Skip to content

chore: sync with internal pipette-clients at 49d8c48b - #27

Merged
m3m0ryh0l3 merged 13 commits into
mainfrom
kyle/internal-pipette-clients-sync
Nov 5, 2026
Merged

m3m0ryh0l3 merged 13 commits into
mainfrom
kyle/internal-pipette-clients-sync

Conversation

@m3m0ryh0l3

@m3m0ryh0l3 m3m0ryh0l3 commented Nov 4, 2026 •

Copy link
Copy Markdown
Collaborator

Problem
This public repository lagged 558 commits behind the internal one, from
db6a093c (the tree of the initial public commit) to 49d8c48b. The public
commits since then were already backported to internal; only the public CI
and release setup was public-only.

Solution

  • Brings the tree up to date with internal main at 49d8c48b:
    • pipette-plan renamed to pipette-ctl, with the campaign, catalog,
      account and auth login verbs; release assets are pipette-ctl-*
    • new runtimes and backends: OVMS, ONNX Runtime GenAI and OpenVINO through
      a local OpenAI server, TensorRT Edge-LLM on Jetson Orin, Apple Core AI,
      and vLLM/SGLang/OpenVINO/ONNX GenAI catalog updates
    • new benchmarks: asr_throughput, asr_max_memory, asr_eval,
      vl_spec_decode, and typed pipette-ops measurement entries
    • pipette-contract fixtures that hold the Swift and Kotlin plan-types
      mirrors to the Rust crate
    • tools/pipette-ios-proxy (drives a device lab's iPhones and installs
      the release build a job pins), tools/gpu-keepalive,
      tools/adb-connector-apk and tools/energy-probes
    • iOS: one app that carries the prismml and lfm3 llama.cpp variants
      beside main, eval sampling and checkpointing, vault archives, headless
      worker commands; Android: Nav3 routes, headless verbs, Kotlin engine
    • methodology docs for the new benchmarks, energy measurement, and the
      Pixel 11 Pro, Galaxy S26 Ultra and iPhone 18 Pro thermal studies
  • Takes pipette-plan-types from public pipette-mgmt at 06daa58f over
    HTTPS, so a build needs no credentials. pipette-ios-proxy keeps its own
    copy of the build-flag prefix, which plan-types no longer exports.
  • Leaves out the two llama.cpp variants that pin private forks
    (lfm3-audio, asr): their submodules, patch series and CI steps.
  • Ports the internal workflows (build.yml, codeql.yml, security.yml,
    pr-body.yml, main-ci-watch.yml) with the public changes kept:
    GitHub-hosted runners, ci/version.sh and the GitHub release from
    release/** pushes, and the v2 sccache namespace. No app token, no
    release.yml, and no copy to a releases bucket.
  • Describes private repositories, lab hosts, the releases bucket, private
    model repositories and home directories in the new content in general
    terms or with placeholders.
  • Corrects the docs that this sync publishes, after a review against the
    code: about 250 contradicted claims, contradictions between docs,
    duplicated passages and history narration. For public readers it numbers
    lab hosts, drops device serials, the lab Wi-Fi name, internal release
    versions and docs/pipette-ios/app-review.md, and names only public
    models.
  • Replaces a contributor's home-directory paths, internal host names, a
    private OpenVINO model in two --help examples, and a device serial in
    pipette-ctl tests with placeholders or public equivalents.

Testing

  • cargo nextest run --workspace --locked --no-fail-fast: 6,758 passed,
    17 skipped; cargo test --doc --workspace: passed
  • cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo doc --workspace --no-deps, ./ci/lint.sh,
    cargo deny check (advisories, bans, licenses, sources) and cargo machete are clean; tools/port-vectors checks with --locked
  • actionlint finds nothing new against the internal workflows, other than
    the xcode-27 runner label, which is GitHub's preview Xcode 27 image
  • The tree differs from internal 49d8c48b only by the removed variants,
    the plan-types source, the ios-proxy prefix constant, the CI changes and
    the scrubs above
  • Not run locally: the Android Gradle build, the iOS app build and the
    Windows-only gpu-keepalive job

Second commit: restores ci/version.sh, which the sync deleted while the release-metadata job still runs it.

