0.0.25

更新：

1. 完整模式（mode 2） — 改为从本地 openvino-genai-thinking/dist/*.whl 安装，不再需要 cmake 编译
2. --fix 升级路径 — 检测到 whl 后询问是否升级到完整模式，同样直接装 whl
3. 保留回退 — whl 不存在时仍走源码编译
4. 新增 openvino-genai-thinking 目录 — 包含预编译 whl、pyproject.toml、README 等打包文件

Updates:

1. Full mode (mode 2) — Now installs from local openvino-genai-thinking/dist/*.whl instead of requiring cmake compilation
2. --fix upgrade path — Detects the whl and prompts whether to upgrade to full mode; also installs directly from the whl
3. Fallback retained — Falls back to source compilation when the whl is not found
4. New openvino-genai-thinking directory — Contains prebuilt whl, pyproject.toml, README, and other packaging files

0.0.24

优化chat.py：

1. 修复 </think> 过滤 — thinking_filter 初始化时 in_think 不再默认 True，改为 False，避免孤立的 </think> 漏出
2. 多行输入支持 — Enter 换行，空行提交
   • 当前行有内容 → Enter 插入 \n 继续输入
   • 当前行空白 → Enter 提交消息
   • Backspace 删除 \n 时光标正确上移并重绘
   • _total_width() 只计算当前行宽度，不受 \n 影响
   • _char_width('\n') 返回 0

chat.py optimizations:

1. Fixed </think> filtering — thinking_filter no longer initializes in_think as True by default; changed to False to prevent stray </think> tags from leaking through

2. Multiline input support — Enter inserts a newline; press Enter on an empty line to submit

   • Current line has content → Enter inserts \n to continue typing
   • Current line is empty → Enter submits the message
   • Backspace on \n correctly moves cursor up and redraws
   • _total_width() only calculates the current line's width, unaffected by \n
   • _char_width('\n') returns 0

0.0.23

修复

1. cli.py — 删除冗余 venv 命令 + 修复 bug

• 删除了 venv 子命令（可被 source .venv/bin/activate 替代）
• 修复 image 命令缺少 --seed 参数导致的 AttributeError

2. features.py — 按功能拆分依赖

• 基础依赖不再全装，改为按 --with 选择的功能安装
• wcwidth/PyMuPDF/soxr → 仅 chat 安装
• fastapi/uvicorn → 仅 server 安装
• 添加 sox 到 TTS 依赖（之前运行时缺失）

3. tts.py — 新增 transformers 版本切换 + 恢复

• _ensure_qwen_tts_tf() — TTS 推理前将 transformers 从 5.x 临时降级到 4.57.3
• _restore_tf() — 推理完成后恢复 transformers + huggingface_hub 到最新版
• run_pipe() 和 run_once() 均包含恢复逻辑

4. convert.py — 强化 torch CPU 修复

• _ensure_qwen_tts() 和 _ensure_qwen_asr() 装完 qwen 包后，不仅修复 torchaudio，还新增 torch 的 CPU 强制重装，防止拉入 nvidia CUDA 包

5. setup.py — 新增 apt 系统依赖检测

• _check_apt_deps(features) — 检测 sox、libsndfile1 等系统包是否安装，缺失时提示 sudo apt install
• _install_features() 中装完 qwen 包后增加 torch CPU 重装

fix

1. cli.py — Removed redundant venv command + bug fix

• Removed the venv subcommand (can be replaced by source .venv/bin/activate)
• Fixed AttributeError caused by missing --seed argument in the image command

2. features.py — Split dependencies by feature

• Base dependencies are no longer installed all at once; installed on demand based on --with
• wcwidth/PyMuPDF/soxr → chat only
• fastapi/uvicorn → server only
• Added sox to TTS dependencies (was missing at runtime)

