Manus Stuck At Execution Steps - No Response From Ollama

by ADMIN 57 views

Description of the Issue

I'm experiencing a frustrating issue where Manus gets stuck during execution steps while using Ollama as the LLM provider. The agent fails to proceed properly and continuously repeats similar responses without selecting any tools. Additionally, I'm seeing repeated "Manus selected 0 tools to use" messages, and requests to /v1/chat/completions fail with 500 errors after 1-minute timeouts.

Steps to Reproduce the Issue

To reproduce this issue, follow these steps:

1. Start Ollama

First, start Ollama using the following command:

sh
ollama serve

2. Run Manus

Next, run Manus using the following command:

sh
python main.py

3. Enter a Simple Query

Enter a simple query like:

MANUS 에 대해서 알려줘 (Tell me about MANUS)

4. Manus Enters a Loop

Manus will enter a loop, repeatedly stating its available tools but failing to execute anything.

Expected Behavior

The expected behavior is that Manus should generate responses properly using the qwq model. It should select and execute relevant tools instead of repeating the same response.

Observed Behavior

However, the observed behavior is that Manus continuously outputs:

🛠️ Manus selected 0 tools to use

Each execution step takes 1+ minute to process. Ollama logs show repeated 500 errors on /v1/chat/completions:

[GIN] 2025/03/12 - 15:40:10 | 500 | 1m0s | 127.0.0.1 | POST "/v1/chat/completions"

Ollama also prints:

makefile
llama_model_load: vocab only - skipping tensors

Troubleshooting Attempts

I've tried the following troubleshooting attempts:

Reduced max_tokens in config.toml

I reduced the max_tokens in config.toml to 32:

max_tokens = 32

Restarted Ollama & re-downloaded the model

I restarted Ollama and re-downloaded the model using the following commands:

ollama rm qwq
ollama pull qwq
ollama serve

Tried forcing GPU mode

I tried forcing GPU mode using the following command:

set OLLAMA_USE_CUDA=1
ollama serve

Tested ollama run qwq directly

I tested ollama run qwq directly, and it responds, but slowly.

System Information

Here's the system information:

  • OS: Windows 10 / 11
  • Ollama Version: 0.5.13
  • Model: qwq
  • Hardware: CPU / GPU (Specify)

Request for Help

Is this an issue with how Manus interacts with Ollama? Why is llama_model_load: vocab only - skipping tensors appearing? Any recommendations on optimizing Ollama settings for better response times?

Thanks in advance for any support! 😊 🚀

Optimizing Ollama Settings for Better Response Times

To optimize Ollama settings for better response times, consider the following recommendations:

1. Increase max_tokens

Increase the max_tokens value in config.toml to allow for more tokens to be processed at once.

2. Use a faster model

Consider using a faster model, such as the qwq-fast model, which is optimized for faster response times.

3. Use a GPU

If possible, use a GPU to accelerate the processing of Ollama requests.

4. Optimize Ollama configuration

Optimize the Ollama configuration by adjusting the temperature and max_tokens values to achieve a balance between response time and accuracy.

5. Monitor Ollama logs

Monitor Ollama logs to identify any issues or errors that may be causing the slow response times.

By following these recommendations, you should be able to optimize Ollama settings for better response times and resolve the issue with Manus getting stuck during execution steps.

Q: What is the issue with Manus getting stuck during execution steps?

A: Manus is getting stuck during execution steps because it is unable to select and execute relevant tools. Instead, it repeatedly states its available tools and fails to proceed.

Q: Why is Ollama printing "llama_model_load: vocab only - skipping tensors"?

A: Ollama is printing "llama_model_load: vocab only - skipping tensors" because it is skipping the tensor loading step and only loading the vocabulary. This may be causing the slow response times.

Q: What are the troubleshooting attempts that have been made so far?

A: The troubleshooting attempts that have been made so far include:

  • Reducing the max_tokens value in config.toml to 32
  • Restarting Ollama and re-downloading the model
  • Trying to force GPU mode using the OLLAMA_USE_CUDA environment variable
  • Testing ollama run qwq directly, which responds but slowly

Q: Is this an issue with how Manus interacts with Ollama?

A: It is possible that this is an issue with how Manus interacts with Ollama. However, further investigation is needed to determine the root cause of the problem.

Q: What are some recommendations for optimizing Ollama settings for better response times?

A: Some recommendations for optimizing Ollama settings for better response times include:

  • Increasing the max_tokens value in config.toml to allow for more tokens to be processed at once
  • Using a faster model, such as the qwq-fast model
  • Using a GPU to accelerate the processing of Ollama requests
  • Optimizing the Ollama configuration by adjusting the temperature and max_tokens values
  • Monitoring Ollama logs to identify any issues or errors that may be causing the slow response times

Q: What are some potential causes of the slow response times?

A: Some potential causes of the slow response times include:

  • Insufficient resources (e.g. CPU, memory, or GPU)
  • Poor Ollama configuration
  • Issues with the model or its loading
  • Network connectivity problems

Q: How can I further troubleshoot this issue?

A: To further troubleshoot this issue, you can try:

  • Checking the Ollama logs for any errors or issues
  • Monitoring the system resources (e.g. CPU, memory, or GPU) to see if there are any bottlenecks
  • Trying different Ollama configurations or models to see if the issue persists
  • Consulting the Ollama documentation or seeking help from the Ollama community

Q: Can I use a different LLM provider instead of Ollama?

A: Yes, you can use a different LLM provider instead of Ollama. However, you will need to configure Manus to use the new provider and adjust the settings accordingly.

Q: How can I report this issue to the Ollama team?

A: To report this issue to the Ollama team, you can:

  • File a bug report on the Ollama GitHub repository
  • Send an email to the Ollama support team
  • Join the Ollama community and report the issue there

By following these steps and recommendations, you should be able to troubleshoot and resolve the issue with Manus getting stuck during execution steps.