用llama试用gemma-4-E2B模型量化版本

l1t

466人浏览 · 2026-04-04 09:04:20

l1t · 2026-04-04 09:04:20 发布

下载llama和gguf文件。

旧的llama不支持新模型，所以要下载新版的llama。gguf选择最常用的Q4_K_M量化，比ollama版本的原始模型小了一半。

wget -c https://github.com/ggml-org/llama.cpp/releases/download/b8648/c-b8648-bin-win-cpu-x64.zip

wget -c https://github.com/ggml-org/llama.cpp/releases/download/b8648/llama-b8648-bin-ubuntu-x64.tar.gz

wget -c https://hf-mirror.com/unsloth/gemma-4-E2B-it-GGUF/resolve/main/gemma-4-E2B-it-Q4_K_M.gguf

wget -c https://www.modelscope.cn/models/unsloth/gemma-4-E2B-it-GGUF/resolve/master/mmproj-BF16.gguf

用命令行界面执行，采用默认参数。将一段英文翻译为中文。


C:\d>llama-b8648\llama-cli -m gemma-4-E2B-it-Q4_K_M.gguf
load_backend: loaded RPC backend from C:\d\llama-b8648\ggml-rpc.dll
load_backend: loaded CPU backend from C:\d\llama-b8648\ggml-cpu-zen4.dll

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b8648-f1ac84119
model      : gemma-4-E2B-it-Q4_K_M.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read <file>        add a text file
  /glob <pattern>     add text files using globbing pattern

将下文翻译成中文。At 4 AM on March 31, 2026, I woke up to my phone blowing up with notifications. The Claude Code source had been exposed, and the entire dev community was in a frenzy. My girlfriend in Korea was genuinely worried I might face legal action from Anthropic just for having the code on my machine — so I did what any engineer would do under pressure: I sat down, ported the core features to Python from scratch, and pushed it before the sun came up.

[Start thinking]

Here’s a thinking process for translating the provided English text:

Analyze the Request: The goal is to translate the given English paragraph into natural and accurate Chinese.
- Source Text: “At 4 AM on March 31, 2026, I woke up to my phone blowing up with notifications. The Claude Code source had been exposed, and the entire dev community was in a frenzy. My girlfriend in Korea was genuinely worried I might face legal action from Anthropic just for having the code on my machine — so I did what any engineer would do under pressure: I sat down, ported the core features to Python from scratch, and pushed it before the sun came up.”
Deconstruct the Source Text (Segment by Segment):
- Segment 1: “At 4 AM on March 31, 2026, I woke up to my phone blowing up with notifications.” (Focus on time, date, and the sudden, overwhelming nature of the notifications.)
- Segment 2: “The Claude Code source had been exposed, and the entire dev community was in a frenzy.” (Focus on the event and the resulting community reaction.)
- Segment 3: “My girlfriend in Korea was genuinely worried I might face legal action from Anthropic just for having the code on my machine —” (Focus on the personal stakes and the specific threat/entity.)
- Segment 4: “so I did what any engineer would do under pressure:” (Transition phrase leading to the action.)
- Segment 5: “I sat down, ported the core features to Python from scratch, and pushed it before the sun came up.” (Focus on the specific, high-pressure actions taken.)
Drafting - Focusing on Tone and Idiomatic Chinese:
- Self-Correction/Refinement Strategy: The tone is urgent, stressful, and heroic (the engineer solving a crisis). The Chinese translation should maintain this high-stakes feeling.
Generating the Translation (Iterative Process):
- (Initial draft thoughts for Segment 1): 2026年3月31日凌晨4点，我醒来时手机响个不停。 (A bit flat.)
- (Refining Segment 1): Use stronger verbs for “blowing up with notifications.”
- (Initial draft thoughts for Segment 2): Claude的代码源暴露了，整个开发社区都非常疯狂。 (Accurate, but can be punchier.)
- (Initial draft thoughts for Segment 3 & 4): 我在韩国的女友真的担心我因为机器上有代码会面临Anthropic的法律行动。所以，我做了任何工程师在压力下都会做的事。 (A bit too literal.)
- (Refining Segment 5): “ported the core features to Python from scratch” needs to sound technical and complete. “pushed it before the sun came up” needs to convey urgency.
Final Review (Checking against the original for accuracy and flow): (This leads to the final suggested translation.)
Final Output Generation. (See the suggested translation below.)
[End thinking]

以下提供几种翻译风格，您可以根据具体语境选择最合适的版本：

方案一：流畅自然型 (推荐)

2026年3月31日凌晨4点，我被手机的通知声惊醒。Claude的代码源已经泄露，整个开发社区陷入了疯狂。我在韩国的女友非常担心我仅仅因为机器上存在这段代码，就会面临Anthropic的法律诉讼——于是，我做了一件任何工程师在压力下都会做的事：我坐下来，从零开始将核心功能移植到Python，并在天亮之前完成了部署。

方案二：强调紧迫感和专业性 (更偏向技术叙事)