3. tts.py — Added transformers version switching + restoration

• _ensure_qwen_tts_tf() — Temporarily downgrades transformers from 5.x to 4.57.3 before TTS inference
• _restore_tf() — Restores transformers + huggingface_hub to the latest version after inference
• Both run_pipe() and run_once() include restoration logic

4. convert.py — Strengthened torch CPU fix

• _ensure_qwen_tts() and _ensure_qwen_asr() now not only fix torchaudio after installing qwen packages, but also force-reinstall torch from the CPU index to prevent nvidia CUDA packages from being pulled in

5. setup.py — Added apt system dependency detection

• _check_apt_deps(features) — Detects whether system packages like sox and libsndfile1 are installed; prompts sudo apt install if missing
• Added torch CPU reinstall in _install_features() after installing qwen packages

0.0.22

新增 --with 按需安装

以前 setup 只能全装，现在可以选：

./ov-cli setup --with chat,image           # 只装聊天 + 文生图（2分钟）
./ov-cli setup --with all                  # 全装（10分钟）
./ov-cli setup --fix                       # 只修复已装的模块

8 个可选模块： chat, image, asr, tts, ui, mcp, server, convert

功能门控

没装的模块会友好提示：

$ ov-cli ui
⚠ ui 模块未安装，请运行: ./ov-cli setup --with ui

New: --with On-demand Installation

Previously setup could only install everything. Now you can pick:

./ov-cli setup --with chat,image           # Chat + Image only (~2 min)
./ov-cli setup --with all                  # Everything (~10 min)
./ov-cli setup --fix                       # Fix installed modules only

8 available modules: chat, image, asr, tts, ui, mcp, server, convert

Feature Gating

Uninstalled modules show a friendly hint instead of crashing:

$ ov-cli ui
⚠ ui module not installed. Run: ./ov-cli setup --with ui

0.0.21

管道模式（--mode pipe）

四个命令新增 --mode pipe 管道模式：

命令	管道输入	stdout 输出
asr	音频文件路径	{"text":"...","time":n}
tts	合成文本	{"path":"...","text":"...","time":n}
chat	提示词	{"text":"...","time":n}
image	描述	{"path":"...","time":n}

模型常驻内存，每次推理 0.5~10s，无需重载。外部程序通过 stdin/stdout 即可高效调用。

---

Pipe Mode (--mode pipe)

Four commands now support --mode pipe (pipeline mode):

Command	Pipe Input	stdout Output
asr	audio file path	{"text":"...","time":n}
tts	text to synthesize	{"path":"...","text":"...","time":n}
chat	prompt	{"text":"...","time":n}
image	description	{"path":"...","time":n}

Model stays in memory, each inference takes 0.5~10s without reloading. External programs can efficiently call via stdin/stdout.

0.0.20

新增

mcp — MCP 协议服务器

启动 MCP (Model Context Protocol) 服务器，通过 stdin/stdout JSON-RPC 暴露 LLM 工具。
可被 VS Code Copilot (agent 模式)、Cursor、Claude Desktop 等支持 MCP 的 AI 编程工具调用。

./ov-cli mcp --model ./Qwen3/2B-ov
./ov-cli mcp --model ./deepseek/7B-ov

暴露的工具：

工具	说明
chat	向本地 LLM 发送提示并获取回复
chat_stream	流式聊天，逐块返回文本

VS Code 配置（.vscode/mcp.json）例如（将路径替换为你的实际路径）：

{
  "servers": {
    "ov-cli": {
      "command": "/run/media/amigor/Project/ov-cli/.venv/bin/ov-cli",
      "args": ["mcp", "--model", "/run/media/amigor/Project/ov-cli/model/deepseek/7B-ov"],
      "type": "stdio",
      "description": "本地 LLM 推理（聊天、翻译、问答）"
    }
  }
}

其他平台（Cursor → .cursor/mcp.json，Claude Desktop → claude_desktop_config.json），格式基本一致。

New

mcp — MCP Protocol Server

MCP (Model Context Protocol) server that exposes LLM tools via stdin/stdout JSON-RPC.
Can be used by VS Code Copilot (agent mode), Cursor, Claude Desktop, and other MCP-compatible AI tools.

