release v1.6.6

improve Models page
chore
2023-12-25 21:02:26 +08:00 · 2023-12-25 20:37:40 +08:00 · 2023-12-25 20:35:00 +08:00 · 2023-12-25 20:34:35 +08:00 · 2023-12-25 20:32:54 +08:00 · 2023-12-25 19:30:57 +08:00
47 changed files with 66247 additions and 128 deletions
--- a/.gitattributes
+++ b/.gitattributes
@@ -1,3 +1,5 @@
+* text=auto eol=lf
+
 backend-python/rwkv_pip/** linguist-vendored
 backend-python/wkv_cuda_utils/** linguist-vendored
 backend-python/get-pip.py linguist-vendored
--- a/.github/dependabot.yml
+++ b/.github/dependabot.yml
@@ -0,0 +1,9 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    commit-message:
+      prefix: "chore"
+      include: "scope"
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -98,6 +98,7 @@ jobs:
          rm ./backend-python/get-pip.py
          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

@@ -124,6 +125,7 @@ jobs:
          rm ./backend-python/get-pip.py
          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
--- a/CURRENT_CHANGE.md
+++ b/CURRENT_CHANGE.md
@@ -1,13 +1,17 @@
 ## Changes

- rwkv.cpp python38 compatibility
- improve rwkv.cpp operation prompts
- add load failed traceback
- fix windows cmd waiting
+- 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
--- a/README.md
+++ b/README.md
@@ -47,13 +47,28 @@ English | [简体中文](README_ZH.md) | [日本語](README_JA.md)

 </div>

-#### Tip: 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`.
+## Tips

-#### 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.
+- 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`.

-#### 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`).
+- 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.

-#### 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.
+- 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

@@ -168,6 +183,10 @@ Tip: You can download https://github.com/josStorer/sgm_plus and unzip it to the
 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
@@ -206,12 +225,16 @@ it in the source code directory.

 ## 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

--- a/README_JA.md
+++ b/README_JA.md
@@ -47,13 +47,26 @@

 </div>

-#### ヒント：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。
+## ヒント

-#### 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。
+- 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。

-#### 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`)。
+- 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。

-#### 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。
+- 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

@@ -167,6 +180,10 @@ Tip: You can download https://github.com/josStorer/sgm_plus and unzip it to the
 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
@@ -205,12 +222,16 @@ it in the source code directory.

 ## 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

 ## Preview

--- a/README_ZH.md
+++ b/README_ZH.md
@@ -46,13 +46,22 @@ API兼容的接口，这意味着一切ChatGPT客户端都是RWKV客户端。

 </div>

-#### 小贴士：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
+## 小贴士

-#### 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
+- 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

-#### 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`)
+- 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

-#### 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
+- 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

@@ -161,6 +170,9 @@ for i in np.argsort(embeddings_cos_sim)[::-1]:
 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
@@ -192,12 +204,16 @@ for i in np.argsort(embeddings_cos_sim)[::-1]:

 ## 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

--- a/backend-golang/app.go
+++ b/backend-golang/app.go
@@ -50,9 +50,12 @@ func (a *App) OnStartup(ctx context.Context) {
 	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/backend-golang/rwkv.go
+++ b/backend-golang/rwkv.go
@@ -10,7 +10,7 @@ import (
 	"strings"
 )

-func (a *App) StartServer(python string, port int, host string, webui bool, rwkvBeta bool, rwkvcpp 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()
@@ -28,6 +28,9 @@ func (a *App) StartServer(python string, port int, host string, webui bool, rwkv
 	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...)
 }
@@ -55,6 +58,17 @@ func (a *App) ConvertSafetensors(modelPath string, outPath string) (string, erro
 	return Cmd(args...)
 }

+func (a *App) ConvertSafetensorsWithPython(python string, modelPath string, outPath string) (string, error) {
+	var err error
+	if python == "" {
+		python, err = GetPython()
+	}
+	if err != nil {
+		return "", err
+	}
+	return Cmd(python, "./backend-python/convert_safetensors.py", "--input", modelPath, "--output", outPath)
+}
+
 func (a *App) ConvertGGML(python string, modelPath string, outPath string, Q51 bool) (string, error) {
 	var err error
 	if python == "" {
--- a/backend-golang/utils.go
+++ b/backend-golang/utils.go
@@ -3,6 +3,7 @@ package backend_golang
 import (
 	"archive/zip"
 	"bufio"
+	"crypto/sha256"
 	"embed"
 	"errors"
 	"fmt"
@@ -112,9 +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
--- a/backend-python/convert_safetensors.py
+++ b/backend-python/convert_safetensors.py
@@ -30,6 +30,33 @@ def convert_file(pt_filename: str, sf_filename: str, rename={}, transpose_names=
    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}')
--- a/backend-python/main.py
+++ b/backend-python/main.py
@@ -37,6 +37,11 @@ def get_args(args: Union[Sequence[str], None] = None):
        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
--- a/backend-python/routes/completion.py
+++ b/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
@@ -335,6 +334,8 @@ The following is a coherent verbose detailed conversation between a girl named {
        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(
@@ -396,6 +397,8 @@ class EmbeddingsBody(BaseModel):


 def embedding_base64(embedding: List[float]) -> str:
+    import numpy as np
+
    return base64.b64encode(np.array(embedding).astype(np.float32)).decode("utf-8")


--- a/backend-python/routes/midi.py
+++ b/backend-python/routes/midi.py
@@ -37,10 +37,14 @@ def text_to_midi(body: TextToMidiBody):
 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)
-    text = convert_midi_to_str(cfg, mid)
+    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": text}
+    return {"text": output_list[0]}


 class TxtToMidiBody(BaseModel):
--- a/backend-python/routes/state_cache.py
+++ b/backend-python/routes/state_cache.py
@@ -87,18 +87,34 @@ def add_state(body: AddStateBody):
        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)
-        devices: List[torch.device] = [
-            (tensor.device if hasattr(tensor, "device") else torch.device("cpu"))
-            for tensor in body.state
-        ]
        dtrie[id] = {
            "tokens": copy.deepcopy(body.tokens),
-            "state": [tensor.cpu() for tensor in body.state]
-            if hasattr(body.state[0], "device")
-            else copy.deepcopy(body.state),
+            "state": state,
            "logits": copy.deepcopy(body.logits),
            "devices": devices,
        }
@@ -174,6 +190,7 @@ def longest_prefix_state(body: LongestPrefixStateBody, request: Request):
        raise HTTPException(status.HTTP_400_BAD_REQUEST, "trie not loaded")

    import torch
+    import numpy as np

    id = -1
    try:
@@ -185,14 +202,16 @@ def longest_prefix_state(body: LongestPrefixStateBody, request: Request):
        v = dtrie[id]
        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(devices[i]) for i, tensor in enumerate(v["state"])]
-            if hasattr(v["state"][0], "device")
-            else v["state"],
+            "state": state,
            "logits": v["logits"],
        }
    else:
--- a/backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt
+++ b/backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt
--- a/backend-python/rwkv_pip/utils.py
+++ b/backend-python/rwkv_pip/utils.py
@@ -84,6 +84,8 @@ class PIPELINE:
        return e / e.sum(axis=axis, keepdims=True)

    def sample_logits(self, logits, temperature=1.0, top_p=0.85, top_k=0):
+        if type(logits) == list:
+            logits = np.array(logits)
        np_logits = type(logits) == np.ndarray
        if np_logits:
            probs = self.np_softmax(logits, axis=-1)
--- a/backend-python/rwkv_pip/webgpu/model.py
+++ b/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)
--- a/backend-python/rwkv_pip/webgpu/web_rwkv_py.cp310-win_amd64.pyd
+++ b/backend-python/rwkv_pip/webgpu/web_rwkv_py.cp310-win_amd64.pyd
--- a/backend-python/utils/midi.py
+++ b/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
--- a/backend-python/utils/midi_filter_config.json
+++ b/backend-python/utils/midi_filter_config.json
@@ -0,0 +1,5 @@
+{
+    "deduplicate_md5": true,
+    "piece_split_delay": 10000,
+    "min_piece_length": 0
+}
--- a/backend-python/utils/rwkv.py
+++ b/backend-python/utils/rwkv.py
@@ -8,7 +8,6 @@ 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

