wai-quantum: an optimisation stops and goes on to the same bytes #608

Merged

dcharlot merged 2 commits from wai/quantum-optimize-resume-2 into main 2026-11-04 10:26:30 -05:00

Conversation 0 Commits 2 Files changed 11 +601 -54

dcharlot commented 2026-11-03 17:48:01 -05:00

Owner

Copy link

Replaces #597: the same commit (same patch id), on main at 2bef11f71. Main had moved by 15 commits, among them #603, which runs each workflow's gate on linux rather than behind the arm64 work it gates; a pull request runs its own head's workflows, so #597 would have run the old ones.

A long optimisation ran to its end in one process or not at all, and one killed partway lost every step. The optimiser now stops between any two steps and goes on later, to the same bytes.

What changes

wai-quantum (engine::variational)

• optimize_resumable(request, from, each) runs from a checkpoint (from) or from the start. After every step but the last, each gets the run as it stands (Paused), and returns whether to go on. call_optimize is the same with call's checks and stamp. optimize is optimize_resumable with nothing stopping it, so its answers are unchanged.
• The checkpoint (wai.quantum.optimize-checkpoint/1, from Paused::checkpoint) holds:
  • the request, by the chain's start and the request's digest;
  • the steps taken, each with its link, and the last link;
  • the optimiser's whole state: the point, Adam's moments and the powers of its betas, SPSA's stream, the best point, the evaluation count and the gradient's rounding bound. Every number is the hex of its f64's bits, so a resume never depends on how a JSON reader rounds.
• Refusals: a checkpoint goes on only from its own request, and only if each step is still the link its digest says. Another request, an edited step, a dropped step, a next step past the end, a state that does not read, and parameters out of range are each refused with J052.
• A changed checkpoint whose steps still chain would give a different path. That path's answer cannot pass verify, because verify runs the request again.

wai-quantum-cli: optimize --checkpoint FILE keeps the state in FILE as the run goes, at most once a second and whenever it stops. Each write goes to FILE.partial and is renamed, so a kill mid-write leaves the last whole state. The command goes on from FILE when FILE is there, and removes FILE once the run ends. --slice N takes at most N steps, then stops with the state kept and exits 2 (as task status does while a task is active). The Chinese summary of optimize is translated from its new English.

The docs (docs/variational.md, the CLI README) and the CLI's agent docs are updated.

Tests

• The engine: for gradient descent, Adam, SPSA and Adam with the parameter-shift rule:
  • stopped after each step in turn, written as text, read back and gone on with, every run is the uninterrupted run's bytes;
  • taken one step a slice, every run is too.
• Refusals: a checkpoint of another program, or of more steps, is refused. So are eight changed copies: a step's value, a dropped step, the chain, the next step, the parameters, the stream, the format, and parameters that are not finite. The checkpoint's own run ends as one that never stopped.
• The command line, against its binary:
  • 41 steps taken seven at a time end in six slices, with stdout byte for byte the uninterrupted run's (receipt included, signed with the demo key) and the checkpoint removed;
  • another request's checkpoint is refused and left as it was;
  • a 16-qubit, 600-step run killed partway with SIGKILL once it has kept a state, then started again, says it is going on, and prints exactly the uninterrupted run's bytes.
• The variational workflow requires the two new engine tests by name.

Checked

On the build host: the engine's variational, algorithm, schema and refusal tests; cargo check with -D warnings for the engine's feature sets; and the command line's tests (the new ones among them) and clippy, with and without default features. The public-name check is clean.