./ov-cli mcp --model ./Qwen3/2B-ov
./ov-cli mcp --model ./deepseek/7B-ov

Exposed tools:

Tool	Description
chat	Send a prompt to the local LLM and get a response
chat_stream	Streaming chat, returns text chunks

VS Code config (.vscode/mcp.json) for example (replace paths with yours):

{
  "servers": {
    "ov-cli": {
      "command": "/run/media/amigor/Project/ov-cli/.venv/bin/ov-cli",
      "args": ["mcp", "--model", "/run/media/amigor/Project/ov-cli/model/deepseek/7B-ov"],
      "type": "stdio",
      "description": "Local LLM inference (chat, translation, Q&A)"
    }
  }
}

Other platforms (Cursor → .cursor/mcp.json, Claude Desktop → claude_desktop_config.json), format is similar.

0.0.19

代码审查修复

• 删除 asr.py 重复行
• asr.py transformers 无条件恢复（同 ui.py 模式）
• 删除 benchmark.py 未使用的 _run_genai_bench()
• image.py CLI 传入 --seed 参数
• convert.py 统一用 pip list 查版本（避免 import 副作用）
• pyproject.toml 补充 gradio/soundfile/scipy 依赖
• setup.py 补充 gradio 安装
• 简化版本号跟踪，每次运行提示

Code Review Fixes

• Removed duplicate line in asr.py
• asr.py: unconditional transformers restore (same pattern as ui.py)
• Removed unused _run_genai_bench() in benchmark.py
• image.py: pass --seed from CLI to run_once()
• convert.py: unified version checking via pip list (avoid import side effects)
• pyproject.toml: added gradio, soundfile, scipy dependencies
• setup.py: added gradio to install list
• Simplified version tracking; show --fix hint on every run

0.0.18

修复：

1. FLUX 模型识别为 "image" — _detect_model_type() 之前找不到 config.json 就直接返回 "chat"，导致加载 LLM 时崩溃。新增检测 model_index.json（Diffusers 格式，_class_name: FluxPipeline）。

2. ASR 语言下拉返回错误 — 选项是 "zh"、"en" 等缩写代码，但 qwen_asr 要求全称（"Chinese"、"English"）。已改用全称列表。

3. TTS/ASR 退出后 transformers 没恢复 — Ctrl+C 退出后 _restore_tf() 没被调用，导致后续运行其他模型时 transformers 仍然是 4.x。修复：在 finally 块中无条件恢复。

4. Ctrl+C 退出有杂乱堆栈 — Python 的 _thread_shutdown() 在解释器退出时再次抛出 KeyboardInterrupt。修复：捕获第一次中断后调用 signal.signal(SIGINT, SIG_IGN) 忽略后续信号。

Fixes:

1. FLUX model detected as "image" — _detect_model_type() previously returned "chat" immediately when config.json was missing, causing a crash when trying to load an LLM. Added detection for model_index.json (Diffusers format, _class_name: FluxPipeline).

2. ASR language dropdown returned errors — Options were abbreviated codes like "zh", "en", but qwen_asr expects full names ("Chinese", "English"). Changed to full name list.

3. Transformers not restored after TTS/ASR exit — _restore_tf() was never called after Ctrl+C, leaving transformers at 4.x for subsequent runs with other models. Fixed: unconditional restore in the finally block.

4. Ctrl+C exit left messy traceback — Python's _thread_shutdown() raises another KeyboardInterrupt during interpreter exit. Fixed: call signal.signal(SIGINT, SIG_IGN) after catching the first interrupt to ignore subsequent signals.

0.0.17

新增 ui.py — Gradio Web UI，支持聊天/TTS/ASR/文生图四种界面

启动 Gradio Web UI，自动检测模型类型（chat/tts/asr/image），提供对应的可视化界面。

# 聊天界面
./ov-cli ui --model ./Qwen3/2B-ov
./ov-cli ui --model ./Qwen3/8B-ov --port 7861                      # 指定端口
./ov-cli ui --model ./model-vlm-ov --share                          # 公开链接
./ov-cli ui --model ./deepseek/7B-ov --reasoning off                # 关闭思考

# TTS / ASR / 文生图界面
./ov-cli ui --model ./0.6B-CV-ov                                    # TTS
./ov-cli ui --model ./Qwen3-ASR-0.6B-ov                             # ASR
./ov-cli ui --model ./FLUX/ov-int4                                  # 文生图

聊天界面功能：

• 流式输出、多轮对话
• 图片上传（VLM 模型）
• 对话历史保存 / 加载 / 删除
• Ctrl+C 安全退出

Bug 修复（chat.py）

_is_genai_format() 现在也会检测 openvino_model.xml，支持 ModelScope 预转换模型

setup 改进（setup.py）

安装模式跟踪，--fix 模式下保留已编译的 GenAI

文档更新（README.md/README_EN.md）

同步新命令说明

New: ui.py — Gradio Web UI (Chat/TTS/ASR/Text-to-Image)

Launches a Gradio Web UI that auto-detects the model type (chat/tts/asr/image) and provides a visual interface.

# Chat UI
./ov-cli ui --model ./Qwen3/2B-ov
./ov-cli ui --model ./Qwen3/8B-ov --port 7861                      # Custom port
./ov-cli ui --model ./model-vlm-ov --share                          # Public link
./ov-cli ui --model ./deepseek/7B-ov --reasoning off                # Disable thinking

# TTS / ASR / Text-to-Image UI
./ov-cli ui --model ./0.6B-CV-ov                                    # TTS
./ov-cli ui --model ./Qwen3-ASR-0.6B-ov                             # ASR
./ov-cli ui --model ./FLUX/ov-int4                                  # Text-to-Image

Chat UI features:

• Streaming output, multi-turn conversation
• Image upload (VLM models)
• Chat history save / load / delete
• Ctrl+C safe exit

Bug fix (chat.py)

_is_genai_format() now also checks for openvino_model.xml, supporting ModelScope pre-converted models

Setup improvements (setup.py)

Installation mode tracking; --fix mode preserves compiled GenAI

Documentation (README.md/README_EN.md)

Updated with new command descriptions

0.0.16

🔄 破坏性变更

旧命令	新命令	旧文件	新文件
ov-cli generate	ov-cli image	generate.py	image.py
—	ov-cli tts 🆕	—	tts.py
ov-cli whisper	ov-cli asr	whisper.py	asr.py

✨ 新功能

Qwen3-ASR 语音识别 (asr + convert)

• asr 命令自动识别 Whisper / Qwen3-ASR
• 推理时自动切换 transformers 版本，完成后恢复
• 支持交互式多轮和 --mode once
• Qwen3-ASR 基于语义自动加标点（优于 Whisper）
• convert 支持 Qwen3-ASR 模型转换

📝 文档

• README 全面更新命令名，新增 tts 章节
• ASR 章节推荐 Qwen3-ASR，附双后端对比

🔄 Breaking Changes

Old Command	New Command	Old File	New File
ov-cli generate	ov-cli image	generate.py	image.py
—	ov-cli tts 🆕	—	tts.py
ov-cli whisper	ov-cli asr	whisper.py	asr.py

✨ New Features

Qwen3-ASR Speech Recognition (asr + convert)

• asr auto-detects Whisper / Qwen3-ASR
• Auto-switches transformers version for inference, restores afterward
• Supports interactive multi-turn and --mode once
• Qwen3-ASR adds punctuation based on semantic understanding (beats Whisper on TTS audio)
• convert supports Qwen3-ASR model conversion

📝 Documentation

• README fully updated with new command names, new tts section
• ASR section recommends Qwen3-ASR with dual-backend comparison