ollama run youtu/youtu failed

[xinzhuang]

When run ollama run youtu/youtu, the terminal shows:

Error: 500 Internal Server Error: llama runner process has terminated: error loading model: error loading model hyperparameters: key not found in model: deepseek2.expert_weights_scale
llama_model_load_from_file_impl: failed to load model

environment info

macos 15.7.3
ollama version 0.13.5

[chenruilong]

同样遇到此错误

[xinzhuang]

我想说，腾讯的程序员们，你们真的做个人吧，readme 文档写得稀烂，作为一个编程 0 基础的小白自己摸索到这个地步还遇到这个报错，你对得起谁？

同学，建议发言还是慎重些哈，希望改一下哈，拜托哈，任何项目，都会有大大小小的问题，需要大家一起贡献呀

[svlys]

Thank you very much for your feedback and attention. Currently, the ollama in its present version doesn't support youtu/youtu. This is because the modifications made to llama.cpp over the past three weeks haven't been incorporated yet.

You can keep an eye on the progress of the PR via this link: ollama/ollama#13570.

If you're using a Mac, you can access it in the following way for now:
First, run brew install llama.cpp.
Then, execute llama-server -hf tencent/Youtu-LLM-2B-GGUF:Q8_0 --host 0.0.0.0 --port 8081 --log-disable.
Finally, configure the request address as "http://127.0.0.1:8081/v1".

[lzz773751548]

Thank you very much for your feedback and attention. Currently, the ollama in its present version doesn't support youtu/youtu. This is because the modifications made to llama.cpp over the past three weeks haven't been incorporated yet.

You can keep an eye on the progress of the PR via this link: ollama/ollama#13570.

If you're using a Mac, you can access it in the following way for now: First, run brew install llama.cpp. Then, execute llama-server -hf tencent/Youtu-LLM-2B-GGUF:Q8_0 --host 0.0.0.0 --port 8081 --log-disable. Finally, configure the request address as "http://127.0.0.1:8081/v1".

使用llama-server调用youtu-llm后，使用agent执行功能只能生成回答，并不能像演示视频一样进行指令解析与执行，请问这个问题如何解决？

[xfey]

Hello, thank you for your interest in this work! 👋

The reason may locate at that, LLM outputs are unpredictable, and sometimes the model fails to output the tool call format correctly, causing commands to be unparsed and unexecuted. I've tried reproducing some prompts locally, and a small number of requests do indeed occasionally fail to be parsed. Additionally, the quantized version of the model may exhibit slight differences in performance compared to the original model.

If it's convenient for you, could you please provide the prompt you entered, and whether you adjusted parameters such as top_k and temperature in the model settings? We will investigate further.

---

您好，首先感谢您关注本工作！👋
大模型输出具有不确定性，有时模型未能正确输出工具调用格式，会导致指令无法被解析和执行。我在本地尝试复现了一些 prompt，有少量请求确实会偶发无法被解析的问题。此外，量化版的模型相比于原版模型可能也会有少量表现的区别。
如您方便，是否可以提供您输入的 prompt，以及您是否有在模型设置时调整 top_k 和 temperature 等参数呢？我们将进行进一步地检查。