@@ -68,6 +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
@@ -222,6 +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
@@ -231,7 +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:
@@ -511,6 +514,7 @@ def get_tokenizer(tokenizer_len: int):
 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"
@@ -526,6 +530,11 @@ def RWKV(model: str, strategy: str, tokenizer: Union[str, None]) -> AbstractRWKV
        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,
--- a/build/darwin/Readme_Install.txt
+++ b/build/darwin/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_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)
--- a/build/linux/Readme_Install.txt
+++ b/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。
--- a/build/windows/Readme_Install.txt
+++ b/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。
--- a/exportModelsJson.js
+++ b/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)
 })
--- a/frontend/package-lock.json
+++ b/frontend/package-lock.json
@@ -18,6 +18,7 @@
        "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",
@@ -40,6 +41,7 @@
      },
      "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",
@@ -2533,6 +2535,21 @@
        "hoist-non-react-statics": "^3.3.0"
      }
    },
+    "node_modules/@types/lodash": {
+      "version": "4.14.202",
+      "resolved": "https://registry.npmjs.org/@types/lodash/-/lodash-4.14.202.tgz",
+      "integrity": "sha512-OvlIYQK9tNneDlS0VN54LLd5uiPCBOp7gS5Z0f1mjoJYBrtStzgmJBxONW3U6OZqdtNzZPmn9BS/7WI7BFFcFQ==",
+      "dev": true
+    },
+    "node_modules/@types/lodash-es": {
+      "version": "4.17.12",
+      "resolved": "https://registry.npmjs.org/@types/lodash-es/-/lodash-es-4.17.12.tgz",
+      "integrity": "sha512-0NgftHUcV4v34VhXm8QBSftKVXtbkBG3ViCjs6+eJ5a6y6Mi/jiFGPc1sC7QK+9BFhWrURE3EOggmWaSxL9OzQ==",
+      "dev": true,
+      "dependencies": {
+        "@types/lodash": "*"
+      }
+    },
    "node_modules/@types/long": {
      "version": "4.0.2",
      "resolved": "https://registry.npmjs.org/@types/long/-/long-4.0.2.tgz",
@@ -4210,6 +4227,11 @@
      "integrity": "sha512-7ylylesZQ/PV29jhEDl3Ufjo6ZX7gCqJr5F7PKrqc93v7fzSymt1BpwEU8nAUXs8qzzvqhbjhK5QZg6Mt/HkBg==",
      "dev": true
    },
+    "node_modules/lodash-es": {
+      "version": "4.17.21",
+      "resolved": "https://registry.npmjs.org/lodash-es/-/lodash-es-4.17.21.tgz",
+      "integrity": "sha512-mKnC+QJ9pWVzv+C4/U3rRsHapFfHvQFoFB92e52xeyGMcX6/OlIl78je1u8vePzYZSkkogMPJ2yjxxsb89cxyw=="
+    },
    "node_modules/long": {
      "version": "4.0.0",
      "resolved": "https://registry.npmjs.org/long/-/long-4.0.0.tgz",
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -19,6 +19,7 @@
    "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",
@@ -41,6 +42,7 @@
  },
  "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",
--- a/frontend/src/_locales/ja/main.json
+++ b/frontend/src/_locales/ja/main.json
@@ -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",
@@ -312,6 +312,8 @@
  "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",
@@ -320,5 +322,6 @@
  "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"
+  "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"
 }
--- a/frontend/src/_locales/zh-hans/main.json
+++ b/frontend/src/_locales/zh-hans/main.json
@@ -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",
@@ -312,6 +312,8 @@
  "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",
@@ -320,5 +322,6 @@
  "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"
+  "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"
 }
--- a/frontend/src/components/RunButton.tsx
+++ b/frontend/src/components/RunButton.tsx
@@ -48,6 +48,7 @@ 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 = '';
@@ -77,7 +78,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
        });
      };

-      if (webgpu) {
+      if (webgpu || webgpuPython) {
        if (!['.st', '.safetensors'].some(ext => modelPath.endsWith(ext))) {
          const stModelPath = modelPath.replace(/\.pth$/, '.st');
          if (await FileExists(stModelPath)) {
@@ -92,7 +93,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
            return;
          } else {
            toastWithButton(t('Please convert model to safe tensors format first'), t('Convert'), () => {
-              convertToSt(modelConfig);
+              convertToSt(modelConfig, navigate);
            });
            commonStore.setStatus({ status: ModelStatus.Offline });
            return;
@@ -100,7 +101,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
        }
      }

-      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 });
@@ -176,7 +177,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
      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.enableWebUI, isUsingCudaBeta, cpp
+        !!modelConfig.enableWebUI, isUsingCudaBeta, cpp, webgpuPython
      ).catch((e) => {
        const errMsg = e.message || e;
        if (errMsg.includes('path contains space'))
@@ -216,7 +217,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean

            const strategy = getStrategy(modelConfig);
            let customCudaFile = '';
-            if ((modelConfig.modelParameters.device.includes('CUDA') || modelConfig.modelParameters.device === 'Custom')
+            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') {
@@ -264,7 +265,7 @@ export const RunButton: FC<{ onClickRun?: MouseEventHandler, iconMode?: boolean
                  navigate({ pathname: '/' + buttonName.toLowerCase() });
                };

