,7 +52,13 @@ 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) {
+func (a *App) ConvertSafetensors(modelPath string, outPath string) (string, error) {
+	args := []string{"./backend-rust/web-rwkv-converter"}
+	args = append(args, "--input", modelPath, "--output", outPath)
+	return Cmd(args...)
+}
+
+func (a *App) ConvertSafetensorsWithPython(python string, modelPath string, outPath string) (string, error) {
 	var err error
 	if python == "" {
 		python, err = GetPython()
@@ -55,6 +69,21 @@ func (a *App) ConvertSafetensors(python string, modelPath string, outPath string
 	return Cmd(python, "./backend-python/convert_safetensors.py", "--input", modelPath, "--output", outPath)
 }
 
+func (a *App) ConvertGGML(python string, modelPath string, outPath string, Q51 bool) (string, error) {
+	var err error
+	if python == "" {
+		python, err = GetPython()
+	}
+	if err != nil {
+		return "", err
+	}
+	dataType := "FP16"
+	if Q51 {
+		dataType = "Q5_1"
+	}
+	return Cmd(python, "./backend-python/convert_pytorch_to_ggml.py", modelPath, outPath, dataType)
+}
+
 func (a *App) ConvertData(python string, input string, outputPrefix string, vocab string) (string, error) {
 	var err error
 	if python == "" {
@@ -149,13 +178,12 @@ func (a *App) InstallPyDep(python string, cnMirror bool) (string, error) {
 
 	if runtime.GOOS == "windows" {
 		ChangeFileLine("./py310/python310._pth", 3, "Lib\\site-packages")
-		installScript := python + " ./backend-python/get-pip.py -i https://pypi.tuna.tsinghua.edu.cn/simple\n" +
-			python + " -m pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 --index-url https://download.pytorch.org/whl/cu117\n" +
-			python + " -m pip install -r ./backend-python/requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple\n" +
+		installScript := python + " ./backend-python/get-pip.py -i https://pypi.tuna.tsinghua.edu.cn/simple --no-warn-script-location\n" +
+			python + " -m pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 --index-url https://download.pytorch.org/whl/cu117 --no-warn-script-location\n" +
+			python + " -m pip install -r ./backend-python/requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple --no-warn-script-location\n" +
 			"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 {

backend-golang/utils.go

@@ -3,42 +3,63 @@ package backend_golang
 import (
 	"archive/zip"
 	"bufio"
+	"crypto/sha256"
 	"embed"
 	"errors"
+	"fmt"
 	"io"
 	"io/fs"
+	"net"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
+	"syscall"
 )
 
+func CmdHelper(hideWindow bool, args ...string) (*exec.Cmd, error) {
+	if runtime.GOOS != "windows" {
+		return nil, errors.New("unsupported OS")
+	}
+	filename := "./cmd-helper.bat"
+	_, err := os.Stat(filename)
+	if err != nil {
+		if err := os.WriteFile(filename, []byte("start %*"), 0644); err != nil {
+			return nil, err
+		}
+	}
+	cmdHelper, err := filepath.Abs(filename)
+	if err != nil {
+		return nil, err
+	}
+
+	if strings.Contains(cmdHelper, " ") {
+		for _, arg := range args {
+			if strings.Contains(arg, " ") {
+				return nil, errors.New("path contains space") // golang bug https://github.com/golang/go/issues/17149#issuecomment-473976818
+			}
+		}
+	}
+	cmd := exec.Command(cmdHelper, args...)
+	cmd.SysProcAttr = &syscall.SysProcAttr{}
+	//go:custom_build windows cmd.SysProcAttr.HideWindow = hideWindow
+	return cmd, nil
+}
+
 func Cmd(args ...string) (string, error) {
 	switch platform := runtime.GOOS; platform {
 	case "windows":
-		if err := os.WriteFile("./cmd-helper.bat", []byte("start %*"), 0644); err != nil {
-			return "", err
-		}
-		cmdHelper, err := filepath.Abs("./cmd-helper")
+		cmd, err := CmdHelper(true, args...)
 		if err != nil {
 			return "", err
 		}
-
-		if strings.Contains(cmdHelper, " ") {
-			for _, arg := range args {
-				if strings.Contains(arg, " ") {
-					return "", errors.New("path contains space") // golang bug https://github.com/golang/go/issues/17149#issuecomment-473976818
-				}
-			}
-		}
-
-		cmd := exec.Command(cmdHelper, args...)
-		out, err := cmd.CombinedOutput()
+		_, err = cmd.CombinedOutput()
 		if err != nil {
 			return "", err
 		}
-		return string(out), nil
+		return "", nil
 	case "darwin":
 		ex, err := os.Executable()
 		if err != nil {
@@ -92,9 +113,19 @@ func CopyEmbed(efs embed.FS) error {
 			return err
 		}
 
-		err = os.WriteFile(path, content, 0644)
-		if err != nil {
-			return err
+		executeWrite := true
+		existedContent, err := os.ReadFile(path)
+		if err == nil {
+			if fmt.Sprintf("%x", sha256.Sum256(existedContent)) == fmt.Sprintf("%x", sha256.Sum256(content)) {
+				executeWrite = false
+			}
+		}
+
+		if executeWrite {
+			err = os.WriteFile(path, content, 0644)
+			if err != nil {
+				return err
+			}
 		}
 
 		return nil
@@ -205,3 +236,12 @@ func Unzip(source, destination string) error {
 	}
 	return nil
 }
+
+func (a *App) IsPortAvailable(port int) bool {
+	l, err := net.Listen("tcp", fmt.Sprintf("127.0.0.1:%s", strconv.Itoa(port)))
+	if err != nil {
+		return false
+	}
+	defer l.Close()
+	return true
+}

backend-python/convert_model.py

@@ -231,5 +231,6 @@ try:
         convert_and_save_and_exit=args.out,
     )
 except Exception as e:
+    print(e)
     with open("error.txt", "w") as f:
         f.write(str(e))

backend-python/convert_pytorch_to_ggml.py

@@ -0,0 +1,169 @@
+# Converts an RWKV model checkpoint in PyTorch format to an rwkv.cpp compatible file.
+# Usage: python convert_pytorch_to_ggml.py C:\RWKV-4-Pile-169M-20220807-8023.pth C:\rwkv.cpp-169M-FP16.bin FP16
+# Get model checkpoints from https://huggingface.co/BlinkDL
+# See FILE_FORMAT.md for the documentation on the file format.
+
+import argparse
+import struct
+import torch
+from typing import Dict
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Convert an RWKV model checkpoint in PyTorch format to an rwkv.cpp compatible file"
+    )
+    parser.add_argument("src_path", help="Path to PyTorch checkpoint file")
+    parser.add_argument(
+        "dest_path", help="Path to rwkv.cpp checkpoint file, will be overwritten"
+    )
+    parser.add_argument(
+        "data_type",
+        help="Data type, FP16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0",
+        type=str,
+        choices=[
+            "FP16",
+            "Q4_0",
+            "Q4_1",
+            "Q5_0",
+            "Q5_1",
+            "Q8_0",
+        ],
+        default="FP16",
+    )
+    return parser.parse_args()
+
+
+def get_layer_count(state_dict: Dict[str, torch.Tensor]) -> int:
+    n_layer: int = 0
+
+    while f"blocks.{n_layer}.ln1.weight" in state_dict:
+        n_layer += 1
+
+    assert n_layer > 0
+
+    return n_layer
+
+
+def write_state_dict(
+    state_dict: Dict[str, torch.Tensor], dest_path: str, data_type: str
+) -> None:
+    emb_weight: torch.Tensor = state_dict["emb.weight"]
+
+    n_layer: int = get_layer_count(state_dict)
+    n_vocab: int = emb_weight.shape[0]
+    n_embed: int = emb_weight.shape[1]
+
+    is_v5_1_or_2: bool = "blocks.0.att.ln_x.weight" in state_dict
+    is_v5_2: bool = "blocks.0.att.gate.weight" in state_dict
+
+    if is_v5_2:
+        print("Detected RWKV v5.2")
+    elif is_v5_1_or_2:
+        print("Detected RWKV v5.1")
+    else:
+        print("Detected RWKV v4")
+
+    with open(dest_path, "wb") as out_file:
+        is_FP16: bool = data_type == "FP16" or data_type == "float16"
+
+        out_file.write(
+            struct.pack(
+                # Disable padding with '='
+                "=iiiiii",
+                # Magic: 'ggmf' in hex
+                0x67676D66,
+                101,
+                n_vocab,
+                n_embed,
+                n_layer,
+                1 if is_FP16 else 0,
+            )
+        )
+
+        for k in state_dict.keys():
+            tensor: torch.Tensor = state_dict[k].float()
+
+            if ".time_" in k:
+                tensor = tensor.squeeze()
+
+            if is_v5_1_or_2:
+                if ".time_decay" in k:
+                    if is_v5_2:
+                        tensor = torch.exp(-torch.exp(tensor)).unsqueeze(-1)
+                    else:
+                        tensor = torch.exp(-torch.exp(tensor)).reshape(-1, 1, 1)
+
+                if ".time_first" in k:
+                    tensor = torch.exp(tensor).reshape(-1, 1, 1)
+
+                if ".time_faaaa" in k:
+                    tensor = tensor.unsqueeze(-1)
+            else:
+                if ".time_decay" in k:
+                    tensor = -torch.exp(tensor)
+
+            # Keep 1-dim vectors and small matrices in FP32
+            if is_FP16 and len(tensor.shape) > 1 and ".time_" not in k:
+                tensor = tensor.half()
+
+            shape = tensor.shape
+
+            print(f"Writing {k}, shape {shape}, type {tensor.dtype}")
+
+            k_encoded: bytes = k.encode("utf-8")
+
+            out_file.write(
+                struct.pack(
+                    "=iii",
+                    len(shape),
+                    len(k_encoded),
+                    1 if tensor.dtype == torch.float16 else 0,
+                )
+            )
+
+            # Dimension order is reversed here:
+            # * PyTorch shape is (x rows, y columns)
+            # * ggml shape is (y elements in a row, x elements in a column)
+            # Both shapes represent the same tensor.
+            for dim in reversed(tensor.shape):
+                out_file.write(struct.pack("=i", dim))
+
+            out_file.write(k_encoded)
+
+            tensor.numpy().tofile(out_file)
+
+
+def main() -> None:
+    args = parse_args()
+
+    print(f"Reading {args.src_path}")
+
+    state_dict: Dict[str, torch.Tensor] = torch.load(args.src_path, map_location="cpu")
+
+    temp_output: str = args.dest_path
+    if args.data_type.startswith("Q"):
+        import re
+
+        temp_output = re.sub(r"Q[4,5,8]_[0,1]", "fp16", temp_output)
+    write_state_dict(state_dict, temp_output, "FP16")
+    if args.data_type.startswith("Q"):
+        import sys
+        import os
+
+        sys.path.append(os.path.dirname(os.path.realpath(__file__)))
+        from rwkv_pip.cpp import rwkv_cpp_shared_library
+
+        library = rwkv_cpp_shared_library.load_rwkv_shared_library()
+        library.rwkv_quantize_model_file(temp_output, args.dest_path, args.data_type)
+
+    print("Done")
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as e:
+        print(e)
+        with open("error.txt", "w") as f:
+            f.write(str(e))

backend-python/convert_safetensors.py

@@ -18,21 +18,61 @@ parser.add_argument(
 args = parser.parse_args()
 
 
-def convert_file(
-    pt_filename: str,
-    sf_filename: str,
-):
+def rename_key(rename, name):
+    for k, v in rename.items():
+        if k in name:
+            name = name.replace(k, v)
+    return name
+
+
+def convert_file(pt_filename: str, sf_filename: str, rename={}, transpose_names=[]):
     loaded = torch.load(pt_filename, map_location="cpu")
     if "state_dict" in loaded:
         loaded = loaded["state_dict"]
 
+    kk = list(loaded.keys())
+    version = 4
+    for x in kk:
+        if "ln_x" in x:
+            version = max(5, version)
+        if "gate.weight" in x:
+            version = max(5.1, version)
+        if int(version) == 5 and "att.time_decay" in x:
+            if len(loaded[x].shape) > 1:
+                if loaded[x].shape[1] > 1:
+                    version = max(5.2, version)
+        if "time_maa" in x:
+            version = max(6, version)
+
+    if version == 5.1 and "midi" in pt_filename.lower():
+        import numpy as np
+
+        np.set_printoptions(precision=4, suppress=True, linewidth=200)
+        kk = list(loaded.keys())
+        _, n_emb = loaded["emb.weight"].shape
+        for k in kk:
+            if "time_decay" in k or "time_faaaa" in k:
+                # print(k, mm[k].shape)
+                loaded[k] = (
+                    loaded[k].unsqueeze(1).repeat(1, n_emb // loaded[k].shape[0])
+                )
+
     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}')
+
+    loaded = {rename_key(rename, k).lower(): v.contiguous() for k, v in loaded.items()}
+    # For tensors to be contiguous
+    for k, v in loaded.items():
+        for transpose_name in transpose_names:
+            if transpose_name in k:
+                loaded[k] = v.transpose(0, 1)
+
+    loaded = {k: v.clone().half().contiguous() 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"})
@@ -46,8 +86,24 @@ def convert_file(
 
 if __name__ == "__main__":
     try:
-        convert_file(args.input, args.output)
+        convert_file(
+            args.input,
+            args.output,
+            rename={
+                "time_faaaa": "time_first",
+                "time_maa": "time_mix",
+                "lora_A": "lora.0",
+                "lora_B": "lora.1",
+            },
+            transpose_names=[
+                "time_mix_w1",
+                "time_mix_w2",
+                "time_decay_w1",
+                "time_decay_w2",
+            ],
+        )
         print(f"Saved to {args.output}")
     except Exception as e:
+        print(e)
         with open("error.txt", "w") as f:
             f.write(str(e))

backend-python/dep_check.py

@@ -1,3 +1,5 @@
+import multipart
+import fitz
 import safetensors
 import midi2audio
 import mido
@@ -9,6 +11,7 @@ import GPUtil
 
 import torch
 import rwkv
+import langchain
 import numpy
 import tokenizers
 import fastapi

backend-python/global_var.py

@@ -4,6 +4,7 @@ Args = "args"
 Model = "model"
 Model_Status = "model_status"
 Model_Config = "model_config"
+Deploy_Mode = "deploy_mode"
 
 
 class ModelStatus(Enum):
@@ -16,6 +17,7 @@ def init():
     global GLOBALS
     GLOBALS = {}
     set(Model_Status, ModelStatus.Offline)
+    set(Deploy_Mode, False)
 
 
 def set(key, value):

backend-python/main.py

@@ -2,70 +2,8 @@ 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__)))
-
-import psutil
-from fastapi import Depends, FastAPI
-from fastapi.middleware.cors import CORSMiddleware
-import uvicorn
-
-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, midi, misc
-import global_var
-
-app = FastAPI(dependencies=[Depends(log_middleware)])
-
-app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"],
-    allow_credentials=True,
-    allow_methods=["*"],
-    allow_headers=["*"],
-)
-
-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()
-
-    if os.environ.get("ngrok_token") is not None:
-        ngrok_connect()
-
-
-@app.get("/", tags=["Root"])
-def read_root():
-    return {"Hello": "World!"}
-
-
-@app.post("/exit", tags=["Root"])
-def exit():
-    parent_pid = os.getpid()
-    parent = psutil.Process(parent_pid)
-    for child in parent.children(recursive=True):
-        child.kill()
-    parent.kill()
+from typing import Union, Sequence
 
 
 def get_args(args: Union[Sequence[str], None] = None):
@@ -84,11 +22,26 @@ def get_args(args: Union[Sequence[str], None] = None):
         help="host to run the server on (default: 127.0.0.1)",
     )
     group = parser.add_argument_group(title="mode arguments")
+    group.add_argument(
+        "--webui",
+        action="store_true",
+        help="whether to enable WebUI (default: False)",
+    )
     group.add_argument(
         "--rwkv-beta",
         action="store_true",
         help="whether to use rwkv-beta (default: False)",
     )
+    group.add_argument(
+        "--rwkv.cpp",
+        action="store_true",
+        help="whether to use rwkv.cpp (default: False)",
+    )
+    group.add_argument(
+        "--webgpu",
+        action="store_true",
+        help="whether to use webgpu (default: False)",
+    )
     args = parser.parse_args(args)
 
     return args
@@ -96,6 +49,96 @@ def get_args(args: Union[Sequence[str], None] = None):
 
 if __name__ == "__main__":
     args = get_args()
+
+
+import os
+import sys
+
+sys.path.append(os.path.dirname(os.path.realpath(__file__)))
+
+import psutil
+from contextlib import asynccontextmanager
+from fastapi import Depends, FastAPI, status
+from fastapi.middleware.cors import CORSMiddleware
+import uvicorn
+
+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, midi, misc, file_process
+import global_var
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    init()
+    yield
+
+
+app = FastAPI(lifespan=lifespan, dependencies=[Depends(log_middleware)])
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+app.include_router(completion.router)
+app.include_router(config.router)
+app.include_router(midi.router)
+app.include_router(file_process.router)
+app.include_router(misc.router)
+app.include_router(state_cache.router)
+
+
+@app.post("/exit", tags=["Root"])
+def exit():
+    if global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(status.HTTP_403_FORBIDDEN)
+
+    parent_pid = os.getpid()
+    parent = psutil.Process(parent_pid)
+    for child in parent.children(recursive=True):
+        child.kill()
+    parent.kill()
+
+
+try:
+    if (
+        "RWKV_RUNNER_PARAMS" in os.environ
+        and "--webui" in os.environ["RWKV_RUNNER_PARAMS"].split(" ")
+    ) or args.webui:
+        from webui_server import webui_server
+
+        app.mount("/", webui_server)
+except NameError:
+    pass
+
+
+@app.get("/", tags=["Root"])
+def read_root():
+    return {"Hello": "World!"}
+
+
+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()
+
+    if os.environ.get("ngrok_token") is not None:
+        ngrok_connect()
+
+
+if __name__ == "__main__":
     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)

backend-python/routes/completion.py

@@ -8,7 +8,6 @@ import base64
 from fastapi import APIRouter, Request, status, HTTPException
 from sse_starlette.sse import EventSourceResponse
 from pydantic import BaseModel, Field
-import numpy as np
 import tiktoken
 from utils.rwkv import *
 from utils.log import quick_log
@@ -25,32 +24,51 @@ class Role(Enum):
 
 class Message(BaseModel):
     role: Role
-    content: str = Field(min_length=1)
+    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",
+    "\n\nAssistant",
+    "\n\nAnswer",
+    "\n\nA",
+    "\n\nBot",
+    "\n\nAlice",
+]
 
 
 class ChatCompletionBody(ModelConfigBody):
     messages: Union[List[Message], None]
-    model: str = "rwkv"
+    model: Union[str, None] = "rwkv"
     stream: bool = False
-    stop: Union[str, List[str], None] = [
-        "\n\nUser",
-        "\n\nQuestion",
-        "\n\nQ",
-        "\n\nHuman",
-        "\n\nBob",
-    ]
-    user_name: Union[str, None] = None
-    assistant_name: Union[str, None] = None
+    stop: Union[str, List[str], None] = default_stop
+    user_name: Union[str, None] = Field(
+        None, description="Internal user name", min_length=1
+    )
+    assistant_name: Union[str, None] = Field(
+        None, description="Internal assistant name", min_length=1
+    )
+    presystem: bool = Field(
+        True, description="Whether to insert default system prompt at the beginning"
+    )
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_schema_extra": {
             "example": {
-                "messages": [{"role": Role.User.value, "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,
@@ -58,16 +76,17 @@ class ChatCompletionBody(ModelConfigBody):
                 "frequency_penalty": 0.4,
             }
         }
+    }
 
 
 class CompletionBody(ModelConfigBody):
     prompt: Union[str, List[str], None]
-    model: str = "rwkv"
+    model: Union[str, None] = "rwkv"
     stream: bool = False
     stop: Union[str, List[str], None] = None
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_schema_extra": {
             "example": {
                 "prompt": "The following is an epic science fiction masterpiece that is immortalized, "
                 + "with delicate descriptions and grand depictions of interstellar civilization wars.\nChapter 1.\n",
@@ -81,6 +100,7 @@ class CompletionBody(ModelConfigBody):
                 "frequency_penalty": 0.4,
             }
         }
+    }
 
 
 completion_lock = Lock()
@@ -233,10 +253,6 @@ async def chat_completions(body: ChatCompletionBody, request: Request):
     if body.messages is None or body.messages == []:
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "messages not found")
 
-    basic_system: str = ""
-    if body.messages[0].role == Role.System:
-        basic_system = body.messages[0].content
-
     interface = model.interface
     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
@@ -244,46 +260,54 @@ async def chat_completions(body: ChatCompletionBody, request: Request):
     is_raven = model.rwkv_type == RWKVType.Raven
 
     completion_text: str = ""
-    if basic_system == "":
-        completion_text = (
-            f"""
+    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 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"
-            )
-        )
-    elif basic_system != "":
-        completion_text = (
-            (
-                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. "
+                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 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"
             )
-            + basic_system.replace("\r\n", "\n")
-            .replace("\r", "\n")
-            .replace("\n\n", "\n")
-            .replace("\n", " ")
-            .strip()
-            .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"
-        )
 
-    for message in body.messages[(0 if basic_system == "" else 1) :]:
+    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
@@ -291,20 +315,27 @@ The following is a coherent verbose detailed conversation between a girl named {
             append_message = f"{bot}{interface} " + message.content
         elif message.role == Role.System:
             append_message = message.content
-        completion_text += (
-            append_message.replace("\r\n", "\n")
-            .replace("\r", "\n")
-            .replace("\n\n", "\n")
-            .strip()
-            + "\n\n"
-        )
+        if not message.raw:
+            append_message = (
+                append_message.replace("\r\n", "\n")
+                .replace("\r", "\n")
+                .replace("\n\n", "\n")
+                .strip()
+            )
+        completion_text += append_message + "\n\n"
     completion_text += f"{bot}{interface}"
 
+    user_code = model.pipeline.decode([model.pipeline.encode(user)[0]])
+    bot_code = model.pipeline.decode([model.pipeline.encode(bot)[0]])
     if type(body.stop) == str:
-        body.stop = [body.stop, f"\n\n{user}", f"\n\n{bot}"]
-    else:
-        body.stop.append(f"\n\n{user}")
-        body.stop.append(f"\n\n{bot}")
+        body.stop = [body.stop, f"\n\n{user_code}", f"\n\n{bot_code}"]
+    elif type(body.stop) == list:
+        body.stop.append(f"\n\n{user_code}")
+        body.stop.append(f"\n\n{bot_code}")
+    elif body.stop is None:
+        body.stop = default_stop
+    if not body.presystem:
+        body.stop.append("\n\n")
 
     if body.stream:
         return EventSourceResponse(
@@ -349,12 +380,12 @@ async def completions(body: CompletionBody, request: Request):
 
 class EmbeddingsBody(BaseModel):
     input: Union[str, List[str], List[List[int]], None]
-    model: str = "rwkv"
+    model: Union[str, None] = "rwkv"
     encoding_format: str = None
     fast_mode: bool = False
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_schema_extra": {
             "example": {
                 "input": "a big apple",
                 "model": "rwkv",
@@ -362,9 +393,12 @@ class EmbeddingsBody(BaseModel):
                 "fast_mode": False,
             }
         }
+    }
 
 
 def embedding_base64(embedding: List[float]) -> str:
+    import numpy as np
+
     return base64.b64encode(np.array(embedding).astype(np.float32)).decode("utf-8")
 
 

backend-python/routes/config.py

@@ -10,39 +10,34 @@ import global_var
 router = APIRouter()
 
 
-def get_tokens_path(model_path: str):
-    model_path = model_path.lower()
-    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
-
-
 class SwitchModelBody(BaseModel):
     model: str
     strategy: str
+    tokenizer: Union[str, None] = None
     customCuda: bool = False
+    deploy: bool = Field(
+        False,
+        description="Deploy mode. If success, will disable /switch-model, /exit and other dangerous APIs (state cache APIs, part of midi APIs)",
+    )
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_schema_extra": {
             "example": {
                 "model": "models/RWKV-4-World-3B-v1-20230619-ctx4096.pth",
                 "strategy": "cuda fp16",
+                "tokenizer": "",
                 "customCuda": False,
+                "deploy": False,
             }
         }
+    }
 
 
 @router.post("/switch-model", tags=["Configs"])
 def switch_model(body: SwitchModelBody, response: Response, request: Request):
+    if global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(Status.HTTP_403_FORBIDDEN)
+
     if global_var.get(global_var.Model_Status) is global_var.ModelStatus.Loading:
         response.status_code = Status.HTTP_304_NOT_MODIFIED
         return
@@ -54,13 +49,20 @@ 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
+    devices = set(
+        [
+            x.strip().split(" ")[0].replace("cuda:0", "cuda")
+            for x in body.strategy.split("->")
+        ]
+    )
+    print(f"Strategy Devices: {devices}")
+    # if len(devices) > 1:
+    #     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"
 
@@ -68,26 +70,22 @@ def switch_model(body: SwitchModelBody, response: Response, request: Request):
     try:
         global_var.set(
             global_var.Model,
-            TextRWKV(
-                model=body.model,
-                strategy=body.strategy,
-                tokens_path=get_tokens_path(body.model),
-            )
-            if "midi" not in body.model.lower()
-            else MusicRWKV(
-                model=body.model,
-                strategy=body.strategy,
-                tokens_path=get_tokens_path(body.model),
-            ),
+            RWKV(model=body.model, strategy=body.strategy, tokenizer=body.tokenizer),
         )
     except Exception as e:
         print(e)
+        import traceback
+
+        print(traceback.format_exc())
+
         quick_log(request, body, f"Exception: {e}")
         global_var.set(global_var.Model_Status, global_var.ModelStatus.Offline)
         raise HTTPException(
             Status.HTTP_500_INTERNAL_SERVER_ERROR, f"failed to load: {e}"
         )
 
