Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
65 commits
Select commit Hold shift + click to select a range
eefbac5
Merge pull request #305 from T-Crypt/T-Crypt/main-into-next-v0.5.0
T-Crypt Nov 8, 2026
f0003de
feat(notify): Windows tray with native toasts
T-Crypt Oct 23, 2026
306a949
Merge pull request #307 from T-Crypt/T-Crypt/windows-tray-rebase
T-Crypt Nov 12, 2026
2665aa4
feat(gateway): a ComfyUI job can yield the card to the LLM it calls
T-Crypt Nov 12, 2026
416ea98
Merge remote-tracking branch 'origin/main' into T-Crypt/main-into-nex…
T-Crypt Nov 12, 2026
402e55b
Merge pull request #309 from T-Crypt/T-Crypt/main-into-next-orphan-wr…
T-Crypt Nov 12, 2026
fdccbc3
Merge remote-tracking branch 'origin/next' into T-Crypt/comfyui-yield
T-Crypt Nov 12, 2026
31cb378
docs(roadmap): the yield and the Windows tray are in next
T-Crypt Nov 12, 2026
3b5b0d4
fix(gateway): a yield's end answers only once ComfyUI holds the card
T-Crypt Nov 12, 2026
a60a2d3
Merge pull request #308 from T-Crypt/T-Crypt/comfyui-yield
T-Crypt Nov 12, 2026
f1b1074
fix(broker): a model that fills the card evicts a parked engine it ca…
T-Crypt Nov 12, 2026
ec74904
Merge pull request #310 from T-Crypt/T-Crypt/beside-parked-fit
T-Crypt Nov 12, 2026
aea4147
fix(broker): keep 512 MiB of load headroom when a full-card model fit…
T-Crypt Nov 12, 2026
1c12995
Merge pull request #312 from T-Crypt/T-Crypt/beside-parked-headroom
T-Crypt Nov 12, 2026
ae5e0df
feat(broker): size a lease from the engine's measured peak with learn…
T-Crypt Nov 12, 2026
9e288c4
Merge pull request #313 from T-Crypt/T-Crypt/learn-size
T-Crypt Nov 12, 2026
9da5a20
feat(status): show every engine's window in berth status and the harbor
T-Crypt Nov 13, 2026
b1131db
feat(broker): berth train locks its cards for a training run
T-Crypt Nov 13, 2026
ab79f4f
docs: refresh the harbor screenshots, How it works and the README
T-Crypt Nov 13, 2026
03e1579
Merge pull request #314 from T-Crypt/T-Crypt/train-lock
T-Crypt Nov 13, 2026
91a3024
Merge remote-tracking branch 'origin/next' into T-Crypt/readme-refresh
T-Crypt Nov 13, 2026
6128976
Merge pull request #315 from T-Crypt/T-Crypt/readme-refresh
T-Crypt Nov 13, 2026
b969900
Merge remote-tracking branch 'origin/next' into T-Crypt/status-schedules
T-Crypt Nov 13, 2026
a42821e
Merge pull request #316 from T-Crypt/T-Crypt/status-schedules
T-Crypt Nov 13, 2026
bdb9ff2
feat(broker): a training keeps its lock while berthd is away
T-Crypt Nov 13, 2026
6e2dc74
Merge pull request #317 from T-Crypt/T-Crypt/train-fail-closed
T-Crypt Nov 13, 2026
383c09d
feat(runner): berth train --exclusive locks the card in the driver
T-Crypt Nov 13, 2026
1bfd344
Merge pull request #318 from T-Crypt/T-Crypt/train-exclusive
T-Crypt Nov 13, 2026
50d0da4
feat(wire): name the llama-swap groups that swap wired models themselves
T-Crypt Nov 13, 2026
82f0399
Merge pull request #319 from T-Crypt/T-Crypt/llama-swap-berth-turns
T-Crypt Nov 13, 2026
8f0d44f
fix(wire): --update keeps the words before the berth binary
T-Crypt Nov 13, 2026
e59f109
Merge pull request #320 from T-Crypt/T-Crypt/wire-update-keeps-prefix
T-Crypt Nov 13, 2026
40a27ed
