[gnezim]

Summary

Raise the calculated GPU wired-memory ceiling introduced in #5 so a 128 GiB inference node defaults to 125 GiB instead of 118 GiB.

- default = physical RAM - 10 GiB  # 128 GiB -> 118 GiB / 120832 MiB
+ default = physical RAM - 3 GiB   # 128 GiB -> 125 GiB / 128000 MiB

Explicit gpu_memory_gb overrides remain unchanged. Update the configuration error, README and example configuration to match.

Evidence

• Before: the generated daemon startup script subtracts 10240 MiB; the 128/64 GiB test cases expect 118/54 GiB.
  After: executable shell fixtures verify 125/61 GiB, an unchanged 125 GiB limit without a redundant write, an 8 GiB host, and rejection at or below 3 GiB. Existing explicit-override and write/readback failure coverage remains intact.
• make dev-check: 407 passed, Ruff passed, using an ignored local copy of tfconfig.example.yaml as tfconfig.yaml for the status tests.
• git diff --check passed.
• No live-node sysctl write, daemon restart, reboot or inference load test was performed.

Merge Danger

Door: two-way

To restore the previous ceiling on a 128 GiB node, set gpu_memory_gb: 118 and reinstall/restart the managed daemon.

Blast Radius: inference-hosts

The next managed daemon installation/restart applies RAM minus 3 GiB wherever the calculated default is selected, including nodes of other sizes. At full GPU use, only 3 GiB remains outside this host-wide ceiling, increasing memory-pressure/swap risk. This is a ceiling, not a reserved allocation; oMLX memory-guard settings remain independent and may impose a lower limit. Merging alone does not change installed daemons.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: 5c5eb538-b189-41e7-b6fd-85b7197553b0

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}