1.4.4

完善1.4.3

Improved 1.4.3

1.4.3

fix

• 多GPU系统（GPU.0/GPU.1）设备选择错误

Fixes

• Fixed device selection failing on multi-GPU systems (GPU.0/GPU.1)

1.4.2

变更

• 固定 openvino==2026.2.0 和 openvino-genai==2026.2.0.0，避免版本不匹配导致 libopenvino.so 找不到

Changes

• Pinned openvino==2026.2.0 and openvino-genai==2026.2.0.0 to prevent version mismatch causing libopenvino.so not found

1.4.1

修复：

翻译模式缺少 reasoning 变量导致 NameError

Fix:

The lack of reasoning variable in translation mode leads to NameError

1.4.0

🚀 新功能

• 移除了 convert 模块 — 不再维护 torch/optimum-intel/transformers 的版本兼容性问题，转换请直接用 optimum-cli
• 移除了 Optimum 格式支持 — Gemma-4 系列现在直接用 VLMPipeline 加载，性能提升 16%（E2B: 16.4→19.1 tok/s），不再需要 optimum-intel

📖 文档

• Xe 驱动 benchmark 补全 — 新增 Qwen3/14B、Gemma-4 31B 数据
• README 大幅清理，移除了 convert 相关文档
• 更新了 benchmark 表格

🚀 New Features

• Removed convert module — No longer maintaining torch/optimum-intel/transformers version compatibility issues. Use optimum-cli directly for conversions.
• Removed Optimum format support — Gemma-4 models now load directly via VLMPipeline, achieving 16% performance improvement (E2B: 16.4→19.1 tok/s). No optimum-intel required.

📖 Documentation

• Xe driver benchmark completed — Added Qwen3/14B and Gemma-4 31B data.
• README overhauled, removed all convert-related documentation.
• Updated benchmark tables.

1.3.0

暂时禁用pdf功能，见 #36386

PDF temporarily disabled — see #36386

1.2.0

更新

1. 添加 _get_gpu_driver() — 检测 /sys/class/drm/card*/device/driver/module，返回 "i915" / "xe" / None
2. _pdf_to_images() 改造：
   • 解除 24 页上限 → 全页转换
   • 300 DPI 渲染（原 120/150）
   • i915 阻断：检测到 i915 驱动时打印警告并返回 None，提示引导时切换到 Xe 驱动
3. 删除 /file 命令 — GenAI 和 Optimum 两处的 handler、帮助文本全部移除
4. 删除 /clear 命令 — /clear all 和 /clear N 两处 handler 全部移除
5. 移除文件持久化支持（见readme）
6. 添加 loaded_files.clear() — GenAI（行 1109）和 Optimum（行 1372）每轮对话后自动清除文件
7. 帮助文本更新 — print_commands_common()、GenAI /help、Optimum /help 中的 /file /clear 条目已删除

New

1. Added _get_gpu_driver() — Detects the kernel driver via /sys/class/drm/card*/device/driver/module, returns "i915" / "xe" / None
2. _pdf_to_images() overhaul:
   • Removed 24-page limit → full page conversion
   • 300 DPI rendering (was 120/150)
   • i915 block: prints a warning and returns None when i915 is detected, advises switching to the Xe driver at boot
3. Removed /file command — Handlers and help text removed from both GenAI and Optimum modes
4. Removed /clear command — Both /clear all and /clear N handlers removed
5. Removed file persistence (see README)
6. Added loaded_files.clear() — GenAI (line 1109) and Optimum (line 1372), clears loaded files after each turn
7. Updated help text — /file and /clear entries removed from print_commands_common(), GenAI /help, and Optimum /help

1.1.0

新增

./ov-cli setup --remove

卸载指定模块的独有 pip 包。逻辑：

• 自动计算哪些包是"独有"的（不被其他已装模块需要）
• 共享包不动（如 soundfile 同时被 asr 和 tts 需要，只删一个时保留）
• 基础依赖（openvino 等）永远不动
• 卸载前会确认

New

./ov-cli setup --remove

Uninstalls exclusive pip packages of specified modules. How it works:

• Automatically determines which packages are "exclusive" to the removed modules (not needed by remaining installed modules)
• Shared packages are kept (e.g. soundfile is used by both asr and tts — removing one keeps it)
• Base dependencies (openvino, etc.) are never touched
• Prompts for confirmation before uninstalling

1.0.1

修复：

--reasoning off 过滤不彻底 — GenAI 交互/once/pipe 模式的思考过滤只去掉了 <think> 标签，但留下了思考内容前缀（如 Here、Alright）。修复后：

• 初始假设在 think 块内 → Here</think> 中的 "Here" 和 </think> 全部丢弃
• 正确处理 <think></think> 完整块的过滤
• 覆盖 GenAI 和 Optimum 两种格式的所有模式（interactive/once/pipe）

Fix:

--reasoning off filtering was incomplete — The thinking filter for GenAI interactive/once/pipe modes only removed <think> tags, but left thinking content prefixes (e.g., Here, Alright) in the output. After the fix:

• Initially assumes we're inside a think block → both Here and </think> from Here</think> are discarded
• Correctly handles full <think></think> blocks
• Covers both GenAI and Optimum formats across all modes (interactive/once/pipe)

1.0.0

哈哈，功能都差不多了吧，最后让 AI 通篇检查了一遍，修复了一个 bug，发布 1.0.0

Features are about there, right? Had the AI do a final pass, fixed one bug — let's ship 1.0.0