-                if ((modelConfig.modelParameters.device === 'CUDA' || modelConfig.modelParameters.device === 'CUDA-Beta') &&
+                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)
@@ -279,7 +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.'
                };
--- a/frontend/src/pages/Composition.tsx
+++ b/frontend/src/pages/Composition.tsx
@@ -152,10 +152,14 @@ const CompositionPanel: FC = observer(() => {
        if (autoPlay) {
          if (commonStore.compositionParams.externalPlay)
            externalPlayListener();
-          else
+          else {
+            if (commonStore.compositionParams.playOnlyGeneratedContent && playerRef.current) {
+              playerRef.current.currentTime = Math.max(commonStore.compositionParams.generationStartTime - 1, 0);
+            }
            setTimeout(() => {
              playerRef.current?.start();
            });
+          }
        }
      });
    });
@@ -314,6 +318,14 @@ const CompositionPanel: FC = observer(() => {
                autoPlay: data.checked as boolean
              });
            }} />
+            <Checkbox className="select-none"
+              size="large" label={t('Only Auto Play Generated Content')} checked={params.playOnlyGeneratedContent}
+              onChange={async (_, data) => {
+                setParams({
+                  autoPlay: data.checked as boolean || commonStore.compositionParams.autoPlay,
+                  playOnlyGeneratedContent: data.checked as boolean
+                });
+              }} />
            <Labeled flex breakline label={t('MIDI Input')}
              desc={t('Select the MIDI input device to be used.')}
              content={
@@ -359,6 +371,9 @@ const CompositionPanel: FC = observer(() => {
              contentText={t('Are you sure you want to reset this page? It cannot be undone.')}
              onConfirm={() => {
                commonStore.setCompositionSubmittedPrompt(defaultCompositionPrompt);
+                setParams({
+                  generationStartTime: 0
+                });
                setPrompt(defaultCompositionPrompt);
              }} />
            <Button className="grow" appearance="primary" onClick={() => {
@@ -368,6 +383,9 @@ const CompositionPanel: FC = observer(() => {
                generateNs(params.autoPlay);
              } else {
                commonStore.setCompositionGenerating(true);
+                setParams({
+                  generationStartTime: playerRef.current ? playerRef.current.duration : 0
+                });
                onSubmit(params.prompt);
              }
            }}>{!commonStore.compositionGenerating ? t('Generate') : t('Stop')}</Button>
--- a/frontend/src/pages/Configs.tsx
+++ b/frontend/src/pages/Configs.tsx
@@ -228,9 +228,18 @@ const Configs: FC = observer(() => {
                    <Select style={{ minWidth: 0 }} className="grow"
                      value={selectedConfig.modelParameters.modelName}
                      onChange={(e, data) => {
-                        setSelectedConfigModelParams({
-                          modelName: data.value
-                        });
+                        const modelSource = commonStore.modelSourceList.find(item => item.name === data.value);
+                        if (modelSource?.customTokenizer)
+                          setSelectedConfigModelParams({
+                            modelName: data.value,
+                            useCustomTokenizer: true,
+                            customTokenizer: modelSource?.customTokenizer
+                          });
+                        else // prevent customTokenizer from being overwritten
+                          setSelectedConfigModelParams({
+                            modelName: data.value,
+                            useCustomTokenizer: false
+                          });
                      }}>
                      {!commonStore.modelSourceList.find(item => item.name === selectedConfig.modelParameters.modelName)?.isComplete
                        && <option key={-1}
@@ -246,7 +255,7 @@ const Configs: FC = observer(() => {
                  </div>
                } />
                {
-                  selectedConfig.modelParameters.device !== 'WebGPU' ?
+                  !selectedConfig.modelParameters.device.startsWith('WebGPU') ?
                    (selectedConfig.modelParameters.device !== 'CPU (rwkv.cpp)' ?
                      <ToolTipButton text={t('Convert')}
                        desc={t('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.')}
@@ -256,7 +265,7 @@ const Configs: FC = observer(() => {
                        onClick={() => convertToGGML(selectedConfig, navigate)} />)
                    : <ToolTipButton text={t('Convert To Safe Tensors Format')}
                      desc=""
-                      onClick={() => convertToSt(selectedConfig)} />
+                      onClick={() => convertToSt(selectedConfig, navigate)} />
                }
                <Labeled label={t('Strategy')} content={
                  <Dropdown style={{ minWidth: 0 }} className="grow" value={t(selectedConfig.modelParameters.device)!}
@@ -274,6 +283,7 @@ const Configs: FC = observer(() => {
                    <Option value="CUDA">CUDA</Option>
                    <Option value="CUDA-Beta">{t('CUDA (Beta, Faster)')!}</Option>
                    <Option value="WebGPU">WebGPU</Option>
+                    <Option value="WebGPU (Python)">WebGPU (Python)</Option>
                    <Option value="Custom">{t('Custom')!}</Option>
                  </Dropdown>
                } />
@@ -281,7 +291,8 @@ const Configs: FC = observer(() => {
                  selectedConfig.modelParameters.device !== 'Custom' && <Labeled label={t('Precision')}
                    desc={t('int8 uses less VRAM, but has slightly lower quality. fp16 has higher quality.')}
                    content={
-                      <Dropdown style={{ minWidth: 0 }} className="grow"
+                      <Dropdown
+                        style={{ minWidth: 0 }} className="grow"
                        value={selectedConfig.modelParameters.precision}
                        selectedOptions={[selectedConfig.modelParameters.precision]}
                        onOptionSelect={(_, data) => {
@@ -294,20 +305,20 @@ const Configs: FC = observer(() => {
                        {selectedConfig.modelParameters.device !== 'CPU' && selectedConfig.modelParameters.device !== 'MPS' &&
                          <Option>fp16</Option>}
                        {selectedConfig.modelParameters.device !== 'CPU (rwkv.cpp)' && <Option>int8</Option>}
-                        {selectedConfig.modelParameters.device === 'WebGPU' && <Option>nf4</Option>}
-                        {selectedConfig.modelParameters.device !== 'CPU (rwkv.cpp)' && selectedConfig.modelParameters.device !== 'WebGPU' &&
+                        {selectedConfig.modelParameters.device.startsWith('WebGPU') && <Option>nf4</Option>}
+                        {selectedConfig.modelParameters.device !== 'CPU (rwkv.cpp)' && !selectedConfig.modelParameters.device.startsWith('WebGPU') &&
                          <Option>fp32</Option>}
                        {selectedConfig.modelParameters.device === 'CPU (rwkv.cpp)' && <Option>Q5_1</Option>}
                      </Dropdown>
                    } />
                }
                {
-                  selectedConfig.modelParameters.device.includes('CUDA') &&
+                  selectedConfig.modelParameters.device.startsWith('CUDA') &&
                  <Labeled label={t('Current Strategy')}
                    content={<Text> {getStrategy(selectedConfig)} </Text>} />
                }
                {
-                  selectedConfig.modelParameters.device.includes('CUDA') &&
+                  selectedConfig.modelParameters.device.startsWith('CUDA') &&
                  <Labeled label={t('Stored Layers')}
                    desc={t('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)')}
                    content={
@@ -320,7 +331,7 @@ const Configs: FC = observer(() => {
                        }} />
                    } />
                }
-                {selectedConfig.modelParameters.device.includes('CUDA') && <div />}
+                {selectedConfig.modelParameters.device.startsWith('CUDA') && <div />}
                {
                  displayStrategyImg &&
                  <img style={{ width: '80vh', height: 'auto', zIndex: 100 }}
@@ -345,7 +356,7 @@ const Configs: FC = observer(() => {
                }
                {selectedConfig.modelParameters.device === 'Custom' && <div />}
                {
-                  (selectedConfig.modelParameters.device.includes('CUDA') || selectedConfig.modelParameters.device === 'Custom') &&
+                  (selectedConfig.modelParameters.device.startsWith('CUDA') || selectedConfig.modelParameters.device === 'Custom') &&
                  <Labeled label={t('Use Custom CUDA kernel to Accelerate')}
                    desc={t('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.')}
                    content={
@@ -394,6 +405,7 @@ const Configs: FC = observer(() => {
              </div>
            }
          />
+          {mq && <div style={{ minHeight: '30px' }} />}
        </div>
        <div className="flex flex-row-reverse sm:fixed bottom-2 right-2">
          <div className="flex gap-2">
--- a/frontend/src/pages/Models.tsx
+++ b/frontend/src/pages/Models.tsx
@@ -153,23 +153,32 @@ const columns: TableColumnDefinition<ModelSourceItem>[] = [
  })
 ];

+const getTags = () => {
+  return Array.from(new Set(
+    ['Recommended',
+      ...commonStore.modelSourceList.map(item => item.tags || []).flat()
+      .filter(i => !i.includes('Other') && !i.includes('Local'))
+      , 'Other', 'Local']));
+};
+
+const getCurrentModelList = () => {
+  if (commonStore.activeModelListTags.length === 0)
+    return commonStore.modelSourceList;
+  else
+    return commonStore.modelSourceList.filter(item => commonStore.activeModelListTags.some(tag => item.tags?.includes(tag)));
+};
+
 const Models: FC = observer(() => {
  const { t } = useTranslation();
-  const [tags, setTags] = useState<Array<string>>([]);
-  const [modelSourceList, setModelSourceList] = useState<ModelSourceItem[]>(commonStore.modelSourceList);
+  const [tags, setTags] = useState<Array<string>>(getTags());
+  const [modelSourceList, setModelSourceList] = useState<ModelSourceItem[]>(getCurrentModelList());

  useEffect(() => {
-    setTags(Array.from(new Set(
-      [...commonStore.modelSourceList.map(item => item.tags || []).flat()
-      .filter(i => !i.includes('Other') && !i.includes('Local'))
-        , 'Other', 'Local'])));
+    setTags(getTags());
  }, [commonStore.modelSourceList]);

  useEffect(() => {
-    if (commonStore.activeModelListTags.length === 0)
-      setModelSourceList(commonStore.modelSourceList);
-    else
-      setModelSourceList(commonStore.modelSourceList.filter(item => commonStore.activeModelListTags.some(tag => item.tags?.includes(tag))));
+    setModelSourceList(getCurrentModelList());
  }, [commonStore.modelSourceList, commonStore.activeModelListTags]);

  return (
--- a/frontend/src/pages/defaultConfigs.ts
+++ b/frontend/src/pages/defaultConfigs.ts
@@ -8,10 +8,10 @@ export const defaultPresets: CompletionPreset[] = [{
  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',
  params: {
    maxResponseToken: 500,
-    temperature: 1.2,
-    topP: 0.5,
-    presencePenalty: 0.4,
-    frequencyPenalty: 0.4,
+    temperature: 1,
+    topP: 0.3,
+    presencePenalty: 0,
+    frequencyPenalty: 1,
    stop: '\\n\\nUser',
    injectStart: '',
    injectEnd: ''
--- a/frontend/src/stores/commonStore.ts
+++ b/frontend/src/stores/commonStore.ts
@@ -96,7 +96,9 @@ class CommonStore {
    useLocalSoundFont: false,
    externalPlay: false,
    midi: null,
-    ns: null
+    ns: null,
+    generationStartTime: 0,
+    playOnlyGeneratedContent: true
  };
  compositionGenerating: boolean = false;
  compositionSubmittedPrompt: string = defaultCompositionPrompt;
--- a/frontend/src/types/composition.ts
+++ b/frontend/src/types/composition.ts
@@ -11,7 +11,9 @@ export type CompositionParams = {
  useLocalSoundFont: boolean,
  externalPlay: boolean,
  midi: ArrayBuffer | null,
-  ns: NoteSequence | null
+  ns: NoteSequence | null,
+  generationStartTime: number,
+  playOnlyGeneratedContent: boolean,
 }
 export type Track = {
  id: string;
--- a/frontend/src/types/configs.ts
+++ b/frontend/src/types/configs.ts
@@ -6,7 +6,7 @@ export type ApiParameters = {
  presencePenalty: number;
  frequencyPenalty: number;
 }
-export type Device = 'CPU' | 'CPU (rwkv.cpp)' | 'CUDA' | 'CUDA-Beta' | 'WebGPU' | 'MPS' | 'Custom';
+export type Device = 'CPU' | 'CPU (rwkv.cpp)' | 'CUDA' | 'CUDA-Beta' | 'WebGPU' | 'WebGPU (Python)' | 'MPS' | 'Custom';
 export type Precision = 'fp16' | 'int8' | 'fp32' | 'nf4' | 'Q5_1';
 export type ModelParameters = {
  // different models can not have the same name
--- a/frontend/src/types/models.ts
+++ b/frontend/src/types/models.ts
@@ -1,15 +1,17 @@
 export type ModelSourceItem = {
  name: string;
-  size: number;
-  lastUpdated: string;
  desc?: { [lang: string]: string | undefined; };
+  size: number;
  SHA256?: string;
+  lastUpdated: string;
  url?: string;
  downloadUrl?: string;
+  tags?: string[];
+  customTokenizer?: string;
+  hide?: boolean;
+
+  lastUpdatedMs?: number;
  isComplete?: boolean;
  isLocal?: boolean;
  localSize?: number;
-  lastUpdatedMs?: number;
-  tags?: string[];
-  hide?: boolean;
 };
--- a/frontend/src/utils/convert-model.ts
+++ b/frontend/src/utils/convert-model.ts
@@ -5,6 +5,7 @@ import {
  ConvertGGML,
  ConvertModel,
  ConvertSafetensors,
+  ConvertSafetensorsWithPython,
  FileExists,
  GetPyError
 } from '../../wailsjs/go/backend_golang/App';
@@ -51,12 +52,22 @@ export const convertModel = async (selectedConfig: ModelConfig, navigate: Naviga
 };


-export const convertToSt = async (selectedConfig: ModelConfig) => {
+export const convertToSt = async (selectedConfig: ModelConfig, navigate: NavigateFunction) => {
+  const webgpuPython = selectedConfig.modelParameters.device === 'WebGPU (Python)';
+  if (webgpuPython) {
+    const ok = await checkDependencies(navigate);
+    if (!ok)
+      return;
+  }
+
  const modelPath = `${commonStore.settings.customModelsPath}/${selectedConfig.modelParameters.modelName}`;
  if (await FileExists(modelPath)) {
    toast(t('Start Converting'), { autoClose: 2000, type: 'info' });
    const newModelPath = modelPath.replace(/\.pth$/, '.st');
-    ConvertSafetensors(modelPath, newModelPath).then(async () => {
+    const convert = webgpuPython ?
+      (input: string, output: string) => ConvertSafetensorsWithPython(commonStore.settings.customPythonPath, input, output)
+      : ConvertSafetensors;
+    convert(modelPath, newModelPath).then(async () => {
      if (!await FileExists(newModelPath)) {
        if (commonStore.platform === 'windows' || commonStore.platform === 'linux')
          toast(t('Convert Failed') + ' - ' + await GetPyError(), { type: 'error' });
--- a/frontend/src/utils/index.tsx
+++ b/frontend/src/utils/index.tsx
@@ -27,6 +27,7 @@ import logo from '../assets/images/logo.png';
 import { Preset } from '../types/presets';
 import { botName, Conversation, MessageType, userName } from '../types/chat';
 import { v4 as uuid } from 'uuid';
+import { findLastIndex } from 'lodash-es';

 export type Cache = {
  version: string
@@ -51,11 +52,11 @@ export async function refreshBuiltInModels(readCache: boolean = false) {
    await ReadJson('cache.json').then((cacheData: Cache) => {
      if (cacheData.models)
        cache.models = cacheData.models;
-      else cache.models = manifest.models;
+      else cache.models = manifest.models.slice();
    }).catch(() => {
-      cache.models = manifest.models;
+      cache.models = manifest.models.slice();
    });
-  else cache.models = manifest.models;
+  else cache.models = manifest.models.slice();

  commonStore.setModelSourceList(cache.models);
  await saveCache().catch(() => {
@@ -90,7 +91,7 @@ export async function refreshLocalModels(cache: {
  for (let i = 0; i < cache.models.length; i++) {
    if (!cache.models[i].lastUpdatedMs)
      cache.models[i].lastUpdatedMs = Date.parse(cache.models[i].lastUpdated);
-    if (!cache.models[i].tags)
+    if (!cache.models[i].tags || !Array.isArray(cache.models[i].tags) || cache.models[i].tags?.length === 0)
      cache.models[i].tags = ['Other'];

    for (let j = i + 1; j < cache.models.length; j++) {
@@ -145,7 +146,7 @@ function initLastUnfinishedModelDownloads() {

 export async function refreshRemoteModels(cache: {
  models: ModelSourceItem[]
-}) {
+}, filter: boolean = true, initUnfinishedModels: boolean = false) {
  const manifestUrls = commonStore.modelSourceManifestList.split(/[,，;；\n]/);
  const requests = manifestUrls.filter(url => url.endsWith('.json')).map(
    url => fetch(url, { cache: 'no-cache' }).then(r => r.json()));
@@ -162,18 +163,16 @@ export async function refreshRemoteModels(cache: {
  });
  cache.models = cache.models.filter((model, index, self) => {
    return modelSuffix.some((ext => model.name.endsWith(ext)))
-      && index === self.findIndex(
-        m => m.name === model.name || (m.SHA256 && m.SHA256 === model.SHA256 && m.size === model.size));
-  });
-  commonStore.setModelSourceList(cache.models);
-  await saveCache().catch(() => {
+      && index === findLastIndex(self,
+        m => m.name === model.name || (!!m.SHA256 && m.SHA256 === model.SHA256 && m.size === model.size));
  });
+  await refreshLocalModels(cache, filter, initUnfinishedModels);
 }

