Skip to content

drift doctor + hivemind patch-level in the startup banner #4

Description

@ApexDevelopment

Half of a recent homelab debugging session was environment forensics that a drift doctor command could do in seconds:

  • Is the installed hivemind wheel current? A desktop was running a wheel predating the per-thread replicate_p2p fix from scripts/hivemind-win32.patch; every server start deadlocked in handler registration. There was no way to tell short of grepping site-packages. Suggest: build_hivemind_windows.py stamps a patch revision into the wheel, the server logs it in the startup banner, and doctor compares it against the in-repo patch.
  • Stray processes: orphaned p2pd.exe daemons (see Tie p2pd child lifetime to the server process #5) and leftover run_server pythons squatting ports.
  • GPU sanity: a 5-second SDPA numeric check (GPU vs CPU max-diff at a few head_dims/dtypes) to rule out backend kernel bugs early on ROCm/MPS/XPU.
  • Torch/device/dtype summary.

🤖 Generated with Claude Code

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions