Every mention of another inference engine is rewritten to say what our code does, or the math it computes. Reference values are "the reference value from the published config" or "the recorded reference output", line citations into another engine's source become the rule itself, comparisons with other engines are dropped, and asides about other engines' bugs go. Identifiers follow: - CUDA: hanzo-paged-attn's `namespace vllm` is `pagedattn`, hanzo-engine's MoE helpers `moe_dtype` and `moe_detail`, the rotary kernels' `hanzo_rotary`; the shuffle/ldg macros are HANZO_*. - GPTQ: the shuffled-layout path (`reconstruct_exllama`, `use_exllama`) is `reconstruct_shuffled` / `is_shuffled`, on the kernel and the FFI side. - Tests: `*_matches_vllm` and friends name what they compare against (the reference fixture, the tile rule). - MiniCPM-o's vision-LLM embedding is `vlm_embedding`; the rotary custom op is `rotary-embedding`. - hanzo-bench's board scores ours against a "rival" whose name comes from the manifest: `rival_<tag>.json`, `--rival*` manifest flags, `rival t/s` columns; dossier.sh and spec_decode.sh take `--rival-bench`. - install.ps1 installs this repository's hanzo-cli, as install.sh does, with HANZO_INSTALL_NCCL / HANZO_INSTALL_NO_NCCL. - bench_http.py reads only the hanzo: metric families. Kept, because they are what they are: LICENSE, LICENSE-APACHE, NOTICE and tools/upstream-map.sh; each ported file's copyright/licence/provenance header; Cargo.toml dependency names; data identifiers (Hugging Face repo IDs, fixture file names and fixture metadata keys the tests read, on-disk paths, executable names the GPU and engine guards match on a host); model and format names. LLM.md says the rule once, under "One name: Hanzo Engine". Co-authored-by: Hanzo Dev <[email protected]>
hanzo-engine-guard
Keeps the local Hanzo engine (hanzo serve) alive safely — it self-heals a
crashed engine but can never DoS the machine. One portable daemon; native
service install on Linux, macOS, and Windows (all user-level, no admin).
hanzo-engine-guard install # register as a native service for this OS
hanzo-engine-guard status # show engine ports, health, desktop state
hanzo-engine-guard uninstall # remove the service
hanzo-engine-guard run # run the daemon in the foreground (what the service runs)
| OS | Mechanism (user-level) |
|---|---|
| Linux | systemd --user service (~/.config/systemd/user/) + enable-linger |
| macOS | launchd LaunchAgent (~/Library/LaunchAgents/ai.hanzo.engine-guard.plist) |
| Windows | Task Scheduler HanzoEngineGuard (/SC ONLOGON, runs pythonw, no console) |
Only dependency: Python 3.8+ and (recommended, cross-platform) psutil. Without
psutil it falls back to pgrep/pkill on Unix.
What it does
-
Circuit breaker (always): kills any process that is the legacy server binary named by
BOMB_PAT(matched precisely by exe/argv0/comm — never a substring of some innocent command), and ifhanzo serveprocesses ever exceedRUNAWAY_CEILING(default 256) it kills them all. This is the backstop against the historical recursion fork-bomb. -
Keepalive (only while a desktop/node is running): if an engine port that was seen healthy goes down with no live process, it respawns it by replaying the exact command + working dir + environment captured while it was healthy (the command alone isn't enough — the engine needs its CUDA/HF env).
Why it cannot DoS the machine
- Backoff between respawns: 15s → 30s → 60s → 120s → 300s. Never a tight loop.
- Give-up cap: after
MAX_FAILS(8) consecutive failed respawns of a port it stops and logs — a broken engine is never hammered forever. - Gated on a desktop/node actually running — no zombie engines after you quit.
- Snapshot-replay only: it only respawns a port it has seen healthy; it never guesses a command/model.
- Runaway ceiling: the hard backstop kills any exponential spawn.
- The systemd unit uses
KillMode=process(a guard restart never kills the engines it spawned) and sets noMemoryMax(a cap would count a respawned engine's multi-GB RAM and cgroup-OOM it).
Tuning (env)
HANZO_ENGINE_RUNAWAY_CEILING— kill-all threshold (default 256).
Known limitation
The keepalive only recovers ports it has seen healthy (it holds the command/cwd/env snapshot in memory, never on disk — no secrets persisted). If the guard restarts while a port is already down, it can't recover that port until the desktop (or a sibling engine) brings it up once. This is the safe trade-off.