 export const refreshModels = async (readCache: boolean = false, initUnfinishedModels: boolean = false) => {
  const cache = await refreshBuiltInModels(readCache);
  await refreshLocalModels(cache, false, initUnfinishedModels);
-  await refreshRemoteModels(cache);
+  await refreshRemoteModels(cache, false, initUnfinishedModels);
 };

 export const getStrategy = (modelConfig: ModelConfig | undefined = undefined) => {
@@ -192,6 +191,7 @@ export const getStrategy = (modelConfig: ModelConfig | undefined = undefined) =>
      strategy += params.precision === 'int8' ? 'fp32i8' : 'fp32';
      break;
    case 'WebGPU':
+    case 'WebGPU (Python)':
      strategy += params.precision === 'nf4' ? 'fp16i4' : params.precision === 'int8' ? 'fp16i8' : 'fp16';
      break;
    case 'CUDA':
@@ -202,6 +202,8 @@ export const getStrategy = (modelConfig: ModelConfig | undefined = undefined) =>
      strategy += params.precision === 'int8' ? 'fp16i8' : params.precision === 'fp32' ? 'fp32' : 'fp16';
      if (params.storedLayers < params.maxStoredLayers)
        strategy += ` *${params.storedLayers}+`;
+      else
+        strategy += ` -> cuda fp16 *1`;
      break;
    case 'MPS':
      if (avoidOverflow)
@@ -307,7 +309,7 @@ export function getServerRoot(defaultLocalPort: number, isCore: boolean = false)
  const coreCustomApiUrl = commonStore.settings.coreApiUrl.trim().replace(/\/$/, '');
  if (isCore && coreCustomApiUrl)
    return coreCustomApiUrl;
-  
+
  const defaultRoot = `http://127.0.0.1:${defaultLocalPort}`;
  if (commonStore.status.status !== ModelStatus.Offline)
    return defaultRoot;
--- a/frontend/vite.config.ts
+++ b/frontend/vite.config.ts
@@ -12,7 +12,7 @@ const vendor = [
  'mobx', 'mobx-react-lite',
  'i18next', 'react-i18next',
  'usehooks-ts', 'react-toastify',
-  'classnames'
+  'classnames', 'lodash-es'
 ];

 const embedded = [
--- a/frontend/wailsjs/go/backend_golang/App.d.ts
+++ b/frontend/wailsjs/go/backend_golang/App.d.ts
@@ -16,6 +16,8 @@ export function ConvertModel(arg1:string,arg2:string,arg3:string,arg4:string):Pr

 export function ConvertSafetensors(arg1:string,arg2:string):Promise<string>;

+export function ConvertSafetensorsWithPython(arg1:string,arg2:string,arg3:string):Promise<string>;
+
 export function CopyFile(arg1:string,arg2:string):Promise<void>;

 export function DeleteFile(arg1:string):Promise<void>;
@@ -64,7 +66,7 @@ export function SaveJson(arg1:string,arg2:any):Promise<void>;

 export function StartFile(arg1:string):Promise<void>;

-export function StartServer(arg1:string,arg2:number,arg3:string,arg4:boolean,arg5:boolean,arg6:boolean):Promise<string>;
+export function StartServer(arg1:string,arg2:number,arg3:string,arg4:boolean,arg5:boolean,arg6:boolean,arg7:boolean):Promise<string>;

 export function StartWebGPUServer(arg1:number,arg2:string):Promise<string>;

--- a/frontend/wailsjs/go/backend_golang/App.js
+++ b/frontend/wailsjs/go/backend_golang/App.js
@@ -30,6 +30,10 @@ export function ConvertSafetensors(arg1, arg2) {
  return window['go']['backend_golang']['App']['ConvertSafetensors'](arg1, arg2);
 }

+export function ConvertSafetensorsWithPython(arg1, arg2, arg3) {
+  return window['go']['backend_golang']['App']['ConvertSafetensorsWithPython'](arg1, arg2, arg3);
+}
+
 export function CopyFile(arg1, arg2) {
  return window['go']['backend_golang']['App']['CopyFile'](arg1, arg2);
 }
@@ -126,8 +130,8 @@ export function StartFile(arg1) {
  return window['go']['backend_golang']['App']['StartFile'](arg1);
 }

-export function StartServer(arg1, arg2, arg3, arg4, arg5, arg6) {
-  return window['go']['backend_golang']['App']['StartServer'](arg1, arg2, arg3, arg4, arg5, arg6);
+export function StartServer(arg1, arg2, arg3, arg4, arg5, arg6, arg7) {
+  return window['go']['backend_golang']['App']['StartServer'](arg1, arg2, arg3, arg4, arg5, arg6, arg7);
 }