Later commits: answers to the review's open questions

  • Top-level docs:
    • credits the Material icons
    • states the model-provider logos as trademarks of their owners
    • points the next parity pass at this PR's merge commit
    • says which license checks CI runs
  • Removes TensorRT Edge-LLM from the public docs, because its server archive comes from a non-public fork.
  • Removes powerd and the in-development meter setup (energy/03-proposal.md), which are not part of the testing conditions.
  • Names each boundary for reading older result rows as a date, not an internal release.
  • Settles facts from the code, git history, the agent-wiki and upstream sources:
    • DSpark's standalone GGUF drafter
    • the CUDA wheel size
    • the VL memory quantities
    • the one-cooler rig: a Black Shark on the back of each phone
    • the MacBook Neo's 20 W adapter
    • the ASR datasets' licenses
    • the Android engine decisions
  • Leaves unchanged the claims that no source settles.

The docs review behind the doc corrections also produced the internal counterpart, Liquid4All/internal-pipette-clients#1851. The code problems it found, which exist in both repositories, are in PIP-864.

Closes PIP-854
Closes PIP-861
Refs PIP-612

m3m0ryh0l3 and others added 2 commits November 4, 2026 15:22
**Problem**
This public repository lagged 558 commits behind the internal one, from
db6a093c (the tree of the initial public commit) to 49d8c48b. The public
commits since then were already backported to internal; only the public CI
and release setup was public-only.

**Solution**
- Brings the tree up to date with internal `main` at `49d8c48b`:
  - `pipette-plan` renamed to `pipette-ctl`, with the campaign, catalog,
    account and `auth login` verbs; release assets are `pipette-ctl-*`
  - new runtimes and backends: OVMS, ONNX Runtime GenAI and OpenVINO through
    a local OpenAI server, TensorRT Edge-LLM on Jetson Orin, Apple Core AI,
    and vLLM/SGLang/OpenVINO/ONNX GenAI catalog updates
  - new benchmarks: `asr_throughput`, `asr_max_memory`, `asr_eval`,
    `vl_spec_decode`, and typed `pipette-ops` measurement entries
  - `pipette-contract` fixtures that hold the Swift and Kotlin plan-types
    mirrors to the Rust crate
  - `tools/pipette-ios-proxy` (drives a device lab's iPhones and installs
    the release build a job pins), `tools/gpu-keepalive`,
    `tools/adb-connector-apk` and `tools/energy-probes`
  - iOS: one app that carries the `prismml` and `lfm3` llama.cpp variants
    beside main, eval sampling and checkpointing, vault archives, headless
    worker commands; Android: Nav3 routes, headless verbs, Kotlin engine
  - methodology docs for the new benchmarks, energy measurement, and the
    Pixel 11 Pro, Galaxy S26 Ultra and iPhone 18 Pro thermal studies
- Takes `pipette-plan-types` from public `pipette-mgmt` at `06daa58f` over
  HTTPS, so a build needs no credentials. `pipette-ios-proxy` keeps its own
  copy of the build-flag prefix, which plan-types no longer exports.
- Leaves out the two llama.cpp variants that pin private forks
  (`lfm3-audio`, `asr`): their submodules, patch series and CI steps.
- Ports the internal workflows (`build.yml`, `codeql.yml`, `security.yml`,
  `pr-body.yml`, `main-ci-watch.yml`) with the public changes kept:
  GitHub-hosted runners, `ci/version.sh` and the GitHub release from
  `release/**` pushes, and the v2 sccache namespace. No app token, no
  `release.yml`, and no copy to a releases bucket.
- Describes private repositories, lab hosts, the releases bucket, private
  model repositories and home directories in the new content in general
  terms or with placeholders.
- Corrects the docs that this sync publishes, after a review against the
  code: about 250 contradicted claims, contradictions between docs,
  duplicated passages and history narration. For public readers it numbers
  lab hosts, drops device serials, the lab Wi-Fi name, internal release
  versions and `docs/pipette-ios/app-review.md`, and names only public
  models.
- Replaces a contributor's home-directory paths, internal host names, a
  private OpenVINO model in two `--help` examples, and a device serial in
  `pipette-ctl` tests with placeholders or public equivalents.