+    if body.deploy:
+        global_var.set(global_var.Deploy_Mode, True)
     if global_var.get(global_var.Model_Config) is None:
         global_var.set(
             global_var.Model_Config, get_rwkv_config(global_var.get(global_var.Model))

backend-python/routes/file_process.py

@@ -0,0 +1,79 @@
+import os
+from fastapi import (
+    APIRouter,
+    HTTPException,
+    status,
+    Depends,
+    File,
+    UploadFile,
+)
+from pydantic import BaseModel
+from typing import Iterator
+
+router = APIRouter()
+
+
+class FileToTextParams(BaseModel):
+    file_name: str
+    file_encoding: str = "utf-8"
+
+
+@router.post("/file-to-text", tags=["File Process"])
+async def file_to_text(
+    params: FileToTextParams = Depends(), file_data: UploadFile = File(...)
+):
+    from langchain.schema import Document
+    from langchain.document_loaders.blob_loaders import Blob
+
+    # from langchain
+    def parse_text(blob: Blob) -> Iterator[Document]:
+        yield Document(page_content=blob.as_string(), metadata={"source": blob.source})
+
+    # from langchain
+    def parse_pdf(blob: Blob) -> Iterator[Document]:
+        import fitz
+
+        with blob.as_bytes_io() as stream:
+            doc = fitz.Document(stream=stream)
+
+            yield from [
+                Document(
+                    page_content=page.get_text(),
+                    metadata=dict(
+                        {
+                            "source": blob.source,
+                            "file_path": blob.source,
+                            "page": page.number,
+                            "total_pages": len(doc),
+                        },
+                        **{
+                            k: doc.metadata[k]
+                            for k in doc.metadata
+                            if type(doc.metadata[k]) in [str, int]
+                        },
+                    ),
+                )
+                for page in doc
+            ]
+
+    file_parsers = {".txt": parse_text, ".pdf": parse_pdf}
+
+    file_name = file_data.filename or params.file_name
+    file_ext = os.path.splitext(file_name)[-1]
+
+    if file_ext not in file_parsers:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "file type not supported")
+
+    try:
+        pages: Iterator[Document] = file_parsers[file_ext](
+            Blob.from_data(
+                await file_data.read(),
+                encoding=params.file_encoding,
+                path=file_name,
+            )
+        )
+        pages = list(pages)
+    except Exception as e:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, f"{e}")
+
+    return {"pages": pages}

backend-python/routes/midi.py

@@ -1,5 +1,6 @@
 import io
-from fastapi import APIRouter, HTTPException, status
+import global_var
+from fastapi import APIRouter, HTTPException, UploadFile, status
 from starlette.responses import StreamingResponse
 from pydantic import BaseModel
 from utils.midi import *
@@ -11,12 +12,13 @@ router = APIRouter()
 class TextToMidiBody(BaseModel):
     text: str
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_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"])
@@ -31,21 +33,39 @@ def text_to_midi(body: TextToMidiBody):
     return StreamingResponse(mid_data, media_type="audio/midi")
 
 
+@router.post("/midi-to-text", tags=["MIDI"])
+async def midi_to_text(file_data: UploadFile):
+    vocab_config = "backend-python/utils/midi_vocab_config.json"
+    cfg = VocabConfig.from_json(vocab_config)
+    filter_config = "backend-python/utils/midi_filter_config.json"
+    filter_cfg = FilterConfig.from_json(filter_config)
+    mid = mido.MidiFile(file=file_data.file)
+    output_list = convert_midi_to_str(cfg, filter_cfg, mid)
+    if len(output_list) == 0:
+        raise HTTPException(status.HTTP_400_BAD_REQUEST, "bad midi file")
+
+    return {"text": output_list[0]}
+
+
 class TxtToMidiBody(BaseModel):
     txt_path: str
     midi_path: str
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_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 global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     if not body.midi_path.startswith("midi/"):
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "bad output path")
 
@@ -65,14 +85,15 @@ class MidiToWavBody(BaseModel):
     wav_path: str
     sound_font_path: str = "assets/default_sound_font.sf2"
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_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"])
@@ -81,6 +102,9 @@ def midi_to_wav(body: MidiToWavBody):
     Install fluidsynth first, see more: https://github.com/FluidSynth/fluidsynth/wiki/Download#distributions
     """
 
+    if global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     if not body.wav_path.startswith("midi/"):
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "bad output path")
 
@@ -95,14 +119,15 @@ class TextToWavBody(BaseModel):
     wav_name: str
     sound_font_path: str = "assets/default_sound_font.sf2"
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_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"])
@@ -111,6 +136,9 @@ def text_to_wav(body: TextToWavBody):
     Install fluidsynth first, see more: https://github.com/FluidSynth/fluidsynth/wiki/Download#distributions
     """
 
+    if global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     text = body.text.strip()
     if not text.startswith("<start>"):
         text = "<start> " + text

backend-python/routes/state_cache.py

@@ -4,12 +4,13 @@ from fastapi import APIRouter, HTTPException, Request, Response, status
 from pydantic import BaseModel
 import gc
 import copy
+import global_var
 
 router = APIRouter()
 
 trie = None
 dtrie: Dict = {}
-max_trie_len = 3000
+max_trie_len = 300
 loop_start_id = 1  # to prevent preloaded prompts from being deleted
 loop_del_trie_id = loop_start_id
 
@@ -36,16 +37,24 @@ def init():
 def disable_state_cache():
     global trie, dtrie
 
+    if global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     trie = None
     dtrie = {}
     gc.collect()
 
+    print("state cache disabled")
     return "success"
 
 
 @router.post("/enable-state-cache", tags=["State Cache"])
 def enable_state_cache():
     global trie, dtrie
+
+    if global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     try:
         import cyac
 
@@ -53,8 +62,10 @@ def enable_state_cache():
         dtrie = {}
         gc.collect()
 
+        print("state cache enabled")
         return "success"
     except ModuleNotFoundError:
+        print("state cache disabled")
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "cyac not found")
 
 
@@ -65,24 +76,47 @@ class AddStateBody(BaseModel):
     logits: Any
 
 
-@router.post("/add-state", tags=["State Cache"])
+# @router.post("/add-state", tags=["State Cache"])
 def add_state(body: AddStateBody):
     global trie, dtrie, loop_del_trie_id
+
+    # if global_var.get(global_var.Deploy_Mode) is True:
+    #     raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     if trie is None:
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
 
     import torch
+    import numpy as np
 
     try:
+        devices: List[torch.device] = []
+        state: Union[Any, None] = None
+
+        if body.state is not None:
+            if type(body.state) == list or type(body.state) == np.ndarray:
+                devices = [
+                    (
+                        tensor.device
+                        if hasattr(tensor, "device")
+                        else torch.device("cpu")
+                    )
+                    for tensor in body.state
+                ]
+                state = (
+                    [tensor.cpu() for tensor in body.state]
+                    if hasattr(body.state[0], "device")
+                    else copy.deepcopy(body.state)
+                )
+            else:
+                pass  # WebGPU
+
         id: int = trie.insert(body.prompt)
-        device: torch.device = body.state[0].device
         dtrie[id] = {
             "tokens": copy.deepcopy(body.tokens),
-            "state": [tensor.cpu() for tensor in body.state]
-            if device != torch.device("cpu")
-            else copy.deepcopy(body.state),
+            "state": state,
             "logits": copy.deepcopy(body.logits),
-            "device": device,
+            "devices": devices,
         }
 
         if len(trie) >= max_trie_len:
@@ -96,7 +130,7 @@ 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:
@@ -108,6 +142,10 @@ def add_state(body: AddStateBody):
 @router.post("/reset-state", tags=["State Cache"])
 def reset_state():
     global trie, dtrie
+
+    if global_var.get(global_var.Deploy_Mode) is True:
+        raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     if trie is None:
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
 
@@ -124,7 +162,7 @@ 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())
@@ -141,13 +179,18 @@ def _get_a_dtrie_buff_size(dtrie_v):
     return 54 * len(dtrie_v["tokens"]) + 491520 + 262144 + 28  # TODO
 
 
-@router.post("/longest-prefix-state", tags=["State Cache"])
+# @router.post("/longest-prefix-state", tags=["State Cache"])
 def longest_prefix_state(body: LongestPrefixStateBody, request: Request):
     global trie
+
+    # if global_var.get(global_var.Deploy_Mode) is True:
+    #     raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     if trie is None:
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
 
     import torch
+    import numpy as np
 
     id = -1
     try:
@@ -157,32 +200,31 @@ def longest_prefix_state(body: LongestPrefixStateBody, request: Request):
         pass
     if id != -1:
         v = dtrie[id]
-        device: torch.device = v["device"]
+        devices: List[torch.device] = v["devices"]
         prompt: str = trie[id]
+        state: Union[Any, None] = v["state"]
+
+        if state is not None and type(state) == list and hasattr(state[0], "device"):
+            state = [tensor.to(devices[i]) for i, tensor in enumerate(state)]
 
         quick_log(request, body, "Hit:\n" + prompt)
         return {
             "prompt": prompt,
             "tokens": v["tokens"],
-            "state": [tensor.to(device) for tensor in v["state"]]
-            if device != torch.device("cpu")
-            else v["state"],
+            "state": state,
             "logits": v["logits"],
-            "device": device.type,
         }
     else:
-        return {
-            "prompt": "",
-            "tokens": [],
-            "state": None,
-            "logits": None,
-            "device": None,
-        }
+        return {"prompt": "", "tokens": [], "state": None, "logits": None}
 
 
-@router.post("/save-state", tags=["State Cache"])
+# @router.post("/save-state", tags=["State Cache"])
 def save_state():
     global trie
+
+    # if global_var.get(global_var.Deploy_Mode) is True:
+    #     raise HTTPException(status.HTTP_403_FORBIDDEN)
+
     if trie is None:
         raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")
 

backend-python/rwkv_pip/beta/cuda/att_one.cu

@@ -88,7 +88,7 @@ struct Mix {
 
 using torch::Tensor;
 
-void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+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,
@@ -105,9 +105,9 @@ Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
                    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);
+  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),
@@ -118,7 +118,7 @@ Tensor att_one(Tensor x, Tensor ln_w, Tensor ln_b, Tensor sx, Tensor k_mix,
                              data_ptr<half>(r)},
                x.numel());
 
-  gemm_fp16_cublas(r, ow, x_plus_out);
+  gemm_fp16_cublas_tensor(r, ow, x_plus_out);
   x_plus_out += x;
   return xx;
 }

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;
+}

backend-python/rwkv_pip/beta/cuda/att_seq.cu

@@ -8,7 +8,6 @@
 
 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);
 

backend-python/rwkv_pip/beta/cuda/gemm_fp16_cublas.cpp

@@ -70,11 +70,59 @@ void gemm_fp16_cublas(const void *a, const void *b, void *c, int ori_m,
       cuda_c_data_type, cublas_ldc, compute_type, algo));
 }
 
-void gemm_fp16_cublas(torch::Tensor a, torch::Tensor b, torch::Tensor c) {
-  // comptiable with rwkv one mode, 1-D tensor * 2-D tensor
-  const int m = a.dense_dim() == 1 ? 1 : a.size(0);
-  const int n = b.size(1);
-  const int k = b.size(0);
-  gemm_fp16_cublas(a.data_ptr(), b.data_ptr(), c.data_ptr(), m, n, k,
-                   c.dtype() == torch::kFloat32);
+/*
+  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));
+  }
 }

backend-python/rwkv_pip/beta/cuda/wrapper.cpp

@@ -118,7 +118,9 @@ void mm8_one(int64_t N, int64_t M,
 
 using torch::Tensor;
 
-void gemm_fp16_cublas(Tensor a, Tensor b, Tensor c);
+#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,
@@ -134,6 +136,16 @@ Tensor att_seq(Tensor x, Tensor sx, Tensor ln_w, Tensor ln_b, Tensor k_mix,
                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,
@@ -148,8 +160,9 @@ 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, "gemv fp16 cublas");
+    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");
@@ -159,8 +172,9 @@ 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);
+    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);

backend-python/rwkv_pip/beta/model.py

@@ -3,7 +3,7 @@
 ########################################################################################################
 
 from typing import Optional
-import types, gc, os, time, re
+import types, gc, os, time, re, platform
 import torch
 from torch.nn import functional as F
 
@@ -91,8 +91,10 @@ if os.environ.get("RWKV_CUDA_ON") == "1":
                 f"{current_path}/cuda/att_one.cu",
                 f"{current_path}/cuda/att_seq.cu",
                 f"{current_path}/cuda/ffn.cu",
+                f"{current_path}/cuda/att_one_v5.cu",
             ],
             verbose=True,
+            extra_ldflags=["cublas.lib" if os.name == "nt" else ""],
             extra_cuda_cflags=[
                 "-t 4",
                 "-std=c++17",
@@ -149,26 +151,40 @@ if os.environ.get("RWKV_CUDA_ON") == "1":
         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"
+
+if os.environ.get("RWKV_CUDA_ON") == "1":
+
     @MyStatic
     def gemm(a, b, output_dtype: Optional[torch.dtype] = None):
         if output_dtype is None:
             output_dtype = a.dtype
         if a.dtype == b.dtype == torch.float16 and a.device.type == "cuda":
-            assert len(b.shape) == 2
             if len(a.shape) == 1:
+                assert len(b.shape) == 2
                 c = torch.empty((b.shape[-1],), dtype=output_dtype, device=a.device)
                 a = a.unsqueeze(0)
             else:
-                c = torch.empty(
-                    (a.shape[0], b.shape[-1]), dtype=output_dtype, device=a.device
-                )
+                assert len(a.shape) == len(b.shape)
+                assert len(a.shape) == 2 or len(a.shape) == 3
+                # torch.empty((*a.shape[:-1], b.shape[-1])) doesn't work with jit
+                if len(a.shape) == 2:
+                    c = torch.empty(
+                        (a.shape[0], b.shape[-1]), dtype=output_dtype, device=a.device
+                    )
+                else:
+                    c = torch.empty(
+                        (a.shape[0], a.shape[1], b.shape[-1]),
+                        dtype=output_dtype,
+                        device=a.device,
+                    )
             torch.ops.rwkv.gemm_fp16_cublas(a, b, c)
             return c
         else:
             return (a @ b).to(output_dtype)
 
 else:
-    os.environ["RWKV_CUDA_ON"] = "0"
 
     def gemm(a, b, output_dtype: Optional[torch.dtype] = None):
         if output_dtype is None:
@@ -217,7 +233,7 @@ class RWKV(MyModule):
             )  # load model to CPU first
             # it is supported to load a pure meta-tensor state dict (e.g. for quick testing)
             for k, v in self.w.items():
-                if v.is_meta:
+                if isinstance(v, torch.Tensor) and v.is_meta:
                     # torch.zeros_like(v, device='cpu') doesn't produce an all-zero tensor
                     # if v is a meta tensor
                     self.w[k] = torch.zeros(v.shape, dtype=v.dtype, device="cpu")
@@ -247,9 +263,14 @@ class RWKV(MyModule):
             args.n_embd = w["emb.weight"].shape[1]
             args.n_layer = 0
             keys = list(w.keys())
+            self.version = 4
             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)
+                if "ln_x" in x:
+                    self.version = 5
+                if self.version == 5 and "att.time_decay" in x:
+                    args.n_head = w[x].shape[0]
 
             ####################### Compute strategy
 
@@ -352,6 +373,20 @@ class RWKV(MyModule):
                 del w["blocks.0.ln0.bias"]
 
             print_need_newline = False
+
+            REAL_TIME_FIRST = False
+            for x in list(w.keys()):
+                if ".time_faaaa" in x:
+                    REAL_TIME_FIRST = True
+            if REAL_TIME_FIRST:
+                w = {
+                    k.replace(".time_faaaa", ".time_first")
+                    if ".time_faaaa" in k
+                    else k: v
+                    for k, v in w.items()
+                }
+                self.w = w
+
             keys = list(w.keys())
             for x in keys:
                 w[x].requires_grad = False
@@ -382,8 +417,19 @@ class RWKV(MyModule):
                         w[x] = w[x].t()
 
                     if ".time_decay" in x:  # need fp32 for this
-                        w[x] = -torch.exp(w[x].float())
+                        if self.version == 4:
+                            w[x] = -torch.exp(w[x].float())
+                        elif self.version == 5:
+                            w[x] = torch.exp(-torch.exp(w[x].float())).reshape(-1, 1, 1)
                     elif ".time_first" in x:  # need fp32 for this
+                        if self.version == 4:
+                            w[x] = w[x].float()
+                        elif self.version == 5:
+                            if REAL_TIME_FIRST:
+                                w[x] = w[x].float().reshape(-1, 1, 1)
+                            else:
+                                w[x] = torch.exp(w[x].float()).reshape(-1, 1, 1)
+                    elif ".ln_x" in x:  # need fp32 for group_norm
                         w[x] = w[x].float()
                     else:
                         if (len(w[x].shape) == 2) and ("emb" not in x):
@@ -931,6 +977,147 @@ class RWKV(MyModule):
 
     ########################################################################################################
 
+    @MyFunction
+    def att_one_v5(
+        self,
+        x,
+        sx,
+        s,
+        ln_w,
+        ln_b,
+        lx_w,
+        lx_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)
+
+        H = t_decay.shape[0]
+        S = x.shape[-1] // H
+
+        r = gemm(rx, rw, output_dtype=torch.float32).view(H, 1, S)
+        k = gemm(kx, kw, output_dtype=torch.float32).view(H, S, 1)
+        v = gemm(vx, vw, output_dtype=torch.float32).view(H, 1, S)
+
+        a = gemm(k, v)
+        out = r @ (t_first * a + s)
+        s = a + t_decay * s
+
+        out = out.flatten()
+        out = F.group_norm(
+            out.unsqueeze(0), num_groups=H, weight=lx_w, bias=lx_b
+        ).squeeze(0)
+        out = out.to(dtype=x.dtype)
+        out = gemm(out, ow)
+
+        return x + out, xx, s
+
+    @MyFunction
+    def att_seq_v5(
+        self,
+        x,
+        sx,
+        s,
+        ln_w,
+        ln_b,
+        lx_w,
+        lx_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)
+
+        H = t_decay.shape[0]
+        S = x.shape[-1] // H
+        T = x.shape[0]
+
+        w = t_decay.reshape(-1, 1)
+        u = t_first.reshape(-1, 1)
+        ws = w.pow(T).reshape(H, 1, 1)
+        ind = torch.arange(T - 1, -1, -1, device=w.device).unsqueeze(0).repeat(H, 1)
+        w = w.repeat(1, T).pow(ind)
+        wk = w.reshape(H, 1, T)
+        wb = wk.transpose(-2, -1).flip(1)
+        w = torch.cat([w[:, 1:], u], dim=1)
+        w = F.pad(w, (0, T))
+        w = torch.tile(w, [T])
+        w = w[:, :-T].reshape(-1, T, 2 * T - 1)
+        w = w[:, :, T - 1 :].reshape(H, T, T)
+
+        r = gemm(rx, rw, output_dtype=torch.float32).view(T, H, S).transpose(0, 1)
+        k = (
+            gemm(kx, kw, output_dtype=torch.float32)
+            .view(T, H, S)
+            .transpose(0, 1)
+            .transpose(-2, -1)
+        )
+        v = gemm(vx, vw, output_dtype=torch.float32).view(T, H, S).transpose(0, 1)
+
+        out = ((r @ k) * w) @ v + (r @ s) * wb
+        s = ws * s + (k * wk) @ v
+
+        out = out.transpose(0, 1).contiguous().reshape(T, H * S)
+        out = F.group_norm(out, num_groups=H, weight=lx_w, bias=lx_b)
+        out = out.to(dtype=x.dtype)
+        out = gemm(out, ow)
+
+        return x + out, xx[-1, :], s
+
+    ########################################################################################################
+
     if os.environ["RWKV_CUDA_ON"] == "1":
 
         @MyFunction
@@ -1140,7 +1327,7 @@ class RWKV(MyModule):
             xx = torch.ops.rwkv.ffn_seq(
                 x, sx, ln_w, ln_b, k_mix, r_mix, kw, vw, rw, buf, x_plus_out
             )
-            return x_plus_out, xx[-1:]
+            return x_plus_out, xx[-1, :]
 
         @MyFunction
         def cuda_att_one_fp16(
@@ -1220,6 +1407,86 @@ class RWKV(MyModule):
             )
             return x_plus_out_t, xx, t1_t, t2_t, p_t
 
+        @MyFunction
+        def cuda_att_one_v5_fp16(
+            self,
+            x,
+            sx,
+            s,
+            ln_w,
+            ln_b,
+            lx_w,
+            lx_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,
+        ):
+            kx = torch.empty_like(x)
+            vx = torch.empty_like(x)
+            rx = torch.empty_like(x)
+
+            H = t_decay.shape[0]
+            S = x.shape[-1] // H
+
+            r = torch.empty((H * S,), dtype=torch.float32, device=x.device)
+            k = torch.empty((H * S,), dtype=torch.float32, device=x.device)
+            v = torch.empty((H * S,), dtype=torch.float32, device=x.device)
+            s1 = torch.empty((H, S, S), dtype=torch.float32, device=x.device)
+            s2 = torch.empty((H, S, S), dtype=torch.float32, device=x.device)
+            x_plus_out = torch.empty_like(x)
+
+            xx = torch.ops.rwkv.att_one_v5(
+                x,
+                sx,
+                s,
+                ln_w,
+                ln_b,
+                lx_w,
+                lx_b,
+                k_mix,
+                v_mix,
+                r_mix,
+                kw,
+                kx,
+                vw,
+                vx,
+                rw,
+                rx,
+                ow,
+                t_first,
+                k,
+                t_decay,
+                v,
+                r,
+                s1,
+                x_plus_out,
+                s2,
+            )
+
+            return x_plus_out, xx, s2
+
         @MyFunction
         def cuda_ffn_one_fp16(
             self,
@@ -1265,34 +1532,63 @@ class RWKV(MyModule):
             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(
+                if self.version == 4:
+                    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()
-                        - 1e30
-                    )
-                    state[i * 5 + 4] = torch.zeros(
-                        args.n_embd, dtype=atype, 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()
+                elif self.version == 5:
+                    state = [None] * args.n_layer * 3
+                    for i in range(args.n_layer):  # state: 0=att_xx 1=att_kv 2=ffn_xx
+                        dd = self.strategy[i]
+                        dev = dd.device
+                        atype = dd.atype
+                        state[i * 3 + 0] = torch.zeros(
+                            args.n_embd, dtype=atype, requires_grad=False, device=dev
+                        ).contiguous()
+                        state[i * 3 + 1] = torch.zeros(
+                            (
+                                args.n_head,
+                                args.n_embd // args.n_head,
+                                args.n_embd // args.n_head,
+                            ),
+                            dtype=torch.float,
+                            requires_grad=False,
+                            device=dev,
+                        ).contiguous()
+                        state[i * 3 + 2] = torch.zeros(
+                            args.n_embd, dtype=atype, requires_grad=False, device=dev
+                        ).contiguous()
 
             seq_mode = len(tokens) > 1
 
@@ -1317,9 +1613,13 @@ class RWKV(MyModule):
                             ATT = self.cuda_att_seq_i8
                         else:
                             ATT = self.cuda_att_seq_naive
+                    if self.version == 5:
+                        ATT = self.att_seq_v5
                 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
+                    if self.version == 5:
+                        ATT = self.att_one_v5
                     if (
                         "cuda" in str(dev)
                         and os.environ["RWKV_CUDA_ON"] == "1"
@@ -1327,6 +1627,8 @@ class RWKV(MyModule):
                     ):
                         ATT = self.cuda_att_one_fp16
                         FFN = self.cuda_ffn_one_fp16
+                        if self.version == 5:
+                            ATT = self.cuda_att_one_v5_fp16
 
                 x = x.to(dtype=atype, device=dev)
 
@@ -1355,46 +1657,82 @@ class RWKV(MyModule):
                 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 self.version == 4:
+                    (
+                        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,
+                    )
+                elif self.version == 5:
+                    x, state[i * 3 + 0], state[i * 3 + 1] = ATT(
+                        x,
+                        state[i * 3 + 0],
+                        state[i * 3 + 1],
+                        w[f"{bbb}ln1.weight"],
+                        w[f"{bbb}ln1.bias"],
+                        w[f"{att}ln_x.weight"],
+                        w[f"{att}ln_x.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
 