 export function StartWebGPUServer(arg1, arg2) {
--- a/frontend/wailsjs/go/models.ts
+++ b/frontend/wailsjs/go/models.ts
--- a/main.go
+++ b/main.go
@@ -109,7 +109,7 @@ func main() {
 	err = wails.Run(&options.App{
 		Title:                    "RWKV-Runner",
 		Width:                    1024,
-		Height:                   680,
+		Height:                   700,
 		MinWidth:                 375,
 		MinHeight:                640,
 		EnableDefaultContextMenu: true,
--- a/manifest.json
+++ b/manifest.json
@@ -1,12 +1,12 @@
 {
-  "version": "1.6.1",
+  "version": "1.6.5",
  "introduction": {
    "en": "RWKV is an open-source, commercially usable large language model with high flexibility and great potential for development.\n### About This Tool\nThis tool aims to lower the barrier of entry for using large language models, making it accessible to everyone. It provides fully automated dependency and model management. You simply need to click and run, following the instructions, to deploy a local large language model. The tool itself is very compact and only requires a single executable file for one-click deployment.\nAdditionally, this tool offers an interface that is fully compatible with the OpenAI API. This means you can use any ChatGPT client as a client for RWKV, enabling capability expansion beyond just chat functionality.\n### Preset Configuration Rules at the Bottom\nThis tool comes with a series of preset configurations to reduce complexity. The naming rules for each configuration represent the following in order: device - required VRAM/memory - model size - model language.\nFor example, \"GPU-8G-3B-EN\" indicates that this configuration is for a graphics card with 8GB of VRAM, a model size of 3 billion parameters, and it uses an English language model.\nLarger model sizes have higher performance and VRAM requirements. Among configurations with the same model size, those with higher VRAM usage will have faster runtime.\nFor example, if you have 12GB of VRAM but running the \"GPU-12G-7B-EN\" configuration is slow, you can downgrade to \"GPU-8G-3B-EN\" for a significant speed improvement.\n### About RWKV\nRWKV is an RNN with Transformer-level LLM performance, which can also be directly trained like a GPT transformer (parallelizable). And it's 100% attention-free. You only need the hidden state at position t to compute the state at position t+1. You can use the \"GPT\" mode to quickly compute the hidden state for the \"RNN\" mode.<br/>So it's combining the best of RNN and transformer - great performance, fast inference, saves VRAM, fast training, \"infinite\" ctx_len, and free sentence embedding (using the final hidden state).",
    "zh": "RWKV是一个开源且允许商用的大语言模型，灵活性很高且极具发展潜力。\n### 关于本工具\n本工具旨在降低大语言模型的使用门槛，做到人人可用，本工具提供了全自动化的依赖和模型管理，你只需要直接点击运行，跟随引导，即可完成本地大语言模型的部署，工具本身体积极小，只需要一个exe即可完成一键部署。\n此外，本工具提供了与OpenAI API完全兼容的接口，这意味着你可以把任意ChatGPT客户端用作RWKV的客户端，实现能力拓展，而不局限于聊天。\n### 底部的预设配置规则\n本工具内置了一系列预设配置，以降低使用难度，每个配置名的规则，依次代表着：设备-所需显存/内存-模型规模-模型语言。\n例如，GPU-8G-3B-CN，表示该配置用于显卡，需要8G显存，模型规模为30亿参数，使用的是中文模型。\n模型规模越大，性能要求越高，显存要求也越高，而同样模型规模的配置中，显存占用越高的，运行速度越快。\n例如当你有12G显存，但运行GPU-12G-7B-CN配置速度比较慢，可降级成GPU-8G-3B-CN，将会大幅提速。\n### 关于RWKV\nRWKV是具有Transformer级别LLM性能的RNN，也可以像GPT Transformer一样直接进行训练（可并行化）。而且它是100% attention-free的。你只需在位置t处获得隐藏状态即可计算位置t + 1处的状态。你可以使用“GPT”模式快速计算用于“RNN”模式的隐藏状态。\n因此，它将RNN和Transformer的优点结合起来 - 高性能、快速推理、节省显存、快速训练、“无限”上下文长度以及免费的语句嵌入（使用最终隐藏状态）。"
  },
  "about": {
-    "en": "<div align=\"center\">\n\nProject Source Code:\nhttps://github.com/josStorer/RWKV-Runner\nAuthor: [@josStorer](https://github.com/josStorer)\nFAQs: https://github.com/josStorer/RWKV-Runner/wiki/FAQs\n\nRelated Repositories:\nRWKV-4-World: https://huggingface.co/BlinkDL/rwkv-4-world/tree/main\nRWKV-4-Raven: https://huggingface.co/BlinkDL/rwkv-4-raven/tree/main\nChatRWKV: https://github.com/BlinkDL/ChatRWKV\nRWKV-LM: https://github.com/BlinkDL/RWKV-LM\nRWKV-LM-LoRA: https://github.com/Blealtan/RWKV-LM-LoRA\nMIDI-LLM-tokenizer: https://github.com/briansemrau/MIDI-LLM-tokenizer\n\n</div>",
-    "zh": "<div align=\"center\">\n\n本项目源码:\nhttps://github.com/josStorer/RWKV-Runner\n作者: [@josStorer](https://github.com/josStorer)\n演示与常见问题说明视频: https://www.bilibili.com/video/BV1hM4y1v76R\n疑难解答: https://www.bilibili.com/read/cv23921171\n\n相关仓库:\nRWKV-4-World: https://huggingface.co/BlinkDL/rwkv-4-world/tree/main\nRWKV-4-Raven: https://huggingface.co/BlinkDL/rwkv-4-raven/tree/main\nChatRWKV: https://github.com/BlinkDL/ChatRWKV\nRWKV-LM: https://github.com/BlinkDL/RWKV-LM\nRWKV-LM-LoRA: https://github.com/Blealtan/RWKV-LM-LoRA\nMIDI-LLM-tokenizer: https://github.com/briansemrau/MIDI-LLM-tokenizer\n\n</div>"
+    "en": "<div align=\"center\">\n\nProject Source Code and Introduction:\nhttps://github.com/josStorer/RWKV-Runner\nAuthor: [@josStorer](https://github.com/josStorer)\n\nRelated Repositories:\nRWKV-5-World: https://huggingface.co/BlinkDL/rwkv-5-world/tree/main\nRWKV-4-World: https://huggingface.co/BlinkDL/rwkv-4-world/tree/main\nRWKV-4-Raven: https://huggingface.co/BlinkDL/rwkv-4-raven/tree/main\nChatRWKV: https://github.com/BlinkDL/ChatRWKV\nRWKV-LM: https://github.com/BlinkDL/RWKV-LM\nRWKV-LM-LoRA: https://github.com/Blealtan/RWKV-LM-LoRA\nMIDI-LLM-tokenizer: https://github.com/briansemrau/MIDI-LLM-tokenizer\nai00_rwkv_server: https://github.com/cgisky1980/ai00_rwkv_server\nrwkv.cpp: https://github.com/saharNooby/rwkv.cpp\nweb-rwkv-py: https://github.com/cryscan/web-rwkv-py\n\n</div>",
+    "zh": "<div align=\"center\">\n\n本项目源码及介绍页:\nhttps://github.com/josStorer/RWKV-Runner\n作者: [@josStorer](https://github.com/josStorer)\n演示与常见问题说明视频: https://www.bilibili.com/video/BV1hM4y1v76R\n\n相关仓库:\nRWKV-5-World: https://huggingface.co/BlinkDL/rwkv-5-world/tree/main\nRWKV-4-World: https://huggingface.co/BlinkDL/rwkv-4-world/tree/main\nRWKV-4-Raven: https://huggingface.co/BlinkDL/rwkv-4-raven/tree/main\nChatRWKV: https://github.com/BlinkDL/ChatRWKV\nRWKV-LM: https://github.com/BlinkDL/RWKV-LM\nRWKV-LM-LoRA: https://github.com/Blealtan/RWKV-LM-LoRA\nMIDI-LLM-tokenizer: https://github.com/briansemrau/MIDI-LLM-tokenizer\nai00_rwkv_server: https://github.com/cgisky1980/ai00_rwkv_server\nrwkv.cpp: https://github.com/saharNooby/rwkv.cpp\nweb-rwkv-py: https://github.com/cryscan/web-rwkv-py\n\n</div>"
  },
  "programFiles": [
    {
@@ -25,8 +25,8 @@
      "size": 385598386,
      "SHA256": "c844a3ee05bcb9065848cb05b10c48a3f381f5ac1953aad89e156ecdf31d7703",
      "lastUpdated": "2023-08-03T15:18:46",
-      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-0.1B-v1-20230803-ctx4096.pth?download=true",
-      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-0.1B-v1-20230803-ctx4096.pth?download=true",
+      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-0.1B-v1-20230803-ctx4096.pth",
+      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-0.1B-v1-20230803-ctx4096.pth",
      "tags": [
        "Main",
        "RWKV-5",
@@ -43,8 +43,8 @@
      "size": 923523954,
      "SHA256": "5a288c54c7f30b0e2d4af23991133fad2af2d5e59ec7ad850ffe78054a5e4f92",