feat(contract): berth contract ENGINE shows one engine's terms
T-Crypt Nov 13, 2026
a7eb6f2
Merge pull request #321 from T-Crypt/T-Crypt/berth-contract
T-Crypt Nov 13, 2026
074de41
refactor(fit): move the gateway's reach into internal/fit
T-Crypt Nov 13, 2026
78e6633
Merge pull request #322 from T-Crypt/T-Crypt/fit-read-only
T-Crypt Nov 13, 2026
0664338
feat(fit): berth fit NAME shows where a request would run and why
T-Crypt Nov 13, 2026
098ef3c
Merge pull request #323 from T-Crypt/T-Crypt/berth-fit
T-Crypt Nov 13, 2026
71b5208
feat(queue): berth queue lets the operator bump, re-rank or drop a wa…
T-Crypt Nov 13, 2026
f8ad189
Merge pull request #324 from T-Crypt/T-Crypt/queue-override
T-Crypt Nov 13, 2026
718eea4
feat(harbor): a QUEUE panel with FIRST, priority and DROP for a paire…
T-Crypt Nov 13, 2026
2391a9b
Merge pull request #325 from T-Crypt/T-Crypt/harbor-queue
T-Crypt Nov 13, 2026
564ce08
fix(fit,contract): read a llama-swap model's terms, list and running …
T-Crypt Nov 13, 2026
b333879
Merge pull request #326 from T-Crypt/T-Crypt/llama-swap-model-terms
T-Crypt Nov 13, 2026
c6a7ac2
feat(gateway): a not_ready 503 says the lease's place in line
T-Crypt Nov 13, 2026
284f233
Merge pull request #327 from T-Crypt/T-Crypt/queue-position-headers
T-Crypt Nov 13, 2026
64c34da
feat(schedule): book each card's hours: runs, priority changes and do…
T-Crypt Nov 14, 2026
5730483
Merge pull request #328 from T-Crypt/T-Crypt/gpu-scheduler
T-Crypt Nov 14, 2026
d78f0ae
fix(runs): a re-attach goes on with the turn it was in
T-Crypt Nov 14, 2026
20c64a5
Merge pull request #329 from T-Crypt/T-Crypt/reattach-keeps-turn
T-Crypt Nov 14, 2026
3e1da2b
feat(events): keep the event stream across restarts, and berth why ID
T-Crypt Nov 14, 2026
1c0031c
fix(cli): a subcommand's help names its group
T-Crypt Nov 14, 2026
19c4fc5
style(cli): gofmt the Context fields
T-Crypt Nov 14, 2026
53370a0
Merge pull request #330 from T-Crypt/T-Crypt/decision-log
T-Crypt Nov 14, 2026
b7834b5
Merge remote-tracking branch 'origin/next' into T-Crypt/subcommand-he…
T-Crypt Nov 14, 2026
ecf2a27
Merge pull request #331 from T-Crypt/T-Crypt/subcommand-help-path
T-Crypt Nov 14, 2026
d2a6b5a
fix(bench): the swap bench compares reasoning and never calls empty a…
T-Crypt Nov 14, 2026
c6307d3
test(gateway): end each test's berth run before its temp dir goes
T-Crypt Nov 14, 2026
1ab9fd7
fix(runner): a drain waits for a resume in flight
T-Crypt Nov 14, 2026
6276da7
Merge pull request #333 from T-Crypt/T-Crypt/gateway-test-run-cleanup
T-Crypt Nov 14, 2026
9501cdc
Merge remote-tracking branch 'origin/next' into T-Crypt/swap-bench-re…
T-Crypt Nov 14, 2026
8a862ee
Merge pull request #332 from T-Crypt/T-Crypt/swap-bench-reasoning
T-Crypt Nov 14, 2026
5a43998
Merge remote-tracking branch 'origin/next' into T-Crypt/drain-waits-f…
T-Crypt Nov 14, 2026
c66c93c
Merge pull request #334 from T-Crypt/T-Crypt/drain-waits-for-resume
T-Crypt Nov 14, 2026
67ec02b
chore(release): v0.6.0
T-Crypt Nov 14, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
129 changes: 129 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,135 @@ has a matching GitHub release whose body comes from its section below.

## [Unreleased]

## [v0.6.0] - 2026-11-14

v0.6.0 gives each card a schedule and makes Berth explain its choices. `berth schedule` books an engine's hours,
a priority change or downtime up to 14 days ahead, agents book their own runs through the gateway, and the
harbor's SCHEDULE tab shows the week. `berth train` runs a training that nothing automated can interrupt.
`berth fit`, `berth contract`, `berth queue` and `berth why` say where a request would run, what an engine costs,
who is next and why a lease waited or was drained. The event log now survives restarts. It was tagged first as
`v0.6.0-rc.1`.

### Added

- `berth why ID` tells one lease's story in plain words: when it asked for the card, each reason it waited, the
grant, each drain and who it was for, the release, and the berthd restarts it lived through. `berth why ENGINE`
reads that engine's newest lease. A lease berthd still holds ends with where it stands now, and one a restart
cut off names the lease its engine re-attached as.
- berthd keeps its event stream in `events.jsonl` in the state directory (`--events`, `""` for memory only). A
restart reads the last 1000 events back, goes on numbering after them, and goes on with the lease numbers, so
`berth events --since N` and `berth why` read across restarts and deploys. The file moves to `events.jsonl.1`
at 8 MiB. Each start is a `node.started` event that names the version.

- A schedule for each card, up to 14 days ahead. `berth schedule` shows what holds each card, the waiters in grant
order and every booking. `berth schedule book` books an engine's run (it goes first, runs at the booking's
priority, and keeps the card against everything but a hold), a priority change for set hours, or downtime (no
new work, holders drain at the start). `move` and `cancel` change a booking, and `--days` and `--earliest`
repeat one or find free hours. Hours nobody booked run by priority, then arrival, as before. berthd keeps the
bookings in `schedule.json`.
- The gateway's `GET`, `POST`, `PATCH` and `DELETE /v1/berth/schedule` let agents and devices read the scheduler
and book, move or cancel their own runs. Downtime and priority changes stay the operator's.
- The harbor's SCHEDULE tab: what runs and waits now, a 7-day board of 24-hour tracks with every booking to the
minute, each engine's window and the live holder, and a list of bookings. A paired device books by clicking an
empty hour and moves or cancels a booking by clicking it. Each booking form says what the booking will do before
you save it.

- `berth train --as NAME [--gpu CARDS] -- CMD` runs a training or fine-tune that nothing automated can interrupt.
It takes the whole of each card it names, waits for them with no deadline, and keeps them until the command
exits. While it holds a card or waits for it, berthd refuses every other request on that card at once: gateway
requests, `berth run`, llama-swap models, scheduled jobs and `berth hold`. Clients get the refusal right away
instead of waiting out their timeout. A second `berth train` waits in line and starts when the first one ends.
Other cards keep serving. `berth train` refuses to start without berthd, and `berth status` shows the lease as
`hold,pinned,training`.
- `berth train --exclusive` also locks the card in the driver on Linux with an NVIDIA card. It puts the granted cards
in exclusive-process compute mode for the run's life, so a program that opens CUDA without Berth cannot open
them, and puts them back to the default mode when the run ends. It needs a sudo rule for `nvidia-smi -c`, and
refuses to start without one, or when another compute process is already on the card.
- A `berth train` run keeps its lock while berthd is away. With berthd down, `berth run` refuses to start an engine
while a training runs on the node, instead of starting it without a lease. A berthd that starts while a training
runs refuses every other request until the training re-attaches, with no time limit, where other engines get
15 seconds.