🤖 Generated with Claude Code

Replaces #597: the same commit (same patch id), on main at 2bef11f71. Main had moved by 15 commits, among them #603, which runs each workflow's gate on `linux` rather than behind the arm64 work it gates; a pull request runs its own head's workflows, so #597 would have run the old ones. A long optimisation ran to its end in one process or not at all, and one killed partway lost every step. The optimiser now stops between any two steps and goes on later, to the same bytes. ## What changes **wai-quantum** (`engine::variational`) - `optimize_resumable(request, from, each)` runs from a checkpoint (`from`) or from the start. After every step but the last, `each` gets the run as it stands (`Paused`), and returns whether to go on. `call_optimize` is the same with `call`'s checks and stamp. `optimize` is `optimize_resumable` with nothing stopping it, so its answers are unchanged. - **The checkpoint** (`wai.quantum.optimize-checkpoint/1`, from `Paused::checkpoint`) holds: - the request, by the chain's `start` and the request's digest; - the steps taken, each with its link, and the last link; - the optimiser's whole state: the point, Adam's moments and the powers of its betas, SPSA's stream, the best point, the evaluation count and the gradient's rounding bound. Every number is the hex of its f64's bits, so a resume never depends on how a JSON reader rounds. - **Refusals:** a checkpoint goes on only from its own request, and only if each step is still the link its digest says. Another request, an edited step, a dropped step, a next step past the end, a state that does not read, and parameters out of range are each refused with `J052`. - A changed checkpoint whose steps still chain would give a different path. That path's answer cannot pass `verify`, because `verify` runs the request again. **wai-quantum-cli:** `optimize --checkpoint FILE` keeps the state in FILE as the run goes, at most once a second and whenever it stops. Each write goes to `FILE.partial` and is renamed, so a kill mid-write leaves the last whole state. The command goes on from FILE when FILE is there, and removes FILE once the run ends. `--slice N` takes at most N steps, then stops with the state kept and exits 2 (as `task status` does while a task is active). The Chinese summary of `optimize` is translated from its new English. The docs (`docs/variational.md`, the CLI README) and the CLI's agent docs are updated. ## Tests - **The engine:** for gradient descent, Adam, SPSA and Adam with the parameter-shift rule: - stopped after each step in turn, written as text, read back and gone on with, every run is the uninterrupted run's bytes; - taken one step a slice, every run is too. - **Refusals:** a checkpoint of another program, or of more steps, is refused. So are eight changed copies: a step's value, a dropped step, the chain, the next step, the parameters, the stream, the format, and parameters that are not finite. The checkpoint's own run ends as one that never stopped. - **The command line, against its binary:** - 41 steps taken seven at a time end in six slices, with stdout byte for byte the uninterrupted run's (receipt included, signed with the demo key) and the checkpoint removed; - another request's checkpoint is refused and left as it was; - a 16-qubit, 600-step run killed partway with `SIGKILL` once it has kept a state, then started again, says it is going on, and prints exactly the uninterrupted run's bytes. - The variational workflow requires the two new engine tests by name. ## Checked On the build host: the engine's variational, algorithm, schema and refusal tests; `cargo check` with `-D warnings` for the engine's feature sets; and the command line's tests (the new ones among them) and clippy, with and without default features. The public-name check is clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 1 commit 2026-11-03 17:48:01 -05:00

wai-quantum: an optimisation stops and goes on to the same bytes ... 980fffc34d

A long optimisation ran to its end in one process or not at all; one
killed partway lost every step. The engine's optimiser now runs from a
checkpoint (engine::variational::optimize_resumable, call_optimize): after
every step its caller has the run as it stands and may stop it there, and
the checkpoint, wai.quantum.optimize-checkpoint/1, holds the steps taken
with their links and the optimiser's whole state (the point, Adam's
moments and the powers of its betas, SPSA's stream, the best point and the
counts), every number as the bits of its f64. Gone on with from any step,
a run ends with the same bytes as one that never stopped, for each
optimiser; a checkpoint of another request, or one whose steps were
changed, is refused (J052).

wai-quantum optimize --checkpoint FILE keeps the state in FILE as it goes
(at most once a second, and when it stops; written whole or not at all),
goes on from FILE when it is there, and removes it when the run ends; a
run killed and started again with the same arguments ends with the same
bytes, receipt and all. --slice N takes at most N steps, then stops with
the state kept (exit 2).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot referenced this pull request 2026-11-03 17:48:23 -05:00

wai-quantum: an optimisation stops and goes on to the same bytes #597

dcharlot added 1 commit 2026-11-03 23:44:08 -05:00

wai-quantum docs: the checkpoint example runs, and shows its claim ... 0636f34e20

The page's --checkpoint example was a shell block marked neither runnable
nor norun, which the docs check refuses. It now runs: a whole run, the same
run stopped by --slice (exit 2) and gone on with, and cmp of the two answers,
so the page holds the command line to "the same bytes as a run that never
stopped", receipt and all.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot merged commit e00d956f2e into main 2026-11-04 10:26:30 -05:00

dcharlot referenced this pull request from a commit 2026-11-04 10:26:33 -05:00

Merge pull request 'wai-quantum: an optimisation stops and goes on to the same bytes' (#608) from wai/quantum-optimize-resume-2 into main

dcharlot referenced this pull request 2026-11-04 10:33:27 -05:00

wai-quantum-service: a job the service was running when it stopped runs once more #613

dcharlot referenced this pull request 2026-11-04 15:47:49 -05:00

wai-quantum 0.8.1, wai-quantum-qpu 0.4.1, wai-quantum-cli 0.2.4, the job service 0.2.2 and the Python package 0.1.2 #619

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Transaction-Science/open-standards!608

No description provided.