**Testing**
- `cargo nextest run --workspace --locked --no-fail-fast`: 6,758 passed,
  17 skipped; `cargo test --doc --workspace`: passed
- `cargo fmt --all --check`, `cargo clippy --workspace --all-targets --locked
  -- -D warnings`, `cargo doc --workspace --no-deps`, `./ci/lint.sh`,
  `cargo deny check` (advisories, bans, licenses, sources) and `cargo
  machete` are clean; `tools/port-vectors` checks with `--locked`
- actionlint finds nothing new against the internal workflows, other than
  the `xcode-27` runner label, which is GitHub's preview Xcode 27 image
- The tree differs from internal `49d8c48b` only by the removed variants,
  the plan-types source, the ios-proxy prefix constant, the CI changes and
  the scrubs above
- Not run locally: the Android Gradle build, the iOS app build and the
  Windows-only `gpu-keepalive` job

Co-authored-by: Yuri Khrustalev <ykhrustalev@users.noreply.github.com>
Co-authored-by: David Berrios <iamstuffed@gmail.com>
Co-authored-by: Rouzbeh Amin Tafreshi <rouzbeh.tafreshi@liquid.ai>
Co-authored-by: Roman Khrupa <hruparoman@gmail.com>
Co-authored-by: Nick Prokopiev <46002699+nickprokopiev@users.noreply.github.com>
Co-authored-by: MwC-Trexx <MwC-Trexx@users.noreply.github.com>
**Problem**
The sync commit deleted `ci/version.sh`. The `release-metadata` job in
`build.yml` still runs it, and every release, `android-app`, `ios-archive` and
`build-artifact` run depends on that job, so each of them would fail.
`CONTRIBUTING.md` also links to the script.

**Solution**
- Restores `ci/version.sh` unchanged from `main`.

**Testing**
- `ci/version.sh` prints a `-dev` version for this branch.

Closes PIP-854
**Problem**
Four statements in the top-level docs were unconfirmed or pointed public
readers at things they cannot reach: an internal commit as the last parity
pass, the four tab icons with no attribution, Rust as a need of the Android
assemble tasks, and a license check "at submodule bump" that CI only partly
runs.

**Solution**
- `client-parity.md`: points the next parity pass at the commit that merges
  #27 instead of an internal SHA.
- `THIRD_PARTY_NOTICES.md`: credits the four `ic_tab_*` drawables to Google's
  Material Symbols and Material Icons (Apache-2.0). Their path data matches the
  published icons.
- `CONTRIBUTING.md`: the `kotlinEngine` assemble tasks need the NDK and the
  CMake build, not Rust.
- `licensing.md`: the Android CI job checks the bundled vendored license texts;
  on iOS, `ios/generate-third-party-licenses.py` is a manual step after a
  submodule bump.
**Problem**
THIRD_PARTY_NOTICES.md listed the model provider logos in both apps as
"Trademark/provenance review needed", an open item in a published notice.

**Solution**
- Names the third-party logos (Google Gemma, IBM Granite, Meta, Mistral AI,
  Qwen) and their asset paths, and states that they are trademarks of their
  owners, shown only to identify a model's provider and not covered by the
  repository license. The Liquid logos are marked as Liquid AI's own.
**Problem**
The pipette-cli docs still held claims that the review could not confirm, and
documented TensorRT Edge-LLM, whose server archive comes from a fork that
public readers cannot get.

