Chasing a Standalone Boot and Fixing a Docker Leak Along the Way

The Goal: Boot Without the Real Library

The local dispatch was working, but it still linked against the real libggml-sycl. I wanted a standalone boot—my own backend with no proprietary binary—so I could reason about and ship the implementation cleanly. Today was an exercise in removing that dependency one interface stub at a time.

A Methodical Stub Series

Each stub was regression-checked with the real library so I couldn’t accidentally break the working path:

Boot probe without the real library, blocked on SYCL init (45358683).
Registry interface stubs (621ece22) and device interface stubs (0ca6707f).
Standalone SYCL init (3cd1cb86), blocked on buffer-type allocation.
Standalone buffer type using SYCL USM (cfd4f403), then the buft-name-as-function-pointer fix with host buft stubs (1a1ff0bf).
Splitting the buffer-type and host buft stubs (eaa70549).
Disabling next<> fallthrough when there’s no real library (9c03a7de), then adding next_sycl for device and registry functions (4520558e)—and reverting it after a static init order issue surfaced (f4f9dea5).
Guarding dev_get/count/by_name behind a real-registry check (2270accc) and adding a crash handler for standalone debugging (fed62fba).
An offload_op stub plus an initialize_local_devices() guard (6a9c6099), then fixing the missing forward declaration (a506c6c4).

Each step moved the crash slightly later in the startup sequence. Following it that way is the only sane approach when there’s no debugger story for the interposed path.

Buffer Ownership Audit

I also did a buffer, device, and context ownership audit (a2893d3d) and passed with no real-library pointer leaks—important, because a leak in the interposition layer would be nearly impossible to find later.

Swarm Automation Overhaul

Between boot attempts I overhauled the swarm automation: a daemon with session persistence and auto-completion (a223b4f0). I also fixed a Docker container leak in the bench scripts by adding a cleanup trap (d8d226b9) and added a throughput floor to the bench gate so a regression below the floor is rejected automatically (3311f1e2). The leak was small but it would eventually exhaust the machine during long benchmark runs.

Where It Stands

Five blockers identified and steadily eliminated. Standalone boot isn’t stable yet, but the failure point keeps moving deeper, which is exactly the signal I want.