      "lastUpdated": "2023-11-14T01:23:49",
-      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-0.4B-v2-20231113-ctx4096.pth?download=true",
-      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-0.4B-v2-20231113-ctx4096.pth?download=true",
+      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-0.4B-v2-20231113-ctx4096.pth",
+      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-0.4B-v2-20231113-ctx4096.pth",
      "tags": [
        "Main",
        "RWKV-5",
@@ -69,6 +69,45 @@
        "Global"
      ]
    },
+    {
+      "name": "RWKV-5-1B5-one-state-slim.pth",
+      "desc": {
+        "en": "RWKV-5 Global Languages 1.5B v2 Ctx16k Role Play",
+        "zh": "RWKV-5 全球语言 1.5B v2 16k上下文 角色扮演",
+        "ja": "RWKV-5 グローバル言語 1.5B v2 16kコンテキスト ロールプレイ"
+      },
+      "size": 3155589871,
+      "SHA256": "43e7b922d7ad49eafa17f8909c2813c91394925bc7f24caf0e19a91aa3281273",
+      "lastUpdated": "2023-11-02T04:03:27",
+      "url": "https://huggingface.co/xiaol/RWKV-v5-world-v2-1.5B-one-state-slim-16k/blob/main/RWKV-5-1B5-one-state-slim.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-v5-world-v2-1.5B-one-state-slim-16k/resolve/main/RWKV-5-1B5-one-state-slim.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-5",
+        "Global",
+        "Role Play"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
+    {
+      "name": "RWKV-5-1B5-one-state-slim-novel-tuned.pth",
+      "desc": {
+        "en": "RWKV-5 Global Languages 1.5B v2 Ctx16k Novel",
+        "zh": "RWKV-5 全球语言 1.5B v2 16k上下文 小说",
+        "ja": "RWKV-5 グローバル言語 1.5B v2 16kコンテキスト 小説"
+      },
+      "size": 3155589871,
+      "SHA256": "4f0aaecdce676e5236018ebd63e3d37c2f300fbac04001ee3a9c00d2f4244d0f",
+      "lastUpdated": "2023-11-03T02:45:52",
+      "url": "https://huggingface.co/xiaol/RWKV-v5-world-v2-1.5B-one-state-slim-16k-novel-tuned/blob/main/RWKV-5-1B5-one-state-slim-novel-tuned.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-v5-world-v2-1.5B-one-state-slim-16k-novel-tuned/resolve/main/RWKV-5-1B5-one-state-slim-novel-tuned.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-5",
+        "Global"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
    {
      "name": "RWKV-5-World-3B-v2-20231113-ctx4096.pth",
      "desc": {
@@ -79,12 +118,13 @@
      "size": 6126106674,
      "SHA256": "a4bd430343c6fd138b85bbc68bb20262d3a2f053ea57dc4b41078269af68ff9c",
      "lastUpdated": "2023-11-14T01:23:49",
-      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-3B-v2-20231113-ctx4096.pth?download=true",
-      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-3B-v2-20231113-ctx4096.pth?download=true",
+      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-3B-v2-20231113-ctx4096.pth",
+      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-3B-v2-20231113-ctx4096.pth",
      "tags": [
        "Main",
        "RWKV-5",
-        "Global"
+        "Global",
+        "Recommended"
      ]
    },
    {
@@ -97,14 +137,94 @@
      "size": 6126106467,
      "SHA256": "efa5178d1c824b94ef17c6c9a456674e5581a8be832becbda9aba4dc533f88c2",
      "lastUpdated": "2023-11-19T04:21:04",
-      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-3B-v2-20231118-ctx16k.pth?download=true",
-      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-3B-v2-20231118-ctx16k.pth?download=true",
+      "url": "https://huggingface.co/BlinkDL/rwkv-5-world/blob/main/RWKV-5-World-3B-v2-20231118-ctx16k.pth",
+      "downloadUrl": "https://huggingface.co/BlinkDL/rwkv-5-world/resolve/main/RWKV-5-World-3B-v2-20231118-ctx16k.pth",
      "tags": [
        "Main",
        "RWKV-5",
-        "Global"
+        "Global",
+        "Recommended"
      ]
    },
+    {
+      "name": "rwkv-v5-7B-0.4-long-ctx-16k.pth",
+      "desc": {
+        "en": "RWKV-5 Global Languages 7B v2 40% Ctx300k Document Reader",
+        "zh": "RWKV-5 全球语言 7B v2 40% 300k上下文 文档阅读",
+        "ja": "RWKV-5 グローバル言語 7B v2 40% 300kコンテキスト ドキュメントリーダー"
+      },
+      "size": 15036198115,
+      "SHA256": "5888471a45caab903c1bd9c35af1c639ac8d03be6ee6eb39fa9fd3194fa6d437",
+      "lastUpdated": "2023-11-10T17:12:04",
+      "url": "https://huggingface.co/xiaol/RWKV-5-world-v2-7B-0.4-300k/blob/main/rwkv-v5-7B-0.4-long-ctx-16k.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-5-world-v2-7B-0.4-300k/resolve/main/rwkv-v5-7B-0.4-long-ctx-16k.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-5",
+        "Global"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
+    {
+      "name": "rwkv-v5.2-7B-horror-16k.pth",
+      "desc": {
+        "en": "RWKV-5 Global Languages 7B v2 40% Ctx16k Horror",
+        "zh": "RWKV-5 全球语言 7B v2 40% 16k上下文 恐怖",
+        "ja": "RWKV-5 グローバル言語 7B v2 40% 16kコンテキスト ホラー"
+      },
+      "size": 15036198115,
+      "SHA256": "3b36ce99bef06627dcb5d860972e2c1515327afe7db415b8c82dd5c3b926b52f",
+      "lastUpdated": "2023-11-13T15:21:25",
+      "url": "https://huggingface.co/xiaol/RWKV-v5.2-7B-horror-16k/blob/main/rwkv-v5.2-7B-horror-16k.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-v5.2-7B-horror-16k/resolve/main/rwkv-v5.2-7B-horror-16k.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-5",
+        "Global"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
+    {
+      "name": "rwkv_v5.2_7B_role_play_16k.pth",
+      "desc": {
+        "en": "RWKV-5 Global Languages 7B v2 Ctx16k Claude Like",
+        "zh": "RWKV-5 全球语言 7B v2 16k上下文 Claude功能",
+        "ja": "RWKV-5 グローバル言語 7B v2 16kコンテキスト Claude機能"
+      },
+      "size": 15036198115,
+      "SHA256": "6fe8a7bf06b9f5e5b740cd87e24bff91325518ad19bf92bf5c75799b3c24b150",
+      "lastUpdated": "2023-11-14T04:18:16",
+      "url": "https://huggingface.co/xiaol/RWKV-v5.2-7B-Role-play-16k/blob/main/rwkv_v5.2_7B_role_play_16k.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-v5.2-7B-Role-play-16k/resolve/main/rwkv_v5.2_7B_role_play_16k.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-5",
+        "Global",
+        "Role Play",
+        "Recommended"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
+    {
+      "name": "RWKV-5-12B-one-state-chat-16k.pth",
+      "desc": {
+        "en": "RWKV-5 Global Languages 12B Ctx16k",
+        "zh": "RWKV-5 全球语言 12B 16k上下文",
+        "ja": "RWKV-5 グローバル言語 12B 16kコンテキスト"
+      },
+      "size": 23157296483,
+      "SHA256": "330be74738d3936f4c9bd6caf838db11c96f52ff360d0f4fa5401d9bafc898ab",
+      "lastUpdated": "2023-12-16T16:34:30",
+      "url": "https://huggingface.co/xiaol/RWKV-v5-12B-one-state-chat-16k/blob/main/RWKV-5-12B-one-state-chat-16k.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-v5-12B-one-state-chat-16k/resolve/main/RWKV-5-12B-one-state-chat-16k.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-5",
+        "Global",
+        "Recommended"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
    {
      "name": "RWKV-4-World-CHNtuned-0.1B-v1-20230617-ctx4096.pth",
      "desc": {
@@ -257,6 +377,25 @@
        "Global"
      ]
    },
+    {
+      "name": "RWKV-for-mobile-4-world-1.5B-20230906-ctx16k.pth",