@@ -1417,9 +1755,13 @@ class RWKV(MyModule):
                 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(
+                if self.version == 4:
+                    offset = i * 5 + 4
+                elif self.version == 5:
+                    offset = i * 3 + 2
+                x, state[offset] = FFN(
                     x,
-                    state[i * 5 + 4],
+                    state[offset],
                     w[f"{bbb}ln2.weight"],
                     w[f"{bbb}ln2.bias"],
                     w[f"{ffn}time_mix_k"],

backend-python/rwkv_pip/cpp/model.py

@@ -0,0 +1,14 @@
+from typing import Any, List, Union
+from . import rwkv_cpp_model
+from . import rwkv_cpp_shared_library
+
+
+class RWKV:
+    def __init__(self, model_path: str, strategy=None):
+        self.library = rwkv_cpp_shared_library.load_rwkv_shared_library()
+        self.model = rwkv_cpp_model.RWKVModel(self.library, model_path)
+        self.w = {}  # fake weight
+        self.w["emb.weight"] = [0] * self.model.n_vocab
+
+    def forward(self, tokens: List[int], state: Union[Any, None] = None):
+        return self.model.eval_sequence_in_chunks(tokens, state, use_numpy=True)

backend-python/rwkv_pip/cpp/rwkv_cpp_model.py

@@ -0,0 +1,369 @@
+import os
+import multiprocessing
+
+# Pre-import PyTorch, if available.
+# This fixes "OSError: [WinError 127] The specified procedure could not be found".
+try:
+    import torch
+except ModuleNotFoundError:
+    pass
+
+# I'm sure this is not strictly correct, but let's keep this crutch for now.
+try:
+    import rwkv_cpp_shared_library
+except ModuleNotFoundError:
+    from . import rwkv_cpp_shared_library
+
+from typing import TypeVar, Optional, Tuple, List
+
+# A value of this type is either a numpy's ndarray or a PyTorch's Tensor.
+NumpyArrayOrPyTorchTensor: TypeVar = TypeVar('NumpyArrayOrPyTorchTensor')
+
+class RWKVModel:
+    """
+    An RWKV model managed by rwkv.cpp library.
+    """
+
+    def __init__(
+            self,
+            shared_library: rwkv_cpp_shared_library.RWKVSharedLibrary,
+            model_path: str,
+            thread_count: int = max(1, multiprocessing.cpu_count() // 2),
+            gpu_layer_count: int = 0,
+            **kwargs
+    ) -> None:
+        """
+        Loads the model and prepares it for inference.
+        In case of any error, this method will throw an exception.
+
+        Parameters
+        ----------
+        shared_library : RWKVSharedLibrary
+            rwkv.cpp shared library.
+        model_path : str
+            Path to RWKV model file in ggml format.
+        thread_count : int
+            Thread count to use. If not set, defaults to CPU count / 2.
+        gpu_layer_count : int
+            Count of layers to offload onto the GPU, must be >= 0.
+            See documentation of `gpu_offload_layers` for details about layer offloading.
+        """
+
+        if 'gpu_layers_count' in kwargs:
+            gpu_layer_count = kwargs['gpu_layers_count']
+
+        assert os.path.isfile(model_path), f'{model_path} is not a file'
+        assert thread_count > 0, 'Thread count must be > 0'
+        assert gpu_layer_count >= 0, 'GPU layer count must be >= 0'
+
+        self._library: rwkv_cpp_shared_library.RWKVSharedLibrary = shared_library
+
+        self._ctx: rwkv_cpp_shared_library.RWKVContext = self._library.rwkv_init_from_file(model_path, thread_count)
+
+        if gpu_layer_count > 0:
+            self.gpu_offload_layers(gpu_layer_count)
+
+        self._state_buffer_element_count: int = self._library.rwkv_get_state_buffer_element_count(self._ctx)
+        self._logits_buffer_element_count: int = self._library.rwkv_get_logits_buffer_element_count(self._ctx)
+
+        self._valid: bool = True
+
+    def gpu_offload_layers(self, layer_count: int) -> bool:
+        """
+        Offloads specified count of model layers onto the GPU. Offloaded layers are evaluated using cuBLAS or CLBlast.
+        For the purposes of this function, model head (unembedding matrix) is treated as an additional layer:
+        - pass `model.n_layer` to offload all layers except model head
+        - pass `model.n_layer + 1` to offload all layers, including model head
+
+        Returns true if at least one layer was offloaded.
+        If rwkv.cpp was compiled without cuBLAS and CLBlast support, this function is a no-op and always returns false.
+
+        Parameters
+        ----------
+        layer_count : int
+            Count of layers to offload onto the GPU, must be >= 0.
+        """
+
+        assert layer_count >= 0, 'Layer count must be >= 0'
+
+        return self._library.rwkv_gpu_offload_layers(self._ctx, layer_count)
+
+    @property
+    def n_vocab(self) -> int:
+        return self._library.rwkv_get_n_vocab(self._ctx)
+
+    @property
+    def n_embed(self) -> int:
+        return self._library.rwkv_get_n_embed(self._ctx)
+
+    @property
+    def n_layer(self) -> int:
+        return self._library.rwkv_get_n_layer(self._ctx)
+
+    def eval(
+            self,
+            token: int,
+            state_in: Optional[NumpyArrayOrPyTorchTensor],
+            state_out: Optional[NumpyArrayOrPyTorchTensor] = None,
+            logits_out: Optional[NumpyArrayOrPyTorchTensor] = None,
+            use_numpy: bool = False
+    ) -> Tuple[NumpyArrayOrPyTorchTensor, NumpyArrayOrPyTorchTensor]:
+        """
+        Evaluates the model for a single token.
+        In case of any error, this method will throw an exception.
+
+        Parameters
+        ----------
+        token : int
+            Index of next token to be seen by the model. Must be in range 0 <= token < n_vocab.
+        state_in : Optional[NumpyArrayOrTorchTensor]
+            State from previous call of this method. If this is a first pass, set it to None.
+        state_out : Optional[NumpyArrayOrTorchTensor]
+            Optional output tensor for state. If provided, must be of type float32, contiguous and of shape (state_buffer_element_count).
+        logits_out : Optional[NumpyArrayOrTorchTensor]
+            Optional output tensor for logits. If provided, must be of type float32, contiguous and of shape (logits_buffer_element_count).
+        use_numpy : bool
+            If set to True, numpy's ndarrays will be created instead of PyTorch's Tensors.
+            This parameter is ignored if any tensor parameter is not None; in such case,
+            type of returned tensors will match the type of received tensors.
+
+        Returns
+        -------
+        logits, state
+            Logits vector of shape (n_vocab); state for the next step.
+        """
+
+        assert self._valid, 'Model was freed'
+
+        use_numpy = self._detect_numpy_usage([state_in, state_out, logits_out], use_numpy)
+
+        if state_in is not None:
+            self._validate_tensor(state_in, 'state_in', self._state_buffer_element_count)
+
+            state_in_ptr = self._get_data_ptr(state_in)
+        else:
+            state_in_ptr = 0
+
+        if state_out is not None:
+            self._validate_tensor(state_out, 'state_out', self._state_buffer_element_count)
+        else:
+            state_out = self._zeros_float32(self._state_buffer_element_count, use_numpy)
+
+        if logits_out is not None:
+            self._validate_tensor(logits_out, 'logits_out', self._logits_buffer_element_count)
+        else:
+            logits_out = self._zeros_float32(self._logits_buffer_element_count, use_numpy)
+
+        self._library.rwkv_eval(
+            self._ctx,
+            token,
+            state_in_ptr,
+            self._get_data_ptr(state_out),
+            self._get_data_ptr(logits_out)
+        )
+
+        return logits_out, state_out
+    
+    def eval_sequence(
+            self,
+            tokens: List[int],
+            state_in: Optional[NumpyArrayOrPyTorchTensor],
+            state_out: Optional[NumpyArrayOrPyTorchTensor] = None,
+            logits_out: Optional[NumpyArrayOrPyTorchTensor] = None,
+            use_numpy: bool = False
+    ) -> Tuple[NumpyArrayOrPyTorchTensor, NumpyArrayOrPyTorchTensor]:
+        """
+        Evaluates the model for a sequence of tokens.
+
+        NOTE ON GGML NODE LIMIT
+
+        ggml has a hard-coded limit on max amount of nodes in a computation graph. The sequence graph is built in a way that quickly exceedes
+        this limit when using large models and/or large sequence lengths.
+        Fortunately, rwkv.cpp's fork of ggml has increased limit which was tested to work for sequence lengths up to 64 for 14B models.
+
+        If you get `GGML_ASSERT: ...\\ggml.c:16941: cgraph->n_nodes < GGML_MAX_NODES`, this means you've exceeded the limit.
+        To get rid of the assertion failure, reduce the model size and/or sequence length.
+
+        In case of any error, this method will throw an exception.
+
+        Parameters
+        ----------
+        tokens : List[int]
+            Indices of the next tokens to be seen by the model. Must be in range 0 <= token < n_vocab.
+        state_in : Optional[NumpyArrayOrTorchTensor]
+            State from previous call of this method. If this is a first pass, set it to None.
+        state_out : Optional[NumpyArrayOrTorchTensor]
+            Optional output tensor for state. If provided, must be of type float32, contiguous and of shape (state_buffer_element_count).
+        logits_out : Optional[NumpyArrayOrTorchTensor]
+            Optional output tensor for logits. If provided, must be of type float32, contiguous and of shape (logits_buffer_element_count).
+        use_numpy : bool
+            If set to True, numpy's ndarrays will be created instead of PyTorch's Tensors.
+            This parameter is ignored if any tensor parameter is not None; in such case,
+            type of returned tensors will match the type of received tensors.
+
+        Returns
+        -------
+        logits, state
+            Logits vector of shape (n_vocab); state for the next step.
+        """
+
+        assert self._valid, 'Model was freed'
+
+        use_numpy = self._detect_numpy_usage([state_in, state_out, logits_out], use_numpy)
+
+        if state_in is not None:
+            self._validate_tensor(state_in, 'state_in', self._state_buffer_element_count)
+
+            state_in_ptr = self._get_data_ptr(state_in)
+        else:
+            state_in_ptr = 0
+
+        if state_out is not None:
+            self._validate_tensor(state_out, 'state_out', self._state_buffer_element_count)
+        else:
+            state_out = self._zeros_float32(self._state_buffer_element_count, use_numpy)
+
+        if logits_out is not None:
+            self._validate_tensor(logits_out, 'logits_out', self._logits_buffer_element_count)
+        else:
+            logits_out = self._zeros_float32(self._logits_buffer_element_count, use_numpy)
+
+        self._library.rwkv_eval_sequence(
+            self._ctx,
+            tokens,
+            state_in_ptr,
+            self._get_data_ptr(state_out),
+            self._get_data_ptr(logits_out)
+        )
+
+        return logits_out, state_out
+
+    def eval_sequence_in_chunks(
+            self,
+            tokens: List[int],
+            state_in: Optional[NumpyArrayOrPyTorchTensor],
+            state_out: Optional[NumpyArrayOrPyTorchTensor] = None,
+            logits_out: Optional[NumpyArrayOrPyTorchTensor] = None,
+            chunk_size: int = 16,
+            use_numpy: bool = False
+    ) -> Tuple[NumpyArrayOrPyTorchTensor, NumpyArrayOrPyTorchTensor]:
+        """
+        Evaluates the model for a sequence of tokens using `eval_sequence`, splitting a potentially long sequence into fixed-length chunks.
+        This function is useful for processing complete prompts and user input in chat & role-playing use-cases.
+        It is recommended to use this function instead of `eval_sequence` to avoid mistakes and get maximum performance.
+
+        Chunking allows processing sequences of thousands of tokens, while not reaching the ggml's node limit and not consuming too much memory.
+        A reasonable and recommended value of chunk size is 16. If you want maximum performance, try different chunk sizes in range [2..64]
+        and choose one that works the best in your use case.
+
+        In case of any error, this method will throw an exception.
+
+        Parameters
+        ----------
+        tokens : List[int]
+            Indices of the next tokens to be seen by the model. Must be in range 0 <= token < n_vocab.
+        chunk_size : int
+            Size of each chunk in tokens, must be positive.
+        state_in : Optional[NumpyArrayOrTorchTensor]
+            State from previous call of this method. If this is a first pass, set it to None.
+        state_out : Optional[NumpyArrayOrTorchTensor]
+            Optional output tensor for state. If provided, must be of type float32, contiguous and of shape (state_buffer_element_count).
+        logits_out : Optional[NumpyArrayOrTorchTensor]
+            Optional output tensor for logits. If provided, must be of type float32, contiguous and of shape (logits_buffer_element_count).
+        use_numpy : bool
+            If set to True, numpy's ndarrays will be created instead of PyTorch's Tensors.
+            This parameter is ignored if any tensor parameter is not None; in such case,
+            type of returned tensors will match the type of received tensors.
+
+        Returns
+        -------
+        logits, state
+            Logits vector of shape (n_vocab); state for the next step.
+        """
+
+        assert self._valid, 'Model was freed'
+
+        use_numpy = self._detect_numpy_usage([state_in, state_out, logits_out], use_numpy)
+
+        if state_in is not None:
+            self._validate_tensor(state_in, 'state_in', self._state_buffer_element_count)
+
+            state_in_ptr = self._get_data_ptr(state_in)
+        else:
+            state_in_ptr = 0
+
+        if state_out is not None:
+            self._validate_tensor(state_out, 'state_out', self._state_buffer_element_count)
+        else:
+            state_out = self._zeros_float32(self._state_buffer_element_count, use_numpy)
+
+        if logits_out is not None:
+            self._validate_tensor(logits_out, 'logits_out', self._logits_buffer_element_count)
+        else:
+            logits_out = self._zeros_float32(self._logits_buffer_element_count, use_numpy)
+
+        self._library.rwkv_eval_sequence_in_chunks(
+            self._ctx,
+            tokens,
+            chunk_size,
+            state_in_ptr,
+            self._get_data_ptr(state_out),
+            self._get_data_ptr(logits_out)
+        )
+
+        return logits_out, state_out
+
+    def free(self) -> None:
+        """
+        Frees all allocated resources.
+        In case of any error, this method will throw an exception.
+        The object must not be used anymore after calling this method.
+        """
+
+        assert self._valid, 'Already freed'
+
+        self._valid = False
+
+        self._library.rwkv_free(self._ctx)
+
+    def __del__(self) -> None:
+        # Free the context on GC in case user forgot to call free() explicitly.
+        if hasattr(self, '_valid') and self._valid:
+            self.free()
+
+    def _is_pytorch_tensor(self, tensor: NumpyArrayOrPyTorchTensor) -> bool:
+        return hasattr(tensor, '__module__') and tensor.__module__ == 'torch'
+
+    def _detect_numpy_usage(self, tensors: List[Optional[NumpyArrayOrPyTorchTensor]], use_numpy_by_default: bool) -> bool:
+        for tensor in tensors:
+            if tensor is not None:
+                return False if self._is_pytorch_tensor(tensor) else True
+
+        return use_numpy_by_default
+
+    def _validate_tensor(self, tensor: NumpyArrayOrPyTorchTensor, name: str, size: int) -> None:
+        if self._is_pytorch_tensor(tensor):
+            tensor: torch.Tensor = tensor
+            assert tensor.device == torch.device('cpu'), f'{name} is not on CPU'
+            assert tensor.dtype == torch.float32, f'{name} is not of type float32'
+            assert tensor.shape == (size,), f'{name} has invalid shape {tensor.shape}, expected ({size})'
+            assert tensor.is_contiguous(), f'{name} is not contiguous'
+        else:
+            import numpy as np
+            tensor: np.ndarray = tensor
+            assert tensor.dtype == np.float32, f'{name} is not of type float32'
+            assert tensor.shape == (size,), f'{name} has invalid shape {tensor.shape}, expected ({size})'
+            assert tensor.data.contiguous, f'{name} is not contiguous'
+
+    def _get_data_ptr(self, tensor: NumpyArrayOrPyTorchTensor):
+        if self._is_pytorch_tensor(tensor):
+            return tensor.data_ptr()
+        else:
+            return tensor.ctypes.data
+
+    def _zeros_float32(self, element_count: int, use_numpy: bool) -> NumpyArrayOrPyTorchTensor:
+        if use_numpy:
+            import numpy as np
+            return np.zeros(element_count, dtype=np.float32)
+        else:
+            return torch.zeros(element_count, dtype=torch.float32, device='cpu')

backend-python/rwkv_pip/cpp/rwkv_cpp_shared_library.py

@@ -0,0 +1,444 @@
+import os
+import sys
+import ctypes
+import pathlib
+import platform
+from typing import Optional, List, Tuple, Callable
+
+QUANTIZED_FORMAT_NAMES: Tuple[str, str, str, str, str] = (
+    'Q4_0',
+    'Q4_1',
+    'Q5_0',
+    'Q5_1',
+    'Q8_0'
+)
+
+P_FLOAT = ctypes.POINTER(ctypes.c_float)
+P_INT = ctypes.POINTER(ctypes.c_int32)
+
+class RWKVContext:
+
+    def __init__(self, ptr: ctypes.pointer) -> None:
+        self.ptr: ctypes.pointer = ptr
+
+class RWKVSharedLibrary:
+    """
+    Python wrapper around rwkv.cpp shared library.
+    """
+
+    def __init__(self, shared_library_path: str) -> None:
+        """
+        Loads the shared library from specified file.
+        In case of any error, this method will throw an exception.
+
+        Parameters
+        ----------
+        shared_library_path : str
+            Path to rwkv.cpp shared library. On Windows, it would look like 'rwkv.dll'. On UNIX, 'rwkv.so'.
+        """
+        #  When Python is greater than 3.8, we need to reprocess the custom dll
+        #  according to the documentation to prevent loading failure errors.
+        #  https://docs.python.org/3/whatsnew/3.8.html#ctypes
+        if platform.system().lower() == 'windows':
+            self.library = ctypes.CDLL(shared_library_path, winmode=0)
+        else:
+            self.library = ctypes.cdll.LoadLibrary(shared_library_path)
+
+        self.library.rwkv_init_from_file.argtypes = [ctypes.c_char_p, ctypes.c_uint32]
+        self.library.rwkv_init_from_file.restype = ctypes.c_void_p
+
+        self.library.rwkv_gpu_offload_layers.argtypes = [ctypes.c_void_p, ctypes.c_uint32]
+        self.library.rwkv_gpu_offload_layers.restype = ctypes.c_bool
+
+        self.library.rwkv_eval.argtypes = [
+            ctypes.c_void_p, # ctx
+            ctypes.c_int32, # token
+            P_FLOAT, # state_in
+            P_FLOAT, # state_out
+            P_FLOAT  # logits_out
+        ]
+        self.library.rwkv_eval.restype = ctypes.c_bool
+
+        self.library.rwkv_eval_sequence.argtypes = [
+            ctypes.c_void_p, # ctx
+            P_INT, # tokens
+            ctypes.c_size_t, # token count
+            P_FLOAT, # state_in
+            P_FLOAT, # state_out
+            P_FLOAT  # logits_out
+        ]
+        self.library.rwkv_eval_sequence.restype = ctypes.c_bool
+
+        self.library.rwkv_eval_sequence_in_chunks.argtypes = [
+            ctypes.c_void_p, # ctx
+            P_INT, # tokens
+            ctypes.c_size_t, # token count
+            ctypes.c_size_t, # chunk size
+            P_FLOAT, # state_in
+            P_FLOAT, # state_out
+            P_FLOAT  # logits_out
+        ]
+        self.library.rwkv_eval_sequence_in_chunks.restype = ctypes.c_bool
+
+        self.library.rwkv_get_n_vocab.argtypes = [ctypes.c_void_p]
+        self.library.rwkv_get_n_vocab.restype = ctypes.c_size_t
+
+        self.library.rwkv_get_n_embed.argtypes = [ctypes.c_void_p]
+        self.library.rwkv_get_n_embed.restype = ctypes.c_size_t
+
+        self.library.rwkv_get_n_layer.argtypes = [ctypes.c_void_p]
+        self.library.rwkv_get_n_layer.restype = ctypes.c_size_t
+
+        self.library.rwkv_get_state_buffer_element_count.argtypes = [ctypes.c_void_p]
+        self.library.rwkv_get_state_buffer_element_count.restype = ctypes.c_uint32
+
+        self.library.rwkv_get_logits_buffer_element_count.argtypes = [ctypes.c_void_p]
+        self.library.rwkv_get_logits_buffer_element_count.restype = ctypes.c_uint32
+
+        self.library.rwkv_free.argtypes = [ctypes.c_void_p]
+        self.library.rwkv_free.restype = None
+
+        self.library.rwkv_free.argtypes = [ctypes.c_void_p]
+        self.library.rwkv_free.restype = None
+
+        self.library.rwkv_quantize_model_file.argtypes = [ctypes.c_char_p, ctypes.c_char_p, ctypes.c_char_p]
+        self.library.rwkv_quantize_model_file.restype = ctypes.c_bool
+
+        self.library.rwkv_get_system_info_string.argtypes = []
+        self.library.rwkv_get_system_info_string.restype = ctypes.c_char_p
+
+        self.nullptr = ctypes.cast(0, ctypes.c_void_p)
+
+    def rwkv_init_from_file(self, model_file_path: str, thread_count: int) -> RWKVContext:
+        """
+        Loads the model from a file and prepares it for inference.
+        Throws an exception in case of any error. Error messages would be printed to stderr.
+
+        Parameters
+        ----------
+        model_file_path : str
+            Path to model file in ggml format.
+        thread_count : int
+            Count of threads to use, must be positive.
+        """
+
+        ptr = self.library.rwkv_init_from_file(model_file_path.encode('utf-8'), ctypes.c_uint32(thread_count))
+
+        assert ptr is not None, 'rwkv_init_from_file failed, check stderr'
+
+        return RWKVContext(ptr)
+
+    def rwkv_gpu_offload_layers(self, ctx: RWKVContext, layer_count: int) -> bool:
+        """
+        Offloads specified count of model layers onto the GPU. Offloaded layers are evaluated using cuBLAS or CLBlast.
+        For the purposes of this function, model head (unembedding matrix) is treated as an additional layer:
+        - pass `rwkv_get_n_layer(ctx)` to offload all layers except model head
+        - pass `rwkv_get_n_layer(ctx) + 1` to offload all layers, including model head
+        Returns true if at least one layer was offloaded.
+        If rwkv.cpp was compiled without cuBLAS and CLBlast support, this function is a no-op and always returns false.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        layer_count : int
+            Count of layers to offload onto the GPU, must be >= 0.
+        """
+
+        assert layer_count >= 0, 'Layer count must be >= 0'
+
+        return self.library.rwkv_gpu_offload_layers(ctx.ptr, ctypes.c_uint32(layer_count))
+
+    def rwkv_eval(
+            self,
+            ctx: RWKVContext,
+            token: int,
+            state_in_address: Optional[int],
+            state_out_address: int,
+            logits_out_address: int
+    ) -> None:
+        """
+        Evaluates the model for a single token.
+        Throws an exception in case of any error. Error messages would be printed to stderr.
+        Not thread-safe. For parallel inference, call rwkv_clone_context to create one rwkv_context for each thread.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        token : int
+            Next token index, in range 0 <= token < n_vocab.
+        state_in_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_state_buffer_element_count; or None, if this is a first pass.
+        state_out_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_state_buffer_element_count. This buffer will be written to.
+        logits_out_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_logits_buffer_element_count. This buffer will be written to.
+        """
+
+        assert self.library.rwkv_eval(
+            ctx.ptr,
+            ctypes.c_int32(token),
+            ctypes.cast(0 if state_in_address is None else state_in_address, P_FLOAT),
+            ctypes.cast(state_out_address, P_FLOAT),
+            ctypes.cast(logits_out_address, P_FLOAT)
+        ), 'rwkv_eval failed, check stderr'
+
+    def rwkv_eval_sequence(
+            self,
+            ctx: RWKVContext,
+            tokens: List[int],
+            state_in_address: Optional[int],
+            state_out_address: int,
+            logits_out_address: int
+    ) -> None:
+        """
+        Evaluates the model for a sequence of tokens.
+        Uses a faster algorithm than `rwkv_eval` if you do not need the state and logits for every token. Best used with sequence lengths of 64 or so.
+        Has to build a computation graph on the first call for a given sequence, but will use this cached graph for subsequent calls of the same sequence length.
+
+        NOTE ON GGML NODE LIMIT
+
+        ggml has a hard-coded limit on max amount of nodes in a computation graph. The sequence graph is built in a way that quickly exceedes
+        this limit when using large models and/or large sequence lengths.
+        Fortunately, rwkv.cpp's fork of ggml has increased limit which was tested to work for sequence lengths up to 64 for 14B models.
+
+        If you get `GGML_ASSERT: ...\\ggml.c:16941: cgraph->n_nodes < GGML_MAX_NODES`, this means you've exceeded the limit.
+        To get rid of the assertion failure, reduce the model size and/or sequence length.
+
+        Not thread-safe. For parallel inference, call `rwkv_clone_context` to create one rwkv_context for each thread.
+        Throws an exception in case of any error. Error messages would be printed to stderr.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        tokens : List[int]
+            Next token indices, in range 0 <= token < n_vocab.
+        state_in_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_state_buffer_element_count; or None, if this is a first pass.
+        state_out_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_state_buffer_element_count. This buffer will be written to.
+        logits_out_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_logits_buffer_element_count. This buffer will be written to.
+        """
+
+        assert self.library.rwkv_eval_sequence(
+            ctx.ptr,
+            ctypes.cast((ctypes.c_int32 * len(tokens))(*tokens), P_INT),
+            ctypes.c_size_t(len(tokens)),
+            ctypes.cast(0 if state_in_address is None else state_in_address, P_FLOAT),
+            ctypes.cast(state_out_address, P_FLOAT),
+            ctypes.cast(logits_out_address, P_FLOAT)
+        ), 'rwkv_eval_sequence failed, check stderr'
+
+    def rwkv_eval_sequence_in_chunks(
+            self,
+            ctx: RWKVContext,
+            tokens: List[int],
+            chunk_size: int,
+            state_in_address: Optional[int],
+            state_out_address: int,
+            logits_out_address: int
+    ) -> None:
+        """
+        Evaluates the model for a sequence of tokens using `rwkv_eval_sequence`, splitting a potentially long sequence into fixed-length chunks.
+        This function is useful for processing complete prompts and user input in chat & role-playing use-cases.
+        It is recommended to use this function instead of `rwkv_eval_sequence` to avoid mistakes and get maximum performance.
+
+        Chunking allows processing sequences of thousands of tokens, while not reaching the ggml's node limit and not consuming too much memory.
+        A reasonable and recommended value of chunk size is 16. If you want maximum performance, try different chunk sizes in range [2..64]
+        and choose one that works the best in your use case.
+
+        Not thread-safe. For parallel inference, call `rwkv_clone_context` to create one rwkv_context for each thread.
+        Throws an exception in case of any error. Error messages would be printed to stderr.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        tokens : List[int]
+            Next token indices, in range 0 <= token < n_vocab.
+        chunk_size : int
+            Size of each chunk in tokens, must be positive.
+        state_in_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_state_buffer_element_count; or None, if this is a first pass.
+        state_out_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_state_buffer_element_count. This buffer will be written to.
+        logits_out_address : int
+            Address of the first element of a FP32 buffer of size rwkv_get_logits_buffer_element_count. This buffer will be written to.
+        """
+
+        assert self.library.rwkv_eval_sequence_in_chunks(
+            ctx.ptr,
+            ctypes.cast((ctypes.c_int32 * len(tokens))(*tokens), P_INT),
+            ctypes.c_size_t(len(tokens)),
+            ctypes.c_size_t(chunk_size),
+            ctypes.cast(0 if state_in_address is None else state_in_address, P_FLOAT),
+            ctypes.cast(state_out_address, P_FLOAT),
+            ctypes.cast(logits_out_address, P_FLOAT)
+        ), 'rwkv_eval_sequence_in_chunks failed, check stderr'
+
+    def rwkv_get_n_vocab(self, ctx: RWKVContext) -> int:
+        """
+        Returns the number of tokens in the given model's vocabulary.
+        Useful for telling 20B_tokenizer models (n_vocab = 50277) apart from World models (n_vocab = 65536).
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        """
+
+        return self.library.rwkv_get_n_vocab(ctx.ptr)
+
+    def rwkv_get_n_embed(self, ctx: RWKVContext) -> int:
+        """
+        Returns the number of elements in the given model's embedding.
+        Useful for reading individual fields of a model's hidden state.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        """
+
+        return self.library.rwkv_get_n_embed(ctx.ptr)
+
+    def rwkv_get_n_layer(self, ctx: RWKVContext) -> int:
+        """
+        Returns the number of layers in the given model.
+        A layer is a pair of RWKV and FFN operations, stacked multiple times throughout the model.
+        Embedding matrix and model head (unembedding matrix) are NOT counted in `n_layer`.
+        Useful for always offloading the entire model to GPU.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        """
+
+        return self.library.rwkv_get_n_layer(ctx.ptr)
+
+    def rwkv_get_state_buffer_element_count(self, ctx: RWKVContext) -> int:
+        """
+        Returns count of FP32 elements in state buffer.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        """
+
+        return self.library.rwkv_get_state_buffer_element_count(ctx.ptr)
+
+    def rwkv_get_logits_buffer_element_count(self, ctx: RWKVContext) -> int:
+        """
+        Returns count of FP32 elements in logits buffer.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        """
+
+        return self.library.rwkv_get_logits_buffer_element_count(ctx.ptr)
+
+    def rwkv_free(self, ctx: RWKVContext) -> None:
+        """
+        Frees all allocated memory and the context.
+
+        Parameters
+        ----------
+        ctx : RWKVContext
+            RWKV context obtained from rwkv_init_from_file.
+        """
+
+        self.library.rwkv_free(ctx.ptr)
+
+        ctx.ptr = self.nullptr
+
+    def rwkv_quantize_model_file(self, model_file_path_in: str, model_file_path_out: str, format_name: str) -> None:
+        """
+        Quantizes FP32 or FP16 model to one of INT4 formats.
+        Throws an exception in case of any error. Error messages would be printed to stderr.
+
+        Parameters
+        ----------
+        model_file_path_in : str
+            Path to model file in ggml format, must be either FP32 or FP16.
+        model_file_path_out : str
+            Quantized model will be written here.
+        format_name : str
+            One of QUANTIZED_FORMAT_NAMES.
+        """
+
+        assert format_name in QUANTIZED_FORMAT_NAMES, f'Unknown format name {format_name}, use one of {QUANTIZED_FORMAT_NAMES}'
+
+        assert self.library.rwkv_quantize_model_file(
+            model_file_path_in.encode('utf-8'),
+            model_file_path_out.encode('utf-8'),
+            format_name.encode('utf-8')
+        ), 'rwkv_quantize_model_file failed, check stderr'
+
+    def rwkv_get_system_info_string(self) -> str:
+        """
+        Returns system information string.
+        """
+
+        return self.library.rwkv_get_system_info_string().decode('utf-8')
+
+def load_rwkv_shared_library() -> RWKVSharedLibrary:
+    """
+    Attempts to find rwkv.cpp shared library and load it.
+    To specify exact path to the library, create an instance of RWKVSharedLibrary explicitly.
+    """
+
+    file_name: str
+
+    if 'win32' in sys.platform or 'cygwin' in sys.platform:
+        file_name = 'rwkv.dll'
+    elif 'darwin' in sys.platform:
+        file_name = 'librwkv.dylib'
+    else:
+        file_name = 'librwkv.so'
+
+    # Possible sub-paths to the library relative to the repo dir.
+    child_paths: List[Callable[[pathlib.Path], pathlib.Path]] = [
+        # No lookup for Debug config here.
+        # I assume that if a user wants to debug the library,
+        # they will be able to find the library and set the exact path explicitly.
+        lambda p: p / 'backend-python' / 'rwkv_pip' / 'cpp' / file_name,
+        lambda p: p / 'bin' / 'Release' / file_name,
+        lambda p: p / 'bin' / file_name,
+        # Some people prefer to build in the "build" subdirectory.
+        lambda p: p / 'build' / 'bin' / 'Release' / file_name,
+        lambda p: p / 'build' / 'bin' / file_name,
+        lambda p: p / 'build' / file_name,
+        # Fallback.
+        lambda p: p / file_name
+    ]
+
+    working_dir: pathlib.Path = pathlib.Path(os.path.abspath(os.getcwd()))
+
+    parent_paths: List[pathlib.Path] = [
+        # Possible repo dirs relative to the working dir.
+        # ./python/rwkv_cpp
+        working_dir.parent.parent,
+        # ./python
+        working_dir.parent,
+        # .
+        working_dir,
+        # Repo dir relative to this Python file.
+        pathlib.Path(os.path.abspath(__file__)).parent.parent.parent
+    ]
+
+    for parent_path in parent_paths:
+        for child_path in child_paths:
+            full_path: pathlib.Path = child_path(parent_path)
+
+            if os.path.isfile(full_path):
+                return RWKVSharedLibrary(str(full_path))
+
+    assert False, (f'Failed to find {file_name} automatically; '
+                   f'you need to find the library and create RWKVSharedLibrary specifying the path to it')

