Skip to content

Plan the live harness link: 4080s, then the 395 - #41

Merged
jmjava merged 1 commit into
mainfrom
cursor/harness-host-plan-937a
Nov 1, 2026
Merged

jmjava merged 1 commit into
mainfrom
cursor/harness-host-plan-937a

Conversation

@jmjava

@jmjava jmjava commented Nov 1, 2026

Copy link
Copy Markdown
Owner

Summary

The live harness (run_harness.py --backend live and run_eval.py --live) follows OLLAMA_BASE_URL. This plan links that URL to the 4080-class cards first, and to the Ryzen AI Max+ 395 after the 4080 link has a fast-model row.

  • IDE 4080 on this machine uses http://127.0.0.1:11434. Another 4080-class machine uses an SSH forward to 127.0.0.1:11438.
  • The 395 uses 127.0.0.1:11440, so the downstairs paper forward on 11436 can stay up.
  • Both hosts start on qwen3.5:9b with OLLAMA_NUM_CTX=16384. The strong tag on a 16 GB 4080 waits until ollama ps shows the fast tag on the GPU.
  • Personal rows stay in gitignored eval-runs/. Phase 3 timings stay on the downstairs forward. Cursor still chooses accept, rewrite, or reject before anything is copied into the repo.

Test plan

  • python -m unittest tests.test_harness_host_plan
  • python scripts/check_deployment_safety.py --skip-listen
Open in Web Open in Cursor 

Personal runs use one loopback URL per GPU host. Phase 3 timings stay on the downstairs forward.

Co-authored-by: jmjava <jmjava@gmail.com>
@jmjava
jmjava merged commit 4e748b6 into main Nov 1, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants