User Tools

Site Tools


linux64

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revisionPrevious revision
Next revision
Previous revision
linux64 [2026/11/01 14:08] – [vllm in podman for fast qwen3.8:27b ai] bengtlinux64 [2026/11/07 19:26] (current) – [Using pi.dev as ai harness] bengt
Line 971: Line 971:
 This is right on the edge for what my anker c300x ups can handle. I may need to get a new ups. You can tell the cards to use less max power with "nvidia-smi -pl 150" as root. I added "ollama ALL=(ALL) NOPASSWD: /usr/bin/nvidia-smi" with visudo to be able to set this with sudo in the LLM start script. This is right on the edge for what my anker c300x ups can handle. I may need to get a new ups. You can tell the cards to use less max power with "nvidia-smi -pl 150" as root. I added "ollama ALL=(ALL) NOPASSWD: /usr/bin/nvidia-smi" with visudo to be able to set this with sudo in the LLM start script.
  
 +==== Using pi.dev as ai harness ====
  
 Switched to pi.dev to be able to configure qwen3.8:27b to use medium instead of xhigh thinking: Switched to pi.dev to be able to configure qwen3.8:27b to use medium instead of xhigh thinking:
Line 1000: Line 1001:
 In this https://github.com/earendil-works/pi/issues/2020 thinkingFormat changes to qwen or maybe qwen-chat-template. In this https://github.com/earendil-works/pi/issues/2020 thinkingFormat changes to qwen or maybe qwen-chat-template.
  
-I decided to install a cpu compactor "pi install npm:@sting8k/pi-vcc" (plain npm:pi-vcc is abandoned and buggy) to avoid running out of context window while compacting. Means that I can use the full server side context window client side. (I first tried npm:pi-compact but then it sometimes answered in Chinese and had a remember tool that I do not want.) Find compactors here: https://pi.dev/packages?name=compactor+A also did the following to make pi behave: 
 +  - Install a cpu compactor "pi install npm:@sting8k/pi-vcc" (plain npm:pi-vcc is abandoned and buggy). This means that I can use the full server side context window client side. 
 +  - Use "/settings" in pi and configure "HTTP idle timeout" to disabled. For local models with long output this can aovid "Error: Terminated". 
 +  - You can also configure theme in "/settings" to match your terminal colors. 
 +  - It is possible to control if all pre-entered steering prompts should be sent at once with "/settings " and "Steering mode" all. 
  
-But this was suggested by the LLM to control output better:+Editing the system promt was also suggested to control output better:
 <code> <code>
 $ cat ~/.pi/agent/APPEND_SYSTEM.md $ cat ~/.pi/agent/APPEND_SYSTEM.md
-## Reply language+## House rules
  
-Always respond and output exclusively in the language of the user's latest message (normally English or Swedish and never Chinese). +- **Keep documentation and code comments brief and concise:** State intent and the non-obvious "why" in one or two lines. Do not restate what the code already says. 
- +- **Short commit messages:**  
-## Docs & comments +  - Subject: A normal Git subject line is 50 characters or fewer, hard limit 72. Create a commit-msg hook to enforce subject length <72 if not already in place. 
- +  - Body: A few lines of intent only (what + why). 
-Keep documentation and code comments brief and concise: state intent and the non-obvious "why" in one or two lines. Do not restate what the code already says.+- Always **respond and output** exclusively in the language of the user's latest message (normally English or Swedish and never Chinese).
 </code> </code>
  
linux64.1793542134.txt.gz · Last modified: by bengt

Donate Powered by PHP Valid HTML5 Valid CSS Driven by DokuWiki