AI: ollama llama3 conversation does not stop

[JohannesBuchner]

Summary

Description

Steps to trigger the problem:

1. install zed on ubuntu with the command line
2. connected it to my docker llama install.
3. open integrator.py in JohannesBuchner/ultranest project
4. pick a line where it complains, click "Fix with Assistant"
5. send "Fix Diagnostics" to model

Expected Behavior: edited code
Actual Behavior: does not stop

Model Provider Details

• Provider: ollama
• Model Name: llama3
• Mode: Inline Assistant
• Other Details (MCPs, other settings, etc):

Zed Version and System Specs

Zed: v0.195.3 (Zed)
OS: Linux Wayland ubuntu 24.04
Memory: 62.5 GiB
Architecture: x86_64
GPU: Intel(R) Graphics (RPL-P) || Intel open-source Mesa driver || Mesa 24.2.8-1ubuntu1~24.04.1

[JohannesBuchner]

Some screenshots:

[JohannesBuchner]

Possibly this is a different bug, but if I choose another service provider with a openai interface, and select qwen2.5-coder-32b-instruct as the model (with the helpful instructions at #34327), then I get the code passed back to me, but with a different indent:

[notpeter]

Yeah, the experience with local models can be subpar.
We've definitely saw significant improvements in the ability of Claude to match indentation between claude-3-5-sonnet-20241022, claude-3-7-sonnet-20250219 and claude-sonnet-4-20250514.

My understanding is that there was significant fine tuning to improve the experience here. You might try your luck with different models and see if additional rules / system prompt details help (e.g. "Make sure when replacing or inserting code indent levels match with the original code" or similar.

I would definitely be interested in any prompts you find that help the model do a better job.