2026年3月31日凌晨4点，我醒来时，手机不断地弹出通知。Claude的代码源暴露了，整个开发社区都陷入了恐慌。我的韩国女友非常担心，我仅仅因为代码在我的机器上，就会被Anthropic追究法律责任。因此，我做出了任何工程师在压力下的反应：我立即着手，从头开始将核心功能移植到Python，并在日出之前完成了推送。

方案三：直译简洁型

2026年3月31日凌晨4点，我被手机的通知淹没。Claude的代码源已经暴露，整个开发社区都陷入了狂热。我韩国的女友非常担心，我因为机器上存在这段代码就会面临Anthropic的法律行动——所以我做了任何工程师在压力下的行为：我坐下来，从零开始将核心功能移植到Python，并在太阳升起前完成了推送。

[ Prompt: 271.0 t/s | Generation: 29.0 t/s ]

The whole thing was orchestrated end-to-end using oh-my-codex (OmX) by @bellman_ych — a workflow layer built on top of OpenAI’s Codex (@OpenAIDevs). I used $team mode for parallel code review and $ralph mode for persistent execution loops with architect-level verification. The entire porting session — from reading the original harness structure to producing a working Python tree with tests — was driven through OmX orchestration.

[Start thinking]

Here’s a thinking process for translating the provided technical paragraph:

Analyze the Source Text and Identify Key Concepts/Jargon:
- “The whole thing was orchestrated end-to-end…” (Focuses on control/management).
- “…using oh-my-codex (OmX) by @bellman_ych…” (Specific tool/person).
- “…a workflow layer built on top of OpenAI’s Codex (@OpenAIDevs).” (Defines the architecture/dependency).
- “I used $team mode for parallel code review…” (Specific feature/mode).
- “…and $ralph mode for persistent execution loops with architect-level verification.” (Another specific feature/mode).
- “The entire porting session — from reading the original harness structure to producing a working Python tree with tests — was driven through OmX orchestration.” (Describes the scope and mechanism of the process).
Determine the Tone and Target Audience:
- The text is highly technical, using specific names (Codex, OmX, $team mode, $ralph mode).
- The tone is confident, detailed, and focused on process and methodology (workflow, orchestration).
- The target audience is likely other developers, engineers, or those familiar with AI/ML tooling.
Drafting - Segment by Segment (Focusing on Clarity and Technical Accuracy):
- Original: The whole thing was orchestrated end-to-end using oh-my-codex (OmX) by @bellman_ych — a workflow layer built on top of OpenAI’s Codex (@OpenAIDevs).
  - Initial thought: 整个过程是通过 @bellman_ych 的 oh-my-codex (OmX) 来端到端编排的，它是一个建立在 OpenAI Codex 上的工作流层。
- Original: I used $team mode for parallel code review and $ralph mode for persistent execution loops with architect-level verification.
  - Initial thought: 我使用了 $team 模式进行并行代码审查，以及 $ralph 模式进行带有架构级验证的持久执行循环。
- Original: The entire porting session — from reading the original harness structure to producing a working Python tree with tests — was driven through OmX orchestration.
  - Initial thought: 整个移植过程——从阅读原始的框架结构到生成带有测试的可用Python树——都是通过 OmX 的编排驱动的。
Review and Refine (Creating a Cohesive Flow): (Checking for smooth transitions and appropriate technical phrasing.)
- Self-Correction during review: Ensure the technical terms sound natural in Chinese. “Orchestrated end-to-end” should be strong. “Workflow layer” needs to be clear.
Final Polish (Selecting the best version for the output). (This leads to the suggested translations below.)

[End thinking]

这是一个高度技术性的段落，翻译时需要保持专业术语的准确性，同时确保句子的逻辑流畅。

以下提供几种翻译方案，您可以根据具体的语境（例如是技术博客、项目报告还是口头介绍）选择最合适的版本：

方案一：专业流畅型（推荐，适合技术文档）

整个过程是通过 @bellman_ych 使用的 oh-my-codex (OmX) 来实现端到端（end-to-end）的编排，它是一个建立在 OpenAI Codex (@OpenAIDevs) 之上的工作流层。我利用了 $team 模式进行并行代码审查，并使用 $ralph 模式进行带有架构级验证的持久执行循环。整个移植过程——从阅读原始的框架结构到生成带有测试的可用 Python 树——都由 OmX 的编排驱动。

方案二：强调工具和流程的简洁型

整个流程均由 @bellman_ych 的 oh-my-codex (OmX) 进行端到端编排，该工具是一个基于 OpenAI Codex (@OpenAIDevs) 的工作流层。我使用了 $team 模式进行并行代码审查，以及 $ralph 模式进行带有架构级验证的持久执行循环。从阅读原始框架结构到生成带测试的 Python 树的整个移植工作，都通过 OmX 的编排得以驱动。

方案三：直译清晰型（更贴近原文结构）

所有这一切都是通过 @bellman_ych 的 oh-my-codex (OmX) 进行端到端编排的，它是一个建立在 OpenAI Codex (@OpenAIDevs) 之上的工作流层。我使用了 $team 模式进行并行代码审查，以及 $ralph 模式进行带有架构级验证的持久执行循环。从阅读原始框架结构到生成带有测试的 Python 树的整个移植环节，都通过 OmX 的编排来驱动。

【术语解释参考】

orchestrated end-to-end: 端到端编排/控制。
workflow layer: 工作流层。
parallel code review: 并行代码审查。
persistent execution loops: 持久执行循环。
architect-level verification: 架构级验证。
harness structure: 框架结构/测试框架结构。
Python tree: Python 树（指项目结构或文件树）。

[ Prompt: 275.1 t/s | Generation: 28.5 t/s ]

The result is a clean-room Python rewrite that captures the architectural patterns of Claude Code’s agent harness without copying any proprietary source. I’m now actively collaborating with @bellman_ych — the creator of OmX himself — to push this further. The basic Python foundation is already in place and functional, but we’re just getting started. Stay tuned — a much more capable version is on the way.

[Start thinking]

Here’s a thinking process for translating the text:

Analyze the Request: The user wants a translation of a short, technical, and forward-looking paragraph about a software project (a Python rewrite of an agent harness).
Deconstruct the Original Text (Sentence by Sentence):
- The result is a clean-room Python rewrite that captures the architectural patterns of Claude Code’s agent harness without copying any proprietary source.
- I’m now actively collaborating with @bellman_ych — the creator of OmX himself — to push this further.
- The basic Python foundation is already in place and functional, but we’re just getting started.
- Stay tuned — a much more capable version is on the way.
Identify Key Technical/Idiomatic Terms and Determine Tone:
- clean-room Python rewrite: This is a crucial concept in software engineering. It means rewriting something based on understanding, not copying. Needs a precise translation.
- architectural patterns: 架构模式。
- agent harness: Agent harness (keep this as is, or clarify if context demands).
- proprietary source: 专有/私有源代码。
- collaborating with… the creator of OmX: 强调合作关系和权威性。
- push this further: 进一步推进/发展。
- basic Python foundation is already in place and functional: 基础已经到位且可用。
- Stay tuned: 保持关注/敬请期待。
- much more capable version is on the way: 更强大/更高级的版本即将推出。
Drafting - Focusing on Flow and Technical Accuracy (Iterative Refinement):
- (Initial thought for S1): 结果是一个干净室的Python重写，它捕获了Claude Code代理框架的架构模式，而没有复制任何专有源。 (A bit stiff.)
- (Refinement for S1): Better to make “clean-room rewrite” sound professional.
Final Translation Options (Presenting various styles): (This leads to the suggested options below.)
Review against the Original Intent: Does the translation accurately convey the achievement (clean rewrite), the current status (collaboration, foundation built), and the future plan (more capable version coming)? Yes. (The selected translations are strong and fit the technical context.)

[End thinking]

这是一个关于项目成果和未来展望的段落，翻译时需要保持专业性，同时让语气既自信又充满期待。

以下提供几种翻译方案，您可以根据具体的语境选择最合适的版本：

方案一：专业流畅型（推荐，适合技术博客或官方发布）

最终成果是一个“干净室”的 Python 重写，它捕捉了 Claude Code 代理框架的架构模式，而没有复制任何专有代码。我现在正积极与 OmX 的创造者 @bellman_ych 合作，以便将项目推向更进一步。基本的 Python 基础已经搭建并可用，但我们才刚刚开始。请持续关注——一个功能更强大的版本即将到来。

方案二：强调成果和行动型（更具行动力）

我们的成果是一个干净的 Python 重构，它保留了 Claude Code 代理框架的架构模式，但完全没有复制任何专有源代码。我现在正与 OmX 的开发者 @bellman_ych 紧密合作，共同推进这项工作。基础的 Python 框架已经完成并可运行，但我们才刚刚起步。敬请期待——一个能力更强的版本正在路上。

方案三：简洁直接型（适合快速更新或社交媒体）

结果是一个干净的 Python 重写，它抓住了 Claude Code 代理框架的架构模式，且没有复制任何私有代码。我现在正与 OmX 的创建者 @bellman_ych 合作，以便进一步发展。基础的 Python 结构已经到位且可用，但我们才刚刚开始。保持关注——一个更强大的版本即将发布。

【关键词解释】

clean-room Python rewrite: “干净室”重写。这是一个工程术语，指在不依赖原始代码的情况下，从头开始重写，确保架构正确但代码是原创的。
architectural patterns: 架构模式。
agent harness: 代理框架/代理控制结构。
proprietary source: 专有源代码/私有代码。
Stay tuned: 保持关注/敬请期待。

[ Prompt: 256.9 t/s | Generation: 28.0 t/s ]

比Ollama版本多了几个选择。可能是不同软件的系统提示词不同。
提示约250 t/s，生成 28 t/s。

用服务器和浏览器执行，识别图像文件中的文字，命令行如下

C:\d>llama-b8648\llama-server -m gemma-4-E2B-it-Q4_K_M.gguf --mmproj mmproj-BF16.gguf --ctx-size 200000 --repeat-penalty 1.15 --repeat-last-n 256 --temp 0.7 --port 8080

中文识别出来的结果非常差。
原文
在这里插入图片描述