- On Windows, `berth notify` is a tray icon in the taskbar. Its notifications are Windows' own toasts, and a click
opens the harbor. The tooltip shows the latest event, and keeps it after explorer restarts. The menu opens the
harbor, turns the toasts off and on, lists the last 10 events and quits. `berth service install --notify`
starts it at logon for the current user ([#186](https://github.com/T-Crypt/berth/pull/186)).
- `POST /v1/berth/yield` lets a ComfyUI node that calls an LLM in the middle of a job give the card to that LLM
and take it back. Before, the LLM's drain of ComfyUI waited for the job to end, the job waited for the LLM, and
neither moved until the node's call timed out. The node frees ComfyUI's models in-process and yields. The drain
then parks ComfyUI as it would between jobs, the LLM runs, and `{"end":true}` answers once ComfyUI holds the card
again. A yield runs out after its seconds (300 by default), so a node that dies mid-yield cannot leave it on
([#299](https://github.com/T-Crypt/berth/issues/299)).
- `learn_size = true` on an engine entry, or `berth run --learn-size`, sizes the lease from the most that engine
has been measured holding plus 512 MiB, in place of the whole card. Engines that fit together then share the card
without anyone typing `needs_mib` or keeping it up to date. An engine not yet measured takes the whole card, and
the grant line in `berth events` names the size it used
([#311](https://github.com/T-Crypt/berth/issues/311)).
- `berth status` has a SCHEDULE table and the harbor a SCHEDULES panel. Each lists every engine that has a
`window`, with its hours, whether they are open, and how long until they close or open. An engine shows before its
window opens, and a `berth run --window` lease shows its own hours. The status JSON carries the same rows as
`schedules`, and each lease carries its `window`.
- `berth contract ENGINE` prints one engine's terms in one place: its priority, the VRAM a lease asks for, the last
run's measured peak, the RAM it needs and holds parked, its minimum turn and window, and how long a switch to it
takes from each tier. Each number says whether berth.toml declares it, Berth's default applies, Berth measured
it, or nothing is known. ENGINE can be an `[[engine]]` name, a discovered engine, or a llama-swap model's
`berth run --as` name. Without berthd it prints the declared terms. The status JSON's `footprints` now carry each
engine's `hot_peak_vram_bytes`, `ram_peak_bytes` and `parked_ram_bytes`. A llama-swap model whose last-run
figure sits on its group's entry is now listed under its own name too. A llama-swap model has no berth.toml
entry, so its priority and minimum turn come from its last run's lease, which carried the flags on its
`berth run` line. Its last peak is the most a run under its name held. The footprint entry carries them as
`priority` and `min_turn_s`.
- `berth fit NAME` says where a request for an engine or a model would run right now and what the start costs. It
ranks each candidate by path: already running on the card, a resume from RAM, a free card, a drain of the holder,
or blocked by a VM, a cordon or a pinned lease. Each candidate names its reason, the tier its weights sit in, and
the switch time the gateway measured from that tier, or unknown. For a model, every engine that serves it is
ranked the way the gateway picks one. A model that no `[[engine]]` names or serves, such as a llama-swap model, is
found the way the gateway finds it, on a live engine whose `/v1/models` lists it. `--explain` lists every
candidate. It is read only, and the gateway now picks through the same ranking.
- `berth queue` lists the waiters in the order berthd will grant them, with their lease IDs. `berth queue bump ID`
moves one to the front of its priority class, `berth queue prio ID batch|interactive|foreground` moves it to
another class, and `berth queue drop ID` refuses it now. Holds and trainings cannot be changed this way. Each
change is a `lease.queue_override` event.
- The harbor's SWITCH tab has a QUEUE panel. It lists the waiters in grant order, and a paired device gets FIRST, a
priority picker and DROP on each one. These buttons go from the harbor straight to berthd, never through the
gateway, because agents hold the gateway's API key and a queue change is the operator's to make.
- A gateway request whose start times out (`not_ready`) while its lease is still in line for the card says where
it stands. The 503 carries `X-Berth-Queue-Position` and, when berthd has measured the starts ahead,
`X-Berth-Start-Estimate: N; lower-bound`. `Retry-After` is the larger of the estimate and 30 s. The estimate
leaves out the holder's time left, so it is a minimum, not a promise.
- `berth wire llama-swap` names each llama-swap group where llama-swap swaps two or more wired models itself.
There llama-swap unloads the old model before Berth sees the request, so `min_turn`, the engine ranking and the
juggle guard never apply between them. The JSON carries the groups as `swapping`. The llama-swap section of
contrib/INTEGRATION.md shows the `swap: false` group and the `healthCheckTimeout` that let Berth decide the turns.

### Fixed

- A drain that arrives while a parked engine is still loading back no longer stops it. `berth run` now waits for
the resume to finish, then parks or drains as usual. Before, the park call reached the engine in the middle of
its load, the engine answered 409, and the run fell back to stopping it. That cold-started the LLM on its next
request and could leave it marked unresponsive ([#275](https://github.com/T-Crypt/berth/issues/275)).
- The swap bench (`contrib/bench/swap`) no longer reports `OUTPUT IDENTICAL yes` for thinking models that ran out
of tokens before answering. It now compares the reasoning a model streams as well as its answer, prints
`n/a (empty)` when every run came back empty, asks for 256 tokens by default (`-max-tokens`), and can ask the
model to skip its thinking (`-no-think`).
- A subcommand's help and error hints name its group. `berth engine show --help` printed `berth show`, which is no
command, and a bad flag said to run `berth show -h`.
- A berthd restart no longer gives every running engine a fresh minimum turn. `berth run` now tells the new berthd
how long its engine had held the card, and the re-attached lease goes on with the turn it was in. Before, each
deploy reset the TURN column to the full `min_turn` and kept waiters out for that long again.
- `berth wire llama-swap` now finds a `berth run` line with a flag written before `--as`, such as
`berth run --min-turn 5m --as qwen ...`. Before, it took the model for unwired, and `--write` added a second
`berth run` in front of the first.
- `berth wire llama-swap --write --update` keeps the words before the berth binary on a `berth run` line, such as
a wrapper script or `env FOO=1`. Before, it rewrote the line from the berth binary on and dropped them.