+      "desc": {
+        "en": "Global Languages 1.5B v1 Ctx16k Claude Like",
+        "zh": "全球语言 1.5B v1 16k上下文 Claude功能",
+        "ja": "グローバル言語 1.5B v1 16kコンテキスト Claude機能"
+      },
+      "size": 3155280301,
+      "SHA256": "20547a6deca32add57c45d2f6cff52c6b59cd3b92676ee369b964affba35619d",
+      "lastUpdated": "2023-09-07T01:35:46",
+      "url": "https://huggingface.co/xiaol/RWKV-claude-for-mobile-v4-world-1.5B-16k/blob/main/RWKV-for-mobile-4-world-1.5B-20230906-ctx16k.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-claude-for-mobile-v4-world-1.5B-16k/resolve/main/RWKV-for-mobile-4-world-1.5B-20230906-ctx16k.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-4",
+        "Global"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
    {
      "name": "RWKV-4-World-3B-v1-OnlyForTest_35%_trained-20230529-ctx4096.pth",
      "desc": {
@@ -513,6 +652,83 @@
        "Global"
      ]
    },
+    {
+      "name": "RWKV-7B-world-one.pth",
+      "desc": {
+        "en": "Global Languages 7B v1 Ctx65k Novel",
+        "zh": "全球语言 7B v1 65k上下文 小说",
+        "ja": "グローバル言語 7B v1 65kコンテキスト 小説"
+      },
+      "size": 15035391533,
+      "SHA256": "7ce95a4b460c3385c75c29b6ebe3cd7db438b1107e85d7d3e42dff85cfaa0b78",
+      "lastUpdated": "2023-10-09T05:23:38",
+      "url": "https://huggingface.co/xiaol/RWKV-v4-world-7B-one-state-65k/blob/main/RWKV-7B-world-one.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-v4-world-7B-one-state-65k/resolve/main/RWKV-7B-world-one.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-4",
+        "Global"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
+    {
+      "name": "rwkv-world-one-novel-cot-ultrachat-novel-instructions.pth",
+      "desc": {
+        "en": "Global Languages 7B v1 Ctx65k Novel Instruction",
+        "zh": "全球语言 7B v1 65k上下文 小说指令",
+        "ja": "グローバル言語 7B v1 65kコンテキスト 小説指示"
+      },
+      "size": 15035391533,
+      "SHA256": "fc2d4643828bb9dfe0733c3b2eb54ba2d996ed3eb6afa051b558da2eb2c1e309",
+      "lastUpdated": "2023-10-22T09:50:39",
+      "url": "https://huggingface.co/xiaol/RWKV-4-world-one-state-ultrachat-COT-65k/blob/main/rwkv-world-one-novel-cot-ultrachat-novel-instructions.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-4-world-one-state-ultrachat-COT-65k/resolve/main/rwkv-world-one-novel-cot-ultrachat-novel-instructions.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-4",
+        "Global"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
+    {
+      "name": "RWKV-world-novel-one-state-ultrachat-cot-tuned-Role-play-65k.pth",
+      "desc": {
+        "en": "Global Languages 7B v1 Ctx65k Role Play",
+        "zh": "全球语言 7B v1 65k上下文 角色扮演",
+        "ja": "グローバル言語 7B v1 65kコンテキスト ロールプレイ"
+      },
+      "size": 15035391533,
+      "SHA256": "2f55b4710dcd360e83b4df9a6358661284d9a6c6108f62c5a30b86df181ed67a",
+      "lastUpdated": "2023-10-22T05:54:27",
+      "url": "https://huggingface.co/xiaol/RWKV-4-world-one-state-ultrachat-COT-65k/blob/main/RWKV-world-novel-one-state-ultrachat-cot-tuned-Role-play-65k.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-4-world-one-state-ultrachat-COT-65k/resolve/main/RWKV-world-novel-one-state-ultrachat-cot-tuned-Role-play-65k.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-4",
+        "Global",
+        "Role Play"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
+    {
+      "name": "RWKV-4-7B-world-one-novel-tuned-65k.pth",
+      "desc": {
+        "en": "Global Languages 7B v1 Ctx65k Chinese Novel Instruction",
+        "zh": "全球语言 7B v1 65k上下文 中文小说指令",
+        "ja": "グローバル言語 7B v1 65kコンテキスト 中国語小説指示"
+      },
+      "size": 15035391533,
+      "SHA256": "e8ff256d74ca404621dcbf87c43c37e25ea745fed30c404fbf45cc5acc7ba2b5",
+      "lastUpdated": "2023-10-15T00:57:53",
+      "url": "https://huggingface.co/xiaol/RWKV-4-world-one-state-novel-tuned-65k/blob/main/RWKV-4-7B-world-one-novel-tuned-65k.pth",
+      "downloadUrl": "https://huggingface.co/xiaol/RWKV-4-world-one-state-novel-tuned-65k/resolve/main/RWKV-4-7B-world-one-novel-tuned-65k.pth",
+      "tags": [
+        "Finetuned",
+        "RWKV-4",
+        "CN"
+      ],
+      "customTokenizer": "backend-python/rwkv_pip/rwkv_vocab_v20230424_special_token.txt"
+    },
    {
      "name": "RWKV-4-World-CHNtuned-7B-v1-20230709-ctx4096.pth",
      "desc": {
Author	SHA1	Message	Date
josc146	8ca920a114	release v1.6.6	2023-12-25 21:02:26 +08:00
josc146	5f3d449a66	improve Models page	2023-12-25 20:37:40 +08:00
josc146	13735e7dfb	chore	2023-12-25 20:35:00 +08:00
josc146	a38d5c3a25	enable web-rwkv-py turbo	2023-12-25 20:34:35 +08:00
josc146	5bae637c67	update Related Repositories	2023-12-25 20:32:54 +08:00
josc146	12e488ba80	improve strategy	2023-12-25 19:30:57 +08:00
josc146	ad30c63c69	update Writer preset params	2023-12-25 19:30:14 +08:00
josc146	a116eff7df	webgpu max_buffer_size	2023-12-25 18:08:13 +08:00
josc146	01bc355dde	allow manifest customTokenizer	2023-12-25 16:57:32 +08:00
josc146	8e05f3c360	chore	2023-12-25 16:56:46 +08:00
josc146	fde988dd4e	update manifest.json	2023-12-25 16:08:20 +08:00
josc146	91401ad14f	* text=auto eol=lf	2023-12-24 22:51:23 +08:00
josc146	280194647c	improve refreshRemoteModels	2023-12-22 14:44:27 +08:00
josc146	2e0a542f33	improve train_log.txt creation	2023-12-22 13:00:13 +08:00
josc146	b988694da7	better CopyEmbed	2023-12-22 12:47:26 +08:00
josc146	512c4d0f73	improve role-playing effect	2023-12-22 10:51:09 +08:00
josc146	5525fb1470	chore	2023-12-22 10:49:28 +08:00
josc146	4db735e026	update readme	2023-12-21 13:46:51 +08:00
josc146	c8c79c39d1	Create dependabot.yml	2023-12-21 12:56:21 +08:00
josc146	bcfb76d8ca	update readme	2023-12-19 14:59:02 +08:00
josc146	2d9aaf8fc9	update readme	2023-12-18 19:55:25 +08:00
josc146	8a3905c09a	reduce precompiled web_rwkv_py size	2023-12-15 16:26:01 +08:00
github-actions[bot]	54cd8a46fa	release v1.6.5	2023-12-14 14:09:13 +00:00
josc146	1b83bf261a	release v1.6.5	2023-12-14 22:07:17 +08:00
josc146	2a7d22dab1	Composition Option: Only Auto Play Generated Content	2023-12-14 22:06:39 +08:00
josc146	f7494b0cfb	update midi_filter_config.json	2023-12-14 21:18:48 +08:00
github-actions[bot]	9ca91d59ec	release v1.6.4	2023-12-14 12:40:56 +00:00
josc146	11feaa6e68	release v1.6.4	2023-12-14 20:40:24 +08:00
josc146	18d4b2304e	WebGPU (Python) strategy	2023-12-14 20:39:42 +08:00
github-actions[bot]	2f45e9c33a	release v1.6.3	2023-12-14 10:43:36 +00:00
josc146	f7df10cb66	release v1.6.3	2023-12-14 18:42:58 +08:00
josc146	46e9a2f5b2	add precompiled web_rwkv_py	2023-12-14 18:42:00 +08:00
josc146	69b8d2e0a1	fix refreshBuiltInModels	2023-12-14 18:37:37 +08:00
josc146	0ddd2e9fea	add WebGPU Python Mode (https://github.com/cryscan/web-rwkv-py )	2023-12-14 18:37:07 +08:00
josc146	01c95f5bc4	chore	2023-12-14 14:13:12 +08:00
josc146	e0bf44d82f	bump MIDI-LLM-tokenizer (fix note off)	2023-12-14 13:33:27 +08:00
josc146	f328e84ea7	update Readme_Install.txt	2023-12-13 15:23:34 +08:00
github-actions[bot]	c81f5015a1	release v1.6.2	2023-12-12 15:51:23 +00:00