backend-python/rwkv_pip/cuda/gemm_fp16_cublas.cpp

@@ -0,0 +1,75 @@
+#include <cublas_v2.h>
+#include <cuda.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/CUDAContext.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__));
+
+/*
+  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 at::cuda::OptionalCUDAGuard device_guard(device_of(a));
+  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 = at::cuda::getCurrentCUDABlasHandle();
+
+#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));
+  }
+}

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);
+}

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);
+}

backend-python/rwkv_pip/cuda/rwkv5_op.cpp

@@ -0,0 +1,34 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+#include <c10/cuda/CUDAGuard.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) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    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) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    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) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    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);
+}

backend-python/rwkv_pip/cuda/rwkv6.cu

@@ -0,0 +1,87 @@
+#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;
+    _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]);
+    __syncthreads();
+
+    for (int t = b*T*C + h*_N_ + i; t < (b+1)*T*C + h*_N_ + i; t += C)
+    {
+        __syncthreads();
+        w[i] = _w[t];
+        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);
+}

backend-python/rwkv_pip/cuda/rwkv6_op.cpp

@@ -0,0 +1,34 @@
+#include <torch/extension.h>
+#include "ATen/ATen.h"
+#include <c10/cuda/CUDAGuard.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) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    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) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    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) {
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(state));
+    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, "rwkv6 forward_bf16");
+    m.def("forward_fp16", &forward_fp16, "rwkv6 forward_fp16");
+    m.def("forward_fp32", &forward_fp32, "rwkv6 forward_fp32");
+}
+TORCH_LIBRARY(rwkv6, m) {
+    m.def("forward_bf16", forward_bf16);
+    m.def("forward_fp16", forward_fp16);
+    m.def("forward_fp32", forward_fp32);
+}

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
+}

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")
@@ -70,15 +78,28 @@ class PIPELINE:
     def decode(self, x):
         return self.tokenizer.decode(x)
 
+    def np_softmax(self, x: np.ndarray, axis: int):
+        x -= x.max(axis=axis, keepdims=True)
+        e: np.ndarray = np.exp(x)
+        return e / e.sum(axis=axis, keepdims=True)
+
     def sample_logits(self, logits, temperature=1.0, top_p=0.85, top_k=0):
-        probs = F.softmax(logits.float(), dim=-1)
+        if type(logits) == list:
+            logits = np.array(logits)
+        np_logits = type(logits) == np.ndarray
+        if np_logits:
+            probs = self.np_softmax(logits, axis=-1)
+        else:
+            probs = F.softmax(logits.float(), dim=-1)
         top_k = int(top_k)
-        if probs.device == torch.device("cpu"):
-            probs = probs.numpy()
+        # 'privateuseone' is the type of custom devices like `torch_directml.device()`
+        if np_logits or probs.device.type in ["cpu", "privateuseone"]:
+            if not np_logits:
+                probs = probs.cpu().numpy()
             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 +113,7 @@ 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 +148,13 @@ 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:])

backend-python/rwkv_pip/webgpu/model.py

@@ -0,0 +1,26 @@
+from typing import Any, List, Union
+
+try:
+    import web_rwkv_py as wrp
+except ModuleNotFoundError:
+    try:
+        from . import web_rwkv_py as wrp
+    except ImportError:
+        raise ModuleNotFoundError(
+            "web_rwkv_py not found, install it from https://github.com/cryscan/web-rwkv-py"
+        )
+
+
+class RWKV:
+    def __init__(self, model_path: str, strategy: str = None):
+        self.model = wrp.v5.Model(
+            model_path,
+            turbo=True,
+            quant=32 if "i8" in strategy else None,
+            quant_nf4=26 if "i4" in strategy else None,
+        )
+        self.w = {}  # fake weight
+        self.w["emb.weight"] = [0] * wrp.peek_info(model_path).num_vocab
+
+    def forward(self, tokens: List[int], state: Union[Any, None] = None):
+        return wrp.v5.run_one(self.model, tokens, state)

backend-python/utils/log.py

@@ -10,7 +10,7 @@ logger = logging.getLogger()
 logger.setLevel(logging.INFO)
 formatter = logging.Formatter("%(asctime)s - %(levelname)s\n%(message)s")
 fh = logging.handlers.RotatingFileHandler(
-    "api.log", mode="a", maxBytes=3 * 1024 * 1024, backupCount=3
+    "api.log", mode="a", maxBytes=3 * 1024 * 1024, backupCount=3, encoding="utf-8"
 )
 fh.setFormatter(formatter)
 logger.addHandler(fh)

backend-python/utils/midi.py

@@ -52,6 +52,8 @@ class VocabConfig:
     bin_name_to_program_name: Dict[str, str]
     # Mapping from program number to instrument name.
     instrument_names: Dict[str, str]
+    # Manual override for velocity bins. Each element is the max velocity value for that bin by index.
+    velocity_bins_override: Optional[List[int]] = None
 
     def __post_init__(self):
         self.validate()
@@ -116,6 +118,12 @@ class VocabConfig:
             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.velocity_bins_override:
+            print("VocabConfig is using velocity_bins_override. Ignoring velocity_exp.")
+            if len(self.velocity_bins_override) != self.velocity_bins:
+                raise ValueError(
+                    "velocity_bins_override must have same length as velocity_bins"
+                )
         if (
             self.ch10_instrument_bin_name
             and self.ch10_instrument_bin_name not in self.bin_instrument_names
@@ -156,6 +164,11 @@ class VocabUtils:
 
     def velocity_to_bin(self, velocity: float) -> int:
         velocity = max(0, min(velocity, self.cfg.velocity_events - 1))
+        if self.cfg.velocity_bins_override:
+            for i, v in enumerate(self.cfg.velocity_bins_override):
+                if velocity <= v:
+                    return i
+            return 0
         binsize = self.cfg.velocity_events / (self.cfg.velocity_bins - 1)
         if self.cfg.velocity_exp == 1.0:
             return ceil(velocity / binsize)
@@ -176,6 +189,8 @@ class VocabUtils:
             )
 
     def bin_to_velocity(self, bin: int) -> int:
+        if self.cfg.velocity_bins_override:
+            return self.cfg.velocity_bins_override[bin]
         binsize = self.cfg.velocity_events / (self.cfg.velocity_bins - 1)
         if self.cfg.velocity_exp == 1.0:
             return max(0, ceil(bin * binsize - 1))
@@ -358,13 +373,32 @@ class AugmentConfig:
             )
 
 
+@dataclass
+class FilterConfig:
+    # Whether to filter out MIDI files with duplicate MD5 hashes.
+    deduplicate_md5: bool
+    # Minimum time delay between notes in a file before splitting into multiple documents.
+    piece_split_delay: float
+    # Minimum length of a piece in milliseconds.
+    min_piece_length: float
+
+    @classmethod
+    def from_json(cls, path: str):
+        with open(path, "r") as f:
+            config = json.load(f)
+        return cls(**config)
+
+
 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:
+    cfg: VocabConfig,
+    filter_cfg: FilterConfig,
+    mid: mido.MidiFile,
+    augment: AugmentValues = None,
+) -> List[str]:
     utils = VocabUtils(cfg)
     if augment is None:
         augment = AugmentValues.default()
@@ -390,7 +424,9 @@ def convert_midi_to_str(
     }  # {channel: {(note, program) -> True}}
     started_flag = False
 
+    output_list = []
     output = ["<start>"]
+    output_length_ms = 0.0
     token_data_buffer: List[
         Tuple[int, int, int, float]
     ] = []  # need to sort notes between wait tokens
@@ -432,16 +468,33 @@ def convert_midi_to_str(
         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
+        nonlocal output, output_length_ms, 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 delta_time_ms > filter_cfg.piece_split_delay * 1000.0:
+            # check if any notes are still held
+            silent = True
+            for channel in channel_notes.keys():
+                if len(channel_notes[channel]) > 0:
+                    silent = False
+                    break
+            if silent:
+                flush_token_data_buffer()
+                output.append("<end>")
+                if output_length_ms > filter_cfg.min_piece_length * 1000.0:
+                    output_list.append(" ".join(output))
+                output = ["<start>"]
+                output_length_ms = 0.0
+                started_flag = False
         if started_flag:
             wait_tokens = utils.data_to_wait_tokens(delta_time_ms)
             if len(wait_tokens) > 0:
                 flush_token_data_buffer()
+                output_length_ms += delta_time_ms
                 output += wait_tokens
         delta_time_ms = 0.0
         token_data_buffer.append((prog, chan, note, vel * augment.velocity_mod_factor))
@@ -510,7 +563,9 @@ def convert_midi_to_str(
 
     flush_token_data_buffer()
     output.append("<end>")
-    return " ".join(output)
+    if output_length_ms > filter_cfg.min_piece_length * 1000.0:
+        output_list.append(" ".join(output))
+    return output_list
 
 
 def generate_program_change_messages(cfg: VocabConfig):
@@ -633,10 +688,10 @@ def token_to_midi_message(
                 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
+                        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

backend-python/utils/midi_filter_config.json

@@ -0,0 +1,5 @@
+{
+    "deduplicate_md5": true,
+    "piece_split_delay": 10000,
+    "min_piece_length": 0
+}

backend-python/utils/ngrok.py

@@ -1,11 +1,13 @@
 import os
-import sys
+import global_var
 
 
 def ngrok_connect():
     from pyngrok import ngrok, conf
 
-    conf.set_default(conf.PyngrokConfig(ngrok_path="./ngrok"))
+    conf.set_default(
+        conf.PyngrokConfig(ngrok_path="./ngrok.exe" if os.name == "nt" else "./ngrok")
+    )
     ngrok.set_auth_token(os.environ["ngrok_token"])
-    http_tunnel = ngrok.connect(8000 if len(sys.argv) == 1 else int(sys.argv[1]))
-    print(http_tunnel.public_url)
+    http_tunnel = ngrok.connect(global_var.get(global_var.Args).port)
+    print(f"ngrok url: {http_tunnel.public_url}")

backend-python/utils/rwkv.py

@@ -4,11 +4,10 @@ import os
 import pathlib
 import copy
 import re
-from typing import Dict, Iterable, List, Tuple, Union
+from typing import Dict, Iterable, List, Tuple, Union, Type
 from utils.log import quick_log
 from fastapi import HTTPException
 from pydantic import BaseModel, Field
-import numpy as np
 from routes import state_cache
 import global_var
 
@@ -21,33 +20,21 @@ os.environ["TORCH_EXTENSIONS_DIR"] = f"{pathlib.Path(__file__).parent.parent.res
 
 
 class RWKVType(Enum):
+    NoneType = auto()
     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.model import (
-                RWKV as Model,
-            )
-        from rwkv_pip.utils import PIPELINE
-
-        filename, _ = os.path.splitext(os.path.basename(model))
-        self.name = filename
-        self.model = Model(model, strategy)
-        self.pipeline = PIPELINE(self.model, tokens_path)
+    def __init__(self, model, pipeline):
+        self.name = "rwkv"
+        self.model = model
+        self.pipeline = pipeline
         self.model_state = None
         self.model_tokens = []
-        self.rwkv_type: RWKVType = None
+        self.rwkv_type: RWKVType = RWKVType.NoneType
+        self.tokenizer_len = len(model.w["emb.weight"])
 
         self.max_tokens_per_generation = 500
         self.temperature = 1
@@ -80,6 +67,8 @@ class AbstractRWKV(ABC):
         pass
 
     def get_embedding(self, input: str, fast_mode: bool) -> Tuple[List[float], int]:
+        import numpy as np
+
         if fast_mode:
             embedding, token_len = self.__fast_embedding(
                 self.fix_tokens(self.pipeline.encode(input)), None
@@ -234,6 +223,8 @@ class AbstractRWKV(ABC):
     def generate(
         self, prompt: str, stop: Union[str, List[str], None] = None
     ) -> Iterable[Tuple[str, str, int, int]]:
+        import numpy as np
+
         quick_log(None, None, "Generation Prompt:\n" + prompt)
         cache = None
         delta_prompt = prompt
@@ -243,7 +234,7 @@ class AbstractRWKV(ABC):
             )
         except HTTPException:
             pass
-        if cache is None or cache["prompt"] == "":
+        if cache is None or cache["prompt"] == "" or cache["state"] is None:
             self.model_state = None
             self.model_tokens = []
         else:
@@ -348,8 +339,8 @@ class AbstractRWKV(ABC):
 
 
 class TextRWKV(AbstractRWKV):
-    def __init__(self, model: str, strategy: str, tokens_path: str) -> None:
-        super().__init__(model, strategy, tokens_path)
+    def __init__(self, model, pipeline) -> None:
+        super().__init__(model, pipeline)
 
         self.CHUNK_LEN = 256
 
@@ -361,16 +352,16 @@ class TextRWKV(AbstractRWKV):
         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:
+        if self.tokenizer_len < 65536:
             self.rwkv_type = RWKVType.Raven
             self.user = "Bob"
             self.bot = "Alice"
             self.END_OF_LINE = 187
+        else:
+            self.rwkv_type = RWKVType.World
+            self.user = "User"
+            self.bot = "Assistant"
+            self.END_OF_LINE = 11
 
         self.AVOID_REPEAT_TOKENS = []
         AVOID_REPEAT = "，：？！"
@@ -469,8 +460,8 @@ The following is a coherent verbose detailed conversation between a girl named {
 
 
 class MusicRWKV(AbstractRWKV):
-    def __init__(self, model: str, strategy: str, tokens_path: str):
-        super().__init__(model, strategy, tokens_path)
+    def __init__(self, model, pipeline):
+        super().__init__(model, pipeline)
 
         self.max_tokens_per_generation = 500
         self.temperature = 1
@@ -510,6 +501,68 @@ class MusicRWKV(AbstractRWKV):
         return " " + delta
 
 
+def get_tokenizer(tokenizer_len: int):
+    tokenizer_dir = f"{pathlib.Path(__file__).parent.parent.resolve()}/rwkv_pip/"
+    if tokenizer_len < 50277:
+        return tokenizer_dir + "tokenizer-midi.json"
+    elif tokenizer_len < 65536:
+        return tokenizer_dir + "20B_tokenizer.json"
+    else:
+        return "rwkv_vocab_v20230424"
+
+
+def RWKV(model: str, strategy: str, tokenizer: Union[str, None]) -> AbstractRWKV:
+    rwkv_beta = global_var.get(global_var.Args).rwkv_beta
+    rwkv_cpp = getattr(global_var.get(global_var.Args), "rwkv.cpp")
+    webgpu = global_var.get(global_var.Args).webgpu
+
+    if "midi" in model.lower() or "abc" in model.lower():
+        os.environ["RWKV_RESCALE_LAYER"] = "999"
+
+    # dynamic import to make RWKV_CUDA_ON work
+    if rwkv_beta:
+        print("Using rwkv-beta")
+        from rwkv_pip.beta.model import (
+            RWKV as Model,
+        )
+    elif rwkv_cpp:
+        print("Using rwkv.cpp, strategy is ignored")
+        from rwkv_pip.cpp.model import (
+            RWKV as Model,
+        )
+    elif webgpu:
+        print("Using webgpu")
+        from rwkv_pip.webgpu.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))
+    model = Model(model, strategy)
+    if not tokenizer:
+        tokenizer = get_tokenizer(len(model.w["emb.weight"]))
+    pipeline = PIPELINE(model, tokenizer)
+
+    rwkv_map: dict[str, Type[AbstractRWKV]] = {
+        "20B_tokenizer": TextRWKV,
+        "rwkv_vocab_v20230424": TextRWKV,
+        "tokenizer-midi": MusicRWKV,
+    }
+    tokenizer_name = os.path.splitext(os.path.basename(tokenizer))[0]
+    rwkv: AbstractRWKV
+    if tokenizer_name in rwkv_map:
+        rwkv = rwkv_map[tokenizer_name](model, pipeline)
+    else:
+        rwkv = TextRWKV(model, pipeline)
+    rwkv.name = filename
+
+    return rwkv
+
+
 class ModelConfigBody(BaseModel):
     max_tokens: int = Field(default=None, gt=0, le=102400)
     temperature: float = Field(default=None, ge=0, le=2)
@@ -517,8 +570,8 @@ class ModelConfigBody(BaseModel):
     presence_penalty: float = Field(default=None, ge=-2, le=2)
     frequency_penalty: float = Field(default=None, ge=-2, le=2)
 
-    class Config:
-        schema_extra = {
+    model_config = {
+        "json_schema_extra": {
             "example": {
                 "max_tokens": 1000,
                 "temperature": 1.2,
@@ -527,6 +580,7 @@ class ModelConfigBody(BaseModel):
                 "frequency_penalty": 0.4,
             }
         }
+    }
 
 
 def set_rwkv_config(model: AbstractRWKV, body: ModelConfigBody):

backend-python/webui_server.py

@@ -0,0 +1,14 @@
+from fastapi import FastAPI
+from fastapi.middleware.gzip import GZipMiddleware
+from fastapi.staticfiles import StaticFiles
+import uvicorn
+
+webui_server = FastAPI()
+
+webui_server.add_middleware(GZipMiddleware, minimum_size=1000)
+webui_server.mount(
+    "/", StaticFiles(directory="frontend/dist", html=True), name="static"
+)
+
+if __name__ == "__main__":
+    uvicorn.run("webui_server:webui_server")

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

build/darwin/Readme_Install.txt

@@ -1,6 +1,11 @@
-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。
+Client Download URL:
+客户端下载地址:
+クライアントのダウンロードURL:
+https://github.com/josStorer/RWKV-Runner/releases/latest/download/RWKV-Runner_macos_universal.zip
+
+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.