- A model that fills the card no longer loads beside a parked engine it cannot fit next to. Berth granted any
whole-card engine the card beside a parked residual under `residual_max_gib`, so a 23.4 GiB model on a 24 GiB
card started beside a kept-up ComfyUI's 768 MiB context and failed to load. Berth now remembers the most each
engine has held, by its own name inside a llama-swap group, and evicts the parked engine first when that, plus
512 MiB the load needs past its measured peak, does not fit beside it. A smaller model still loads beside it,
and an engine never measured is granted as before.

### Changed

- An engine left without a lease after berthd restarts now tells you whether its `berth run` can still re-attach
Expand Down
50 changes: 43 additions & 7 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@ berth status # who holds the card, who

## How it works

<p align="center"><img src="assets/how-it-works.svg" alt="How berth works. Clients ask for the card: agents such as Hermes and ComfyUI workflows through the berth gateway, one endpoint that starts engines on demand; an engine under berth run; a game; and a VM through berth vfio run. berthd queues them by priority, drains the holder at its boundary (saving KV caches or parking it in RAM), waits for the driver to show the memory free, then grants the lease. gpu0 serves strata while gpu1 is passed through to a VM. Below, a timeline of one 24 GiB card: llama-swap, strata, a game, a VM passage and ComfyUI take turns and never overlap; markers 1 to 4 show one handoff." width="100%"></p>
<p align="center"><img src="assets/how-it-works.svg" alt="How berth works. Clients ask for the card: agents such as Hermes and ComfyUI workflows through the berth gateway, one endpoint, with one URL per card, that starts engines on demand; an engine under berth run; a game or a training run under berth train, which locks its cards; and a VM through berth vfio run. berthd queues them by priority, drains the holder at its boundary (saving KV caches or parking it in RAM), waits for the driver to show the memory free, then grants the lease. gpu0 serves strata while gpu1 is passed through to a VM. Below, a timeline of one 24 GiB card: llama-swap, strata, a game, a VM passage and ComfyUI take turns and never overlap; markers 1 to 4 show one handoff." width="100%"></p>

Every engine asks berthd for the card: through `berth run`, the gateway, ComfyUI's API or a game
claim. When another engine holds it, Berth:
Expand All @@ -42,7 +42,7 @@ claim. When another engine holds it, Berth:
The numbered markers on the timeline are those four steps. A process that holds VRAM without a lease
shows up as a stowaway: Berth reports its PID and memory and leaves it running.

> **Status: alpha, v0.5.0.** Runs daily on an RTX 4090 workstation beside llama-swap, NInfer,
> **Status: alpha, v0.6.0.** Runs daily on an RTX 4090 workstation beside llama-swap, NInfer,
> Strata, ComfyUI and games, and testers run it on Windows. The [docs](docs/README.md) cover what works and has tests; the
> [roadmap](docs/roadmap.md) lists the phases and what is still missing.

Expand Down Expand Up @@ -110,8 +110,21 @@ line goes away.
render farm job, a cracking session, by name, command line or launcher.
- **`berth hold`** frees the card for a call or a game you start yourself. Ctrl-C gives it back. On a node with
several cards, `--gpu` names the ones it takes.
- **`berth run --pinned`** for an app or a training run that must never be stopped mid-work, on one card or
several. See [Training runs](docs/leasing.md#training-runs).
- **`berth train`** runs a training or fine-tune that nothing automated can interrupt, on one card or several.
It takes the whole of each card, and berthd refuses every other request on those cards until it exits, so
agents and scheduled jobs get an answer at once instead of waiting. A second training waits in line. See
[Training runs](docs/leasing.md#training-runs).
- **`berth run --pinned`** for an app that must never be stopped mid-work but may share the card, such as Blender.

**Plans the card's hours**
- **A schedule per card.** `berth schedule book` gives an engine the card for set hours, changes its priority for
them, or books downtime, up to 14 days ahead. A booked run goes first and keeps the card against everything but
a hold. Hours nobody booked run by priority, then arrival. See [Book the card](docs/leasing.md#book-the-card).
- **Agents book their own runs** through `/v1/berth/schedule` on the gateway, and the harbor's SCHEDULE tab shows
the week to the minute. A paired phone books by clicking an empty hour.
- **`berth queue`** lists the waiters in grant order. `bump`, `prio` and `drop` change one by hand.
- **`learn_size = true`** sizes a lease from what the engine held last time, so engines that fit together share
the card without anyone typing `needs_mib`.

**One API for your engines** *(preview)*
- **`berth gateway`**: one OpenAI-compatible port. Name a model and Berth starts its engine,
Expand Down Expand Up @@ -158,6 +171,10 @@ line goes away.
leaderboard; 30 days of history on disk.
- **The harbor web UI** (`berth harbor`): a departures board of leases, a collisions-avoided
counter, charts and the event log. Read-only, loopback by default.
- **Explains itself.** `berth why ENGINE` tells that engine's last lease in plain words: each wait and its
reason, the grant, each drain and who it was for, and the restarts it lived through. `berth fit MODEL` says where
a request would run now and what the start costs, and `berth contract ENGINE` prints an engine's VRAM, RAM, turn
and switch times, each marked declared, measured or unknown. The event log is kept on disk across restarts.
- **Alerts** to the desktop or a webhook, `/metrics` for Prometheus, and `berth export` for
CSV or JSON.

Expand Down Expand Up @@ -253,16 +270,16 @@ which `berth service status` prints.

## Preview

The first two screenshots show the harbor web UI with its built-in demo data, on the CARD tab and on
the Engines tab in advanced mode. The third shows the CARD tab of a live node on an iPhone.
The screenshots show the harbor web UI with its built-in demo data: the CARD tab, the Engines tab in advanced
mode, and the CARD tab at phone width.
[The harbor and telemetry](docs/harbor.md) explains every panel and shows how to add the harbor to an
iPhone's Home Screen.

<p align="center"><img src="assets/preview-harbor.png" alt="Berth harbor: the CARD tab with ticket counter, request KPIs, the tide of card memory by holder, and the model timeline" width="900"></p>

<p align="center"><img src="assets/preview-requests.png" alt="Berth harbor Engines tab, advanced mode: decode throughput per engine, the token ledger, the per-model leaderboard, storage and residency" width="900"></p>

<p align="center"><img src="assets/preview-mobile.png" alt="Berth harbor on an iPhone: the CARD tab with the ticket counter, collisions avoided, the collision tiles and the decode and TPOT KPIs, with CARD, ENGINES and ABOUT in a bottom bar" width="300"></p>
<p align="center"><img src="assets/preview-mobile.png" alt="Berth harbor at phone width: the CARD tab with the ticket counter, collisions avoided, the collision tiles and the decode and TPOT KPIs, with CARD, SWITCH, ENGINES and ABOUT in a bottom bar" width="300"></p>

## Install

Expand Down Expand Up @@ -328,6 +345,25 @@ request to the reply, lease included (2026-10-23):

Strata's numbers and the method are in [Benchmarks](docs/benchmarks.md#swap-numbers).

## A month on the reference workstation

berthd counts every handoff it makes, and the harbor's collisions board shows the counts. These are the counts
from one RTX 4090 workstation between 2026-10-12 and 2026-11-13, running llama-swap, NInfer, Strata, ComfyUI,
coding agents and games:

| What Berth did | Count |
|---|---|
| Requests that waited for the card instead of loading on top of the engine holding it | 1,136 |
| Engines drained at a reply or job boundary to make room | 901 |
| VRAM handed from one engine to the next | 23.9 TiB |
| Processes holding VRAM without a lease, reported and left running | 573 |
| Swaps between two engines of equal priority that the ranking stopped | 57 |
| Releases that stalled and were caught before the card stayed full | 4 |
| Requests refused after their wait ran out | 24 |

In the last 7 days, 23 engines and jobs held the card for 143 lease-hours. Leases that share the card count their
own hours each.

## Roadmap

The phases, the shipped releases and what comes next are in the [roadmap](docs/roadmap.md).
Expand Down
Loading
Loading