chore(benchmarks): repoint dead /mnt-1 paths at /var after #3159 decommission (#3162) - #3166
Conversation
…mmission (#3162) /mnt-1 was decommissioned in #3158/#3159; every remaining default in this tree still pointed at it, so scripts would silently write to the root fs instead of failing loud. Sweep replaces all live /mnt-1 defaults with /var across the round-7 corpus/training scripts, disk-layout docs, and the build-supervisor. The only live-breaking mount was unsloth's hf-cache bind, now repointed. Dated historical records (benchmark run logs, plan docs) are left referencing /mnt-1 as-is with a note, since they describe what actually happened at the time. rex86 large-model handling moves to a /var subdir since there is no second volume to split across anymore. Closes #3162
Strix Security ReviewNo security issues found. Updated for Reviewed by Strix |
Dependency Review✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.Scanned FilesNone |
There was a problem hiding this comment.
This PR is a mechanical chore that repoints decommissioned /mnt-1 paths to /var across benchmark/training shell scripts, Docker Compose bind mounts, a Python path constant, and supporting documentation. Every changed file was reviewed. The changes are path-string substitutions with no new input sources, sinks, or removed security controls. The single behavioral edit — build-supervisor.sh's DIR variable becoming environment-overridable via ${DIR:-...} — is a standard configurability idiom whose value is only used to locate a trusted build script in an operator/CI-run context, with no evidence of untrusted reachability. No authentication, authorization, injection, secret-handling, or file-path security regressions were introduced.
Reviewed by Strix
Configure security review settings
Why
/mnt-1 was decommissioned in #3158/#3159. Every remaining default across the
benchmark/training tree still pointed at the dead mount, so scripts would
silently write to the root fs instead of failing loud on a missing volume.
What
Closes #3162