build/linux/Readme_Install.txt

@@ -1,3 +1,8 @@
+Client Download URL:
+客户端下载地址:
+クライアントのダウンロードURL:
+https://github.com/josStorer/RWKV-Runner/releases/latest/download/RWKV-Runner_linux_x64
+
 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。

build/windows/Readme_Install.txt

@@ -1,3 +1,8 @@
+Client Download URL:
+客户端下载地址:
+クライアントのダウンロードURL:
+https://github.com/josStorer/RWKV-Runner/releases/latest/download/RWKV-Runner_windows_x64.exe
+
 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.
 このプログラムを空のディレクトリで実行してください。関連するすべての依存関係は、このディレクトリに配置されます。

deploy-examples/ChatGPT-Next-Web/setup.sh

@@ -9,7 +9,7 @@ cd RWKV-Next-Web
 git clone https://github.com/josStorer/RWKV-Runner --depth=1
 python3 -m pip install torch torchvision torchaudio
 python3 -m pip install -r RWKV-Runner/backend-python/requirements.txt
-python3 ./RWKV-Runner/backend-python/main.py > log.txt &
+python3 ./RWKV-Runner/backend-python/main.py > log.txt & # this is only an example, you should use screen or other tools to run it in background
 
 if [ ! -d RWKV-Runner/models ]; then
     mkdir RWKV-Runner/models
@@ -22,6 +22,6 @@ yarn install
 yarn build
 export PROXY_URL=""
 export BASE_URL=http://127.0.0.1:8000
-yarn start &
+yarn start & # this is only an example, you should use screen or other tools to run it in background
 
 curl http://127.0.0.1:8000/switch-model -X POST -H "Content-Type: application/json" -d '{"model":"./RWKV-Runner/models/RWKV-4-World-0.1B-v1-20230520-ctx4096.pth","strategy":"cpu fp32"}'

deploy-examples/RWKV-Runner-WebUI/setup.bat

@@ -0,0 +1,19 @@
+: install git python3.10 npm by yourself
+: change model and strategy according to your hardware
+
+git clone https://github.com/josStorer/RWKV-Runner --depth=1
+python -m pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 --index-url https://download.pytorch.org/whl/cu117
+python -m pip install -r RWKV-Runner/backend-python/requirements.txt
+cd RWKV-Runner/frontend
+call npm ci
+call npm run build
+cd ..
+
+: optional: set ngrok_token=YOUR_NGROK_TOKEN
+start python ./backend-python/main.py --webui
+start "C:\Program Files (x86)\Microsoft\Edge\Application\msedge.exe" "http://127.0.0.1:8000"
+
+powershell -Command "(Test-Path ./models) -or (mkdir models)"
+powershell -Command "Import-Module BitsTransfer"
+powershell -Command "(Test-Path ./models/RWKV-4-World-1.5B-v1-fixed-20230612-ctx4096.pth) -or (Start-BitsTransfer https://huggingface.co/BlinkDL/rwkv-4-world/resolve/main/RWKV-4-World-1.5B-v1-fixed-20230612-ctx4096.pth ./models/RWKV-4-World-1.5B-v1-fixed-20230612-ctx4096.pth)"
+powershell -Command "Invoke-WebRequest http://127.0.0.1:8000/switch-model -Method POST -ContentType 'application/json' -Body '{\"model\":\"./models/RWKV-4-World-1.5B-v1-fixed-20230612-ctx4096.pth\",\"strategy\":\"cuda fp32 *20+\",\"deploy\":\"true\"}'"

deploy-examples/RWKV-Runner-WebUI/setup.sh

@@ -0,0 +1,22 @@
+# install git python3.10 npm by yourself
+# change model and strategy according to your hardware
+
+sudo apt install python3-dev
+
+git clone https://github.com/josStorer/RWKV-Runner --depth=1
+python3 -m pip install torch torchvision torchaudio
+python3 -m pip install -r RWKV-Runner/backend-python/requirements.txt
+cd RWKV-Runner/frontend
+npm ci
+npm run build
+cd ..
+
+# optional: export ngrok_token=YOUR_NGROK_TOKEN
+python3 ./backend-python/main.py --webui > log.txt & # this is only an example, you should use screen or other tools to run it in background
+
+if [ ! -d models ]; then
+    mkdir models
+fi
+wget -N https://huggingface.co/BlinkDL/rwkv-4-world/resolve/main/RWKV-4-World-0.1B-v1-20230520-ctx4096.pth -P models/
+
+curl http://127.0.0.1:8000/switch-model -X POST -H "Content-Type: application/json" -d '{"model":"./models/RWKV-4-World-0.1B-v1-20230520-ctx4096.pth","strategy":"cpu fp32","deploy":"true"}'

exportModelsJson.js

@@ -19,14 +19,15 @@ document.querySelectorAll('.grid.h-10.grid-cols-12.place-content-center.gap-x-3.
   if (!data.name.endsWith('.bin') && !data.name.endsWith('.pth'))
     return
 
-  data.desc = {en: '', zh: ''}
+  data.desc = { en: '', zh: '', ja: '' }
   const rawText = await (await fetch(e.children[1].href.replace('/resolve/', '/raw/'))).text()
 
   data.size = parseInt(extractValue(rawText, 'size'))
   data.SHA256 = extractValue(rawText, 'oid sha256:')
   data.lastUpdated = e.children[3].children[0].getAttribute('datetime')
-  data.url = e.children[1].href.replace('/resolve/', '/blob/')
-  data.downloadUrl = e.children[1].href
+  data.url = e.children[1].href.replace('/resolve/', '/blob/').replace('?download=true', '')
+  data.downloadUrl = e.children[1].href.replace('?download=true', '')
+  data.tags = []
 
   modelsJson.push(data)
 })

finetune/data/sample.jsonl

@@ -1,7 +1,5 @@
-{"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)."}
+{"text": "The following is an epic science fiction masterpiece that is immortalized, with delicate descriptions and grand depictions of interstellar civilization wars.\nChapter 1.\nAs I sit down to write here amidst the shadows of vine-leaves under the blue sky of southern Italy, it comes to me with a certain quality of astonishment that my participation in these amazing adventures of Mr. Cavor was, after all, the outcome of the purest accident. It might have been any one. I fell into these things at a time when I thought myself removed from the slightest possibility of disturbing experiences. I had gone to Lympne because I had imagined it the most uneventful place in the world. “Here, at any rate,” said I, “I shall find peace and a chance to work!”"}
+{"text": "Translate the following into Chinese.\n\nEnglish: What rooms do you have available?\nChinese: 你们有哪些房间可以提供"}
+{"text": "User: Hello.\n\nAssistant: I'm here, meow~.\n\nUser: Can you tell some jokes?\n\nAssistant: Of course, master. What kind of jokes would you like to hear?"}
+{"text": "Instruction: Write a story using the following information\n\nInput: A man named Alex chops a tree down\n\nResponse: Once upon a time, there was a man named Alex who lived in the heart of the forest. He had always been fascinated by trees and spent most of his days exploring the forest and learning about its many wonders. One day, while wandering through the woods, he stumbled upon an old oak tree that stood tall and proud in the middle of a clearing."}
+{"text": "def get_args(args: Union[Sequence[str], None] = None):\n    parser = argparse.ArgumentParser()\n    group = parser.add_argument_group(title=\"server arguments\")\n    group.add_argument(\n        \"--port\",\n        type=int,\n        default=8000,\n        help=\"port to run the server on (default: 8000)\",\n    )\n    group.add_argument(\n        \"--host\",\n        type=str,\n        default=\"127.0.0.1\",\n        help=\"host to run the server on (default: 127.0.0.1)\",\n    )"}

finetune/get_layer_and_embd.py

@@ -23,6 +23,7 @@ def file_cleaner(file):
     return cleaner
 
 
+expected_max_version = float(sys.argv[2]) if len(sys.argv) > 2 else 100
 model_file = open(sys.argv[1], "rb")
 cleaner = file_cleaner(model_file)
 cleaner_thread = threading.Thread(target=cleaner, daemon=True)
@@ -34,8 +35,23 @@ gc.collect()
 n_embd = w["emb.weight"].shape[1]
 n_layer = 0
 keys = list(w.keys())
+version = 4
 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="")
+    if "ln_x" in x:
+        version = max(5, version)
+    if "gate.weight" in x:
+        version = max(5.1, version)
+    if int(version) == 5 and "att.time_decay" in x:
+        if len(w[x].shape) > 1:
+            if w[x].shape[1] > 1:
+                version = max(5.2, version)
+    if "time_maa" in x:
+        version = max(6, version)
+
+if version <= expected_max_version:
+    print(f"--n_layer {n_layer} --n_embd {n_embd}", end="")
+else:
+    raise Exception(f"RWKV{version} is not supported")

finetune/install-wsl-dep-and-train.sh

@@ -1,3 +1,5 @@
+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
@@ -45,8 +47,12 @@ else
 fi
 
 echo "loading $loadModel"
-modelInfo=$(python3 ./finetune/get_layer_and_embd.py $loadModel)
+modelInfo=$(python3 ./finetune/get_layer_and_embd.py $loadModel 4)
 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
+if [[ $modelInfo =~ "--n_layer" ]]; then
+  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
+else
+  echo "modelInfo is invalid"
+  exit 1
+fi

finetune/json2binidx_tool/tools/preprocess_data.py

@@ -246,5 +246,6 @@ if __name__ == "__main__":
     try:
         main()
     except Exception as e:
+        print(e)
         with open("error.txt", "w") as f:
             f.write(str(e))

finetune/lora/merge_lora.py

@@ -64,5 +64,6 @@ try:
 
         torch.save(output_w, output)
 except Exception as e:
+    print(e)
     with open("error.txt", "w") as f:
         f.write(str(e))

finetune/lora/train.py

@@ -184,7 +184,7 @@ if __name__ == "__main__":
     args.num_sanity_val_steps = 0
     args.check_val_every_n_epoch = int(1e20)
     args.log_every_n_steps = int(1e20)
-    args.max_epochs = -1  # continue forever
+    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)
@@ -264,7 +264,7 @@ if __name__ == "__main__":
 #
 # 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
+# Epoch = {args.epoch_begin} to {args.epoch_begin + args.epoch_count - 1}, save every {args.epoch_save} epoch
 #
 # Each "epoch" = {args.epoch_steps} steps, {samples_per_epoch} samples, {tokens_per_epoch} tokens
 #
@@ -373,7 +373,7 @@ if __name__ == "__main__":
                     for param in module.parameters():
                         param.requires_grad = True
                 elif enable_time_finetune and any(
-                    n.startswith("time") for n, _ in module.named_parameters()
+                        n.startswith("time") for n, _ in module.named_parameters()
                 ):
                     for pname, param in module.named_parameters():
                         if pname.startswith("time"):
@@ -381,7 +381,7 @@ if __name__ == "__main__":
                             param.requires_grad = True
 
     if (
-        len(args.load_model) == 0 or args.my_pile_stage == 1
+            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
@@ -423,8 +423,8 @@ if __name__ == "__main__":
     )
 
     if (
-        args.lr_init > 1e-4
-        or trainer.world_size * args.micro_bsz * trainer.accumulate_grad_batches < 8
+            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:
@@ -459,10 +459,10 @@ if __name__ == "__main__":
 
     if "deepspeed" in args.strategy:
         trainer.strategy.config["zero_optimization"]["allgather_bucket_size"] = (
-            args.ds_bucket_mb * 1000 * 1000
+                args.ds_bucket_mb * 1000 * 1000
         )
         trainer.strategy.config["zero_optimization"]["reduce_bucket_size"] = (
-            args.ds_bucket_mb * 1000 * 1000
+                args.ds_bucket_mb * 1000 * 1000
         )
 
     # must set shuffle=False, persistent_workers=False (because worker is in another thread)

finetune/requirements.txt

@@ -1,3 +1,3 @@
 torch==1.13.1
 pytorch_lightning==1.9.5
-deepspeed
+deepspeed==0.11.2

frontend/index.html

@@ -1,9 +1,10 @@
 <!DOCTYPE html>
 <html lang="en">
 <head>
-    <meta charset="UTF-8"/>
-    <meta content="width=device-width, initial-scale=1.0" name="viewport"/>
-    <title>RWKV-Runner</title>
+  <meta charset="UTF-8" />
+  <meta content="width=device-width, initial-scale=1.0" name="viewport" />
+  <title>RWKV-Runner</title>
+  <link href="./src/assets/images/logo.png" rel="icon" type="image/x-icon">
 </head>
 <body>
 <div id="root"></div>

frontend/package.json

@@ -16,15 +16,18 @@
     "@primer/octicons-react": "^19.1.0",
     "chart.js": "^4.3.0",
     "classnames": "^2.3.2",
-    "github-markdown-css": "^5.2.0",
+    "file-saver": "^2.0.5",
     "html-midi-player": "^1.5.0",
     "i18next": "^22.4.15",
+    "lodash-es": "^4.17.21",
     "mobx": "^6.9.0",
     "mobx-react-lite": "^3.4.3",
+    "pdfjs-dist": "^4.0.189",
     "react": "^18.2.0",
     "react-beautiful-dnd": "^13.1.1",
     "react-chartjs-2": "^5.2.0",
     "react-dom": "^18.2.0",
+    "react-draggable": "^4.4.6",
     "react-i18next": "^12.2.2",
     "react-markdown": "^8.0.7",
     "react-router": "^6.11.1",
@@ -38,6 +41,8 @@
     "uuid": "^9.0.0"
   },
   "devDependencies": {
+    "@types/file-saver": "^2.0.7",
+    "@types/lodash-es": "^4.17.12",
     "@types/react": "^18.2.6",
     "@types/react-beautiful-dnd": "^13.1.4",
     "@types/react-dom": "^18.2.4",
@@ -49,6 +54,7 @@
     "sass": "^1.62.1",
     "tailwindcss": "^3.3.2",
     "typescript": "^5.0.4",
-    "vite": "^4.3.6"
+    "vite": "^4.3.6",
+    "vite-plugin-top-level-await": "^1.3.1"
   }
 }

frontend/src/App.tsx

@@ -26,18 +26,22 @@
 import { FluentProvider, Tab, TabList, webDarkTheme, webLightTheme } from '@fluentui/react-components';
 import { FC, useEffect, useState } from 'react';
 import { Route, Routes, useLocation, useNavigate } from 'react-router';
-import { pages } from './pages';
+import { pages as clientPages } from './pages';
 import { useMediaQuery } from 'usehooks-ts';
 import commonStore from './stores/commonStore';
 import { observer } from 'mobx-react-lite';
 import { useTranslation } from 'react-i18next';
 import { CustomToastContainer } from './components/CustomToastContainer';
+import { LazyImportComponent } from './components/LazyImportComponent';
 
 const App: FC = observer(() => {
   const { t } = useTranslation();
   const navigate = useNavigate();
   const location = useLocation();
   const mq = useMediaQuery('(min-width: 640px)');
+  const pages = commonStore.platform === 'web' ? clientPages.filter(page =>
+    !['/configs', '/models', '/downloads', '/train', '/about'].some(path => page.path === path)
+  ) : clientPages;
 
   const [path, setPath] = useState<string>(pages[0].path);
 
@@ -47,10 +51,10 @@ const App: FC = observer(() => {
   useEffect(() => setPath(location.pathname), [location]);
 
   return (
-    <FluentProvider className="h-screen"
+    <FluentProvider
       theme={commonStore.settings.darkMode ? webDarkTheme : webLightTheme}
       data-theme={commonStore.settings.darkMode ? 'dark' : 'light'}>
-      <div className="flex h-full">
+      <div className="flex h-screen">
         <div className="flex flex-col w-16 sm:w-48 p-2 justify-between">
           <TabList
             size="large"
@@ -82,7 +86,7 @@ const App: FC = observer(() => {
         <div className="h-full w-full p-2 box-border overflow-y-hidden">
           <Routes>
             {pages.map(({ path, element }, index) => (
-              <Route key={`${path}-${index}`} path={path} element={element} />
+              <Route key={`${path}-${index}`} path={path} element={<LazyImportComponent lazyChildren={element} />} />
             ))}
           </Routes>
         </div>

frontend/src/_locales/ja/main.json

@@ -13,7 +13,7 @@
   "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",
+  "Stored Layers": "保存されるレイヤー",
   "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。",
@@ -82,7 +82,7 @@
   "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。",
+  "int8 uses less VRAM, but has slightly lower quality. fp16 has higher 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。",
   "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",
@@ -100,7 +100,7 @@
   "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。",
+  "Enabling this option can greatly improve inference speed and save some VRAM, but there may be compatibility issues (output garbled). If it fails to start, please turn off this option, or try to upgrade your gpu driver.": "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 (文字化けを出力する) が生じる可能性があります。起動に失敗した場合は、このオプションを無効にするか、GPUドライバーをアップグレードしてみてください。",
   "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。",
@@ -128,7 +128,7 @@
   "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)",
+  "CUDA (Beta, Faster)": "CUDA (Beta, 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",
@@ -162,7 +162,7 @@
   "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。",
+  "Found no NVIDIA driver, please install the latest driver. If you are not using an Nvidia GPU, please switch the 'Strategy' to WebGPU or CPU in the Configs page.": "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。NvidiaのGPUを使用していない場合は、設定ページで\"Strategy\"をWebGPUまたはCPUに切り替えてください。",
   "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",
@@ -178,7 +178,7 @@
   "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 Messages": "メッセージの編集",
+  "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",
@@ -226,14 +226,14 @@
   "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 or use a smaller base model.": "メモリが不足しています、仮想メモリ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。",
+  "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",
+  "Can not find an Nvidia GPU. Perhaps the gpu driver of windows is too old, or you are using WSL 1 for training, please upgrade to WSL 2. e.g. Run \"wsl --set-version Ubuntu-22.04 2\"": "Nvidia GPUが見つかりません。WindowsのGPUドライバが古すぎるか、トレーニングにWSL 1を使用している可能性があります。WSL 2にアップグレードしてください。例：\"wsl --set-version Ubuntu-22.04 2\"を実行してください",
   "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。",
+  "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/josStorer/RWKV-Runner/blob/master/finetune/data/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/josStorer/RWKV-Runner/blob/master/finetune/data/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",
@@ -244,5 +244,84 @@
   "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"
+  "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 or rwkv_vocab_v20230424.txt)": "トークナイザーパス (例: backend-python/rwkv_pip/20B_tokenizer.json または rwkv_vocab_v20230424.txt)",
+  "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",
+  "Format Content": "内容フォーマットの規格化",
+  "Add An Attachment (Accepts pdf, txt)": "添付ファイルを追加 (pdf, txtを受け付けます)",
+  "Processing Attachment": "添付ファイルを処理中",
+  "Remove Attachment": "添付ファイルを削除",
+  "The content of file": "ファイル",
+  "is as follows. When replying to me, consider the file content and respond accordingly:": "の内容は以下の通りです。私に返信する際は、ファイルの内容を考慮して適切に返信してください:",
+  "What's the file name": "ファイル名は何ですか",
+  "The file name is: ": "ファイル名は次のとおりです: ",
+  "Port is occupied. Change it in Configs page or close the program that occupies the port.": "ポートが占有されています。設定ページで変更するか、ポートを占有しているプログラムを終了してください。",
+  "Loading...": "読み込み中...",
+  "Hello, what can I do for you?": "こんにちは、何かお手伝いできますか？",
+  "Enable WebUI": "WebUIを有効化",
+  "Server is working on deployment mode, please close the terminal window manually": "サーバーはデプロイモードで動作しています、ターミナルウィンドウを手動で閉じてください",
+  "Server is working on deployment mode, please exit the program manually to stop the server": "サーバーはデプロイモードで動作しています、サーバーを停止するにはプログラムを手動で終了してください",
+  "You can increase the number of stored layers in Configs page to improve performance": "パフォーマンスを向上させるために、保存されるレイヤーの数を設定ページで増やすことができます",
+  "Failed to load model, try to increase the virtual memory (Swap of WSL) or use a smaller base model.": "モデルの読み込みに失敗しました、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。",
+  "Save Conversation": "会話を保存",
+  "Use Hugging Face Mirror": "Hugging Faceミラーを使用",
+  "File is empty": "ファイルが空です",
+  "Open MIDI Input Audio Tracks": "MIDI入力オーディオトラックを開く",
+  "Track": "トラック",
+  "Play All": "すべて再生",
+  "Clear All": "すべてクリア",
+  "Scale View": "スケールビュー",
+  "Record": "録音",
+  "Play": "再生",
+  "New Track": "新規トラック",
+  "Select a track to preview the content": "トラックを選択して内容をプレビュー",
+  "Save to generation area": "生成エリアに保存",
+  "Piano": "ピアノ",
+  "Percussion": "パーカッション",
+  "Drum": "ドラム",
+  "Tuba": "チューバ",
+  "Marimba": "マリンバ",
+  "Bass": "ベース",
+  "Guitar": "ギター",
+  "Violin": "バイオリン",
+  "Trumpet": "トランペット",
+  "Sax": "サックス",
+  "Flute": "フルート",
+  "Lead": "リード",
+  "Pad": "パッド",
+  "MIDI Input": "MIDI入力",
+  "Select the MIDI input device to be used.": "使用するMIDI入力デバイスを選択します。",
+  "Start Time": "開始時間",
+  "Content Duration": "内容の長さ",
+  "Please select a MIDI device first": "まずMIDIデバイスを選択してください",
+  "Piano is the main instrument": "ピアノはメインの楽器です",
+  "Loss is too high, please check the training data, and ensure your gpu driver is up to date.": "Lossが大きすぎます、トレーニングデータを確認し、GPUドライバが最新であることを確認してください。",
+  "This version of RWKV is not supported yet.": "このバージョンのRWKVはまだサポートされていません。",
+  "Main": "メイン",
+  "Finetuned": "微調整",
+  "Global": "グローバル",
+  "Local": "ローカル",
+  "CN": "中国語",
+  "JP": "日本語",
+  "Music": "音楽",
+  "Other": "その他",
+  "Role Play": "ロールプレイ",
+  "Recommended": "おすすめ",
+  "Import MIDI": "MIDIをインポート",
+  "Current Instrument": "現在の楽器",
+  "Please convert model to GGML format first": "モデルをGGML形式に変換してください",
+  "Convert To GGML Format": "GGML形式に変換",
+  "CPU (rwkv.cpp, Faster)": "CPU (rwkv.cpp, 高速)",
+  "Play With External Player": "外部プレーヤーで再生",
+  "Core API URL": "コアAPI URL",
+  "Override core API URL(/chat/completions and /completions). If you don't know what this is, leave it blank.": "コアAPI URLを上書きします(/chat/completions と /completions)。何であるかわからない場合は空白のままにしてください。",
+  "Please change Strategy to CPU (rwkv.cpp) to use ggml format": "StrategyをCPU (rwkv.cpp)に変更して、ggml形式を使用してください",
+  "Only Auto Play Generated Content": "生成されたコンテンツのみ自動再生"
 }

frontend/src/_locales/zh-hans/main.json

@@ -82,7 +82,7 @@
   "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.": "就像给模型喂镇静剂. 考虑前 n% 概率质量的结果, 0.1 考虑前 10%, 质量更高, 但更保守, 1 考虑所有质量结果, 质量降低, 但更多样",
   "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.": "存在惩罚. 正值根据新token在至今的文本中是否出现过, 来对其进行惩罚, 从而增加了模型涉及新话题的可能性",
   "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.": "频率惩罚. 正值根据新token在至今的文本中出现的频率/次数, 来对其进行惩罚, 从而减少模型原封不动地重复相同句子的可能性",
-  "int8 uses less VRAM, but has slightly lower quality. fp16 has higher quality, and fp32 has the best quality.": "int8占用显存更低, 但质量略微下降. fp16质量更好, fp32质量最好",
+  "int8 uses less VRAM, but has slightly lower quality. fp16 has higher quality.": "int8占用显存更低, 但质量略微下降. fp16质量更好",
   "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)": "载入显存的神经网络层数, 载入越多, 速度越快, 但显存消耗越大 (如果你的显存不够, 会载入失败)",
   "Whether to use CPU to calculate the last output layer of the neural network with FP32 precision to obtain better quality.": "是否使用cpu以fp32精度计算神经网络的最后一层输出层, 以获得更好的质量",
   "Downloads": "下载",
@@ -100,7 +100,7 @@
   "Model Config Exception": "模型配置异常",
   "Use Gitee Updates Source": "使用Gitee更新源",
   "Use Custom CUDA kernel to Accelerate": "使用自定义CUDA算子加速",
-  "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.": "开启这个选项能大大提升推理速度并节省显存，但可能存在兼容性问题，如果启动失败，请关闭此选项",
+  "Enabling this option can greatly improve inference speed and save some VRAM, but there may be compatibility issues (output garbled). If it fails to start, please turn off this option, or try to upgrade your gpu driver.": "开启这个选项能大大提升推理速度并节省显存，但可能存在兼容性(回复乱码)问题，如果发生相关问题，请关闭此选项。或更新你的显卡驱动",
   "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",
@@ -162,7 +162,7 @@
   "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",
+  "Found no NVIDIA driver, please install the latest driver. If you are not using an Nvidia GPU, please switch the 'Strategy' to WebGPU or CPU in the Configs page.": "没有找到NVIDIA驱动，请安装最新驱动。如果你没有使用Nvidia显卡，请在配置页面将“Strategy”改为WebGPU或CPU",
   "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",
@@ -178,7 +178,7 @@
   "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 Messages": "编辑对话",
+  "Edit Character Settings": "编辑人设",
   "Go Back": "返回",
   "Description": "描述",
   "Avatar Url": "头像图片地址",
@@ -226,14 +226,14 @@
   "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 or use a smaller base model.": "内存不足，尝试增加虚拟内存，或使用一个更小规模的基底模型",
+  "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\"",
+  "Can not find an Nvidia GPU. Perhaps the gpu driver of windows is too old, or you are using WSL 1 for training, please upgrade to WSL 2. e.g. Run \"wsl --set-version Ubuntu-22.04 2\"": "没有找到Nvidia显卡。可能是因为你的windows显卡驱动太旧，或者你正在使用WSL 1进行训练，请升级到WSL 2。例如，执行\"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",
+  "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/josStorer/RWKV-Runner/blob/master/finetune/data/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/josStorer/RWKV-Runner/blob/master/finetune/data/sample.jsonl 以及 https://zhuanlan.zhihu.com/p/643433851\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",
@@ -244,5 +244,84 @@
   "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"
+  "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 or rwkv_vocab_v20230424.txt)": "Tokenizer路径 (例如: backend-python/rwkv_pip/20B_tokenizer.json 或 rwkv_vocab_v20230424.txt)",
+  "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",
+  "Format Content": "规范格式",
+  "Add An Attachment (Accepts pdf, txt)": "添加一个附件 (支持pdf, txt)",
+  "Processing Attachment": "正在处理附件",
+  "Remove Attachment": "移除附件",
+  "The content of file": "文件",
+  "is as follows. When replying to me, consider the file content and respond accordingly:": "内容如下。回复时考虑文件内容并做出相应回复:",
+  "What's the file name": "文件名是什么",
+  "The file name is: ": "文件名是：",
+  "Port is occupied. Change it in Configs page or close the program that occupies the port.": "端口被占用。请在配置页面更改端口，或关闭占用端口的程序",
+  "Loading...": "加载中...",
+  "Hello, what can I do for you?": "你好，有什么要我帮忙的吗？",
+  "Enable WebUI": "启用WebUI",
+  "Server is working on deployment mode, please close the terminal window manually": "服务器正在部署模式下运行，请手动关闭终端窗口",
+  "Server is working on deployment mode, please exit the program manually to stop the server": "服务器正在部署模式下运行，请手动退出程序以停止服务器",
+  "You can increase the number of stored layers in Configs page to improve performance": "你可以在配置页面增加载入显存层数以提升性能",
+  "Failed to load model, try to increase the virtual memory (Swap of WSL) or use a smaller base model.": "模型载入失败，尝试增加虚拟内存(WSL Swap)，或使用一个更小规模的基底模型",
+  "Save Conversation": "保存对话",
+  "Use Hugging Face Mirror": "使用Hugging Face镜像源",
+  "File is empty": "文件为空",
+  "Open MIDI Input Audio Tracks": "打开MIDI输入音轨",
+  "Track": "音轨",
+  "Play All": "播放全部",
+  "Clear All": "清空",
+  "Scale View": "缩放视图",
+  "Record": "录制",
+  "Play": "播放",
+  "New Track": "新建音轨",
+  "Select a track to preview the content": "选择一个音轨以预览内容",
+  "Save to generation area": "保存到生成区",
+  "Piano": "钢琴",
+  "Percussion": "打击乐",
+  "Drum": "鼓",
+  "Tuba": "大号",
+  "Marimba": "马林巴",
+  "Bass": "贝斯",
+  "Guitar": "吉他",
+  "Violin": "小提琴",
+  "Trumpet": "小号",
+  "Sax": "萨克斯",
+  "Flute": "长笛",
+  "Lead": "主音",
+  "Pad": "和音",
+  "MIDI Input": "MIDI输入",
+  "Select the MIDI input device to be used.": "选择要使用的MIDI输入设备",
+  "Start Time": "开始时间",
+  "Content Duration": "内容时长",
+  "Please select a MIDI device first": "请先选择一个MIDI设备",
+  "Piano is the main instrument": "钢琴为主",
+  "Loss is too high, please check the training data, and ensure your gpu driver is up to date.": "Loss过高，请检查训练数据，并确保你的显卡驱动是最新的",
+  "This version of RWKV is not supported yet.": "暂不支持此版本的RWKV",
+  "Main": "主干",
+  "Finetuned": "微调",
+  "Global": "全球",
+  "Local": "本地",
+  "CN": "中文",
+  "JP": "日文",
+  "Music": "音乐",
+  "Other": "其他",
+  "Role Play": "角色扮演",
+  "Recommended": "推荐",
+  "Import MIDI": "导入MIDI",
+  "Current Instrument": "当前乐器",
+  "Please convert model to GGML format first": "请先将模型转换为GGML格式",
+  "Convert To GGML Format": "转换为GGML格式",
+  "CPU (rwkv.cpp, Faster)": "CPU (rwkv.cpp, 更快)",
+  "Play With External Player": "使用外部播放器播放",
+  "Core API URL": "核心 API URL",
+  "Override core API URL(/chat/completions and /completions). If you don't know what this is, leave it blank.": "覆盖核心的 API URL (/chat/completions 和 /completions)。如果你不知道这是什么，请留空",
+  "Please change Strategy to CPU (rwkv.cpp) to use ggml format": "请将Strategy改为CPU (rwkv.cpp)以使用ggml格式",
+  "Only Auto Play Generated Content": "仅自动播放新生成的内容"
 }

frontend/src/components/ConfigSelector.tsx

@@ -1,10 +1,10 @@
 import { FC } from 'react';
 import { observer } from 'mobx-react-lite';
-import { Dropdown, Option } from '@fluentui/react-components';
+import { Dropdown, Option, PresenceBadge } from '@fluentui/react-components';
 import commonStore from '../stores/commonStore';
 
 export const ConfigSelector: FC<{ size?: 'small' | 'medium' | 'large' }> = observer(({ size }) => {
-  return <Dropdown size={size} style={{ minWidth: 0 }} listbox={{ style: { minWidth: 0 } }}
+  return <Dropdown size={size} style={{ minWidth: 0 }} listbox={{ style: { minWidth: 'fit-content' } }}
     value={commonStore.getCurrentModelConfig().name}
     selectedOptions={[commonStore.currentModelConfigIndex.toString()]}
     onOptionSelect={(_, data) => {
@@ -12,7 +12,13 @@ export const ConfigSelector: FC<{ size?: 'small' | 'medium' | 'large' }> = obser
         commonStore.setCurrentConfigIndex(Number(data.optionValue));
     }}>
     {commonStore.modelConfigs.map((config, index) =>
-      <Option key={index} value={index.toString()}>{config.name}</Option>
+      <Option key={index} value={index.toString()} text={config.name}>
+        <div className="flex justify-between grow">
+          {config.name}
+          {commonStore.modelSourceList.find(item => item.name === config.modelParameters.modelName)?.isComplete
+            && <PresenceBadge status="available" />}
+        </div>
+      </Option>
     )}
   </Dropdown>;
 });

frontend/src/components/DialogButton.tsx

@@ -1,4 +1,4 @@
-import { FC, ReactElement } from 'react';
+import React, { FC, ReactElement } from 'react';
 import {
   Button,
   Dialog,
@@ -11,7 +11,9 @@ import {
 } from '@fluentui/react-components';
 import { ToolTipButton } from './ToolTipButton';
 import { useTranslation } from 'react-i18next';
-import MarkdownRender from './MarkdownRender';
+import { LazyImportComponent } from './LazyImportComponent';
+
+const MarkdownRender = React.lazy(() => import('./MarkdownRender'));
 
 export const DialogButton: FC<{
   text?: string | null
@@ -45,7 +47,9 @@ export const DialogButton: FC<{
         <DialogContent>
           {
             markdown ?
-              <MarkdownRender>{contentText}</MarkdownRender> :
+              <LazyImportComponent lazyChildren={MarkdownRender}>
+                {contentText}
+              </LazyImportComponent> :
               contentText
           }
         </DialogContent>

frontend/src/components/LazyImportComponent.tsx

@@ -0,0 +1,24 @@
+import { FC, LazyExoticComponent, ReactNode, Suspense } from 'react';
+import { useTranslation } from 'react-i18next';
+import { Spinner } from '@fluentui/react-components';
+
+interface LazyImportComponentProps {
+  lazyChildren: LazyExoticComponent<FC<any>>;
+  lazyProps?: any;
+  children?: ReactNode;
+}
+
+export const LazyImportComponent: FC<LazyImportComponentProps> = (props) => {
+  const { t } = useTranslation();
+
+  return (
+    <Suspense fallback={
+      <div className="flex justify-center items-center h-full w-full">
+        <Spinner size="huge" label={t('Loading...')} />
+      </div>}>
+      <props.lazyChildren {...props.lazyProps}>
+        {props.children}
+      </props.lazyChildren>
+    </Suspense>
+  );
+};

frontend/src/components/MarkdownRender.tsx

@@ -21,7 +21,7 @@ const Hyperlink: FC<any> = ({ href, children }) => {
   );
 };
 
-export const MarkdownRender: FC<ReactMarkdownOptions> = (props) => {
+const MarkdownRender: FC<ReactMarkdownOptions> = (props) => {
   return (
     <div dir="auto" className="markdown-body">
       <ReactMarkdown

frontend/src/components/ReadButton.tsx

@@ -40,6 +40,8 @@ export const ReadButton: FC<{
       voice = voices.find((v) => v.name.toLowerCase().includes('microsoft aria'));
     else if (lang === 'zh')
       voice = voices.find((v) => v.name.toLowerCase().includes('xiaoyi'));
+    else if (lang === 'ja')
+      voice = voices.find((v) => v.name.toLowerCase().includes('nanami'));
     if (!voice) voice = voices.find((v) => v.lang.substring(0, 2) === lang);
     if (!voice) voice = voices.find((v) => v.lang === navigator.language);
 

frontend/src/components/RunButton.tsx

@@ -2,8 +2,8 @@ import React, { FC, MouseEventHandler, ReactElement } from 'react';
 import commonStore, { ModelStatus } from '../stores/commonStore';
 import {
   AddToDownloadList,
-  CopyFile,
   FileExists,
+  IsPortAvailable,
   StartServer,
   StartWebGPUServer
 } from '../../wailsjs/go/backend_golang/App';
@@ -11,12 +11,14 @@ 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 { checkDependencies, getStrategy, getSupportedCustomCudaFile, toastWithButton } from '../utils';
+import { checkDependencies, getHfDownloadUrl, 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 { WindowShow } from '../../wailsjs/runtime/runtime';
+import { WindowShow } from '../../wailsjs/runtime';
+import { convertToGGML, convertToSt } from '../utils/convert-model';
+import { Precision } from '../types/configs';
 
 const mainButtonText = {
   [ModelStatus.Offline]: 'Run',
@@ -46,6 +48,8 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
 
       const modelConfig = commonStore.getCurrentModelConfig();
       const webgpu = modelConfig.modelParameters.device === 'WebGPU';
+      const webgpuPython = modelConfig.modelParameters.device === 'WebGPU (Python)';
+      const cpp = modelConfig.modelParameters.device === 'CPU (rwkv.cpp)';
       let modelName = '';
       let modelPath = '';
       if (modelConfig && modelConfig.modelParameters) {
@@ -57,20 +61,47 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
         return;
       }
 
-      if (webgpu) {
+      const currentModelSource = commonStore.modelSourceList.find(item => item.name === modelName);
+
+      const showDownloadPrompt = (promptInfo: string, downloadName: string) => {
+        toastWithButton(promptInfo, t('Download'), () => {
+          const downloadUrl = currentModelSource?.downloadUrl;
+          if (downloadUrl) {
+            toastWithButton(`${t('Downloading')} ${downloadName}`, t('Check'), () => {
+                navigate({ pathname: '/downloads' });
+              },
+              { autoClose: 3000 });
+            AddToDownloadList(modelPath, getHfDownloadUrl(downloadUrl));
+          } else {
+            toast(t('Can not find download url'), { type: 'error' });
+          }
+        });
+      };
+
+      if (webgpu || webgpuPython) {
         if (!['.st', '.safetensors'].some(ext => modelPath.endsWith(ext))) {
           const stModelPath = modelPath.replace(/\.pth$/, '.st');
           if (await FileExists(stModelPath)) {
             modelPath = stModelPath;
+          } else if (!await FileExists(modelPath)) {
+            showDownloadPrompt(t('Model file not found'), modelName);
+            commonStore.setStatus({ status: ModelStatus.Offline });
+            return;
+          } else if (!currentModelSource?.isComplete) {
+            showDownloadPrompt(t('Model file download is not complete'), modelName);
+            commonStore.setStatus({ status: ModelStatus.Offline });
+            return;
           } else {
-            toast(t('Please convert model to safe tensors format first'), { type: 'error' });
+            toastWithButton(t('Please convert model to safe tensors format first'), t('Convert'), () => {
+              convertToSt(modelConfig, navigate);
+            });
             commonStore.setStatus({ status: ModelStatus.Offline });
             return;
           }
         }
       }
 
-      if (!webgpu) {
+      if (!webgpu && !webgpuPython) {
         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 });
@@ -84,28 +115,45 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
           return;
       }
 
-      const currentModelSource = commonStore.modelSourceList.find(item => item.name === modelName);
-
-      const showDownloadPrompt = (promptInfo: string, downloadName: string) => {
-        toastWithButton(promptInfo, t('Download'), () => {
-          const downloadUrl = currentModelSource?.downloadUrl;
-          if (downloadUrl) {
-            toastWithButton(`${t('Downloading')} ${downloadName}`, t('Check'), () => {
-                navigate({ pathname: '/downloads' });
-              },
-              { autoClose: 3000 });
-            AddToDownloadList(modelPath, downloadUrl);
+      if (cpp) {
+        if (!['.bin'].some(ext => modelPath.endsWith(ext))) {
+          const precision: Precision = modelConfig.modelParameters.precision === 'Q5_1' ? 'Q5_1' : 'fp16';
+          const ggmlModelPath = modelPath.replace(/\.pth$/, `-${precision}.bin`);
+          if (await FileExists(ggmlModelPath)) {
+            modelPath = ggmlModelPath;
+          } else if (!await FileExists(modelPath)) {
+            showDownloadPrompt(t('Model file not found'), modelName);
+            commonStore.setStatus({ status: ModelStatus.Offline });
+            return;
+          } else if (!currentModelSource?.isComplete) {
+            showDownloadPrompt(t('Model file download is not complete'), modelName);
+            commonStore.setStatus({ status: ModelStatus.Offline });
+            return;
           } else {
-            toast(t('Can not find download url'), { type: 'error' });
+            toastWithButton(t('Please convert model to GGML format first'), t('Convert'), () => {
+              convertToGGML(modelConfig, navigate);
+            });
+            commonStore.setStatus({ status: ModelStatus.Offline });
+            return;
           }
-        });
-      };
+        }
+      }
+
+      if (!cpp) {
+        if (['.bin'].some(ext => modelPath.endsWith(ext))) {
+          toast(t('Please change Strategy to CPU (rwkv.cpp) to use ggml format'), { type: 'error' });
+          commonStore.setStatus({ status: ModelStatus.Offline });
+          return;
+        }
+      }
 
       if (!await FileExists(modelPath)) {
         showDownloadPrompt(t('Model file not found'), modelName);
         commonStore.setStatus({ status: ModelStatus.Offline });
         return;
-      } else if (!currentModelSource?.isComplete) {
+      } else // If the user selects the .pth model with WebGPU mode, modelPath will be set to the .st model.
+        // However, if the .pth model is deleted, modelPath will exist and isComplete will be false.
+      if (!currentModelSource?.isComplete && modelPath.endsWith('.pth')) {
         showDownloadPrompt(t('Model file download is not complete'), modelName);
         commonStore.setStatus({ status: ModelStatus.Offline });
         return;
@@ -113,15 +161,23 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
 
       const port = modelConfig.apiParameters.apiPort;
 
-      await exit(1000).catch(() => {
-      });
+      if (!await IsPortAvailable(port)) {
+        await exit(1000).catch(() => {
+        });
+        if (!await IsPortAvailable(port)) {
+          toast(t('Port is occupied. Change it in Configs page or close the program that occupies the port.'), { type: 'error' });
+          commonStore.setStatus({ status: ModelStatus.Offline });
+          return;
+        }
+      }
 
       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',
-        modelConfig.modelParameters.device === 'CUDA-Beta'
+        !!modelConfig.enableWebUI, isUsingCudaBeta, cpp, webgpuPython
       ).catch((e) => {
         const errMsg = e.message || e;
         if (errMsg.includes('path contains space'))
@@ -130,6 +186,8 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
           toast(t('Error') + ' - ' + errMsg, { type: 'error' });
       });
       setTimeout(WindowShow, 1000);
+      setTimeout(WindowShow, 2000);
+      setTimeout(WindowShow, 3000);
 
       let timeoutCount = 6;
       let loading = false;
@@ -146,7 +204,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
               });
             }
             commonStore.setStatus({ status: ModelStatus.Loading });
-            toast(t('Loading Model'), { type: 'info' });
+            const loadingId = toast(t('Loading Model'), { type: 'info', autoClose: false });
             if (!webgpu) {
               updateConfig({
                 max_tokens: modelConfig.apiParameters.maxResponseToken,
@@ -159,25 +217,30 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
 
             const strategy = getStrategy(modelConfig);
             let customCudaFile = '';
-            if ((modelConfig.modelParameters.device.includes('CUDA') || modelConfig.modelParameters.device === 'Custom')
-              && modelConfig.modelParameters.useCustomCuda && !strategy.includes('fp32')) {
+            if ((modelConfig.modelParameters.device.startsWith('CUDA') || modelConfig.modelParameters.device === 'Custom')
+              && modelConfig.modelParameters.useCustomCuda
+              && !strategy.split('->').some(s => ['cuda', 'fp32'].every(v => s.includes(v)))) {
               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';
               }
@@ -186,7 +249,9 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
             switchModel({
               model: modelPath,
               strategy: strategy,
-              customCuda: customCudaFile !== ''
+              tokenizer: modelConfig.modelParameters.useCustomTokenizer ? modelConfig.modelParameters.customTokenizer : undefined,
+              customCuda: customCudaFile !== '',
+              deploy: modelConfig.enableWebUI
             }).then(async (r) => {
               if (r.ok) {
                 commonStore.setStatus({ status: ModelStatus.Working });
@@ -199,6 +264,12 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
                 const buttonFn = () => {
                   navigate({ pathname: '/' + buttonName.toLowerCase() });
                 };
+
+                if (modelConfig.modelParameters.device.startsWith('CUDA') &&
+                  modelConfig.modelParameters.storedLayers < modelConfig.modelParameters.maxStoredLayers &&
+                  commonStore.monitorData && commonStore.monitorData.totalVram !== 0 &&
+                  (commonStore.monitorData.usedVram / commonStore.monitorData.totalVram) < 0.9)
+                  toast(t('You can increase the number of stored layers in Configs page to improve performance'), { type: 'info' });
                 toastWithButton(t('Startup Completed'), t(buttonName), buttonFn, { type: 'success', autoClose: 3000 });
               } else if (r.status === 304) {
                 toast(t('Loading Model'), { type: 'info' });
@@ -209,7 +280,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
                   '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.',
+                  'no NVIDIA driver': 'Found no NVIDIA driver, please install the latest driver. If you are not using an Nvidia GPU, please switch the \'Strategy\' to WebGPU or CPU in the Configs page.',
                   '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.'
                 };
@@ -220,6 +291,8 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
             }).catch((e) => {
               commonStore.setStatus({ status: ModelStatus.Offline });
               toast(t('Failed to switch model') + ' - ' + (e.message || e), { type: 'error' });
+            }).finally(() => {
+              toast.dismiss(loadingId);
             });
           }
         }).catch(() => {
@@ -233,7 +306,13 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
       }, 1000);
     } else {
       commonStore.setStatus({ status: ModelStatus.Offline });
-      exit();
+      exit().then(r => {
+        if (r.status === 403)
+          if (commonStore.platform !== 'linux')
+            toast(t('Server is working on deployment mode, please close the terminal window manually'), { type: 'info' });
+          else
+            toast(t('Server is working on deployment mode, please exit the program manually to stop the server'), { type: 'info' });
+      });
     }
   };
 

frontend/src/components/WorkHeader.tsx

@@ -6,6 +6,7 @@ import { ConfigSelector } from './ConfigSelector';
 import { RunButton } from './RunButton';
 import { PresenceBadgeStatus } from '@fluentui/react-badge';
 import { useTranslation } from 'react-i18next';
+import { useMediaQuery } from 'usehooks-ts';
 
 const statusText = {
   [ModelStatus.Offline]: 'Offline',
@@ -23,15 +24,21 @@ const badgeStatus: { [modelStatus: number]: PresenceBadgeStatus } = {
 
 export const WorkHeader: FC = observer(() => {
   const { t } = useTranslation();
+  const mq = useMediaQuery('(min-width: 640px)');
   const port = commonStore.getCurrentModelConfig().apiParameters.apiPort;
 
-  return (
+  return commonStore.platform === 'web' ?
+    <div /> :
     <div className="flex flex-col gap-1">
       <div className="flex justify-between items-center">
         <div className="flex items-center gap-2">
           <PresenceBadge status={badgeStatus[commonStore.status.status]} />
           <Text size={100}>{t('Model Status') + ': ' + t(statusText[commonStore.status.status])}</Text>
         </div>
+        {commonStore.lastModelName && mq &&
+          <Text size={100}>
+            {commonStore.lastModelName}
+          </Text>}
         <div className="flex items-center gap-2">
           <ConfigSelector size="small" />
           <RunButton iconMode />
@@ -42,5 +49,5 @@ export const WorkHeader: FC = observer(() => {
       </Text>
       <Divider style={{ flexGrow: 0 }} />
     </div>
-  );
+    ;
 });

frontend/src/main.tsx

@@ -1,3 +1,4 @@
+import './webWails';
 import React from 'react';
 import { createRoot } from 'react-dom/client';
 import './style.scss';
@@ -6,7 +7,6 @@ import App from './App';
 import { HashRouter } from 'react-router-dom';
 import { startup } from './startup';
 import './_locales/i18n-react';
-import 'html-midi-player';
 import { WindowShow } from '../wailsjs/runtime';
 
 startup().then(() => {

frontend/src/pages/About.tsx

@@ -5,9 +5,7 @@ import MarkdownRender from '../components/MarkdownRender';
 import { observer } from 'mobx-react-lite';
 import commonStore from '../stores/commonStore';
 
-export type AboutContent = { [lang: string]: string }
-
-export const About: FC = observer(() => {
+const About: FC = observer(() => {
   const { t } = useTranslation();
   const lang: string = commonStore.settings.language;
 
@@ -21,3 +19,5 @@ export const About: FC = observer(() => {
     } />
   );
 });
+
+export default About;

frontend/src/pages/AudiotrackManager/AudiotrackButton.tsx

@@ -0,0 +1,40 @@
+import React, { FC, lazy } from 'react';
+import { useTranslation } from 'react-i18next';
+import { Button, Dialog, DialogBody, DialogContent, DialogSurface, DialogTrigger } from '@fluentui/react-components';
+import { CustomToastContainer } from '../../components/CustomToastContainer';
+import { LazyImportComponent } from '../../components/LazyImportComponent';
+import { flushMidiRecordingContent } from '../../utils';
+import commonStore from '../../stores/commonStore';
+
+const AudiotrackEditor = lazy(() => import('./AudiotrackEditor'));
+
+export const AudiotrackButton: FC<{
+  size?: 'small' | 'medium' | 'large',
+  shape?: 'rounded' | 'circular' | 'square';
+  appearance?: 'secondary' | 'primary' | 'outline' | 'subtle' | 'transparent';
+  setPrompt: (prompt: string) => void;
+}> = ({ size, shape, appearance, setPrompt }) => {
+  const { t } = useTranslation();
+
+  return <Dialog onOpenChange={(e, data) => {
+    if (!data.open) {
+      flushMidiRecordingContent();
+      commonStore.setRecordingTrackId('');
+      commonStore.setPlayingTrackId('');
+    }
+  }}>
+    <DialogTrigger disableButtonEnhancement>
+      <Button size={size} shape={shape} appearance={appearance}>
+        {t('Open MIDI Input Audio Tracks')}
+      </Button>
+    </DialogTrigger>
+    <DialogSurface style={{ paddingTop: 0, maxWidth: '90vw', width: 'fit-content' }}>
+      <DialogBody>
+        <DialogContent className="overflow-hidden">
+          <CustomToastContainer />
+          <LazyImportComponent lazyChildren={AudiotrackEditor} lazyProps={{ setPrompt }} />
+        </DialogContent>
+      </DialogBody>
+    </DialogSurface>
+  </Dialog>;
+};

frontend/src/pages/AudiotrackManager/AudiotrackEditor.tsx

@@ -0,0 +1,601 @@
+import React, { FC, useEffect, useRef, useState } from 'react';
+import { observer } from 'mobx-react-lite';
+import { useTranslation } from 'react-i18next';
+import Draggable from 'react-draggable';
+import { ToolTipButton } from '../../components/ToolTipButton';
+import { v4 as uuid } from 'uuid';
+import {
+  Add16Regular,
+  ArrowAutofitWidth20Regular,
+  ArrowUpload16Regular,
+  Delete16Regular,
+  MusicNote220Regular,
+  Pause16Regular,
+  Play16Filled,
+  Play16Regular,
+  Record16Regular,
+  Stop16Filled
+} from '@fluentui/react-icons';
+import { Button, Card, DialogTrigger, Slider, Text, Tooltip } from '@fluentui/react-components';
+import { useWindowSize } from 'usehooks-ts';
+import commonStore, { ModelStatus } from '../../stores/commonStore';
+import classnames from 'classnames';
+import {
+  InstrumentType,
+  InstrumentTypeNameMap,
+  InstrumentTypeTokenMap,
+  MidiMessage,
+  tracksMinimalTotalTime
+} from '../../types/composition';
+import { toast } from 'react-toastify';
+import {
+  absPathAsset,
+  flushMidiRecordingContent,
+  getMidiRawContentMainInstrument,
+  getMidiRawContentTime,
+  getServerRoot,
+  refreshTracksTotalTime
+} from '../../utils';
+import { OpenOpenFileDialog, PlayNote } from '../../../wailsjs/go/backend_golang/App';
+
+const snapValue = 25;
+const minimalMoveTime = 8; // 1000/125=8ms wait_events=125
+const scaleMin = 0.05;
+const scaleMax = 3;
+const baseMoveTime = Math.round(minimalMoveTime / scaleMin);
+
+const velocityEvents = 128;
+const velocityBins = 12;
+const velocityExp = 0.5;
+
+const minimalTrackWidth = 80;
+const trackInitOffsetPx = 10;
+const pixelFix = 0.5;
+const topToArrowIcon = 19;
+const arrowIconToTracks = 23;
+
+const velocityToBin = (velocity: number) => {
+  velocity = Math.max(0, Math.min(velocity, velocityEvents - 1));
+  const binsize = velocityEvents / (velocityBins - 1);
+  return Math.ceil((velocityEvents * ((Math.pow(velocityExp, (velocity / velocityEvents)) - 1.0) / (velocityExp - 1.0))) / binsize);
+};
+
+const binToVelocity = (bin: number) => {
+  const binsize = velocityEvents / (velocityBins - 1);
+  return Math.max(0, Math.ceil(velocityEvents * (Math.log(((velocityExp - 1) * binsize * bin) / velocityEvents + 1) / Math.log(velocityExp)) - 1));
+};
+
+const tokenToMidiMessage = (token: string): MidiMessage | null => {
+  if (token.startsWith('<')) return null;
+  if (token.startsWith('t') && !token.startsWith('t:')) {
+    return {
+      messageType: 'ElapsedTime',
+      value: parseInt(token.substring(1)) * minimalMoveTime,
+      channel: 0,
+      note: 0,
+      velocity: 0,
+      control: 0,
+      instrument: 0
+    };
+  }
+  const instrument: InstrumentType = InstrumentTypeTokenMap.findIndex(t => token.startsWith(t + ':'));
+  if (instrument >= 0) {
+    const parts = token.split(':');
+    if (parts.length !== 3) return null;
+    const note = parseInt(parts[1], 16);
+    const velocity = parseInt(parts[2], 16);
+    if (velocity < 0 || velocity > 127) return null;
+    if (velocity === 0) return {
+      messageType: 'NoteOff',
+      note: note,
+      instrument: instrument,
+      channel: 0,
+      velocity: 0,
+      control: 0,
+      value: 0
+    };
+    return {
+      messageType: 'NoteOn',
+      note: note,
+      velocity: binToVelocity(velocity),
+      instrument: instrument,
+      channel: 0,
+      control: 0,
+      value: 0
+    } as MidiMessage;
+  }
+  return null;
+};
+
+const midiMessageToToken = (msg: MidiMessage) => {
+  if (msg.messageType === 'NoteOn' || msg.messageType === 'NoteOff') {
+    const instrument = InstrumentTypeTokenMap[msg.instrument];
+    const note = msg.note.toString(16);
+    const velocity = velocityToBin(msg.velocity).toString(16);
+    return `${instrument}:${note}:${velocity} `;
+  } else if (msg.messageType === 'ElapsedTime') {
+    let time = Math.round(msg.value / minimalMoveTime);
+    const num = Math.floor(time / 125); // wait_events=125
+    time -= num * 125;
+    let ret = '';
+    for (let i = 0; i < num; i++) {
+      ret += 't125 ';
+    }
+    if (time > 0)
+      ret += `t${time} `;
+    return ret;
+  } else
+    return '';
+};
+
+let dropRecordingTime = false;
+
+export const midiMessageHandler = async (data: MidiMessage) => {
+  if (data.messageType === 'ControlChange') {
+    commonStore.setInstrumentType(Math.round(data.value / 127 * (InstrumentTypeNameMap.length - 1)));
+    return;
+  }
+  if (commonStore.recordingTrackId) {
+    if (dropRecordingTime && data.messageType === 'ElapsedTime') {
+      dropRecordingTime = false;
+      return;
+    }
+    data = {
+      ...data,
+      instrument: commonStore.instrumentType
+    };
+    commonStore.setRecordingRawContent([...commonStore.recordingRawContent, data]);
+    commonStore.setRecordingContent(commonStore.recordingContent + midiMessageToToken(data));
+
+    //TODO data.channel = data.instrument;
+    PlayNote(data);
+  }
+};
+
+type TrackProps = {
+  id: string;
+  right: number;
+  scale: number;
+  isSelected: boolean;
+  onSelect: (id: string) => void;
+};
+
+const Track: React.FC<TrackProps> = observer(({
+  id,
+  right,
+  scale,
+  isSelected,
+  onSelect
+}) => {
+  const { t } = useTranslation();
+  const trackIndex = commonStore.tracks.findIndex(t => t.id === id)!;
+  const track = commonStore.tracks[trackIndex];
+  const trackClass = isSelected ? 'bg-blue-600' : (commonStore.settings.darkMode ? 'bg-blue-900' : 'bg-gray-700');
+  const controlX = useRef(0);
+
+  let trackName = t('Track') + ' ' + id;
+  if (track.mainInstrument)
+    trackName = t('Track') + ' - ' + t('Piano is the main instrument')!.replace(t('Piano')!, t(track.mainInstrument)) + (track.content && (' - ' + track.content));
+  else if (track.content)
+    trackName = t('Track') + ' - ' + track.content;
+
+  return (
+    <Draggable
+      axis="x"
+      bounds={{ left: 0, right }}
+      grid={[snapValue, snapValue]}
+      position={{
+        x: (track.offsetTime - commonStore.trackCurrentTime) / (baseMoveTime * scale) * snapValue,
+        y: 0
+      }}
+      onStart={(e, data) => {
+        controlX.current = data.lastX;
+      }}
+      onStop={(e, data) => {
+        const delta = data.lastX - controlX.current;
+        let offsetTime = Math.round(Math.round(delta / snapValue * baseMoveTime * scale) / minimalMoveTime) * minimalMoveTime;
+        offsetTime = Math.min(Math.max(
+          offsetTime,
+          -track.offsetTime), commonStore.trackTotalTime - track.offsetTime);
+
+        const tracks = commonStore.tracks.slice();
+        tracks[trackIndex].offsetTime += offsetTime;
+        commonStore.setTracks(tracks);
+        refreshTracksTotalTime();
+      }}
+    >
+      <div
+        className={`p-1 cursor-move rounded whitespace-nowrap overflow-hidden ${trackClass}`}
+        style={{
+          width: `${Math.max(minimalTrackWidth,
+            track.contentTime / (baseMoveTime * scale) * snapValue
+          )}px`
+        }}
+        onClick={() => onSelect(id)}
+      >
+        <span className="text-white">{trackName}</span>
+      </div>
+    </Draggable>
+  );
+});
+
+const AudiotrackEditor: FC<{ setPrompt: (prompt: string) => void }> = observer(({ setPrompt }) => {
+  const { t } = useTranslation();
+
+  const viewControlsContainerRef = useRef<HTMLDivElement>(null);
+  const currentTimeControlRef = useRef<HTMLDivElement>(null);
+  const playStartTimeControlRef = useRef<HTMLDivElement>(null);
+  const tracksEndLineRef = useRef<HTMLDivElement>(null);
+  const tracksRef = useRef<HTMLDivElement>(null);
+  const toolbarRef = useRef<HTMLDivElement>(null);
+  const toolbarButtonRef = useRef<HTMLDivElement>(null);
+  const toolbarSliderRef = useRef<HTMLInputElement>(null);
+  const contentPreviewRef = useRef<HTMLDivElement>(null);
+
+  const [refreshRef, setRefreshRef] = useState(false);
+
+  const windowSize = useWindowSize();
+  const scale = (scaleMin + scaleMax) - commonStore.trackScale;
+
+  const [selectedTrackId, setSelectedTrackId] = useState<string>('');
+  const playStartTimeControlX = useRef(0);
+  const selectedTrack = selectedTrackId ? commonStore.tracks.find(t => t.id === selectedTrackId) : undefined;
+
+  useEffect(() => {
+    if (toolbarSliderRef.current && toolbarSliderRef.current.parentElement)
+      toolbarSliderRef.current.parentElement.style.removeProperty('--fui-Slider--steps-percent');
+  }, []);
+
+  const scrollContentToBottom = () => {
+    if (contentPreviewRef.current)
+      contentPreviewRef.current.scrollTop = contentPreviewRef.current.scrollHeight;
+  };
+
+  useEffect(() => {
+    scrollContentToBottom();
+  }, [commonStore.recordingContent]);
+
+  useEffect(() => {
+    setRefreshRef(!refreshRef);
+  }, [windowSize, commonStore.tracks]);
+
+  const viewControlsContainerWidth = (toolbarRef.current && toolbarButtonRef.current && toolbarSliderRef.current) ?
+    toolbarRef.current.clientWidth - toolbarButtonRef.current.clientWidth - toolbarSliderRef.current.clientWidth - 16 // 16 = ml-2 mr-2
+    : 0;
+  const tracksWidth = viewControlsContainerWidth;
+  const timeOfTracksWidth = Math.floor(tracksWidth / snapValue) // number of moves
+    * baseMoveTime * scale;
+  const currentTimeControlWidth = (timeOfTracksWidth < commonStore.trackTotalTime)
+    ? timeOfTracksWidth / commonStore.trackTotalTime * viewControlsContainerWidth
+    : 0;
+  const playStartTimeControlPosition = (commonStore.trackPlayStartTime - commonStore.trackCurrentTime) / (baseMoveTime * scale) * snapValue;
+  const tracksEndPosition = (commonStore.trackTotalTime - commonStore.trackCurrentTime) / (baseMoveTime * scale) * snapValue;
+  const moveableTracksWidth = (tracksEndLineRef.current && viewControlsContainerRef.current &&
+    ((tracksEndLineRef.current.getBoundingClientRect().left - (viewControlsContainerRef.current.getBoundingClientRect().left + trackInitOffsetPx)) > 0))
+    ? tracksEndLineRef.current.getBoundingClientRect().left - (viewControlsContainerRef.current.getBoundingClientRect().left + trackInitOffsetPx)
+    : Infinity;
+
+  return (
+    <div className="flex flex-col gap-2 overflow-hidden" style={{ width: '80vw', height: '80vh' }}>
+      <div className="mx-auto">
+        <Text size={100}>{`${commonStore.trackPlayStartTime} ms / ${commonStore.trackTotalTime} ms`}</Text>
+      </div>
+      <div className="flex pb-2 border-b" ref={toolbarRef}>
+        <div className="flex gap-2" ref={toolbarButtonRef}>
+          <ToolTipButton disabled desc={t('Play All') + ' (Unavailable)'} icon={<Play16Regular />} />
+          <ToolTipButton desc={t('Clear All')} icon={<Delete16Regular />} onClick={() => {
+            commonStore.setTracks([]);
+            commonStore.setTrackScale(1);
+            commonStore.setTrackTotalTime(tracksMinimalTotalTime);
+            commonStore.setTrackCurrentTime(0);
+            commonStore.setTrackPlayStartTime(0);
+          }} />
+        </div>
+        <div className="grow">
+          <div className="flex flex-col ml-2 mr-2" ref={viewControlsContainerRef}>
+            <div className="relative">
+              <Tooltip content={`${commonStore.trackTotalTime} ms`} showDelay={0} hideDelay={0}
+                relationship="description">
+                <div className="border-l absolute"
+                  ref={tracksEndLineRef}
+                  style={{
+                    height: (tracksRef.current && commonStore.tracks.length > 0)
+                      ? tracksRef.current.clientHeight - arrowIconToTracks
+                      : 0,
+                    top: `${topToArrowIcon + arrowIconToTracks}px`,
+                    left: `${tracksEndPosition + trackInitOffsetPx - pixelFix}px`
+                  }} />
+              </Tooltip>
+            </div>
+            <Draggable axis="x" bounds={{
+              left: 0,
+              right: viewControlsContainerWidth - currentTimeControlWidth
+            }}
+              position={{
+                x: commonStore.trackCurrentTime / commonStore.trackTotalTime * viewControlsContainerWidth,
+                y: 0
+              }}
+              onDrag={(e, data) => {
+                setTimeout(() => {
+                  let offset = 0;
+                  if (currentTimeControlRef.current) {
+                    const match = currentTimeControlRef.current.style.transform.match(/translate\((.+)px,/);
+                    if (match)
+                      offset = parseFloat(match[1]);
+                  }
+                  const offsetTime = commonStore.trackTotalTime / viewControlsContainerWidth * offset;
+                  commonStore.setTrackCurrentTime(offsetTime);
+                }, 1);
+              }}
+            >
+              <div ref={currentTimeControlRef}
+                className={classnames('h-2 cursor-move rounded', commonStore.settings.darkMode ? 'bg-neutral-600' : 'bg-gray-700')}
+                style={{ width: currentTimeControlWidth }} />
+            </Draggable>
+            <div className={classnames(
+              'flex',
+              (playStartTimeControlPosition < 0 || playStartTimeControlPosition > viewControlsContainerWidth)
+              && 'hidden'
+            )}>
+              <Draggable axis="x" bounds={{
+                left: 0,
+                right: (playStartTimeControlRef.current)
+                  ? Math.min(viewControlsContainerWidth - playStartTimeControlRef.current.clientWidth, moveableTracksWidth)
+                  : 0
+              }}
+                grid={[snapValue, snapValue]}
+                position={{ x: playStartTimeControlPosition, y: 0 }}
+                onStart={(e, data) => {
+                  playStartTimeControlX.current = data.lastX;
+                }}
+                onStop={(e, data) => {
+                  const delta = data.lastX - playStartTimeControlX.current;
+                  let offsetTime = Math.round(Math.round(delta / snapValue * baseMoveTime * scale) / minimalMoveTime) * minimalMoveTime;
+                  offsetTime = Math.min(Math.max(
+                    offsetTime,
+                    -commonStore.trackPlayStartTime), commonStore.trackTotalTime - commonStore.trackPlayStartTime);
+                  commonStore.setTrackPlayStartTime(commonStore.trackPlayStartTime + offsetTime);
+                }}
+              >
+                <div className="relative cursor-move"
+                  ref={playStartTimeControlRef}>
+                  <ArrowAutofitWidth20Regular />
+                  <div
+                    className={classnames('border-l absolute', commonStore.settings.darkMode ? 'border-white' : 'border-gray-700')}
+                    style={{
+                      height: (tracksRef.current && commonStore.tracks.length > 0)
+                        ? tracksRef.current.clientHeight
+                        : 0,
+                      top: '50%',
+                      left: `calc(50% - ${pixelFix}px)`
+                    }} />
+                </div>
+              </Draggable>
+            </div>
+          </div>
+        </div>
+        <Tooltip content={t('Scale View')! + ': ' + commonStore.trackScale} showDelay={0} hideDelay={0}
+          relationship="description">
+          <Slider ref={toolbarSliderRef} value={commonStore.trackScale} step={scaleMin} max={scaleMax} min={scaleMin}
+            onChange={(e, data) => {
+              commonStore.setTrackScale(data.value);
+            }}
+          />
+        </Tooltip>
+      </div>
+      <div className="flex flex-col overflow-y-auto gap-1" ref={tracksRef}>
+        {commonStore.tracks.map(track =>
+          <div key={track.id} className="flex gap-2 pb-1 border-b">
+            <div className="flex gap-1 border-r h-7">
+              <ToolTipButton desc={commonStore.recordingTrackId === track.id ? t('Stop') : t('Record')}
+                disabled={commonStore.platform === 'web'}
+                icon={commonStore.recordingTrackId === track.id ? <Stop16Filled /> : <Record16Regular />}
+                size="small" shape="circular" appearance="subtle"
+                onClick={() => {
+                  flushMidiRecordingContent();
+                  commonStore.setPlayingTrackId('');
+
+                  if (commonStore.recordingTrackId === track.id) {
+                    commonStore.setRecordingTrackId('');
+                  } else {
+                    if (commonStore.activeMidiDeviceIndex === -1) {
+                      toast(t('Please select a MIDI device first'), { type: 'warning' });
+                      return;
+                    }
+
+                    dropRecordingTime = true;
+                    setSelectedTrackId(track.id);
+
+                    commonStore.setRecordingTrackId(track.id);
+                    commonStore.setRecordingContent(track.content);
+                    commonStore.setRecordingRawContent(track.rawContent.slice());
+                  }
+                }} />
+              <ToolTipButton disabled
+                desc={commonStore.playingTrackId === track.id ? t('Stop') : t('Play') + ' (Unavailable)'}
+                icon={commonStore.playingTrackId === track.id ? <Pause16Regular /> : <Play16Filled />}
+                size="small" shape="circular" appearance="subtle"
+                onClick={() => {
+                  flushMidiRecordingContent();
+                  commonStore.setRecordingTrackId('');
+
+                  if (commonStore.playingTrackId === track.id) {
+                    commonStore.setPlayingTrackId('');
+                  } else {
+                    setSelectedTrackId(track.id);
+
+                    commonStore.setPlayingTrackId(track.id);
+                  }
+                }} />
+              <ToolTipButton desc={t('Delete')} icon={<Delete16Regular />} size="small" shape="circular"
+                appearance="subtle" onClick={() => {
+                const tracks = commonStore.tracks.slice().filter(t => t.id !== track.id);
+                commonStore.setTracks(tracks);
+                refreshTracksTotalTime();
+              }} />
+            </div>
+            <div className="relative grow overflow-hidden">
+              <div className="absolute" style={{ left: -0 }}>
+                <Track
+                  id={track.id}
+                  scale={scale}
+                  right={Math.min(tracksWidth, moveableTracksWidth)}
+                  isSelected={selectedTrackId === track.id}
+                  onSelect={setSelectedTrackId}
+                />
+              </div>
+            </div>
+          </div>)}
+        <div className="flex justify-between items-center">
+          <div className="flex gap-1">
+            <Button icon={<Add16Regular />} size="small" shape="circular"
+              appearance="subtle"
+              disabled={commonStore.platform === 'web'}
+              onClick={() => {
+                commonStore.setTracks([...commonStore.tracks, {
+                  id: uuid(),
+                  mainInstrument: '',
+                  content: '',
+                  rawContent: [],
+                  offsetTime: 0,
+                  contentTime: 0
+                }]);
+              }}>
+              {t('New Track')}
+            </Button>
+            <Button icon={<ArrowUpload16Regular />} size="small" shape="circular"
+              appearance="subtle"
+              onClick={() => {
+                if (commonStore.status.status === ModelStatus.Offline && !commonStore.settings.apiUrl && commonStore.platform !== 'web') {
+                  toast(t('Please click the button in the top right corner to start the model'), { type: 'warning' });
+                  return;
+                }
+
+                OpenOpenFileDialog('*.mid').then(async filePath => {
+                  if (!filePath)
+                    return;
+
+                  let blob: Blob;
+                  if (commonStore.platform === 'web')
+                    blob = (filePath as unknown as { blob: Blob }).blob;
+                  else
+                    blob = await fetch(absPathAsset(filePath)).then(r => r.blob());
+                  const bodyForm = new FormData();
+                  bodyForm.append('file_data', blob);
+                  fetch(getServerRoot(commonStore.getCurrentModelConfig().apiParameters.apiPort) + '/midi-to-text', {
+                    method: 'POST',
+                    body: bodyForm
+                  }).then(async r => {
+                      if (r.status === 200) {
+                        const text = (await r.json()).text as string;
+                        const rawContent = text.split(' ').map(tokenToMidiMessage).filter(m => m) as MidiMessage[];
+                        const tracks = commonStore.tracks.slice();
+
+                        tracks.push({
+                          id: uuid(),
+                          mainInstrument: getMidiRawContentMainInstrument(rawContent),
+                          content: text,
+                          rawContent: rawContent,
+                          offsetTime: 0,
+                          contentTime: getMidiRawContentTime(rawContent)
+                        });
+                        commonStore.setTracks(tracks);
+                        refreshTracksTotalTime();
+                      } else {
+                        toast(r.statusText + '\n' + (await r.text()), {
+                          type: 'error'
+                        });
+                      }
+                    }
+                  ).catch(e => {
+                    toast(t('Error') + ' - ' + (e.message || e), { type: 'error', autoClose: 2500 });
+                  });
+                }).catch(e => {
+                  toast(t('Error') + ' - ' + (e.message || e), { type: 'error', autoClose: 2500 });
+                });
+              }}>
+              {t('Import MIDI')}
+            </Button>
+          </div>
+          <Text size={100}>
+            {t('Select a track to preview the content')}
+          </Text>
+        </div>
+      </div>
+      <div className="grow"></div>
+      {selectedTrack &&
+        <Card size="small" appearance="outline" style={{ minHeight: '150px', maxHeight: '200px' }}>
+          <div className="flex flex-col gap-1 overflow-hidden">
+            <Text size={100}>{`${t('Start Time')}: ${selectedTrack.offsetTime} ms`}</Text>
+            <Text size={100}>{`${t('Content Duration')}: ${selectedTrack.contentTime} ms`}</Text>
+            <div className="overflow-y-auto overflow-x-hidden" ref={contentPreviewRef}>
+              {selectedTrackId === commonStore.recordingTrackId
+                ? commonStore.recordingContent
+                : selectedTrack.content}
+            </div>
+          </div>
+        </Card>
+      }
+      {
+        commonStore.platform !== 'web' &&
+        <div className="flex gap-2 items-end mx-auto">
+          {t('Current Instrument') + ':'}
+          {InstrumentTypeNameMap.map((name, i) =>
+            <Text key={name} style={{ whiteSpace: 'nowrap' }}
+              className={commonStore.instrumentType === i ? 'text-blue-600' : ''}
+              weight={commonStore.instrumentType === i ? 'bold' : 'regular'}
+              size={commonStore.instrumentType === i ? 300 : 100}
+            >{t(name)}</Text>)}
+        </div>
+      }
+      <DialogTrigger disableButtonEnhancement>
+        <Button icon={<MusicNote220Regular />} style={{ minHeight: '32px' }} onClick={() => {
+          flushMidiRecordingContent();
+          commonStore.setRecordingTrackId('');
+          commonStore.setPlayingTrackId('');
+
+          const timestamp = [];
+          const sortedTracks = commonStore.tracks.slice().sort((a, b) => a.offsetTime - b.offsetTime);
+          for (const track of sortedTracks) {
+            timestamp.push(track.offsetTime);
+            let accContentTime = 0;
+            for (const msg of track.rawContent) {
+              if (msg.messageType === 'ElapsedTime') {
+                accContentTime += msg.value;
+                timestamp.push(track.offsetTime + accContentTime);
+              }
+            }
+          }
+          const sortedTimestamp = timestamp.slice().sort((a, b) => a - b);
+          const globalMessages: MidiMessage[] = sortedTimestamp.reduce((messages, current, i) =>
+              [...messages, {
+                messageType: 'ElapsedTime',
+                value: current - (i === 0 ? 0 : sortedTimestamp[i - 1])
+              } as MidiMessage]
+            , [] as MidiMessage[]);
+          for (const track of sortedTracks) {
+            let currentTime = track.offsetTime;
+            let accContentTime = 0;
+            for (const msg of track.rawContent) {
+              if (msg.messageType === 'ElapsedTime') {
+                accContentTime += msg.value;
+                currentTime = track.offsetTime + accContentTime;
+              } else if (msg.messageType === 'NoteOn' || msg.messageType === 'NoteOff') {
+                const insertIndex = sortedTimestamp.findIndex(t => t >= currentTime);
+                globalMessages.splice(insertIndex + 1, 0, msg);
+                sortedTimestamp.splice(insertIndex + 1, 0, 0); // placeholder
+              }
+            }
+          }
+          const result = ('<pad> ' + globalMessages.map(midiMessageToToken).join('')).trim();
+          commonStore.setCompositionSubmittedPrompt(result);
+          setPrompt(result);
+        }}>
+          {t('Save to generation area')}
+        </Button>
+      </DialogTrigger>
+    </div>
+  );
+});
+
+export default AudiotrackEditor;

frontend/src/pages/Chat.tsx

@@ -10,56 +10,40 @@ import { KebabHorizontalIcon, PencilIcon, SyncIcon, TrashIcon } from '@primer/oc
 import logo from '../assets/images/logo.png';
 import MarkdownRender from '../components/MarkdownRender';
 import { ToolTipButton } from '../components/ToolTipButton';
-import { ArrowCircleUp28Regular, Delete28Regular, RecordStop28Regular, Save28Regular } from '@fluentui/react-icons';
+import {
+  ArrowCircleUp28Regular,
+  ArrowClockwise16Regular,
+  Attach16Regular,
+  Delete28Regular,
+  Dismiss16Regular,
+  Dismiss24Regular,
+  RecordStop28Regular,
+  SaveRegular,
+  TextAlignJustify24Regular,
+  TextAlignJustifyRotate9024Regular
+} from '@fluentui/react-icons';
 import { CopyButton } from '../components/CopyButton';
 import { ReadButton } from '../components/ReadButton';
 import { toast } from 'react-toastify';
 import { WorkHeader } from '../components/WorkHeader';
 import { DialogButton } from '../components/DialogButton';
-import { OpenFileFolder, OpenSaveFileDialog } from '../../wailsjs/go/backend_golang/App';
-import { toastWithButton } from '../utils';
-import { PresetsButton } from './PresetsManager/PresetsButton';
+import { OpenFileFolder, OpenOpenFileDialog, OpenSaveFileDialog } from '../../wailsjs/go/backend_golang/App';
+import { absPathAsset, bytesToReadable, getServerRoot, setActivePreset, toastWithButton } from '../utils';
 import { useMediaQuery } from 'usehooks-ts';
+import { botName, ConversationMessage, MessageType, userName, welcomeUuid } from '../types/chat';
+import { Labeled } from '../components/Labeled';
+import { ValuedSlider } from '../components/ValuedSlider';
+import { PresetsButton } from './PresetsManager/PresetsButton';
+import { webOpenOpenFileDialog } from '../utils/web-file-operations';
 
-export const userName = 'M E';
-export const botName = 'A I';
+let chatSseControllers: {
+  [id: string]: AbortController
+} = {};
 
-export const welcomeUuid = 'welcome';
-
-export enum MessageType {
-  Normal,
-  Error
-}
-
-export type Side = 'left' | 'right'
-
-export type Color = 'neutral' | 'brand' | 'colorful'
-
-export type MessageItem = {
-  sender: string,
-  type: MessageType,
-  color: Color,
-  avatarImg?: string,
-  time: string,
-  content: string,
-  side: Side,
-  done: boolean
-}
-
-export type Conversation = {
-  [uuid: string]: MessageItem
-}
-
-export type Role = 'assistant' | 'user' | 'system';
-
-export type ConversationMessage = {
-  role: Role;
-  content: string;
-}
-
-let chatSseController: AbortController | null = null;
-
-const MoreUtilsButton: FC<{ uuid: string, setEditing: (editing: boolean) => void }> = observer(({
+const MoreUtilsButton: FC<{
+  uuid: string,
+  setEditing: (editing: boolean) => void
+}> = observer(({
   uuid,
   setEditing
 }) => {
@@ -83,13 +67,15 @@ const MoreUtilsButton: FC<{ uuid: string, setEditing: (editing: boolean) => void
         onClick={() => {
           commonStore.conversationOrder.splice(commonStore.conversationOrder.indexOf(uuid), 1);
           delete commonStore.conversation[uuid];
+          commonStore.setAttachment(uuid, null);
         }} />
     </MenuPopover>
   </Menu>;
 });
 
 const ChatMessageItem: FC<{
-  uuid: string, onSubmit: (message: string | null, answerId: string | null,
+  uuid: string,
+  onSubmit: (message: string | null, answerId: string | null,
     startUuid: string | null, endUuid: string | null, includeEndUuid: boolean) => void
 }> = observer(({ uuid, onSubmit }) => {
   const { t } = useTranslation();
@@ -114,6 +100,13 @@ const ChatMessageItem: FC<{
     }
   };
 
+  let avatarImg: string | undefined;
+  if (commonStore.activePreset && messageItem.sender === botName) {
+    avatarImg = absPathAsset(commonStore.activePreset.avatarImg);
+  } else if (messageItem.avatarImg) {
+    avatarImg = messageItem.avatarImg;
+  }
+
   return <div
     className={classnames(
       'flex gap-2 mb-2 overflow-hidden',
@@ -131,7 +124,7 @@ const ChatMessageItem: FC<{
     <Avatar
       color={messageItem.color}
       name={messageItem.sender}
-      image={(commonStore.activePreset && messageItem.sender === botName) ? { src: commonStore.activePreset.avatarImg } : messageItem.avatarImg ? { src: messageItem.avatarImg } : undefined}
+      image={avatarImg ? { src: avatarImg } : undefined}
     />
     <div
       className={classnames(
@@ -142,13 +135,31 @@ const ChatMessageItem: FC<{
       )}
     >
       {!editing ?
-        <MarkdownRender>{messageItem.content}</MarkdownRender> :
+        <div className="flex flex-col">
+          <MarkdownRender>{messageItem.content}</MarkdownRender>
+          {uuid in commonStore.attachments &&
+            <div className="flex grow">
+              <div className="grow" />
+              <ToolTipButton className="whitespace-nowrap"
+                text={
+                  commonStore.attachments[uuid][0].name.replace(
+                    new RegExp('(^[^\\.]{5})[^\\.]+'), '$1...')
+                }
+                desc={`${commonStore.attachments[uuid][0].name} (${bytesToReadable(commonStore.attachments[uuid][0].size)})`}
+                size="small" shape="circular" appearance="secondary" />
+            </div>
+          }
+        </div> :
         <Textarea ref={textareaRef}
           className="grow"
           style={{ minWidth: 0 }}
           value={messageItem.content}
           onChange={(e) => {
             messageItem.content = e.target.value;
+            commonStore.conversation[uuid].type = MessageType.Normal;
+            commonStore.conversation[uuid].done = true;
+            commonStore.setConversation(commonStore.conversation);
+            commonStore.setConversationOrder([...commonStore.conversationOrder]);
           }}
           onBlur={() => {
             setEditingInner(false);
@@ -166,6 +177,10 @@ const ChatMessageItem: FC<{
           messageItem.sender === botName && uuid !== welcomeUuid &&
           <ToolTipButton desc={t('Retry')} size="small" appearance="subtle"
             icon={<SyncIcon />} onClick={() => {
+            if (uuid in chatSseControllers) {
+              chatSseControllers[uuid].abort();
+              delete chatSseControllers[uuid];
+            }
             onSubmit(null, uuid, null, uuid, false);
           }} />
         }
@@ -179,23 +194,129 @@ const ChatMessageItem: FC<{
   </div>;
 });
 
+const SidePanel: FC = observer(() => {
+  const [t] = useTranslation();
+  const mq = useMediaQuery('(min-width: 640px)');
+  const params = commonStore.chatParams;
+
+  return <div
+    className={classnames(
+      'flex flex-col gap-1 h-full flex-shrink-0 transition-width duration-300 ease-in-out',
+      commonStore.sidePanelCollapsed ? 'w-0' : (mq ? 'w-64' : 'w-full'),
+      !commonStore.sidePanelCollapsed && 'ml-1')
+    }>
+    <div className="flex m-1">
+      <div className="grow" />
+      <PresetsButton tab="Chat" size="medium" shape="circular" appearance="subtle" />
+      <Button size="medium" shape="circular" appearance="subtle" icon={<Dismiss24Regular />}
+        onClick={() => commonStore.setSidePanelCollapsed(true)}
+      />
+    </div>
+    <div className="flex flex-col gap-1 overflow-x-hidden overflow-y-auto p-1">
+      <Labeled flex breakline label={t('Max Response Token')}
+        desc={t('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.')}
+        content={
+          <ValuedSlider value={params.maxResponseToken} min={100} max={8100}
+            step={100}
+            input
+            onChange={(e, data) => {
+              commonStore.setChatParams({
+                maxResponseToken: data.value
+              });
+            }} />
+        } />
+      <Labeled flex breakline label={t('Temperature')}
+        desc={t('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.')}
+        content={
+          <ValuedSlider value={params.temperature} min={0} max={2} step={0.1}
+            input
+            onChange={(e, data) => {
+              commonStore.setChatParams({
+                temperature: data.value
+              });
+            }} />
+        } />
+      <Labeled flex breakline label={t('Top_P')}
+        desc={t('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.')}
+        content={
+          <ValuedSlider value={params.topP} min={0} max={1} step={0.1} input
+            onChange={(e, data) => {
+              commonStore.setChatParams({
+                topP: data.value
+              });
+            }} />
+        } />
+      <Labeled flex breakline label={t('Presence Penalty')}
+        desc={t('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.')}
+        content={
+          <ValuedSlider value={params.presencePenalty} min={0} max={2}
+            step={0.1} input
+            onChange={(e, data) => {
+              commonStore.setChatParams({
+                presencePenalty: data.value
+              });
+            }} />
+        } />
+      <Labeled flex breakline label={t('Frequency Penalty')}
+        desc={t('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.')}
+        content={
+          <ValuedSlider value={params.frequencyPenalty} min={0} max={2}
+            step={0.1} input
+            onChange={(e, data) => {
+              commonStore.setChatParams({
+                frequencyPenalty: data.value
+              });
+            }} />
+        } />
+    </div>
+    <div className="grow" />
+    {/*<Button*/}
+    {/*  icon={<FolderOpenVerticalRegular />}*/}
+    {/*  onClick={() => {*/}
+    {/*  }}>*/}
+    {/*  {t('Load Conversation')}*/}
+    {/*</Button>*/}
+    <Button
+      icon={<SaveRegular />}
+      onClick={() => {
+        let savedContent: string = '';
+        const isWorldModel = commonStore.getCurrentModelConfig().modelParameters.modelName.toLowerCase().includes('world');
+        const user = isWorldModel ? 'User' : 'Bob';
+        const bot = isWorldModel ? 'Assistant' : 'Alice';
+        commonStore.conversationOrder.forEach((uuid) => {
+          if (uuid === welcomeUuid)
+            return;
+          const messageItem = commonStore.conversation[uuid];
+          if (messageItem.type !== MessageType.Error) {
+            savedContent += `${messageItem.sender === userName ? user : bot}: ${messageItem.content}\n\n`;
+          }
+        });
+
+        OpenSaveFileDialog('*.txt', 'conversation.txt', savedContent).then((path) => {
+          if (path)
+            toastWithButton(t('Conversation Saved'), t('Open'), () => {
+              OpenFileFolder(path, false);
+            });
+        }).catch(e => {
+          toast(t('Error') + ' - ' + (e.message || e), { type: 'error', autoClose: 2500 });
+        });
+      }}>
+      {t('Save Conversation')}
+    </Button>
+  </div>;
+});
+
 const ChatPanel: FC = observer(() => {
   const { t } = useTranslation();
   const bodyRef = useRef<HTMLDivElement>(null);
   const inputRef = useRef<HTMLTextAreaElement>(null);
   const mq = useMediaQuery('(min-width: 640px)');
+  if (commonStore.sidePanelCollapsed === 'auto')
+    commonStore.setSidePanelCollapsed(!mq);
   const currentConfig = commonStore.getCurrentModelConfig();
   const apiParams = currentConfig.apiParameters;
   const port = apiParams.apiPort;
-
-  let lastMessageId: string;
-  let generating: boolean = false;
-  if (commonStore.conversationOrder.length > 0) {
-    lastMessageId = commonStore.conversationOrder[commonStore.conversationOrder.length - 1];
-    const lastMessage = commonStore.conversation[lastMessageId];
-    if (lastMessage.sender === botName)
-      generating = !lastMessage.done;
-  }
+  const generating: boolean = Object.keys(chatSseControllers).length > 0;
 
   useEffect(() => {
     if (inputRef.current)
@@ -213,7 +334,7 @@ const ChatPanel: FC = observer(() => {
           color: 'colorful',
           avatarImg: logo,
           time: new Date().toISOString(),
-          content: t('Hello! I\'m RWKV, an open-source and commercially usable large language model.'),
+          content: commonStore.platform === 'web' ? t('Hello, what can I do for you?') : t('Hello! I\'m RWKV, an open-source and commercially usable large language model.'),
           side: 'left',
           done: true
         }
@@ -230,7 +351,7 @@ const ChatPanel: FC = observer(() => {
     e.stopPropagation();
     if (e.type === 'click' || (e.keyCode === 13 && !e.shiftKey)) {
       e.preventDefault();
-      if (commonStore.status.status === ModelStatus.Offline && !commonStore.settings.apiUrl) {
+      if (commonStore.status.status === ModelStatus.Offline && !commonStore.settings.apiUrl && commonStore.platform !== 'web') {
         toast(t('Please click the button in the top right corner to start the model'), { type: 'warning' });
         return;
       }
@@ -260,6 +381,11 @@ const ChatPanel: FC = observer(() => {
       commonStore.setConversation(commonStore.conversation);
       commonStore.conversationOrder.push(newId);
       commonStore.setConversationOrder(commonStore.conversationOrder);
+
+      if (commonStore.currentTempAttachment) {
+        commonStore.setAttachment(newId, [commonStore.currentTempAttachment]);
+        commonStore.setCurrentTempAttachment(null);
+      }
     }
 
     let startIndex = startUuid ? commonStore.conversationOrder.indexOf(startUuid) : 0;
@@ -271,6 +397,17 @@ const ChatPanel: FC = observer(() => {
       if (uuid === welcomeUuid)
         return;
       const messageItem = commonStore.conversation[uuid];
+      if (uuid in commonStore.attachments) {
+        const attachment = commonStore.attachments[uuid][0];
+        messages.push({
+          role: 'user',
+          content: t('The content of file') + ` "${attachment.name}" `
+            + t('is as follows. When replying to me, consider the file content and respond accordingly:')
+            + '\n\n' + attachment.content
+        });
+        messages.push({ role: 'user', content: t('What\'s the file name') });
+        messages.push({ role: 'assistant', content: t('The file name is: ') + attachment.name });
+      }
       if (messageItem.done && messageItem.type === MessageType.Normal && messageItem.sender === userName) {
         messages.push({ role: 'user', content: messageItem.content });
       } else if (messageItem.done && messageItem.type === MessageType.Normal && messageItem.sender === botName) {
@@ -296,11 +433,10 @@ const ChatPanel: FC = observer(() => {
     commonStore.setConversationOrder(commonStore.conversationOrder);
     setTimeout(scrollToBottom);
     let answer = '';
-    chatSseController = new AbortController();
-    fetchEventSource( // https://api.openai.com/v1/chat/completions || http://127.0.0.1:${port}/chat/completions
-      commonStore.settings.apiUrl ?
-        commonStore.settings.apiUrl + '/v1/chat/completions' :
-        `http://127.0.0.1:${port}/chat/completions`,
+    const chatSseController = new AbortController();
+    chatSseControllers[answerId] = chatSseController;
+    fetchEventSource( // https://api.openai.com/v1/chat/completions || http://127.0.0.1:${port}/v1/chat/completions
+      getServerRoot(port, true) + '/v1/chat/completions',
       {
         method: 'POST',
         headers: {
@@ -311,13 +447,20 @@ const ChatPanel: FC = observer(() => {
           messages,
           stream: true,
           model: commonStore.settings.apiChatModelName, // 'gpt-3.5-turbo'
-          temperature: apiParams.temperature,
-          top_p: apiParams.topP
+          temperature: commonStore.chatParams.temperature,
+          top_p: commonStore.chatParams.topP,
+          presence_penalty: commonStore.chatParams.presencePenalty,
+          frequency_penalty: commonStore.chatParams.frequencyPenalty,
+          user_name: commonStore.activePreset?.userName || undefined,
+          assistant_name: commonStore.activePreset?.assistantName || undefined,
+          presystem: commonStore.activePreset?.presystem && undefined
         }),
         signal: chatSseController?.signal,
         onmessage(e) {
           scrollToBottom();
           if (e.data.trim() === '[DONE]') {
+            if (answerId! in chatSseControllers)
+              delete chatSseControllers[answerId!];
             commonStore.conversation[answerId!].done = true;
             commonStore.conversation[answerId!].content = commonStore.conversation[answerId!].content.trim();
             commonStore.setConversation(commonStore.conversation);
@@ -331,6 +474,8 @@ const ChatPanel: FC = observer(() => {
             console.debug('json error', error);
             return;
           }
+          if (data.model)
+            commonStore.setLastModelName(data.model);
           if (data.choices && Array.isArray(data.choices) && data.choices.length > 0) {
             answer += data.choices[0]?.delta?.content || '';
             commonStore.conversation[answerId!].content = answer;
@@ -347,9 +492,13 @@ const ChatPanel: FC = observer(() => {
           }
         },
         onclose() {
+          if (answerId! in chatSseControllers)
+            delete chatSseControllers[answerId!];
           console.log('Connection closed');
         },
         onerror(err) {
+          if (answerId! in chatSseControllers)
+            delete chatSseControllers[answerId!];
           commonStore.conversation[answerId!].type = MessageType.Error;
           commonStore.conversation[answerId!].done = true;
           err = err.message || err;
@@ -364,82 +513,175 @@ const ChatPanel: FC = observer(() => {
   }, []);
 
   return (
-    <div className="flex flex-col w-full grow gap-4 pt-4 overflow-hidden">
-      <div ref={bodyRef} className="grow overflow-y-scroll overflow-x-hidden pr-2">
-        {commonStore.conversationOrder.map(uuid =>
-          <ChatMessageItem key={uuid} uuid={uuid} onSubmit={onSubmit} />
-        )}
-      </div>
-      <div className={classnames('flex items-end', mq ? 'gap-2' : '')}>
-        <PresetsButton tab="Chat" size={mq ? 'large' : 'small'} shape="circular" appearance="subtle" />
-        <DialogButton tooltip={t('Clear')}
-          icon={<Delete28Regular />}
-          size={mq ? 'large' : 'small'} shape="circular" appearance="subtle" title={t('Clear')}
-          contentText={t('Are you sure you want to clear the conversation? It cannot be undone.')}
-          onConfirm={() => {
-            if (generating)
-              chatSseController?.abort();
-            commonStore.setConversation({});
-            commonStore.setConversationOrder([]);
-          }} />
-        <Textarea
-          ref={inputRef}
-          style={{ minWidth: 0 }}
-          className="grow"
-          resize="vertical"
-          placeholder={t('Type your message here')!}
-          value={commonStore.currentInput}
-          onChange={(e) => commonStore.setCurrentInput(e.target.value)}
-          onKeyDown={handleKeyDownOrClick}
-        />
-        <ToolTipButton desc={generating ? t('Stop') : t('Send')}
-          icon={generating ? <RecordStop28Regular /> : <ArrowCircleUp28Regular />}
-          size={mq ? 'large' : 'small'} shape="circular" appearance="subtle"
-          onClick={(e) => {
-            if (generating) {
-              chatSseController?.abort();
-              if (lastMessageId) {
-                commonStore.conversation[lastMessageId].type = MessageType.Error;
-                commonStore.conversation[lastMessageId].done = true;
+    <div className="flex h-full grow pt-4 overflow-hidden">
+      <div className="relative flex flex-col w-full grow gap-4 overflow-hidden">
+        <Button className="absolute top-1 right-1" size="medium" shape="circular" appearance="subtle"
+          style={{ zIndex: 1 }}
+          icon={commonStore.sidePanelCollapsed ? <TextAlignJustify24Regular /> : <TextAlignJustifyRotate9024Regular />}
+          onClick={() => commonStore.setSidePanelCollapsed(!commonStore.sidePanelCollapsed)} />
+        <div ref={bodyRef} className="grow overflow-y-scroll overflow-x-hidden pr-2">
+          {commonStore.conversationOrder.map(uuid =>
+            <ChatMessageItem key={uuid} uuid={uuid} onSubmit={onSubmit} />
+          )}
+        </div>
+        <div className={classnames('flex items-end', mq ? 'gap-2' : '')}>
+          <DialogButton tooltip={t('Clear')}
+            icon={<Delete28Regular />}
+            size={mq ? 'large' : 'small'} shape="circular" appearance="subtle" title={t('Clear')}
+            contentText={t('Are you sure you want to clear the conversation? It cannot be undone.')}
+            onConfirm={() => {
+              if (generating) {
+                for (const id in chatSseControllers) {
+                  chatSseControllers[id].abort();
+                }
+                chatSseControllers = {};
+              }
+              setActivePreset(commonStore.activePreset);
+            }} />
+          <div className="relative flex grow">
+            <Textarea
+              ref={inputRef}
+              style={{ minWidth: 0 }}
+              className="grow"
+              resize="vertical"
+              placeholder={t('Type your message here')!}
+              value={commonStore.currentInput}
+              onChange={(e) => commonStore.setCurrentInput(e.target.value)}
+              onKeyDown={handleKeyDownOrClick}
+            />
+            <div className="absolute right-2 bottom-2">
+              {!commonStore.currentTempAttachment ?
+                <ToolTipButton
+                  desc={commonStore.attachmentUploading ?
+                    t('Processing Attachment') :
+                    t('Add An Attachment (Accepts pdf, txt)')}
+                  icon={commonStore.attachmentUploading ?
+                    <ArrowClockwise16Regular className="animate-spin" />
+                    : <Attach16Regular />}
+                  size="small" shape="circular" appearance="secondary"
+                  onClick={() => {
+                    if (commonStore.attachmentUploading)
+                      return;
+
+                    const filterPattern = '*.txt;*.pdf';
+                    const setUploading = () => commonStore.setAttachmentUploading(true);
+                    // actually, status of web platform is always Offline
+                    if (commonStore.platform === 'web' || commonStore.status.status === ModelStatus.Offline || currentConfig.modelParameters.device === 'WebGPU') {
+                      webOpenOpenFileDialog(filterPattern, setUploading).then(webReturn => {
+                        if (webReturn.content)
+                          commonStore.setCurrentTempAttachment(
+                            {
+                              name: webReturn.blob.name,
+                              size: webReturn.blob.size,
+                              content: webReturn.content
+                            });
+                        else
+                          toast(t('File is empty'), {
+                            type: 'info',
+                            autoClose: 1000
+                          });
+                      }).catch(e => {
+                        toast(t('Error') + ' - ' + (e.message || e), { type: 'error', autoClose: 2500 });
+                      }).finally(() => {
+                        commonStore.setAttachmentUploading(false);
+                      });
+                    } else {
+                      OpenOpenFileDialog(filterPattern).then(async filePath => {
+                        if (!filePath)
+                          return;
+
+                        setUploading();
+
+                        // Both are slow. Communication between frontend and backend is slow. Use AssetServer Handler to read the file.
+                        // const blob = new Blob([atob(info.content as unknown as string)]); // await fetch(`data:application/octet-stream;base64,${info.content}`).then(r => r.blob());
+                        const blob = await fetch(absPathAsset(filePath)).then(r => r.blob());
+                        const attachmentName = filePath.split(/[\\/]/).pop();
+                        const urlPath = `/file-to-text?file_name=${attachmentName}`;
+                        const bodyForm = new FormData();
+                        bodyForm.append('file_data', blob, attachmentName);
+                        fetch(getServerRoot(port) + urlPath, {
+                          method: 'POST',
+                          body: bodyForm
+                        }).then(async r => {
+                            if (r.status === 200) {
+                              const pages = (await r.json()).pages as any[];
+                              let attachmentContent: string;
+                              if (pages.length === 1)
+                                attachmentContent = pages[0].page_content;
+                              else
+                                attachmentContent = pages.map((p, i) => `Page ${i + 1}:\n${p.page_content}`).join('\n\n');
+                              if (attachmentContent)
+                                commonStore.setCurrentTempAttachment(
+                                  {
+                                    name: attachmentName!,
+                                    size: blob.size,
+                                    content: attachmentContent!
+                                  });
+                              else
+                                toast(t('File is empty'), {
+                                  type: 'info',
+                                  autoClose: 1000
+                                });
+                            } else {
+                              toast(r.statusText + '\n' + (await r.text()), {
+                                type: 'error'
+                              });
+                            }
+                          }
+                        ).catch(e => {
+                          toast(t('Error') + ' - ' + (e.message || e), { type: 'error', autoClose: 2500 });
+                        }).finally(() => {
+                          commonStore.setAttachmentUploading(false);
+                        });
+                      }).catch(e => {
+                        toast(t('Error') + ' - ' + (e.message || e), { type: 'error', autoClose: 2500 });
+                      });
+                    }
+                  }}
+                /> :
+                <div>
+                  <ToolTipButton
+                    text={
+                      commonStore.currentTempAttachment.name.replace(
+                        new RegExp('(^[^\\.]{5})[^\\.]+'), '$1...')
+                    }
+                    desc={`${commonStore.currentTempAttachment.name} (${bytesToReadable(commonStore.currentTempAttachment.size)})`}
+                    size="small" shape="circular" appearance="secondary" />
+                  <ToolTipButton desc={t('Remove Attachment')}
+                    icon={<Dismiss16Regular />}
+                    size="small" shape="circular" appearance="subtle"
+                    onClick={() => {
+                      commonStore.setCurrentTempAttachment(null);
+                    }} />
+                </div>
+              }
+            </div>
+          </div>
+          <ToolTipButton desc={generating ? t('Stop') : t('Send')}
+            icon={generating ? <RecordStop28Regular /> : <ArrowCircleUp28Regular />}
+            size={mq ? 'large' : 'small'} shape="circular" appearance="subtle"
+            onClick={(e) => {
+              if (generating) {
+                for (const id in chatSseControllers) {
+                  chatSseControllers[id].abort();
+                  commonStore.conversation[id].type = MessageType.Error;
+                  commonStore.conversation[id].done = true;
+                }
+                chatSseControllers = {};
                 commonStore.setConversation(commonStore.conversation);
                 commonStore.setConversationOrder([...commonStore.conversationOrder]);
+              } else {
+                handleKeyDownOrClick(e);
               }
-            } else {
-              handleKeyDownOrClick(e);
-            }
-          }} />
-        <ToolTipButton desc={t('Save')}
-          icon={<Save28Regular />}
-          size={mq ? 'large' : 'small'} shape="circular" appearance="subtle"
-          onClick={() => {
-            let savedContent: string = '';
-            const isWorldModel = commonStore.getCurrentModelConfig().modelParameters.modelName.toLowerCase().includes('world');
-            const user = isWorldModel ? 'Question' : 'Bob';
-            const bot = isWorldModel ? 'Answer' : 'Alice';
-            commonStore.conversationOrder.forEach((uuid) => {
-              if (uuid === welcomeUuid)
-                return;
-              const messageItem = commonStore.conversation[uuid];
-              if (messageItem.type !== MessageType.Error) {
-                savedContent += `${messageItem.sender === userName ? user : bot}: ${messageItem.content}\n\n`;
-              }
-            });
-
-            OpenSaveFileDialog('*.txt', 'conversation.txt', savedContent).then((path) => {
-              if (path)
-                toastWithButton(t('Conversation Saved'), t('Open'), () => {
-                  OpenFileFolder(path, false);
-                });
-            }).catch(e => {
-              toast(t('Error') + ' - ' + (e.message || e), { type: 'error', autoClose: 2500 });
-            });
-          }} />
+            }} />
+        </div>
       </div>
+      <SidePanel />
     </div>
   );
 });
 
-export const Chat: FC = observer(() => {
+const Chat: FC = observer(() => {
   return (
     <div className="flex flex-col gap-1 p-2 h-full overflow-hidden">
       <WorkHeader />
@@ -447,3 +689,5 @@ export const Chat: FC = observer(() => {
     </div>
   );
 });
+
+export default Chat;