**Solution**
- Removes TensorRT Edge-LLM from the public docs: its page, its rows in the
  usage and model tables, and its mentions in the methodology, architecture
  and licensing docs.
- DSpark: the public release ships the GGUF drafter on its own, so the
  `gguf-vision-speculative` example pairs `LFM2.5-VL-3B-GGUF` with
  `LFM2.5-VL-3B-DSpark-GGUF` by URL.
- Names release 2026.08.1 as the last one with the older `benchmarks/remote/`
  layout.
- `GGML_CUDA_GRAPH_OPT`: states the measured condition (one visible GPU per
  server); llama.cpp has no single-GPU limit on it.
- torch-oai: Linux is the only host it has run on; the CLI does not refuse
  others.
- Drops the model families from the Windows `STATUS_STACK_OVERFLOW` note.
**Problem**
Three statements could not be confirmed: that every CI build of the iOS app is
a private-API build, the decode tok/s column of the OpenVINO precision table,
and the size of the CUDA wheels for ONNX GenAI.

**Solution**
- job-generation.md: the farm's app builds are private-API builds; the claim
  about CI builds is gone.
- openvino-ir.md: drops the tok/s column. Two of its rows do not follow from
  the decode times beside them, and no source says which figure is right.
- onnx-genai-measurement.md: the seven pinned `nvidia-*` wheels come to about
  1.3 GB (1.2 GiB) on Linux x86-64, for each CUDA row.
**Problem**
About 50 statements in the methodology, iOS, Android, tools and thermal docs
were unconfirmed or contradicted each other, and many notes about older rows
named no boundary.

**Solution**
- Settles each answered question from the agent-wiki, the code, git history,
  the pinned pipette-mgmt, the vLLM and clerk-ios sources, or the Hugging
  Face Hub.
- Names a boundary for each note about older rows. Public docs state it as a
  date; internal docs give the first release tag.
- Corrects claims the sources contradict, such as the cooler setup per phone,
  the S26 resting temperature, the M4 Max in the fleet table, the
  `vl_spec_decode` catalog, the GMKtec power-raise figure, the `backend=` output
  of llama-bench and the gpu-keepalive hosts.
- Separates the measured from the inferred: the VL memory figures name the
  quantity each one measures, and the Android engine port records the Phase 4
  criterion that Section 7 set.
- Leaves unchanged the claims that no source settles.
**Problem**
Three Android statements waited on the owner: whether `kotlinEngine` stays the
default after the Phase 4 finding, why the Compose UI has no Hugging Face token
field, and what staging does with a missing allowlisted variant.

**Solution**
- engine-port.md: `kotlinEngine` stays the default flavor by design.
- implementation.md: the missing token field is deliberate. Lab benchmarks run
  through the CLI, and the app serves public real-world tests that need no
  token.
- build.md: staging leaves out an allowlisted variant that ggml does not build,
  without an error, as long as one variant stages.
**Problem**
asr-eval.md called SPGISpeech "licensed for non-commercial use only". Its
Kensho Terms of Usage say "academic research purposes and internal use only",
with no redistribution. The sentence also left out the URGENT 2024 and
GigaSpeech licenses.

**Solution**
- States each license: SPGISpeech under Kensho's Terms of Usage (academic
  research and internal use only, no redistribution without written consent),
  URGENT 2024 under CC-BY-NC-SA-4.0, GigaSpeech under Apache-2.0 behind a
  Hugging Face access gate.
- Drops "non-commercial" from the usage sentence; only URGENT's license
  uses that word.
…docs

**Problem**
The public energy docs described powerd, the lab's power service, and a
planned setup of several KM003C meters as if they were part of the testing
conditions. powerd is not deployed, and the meter setup is still in
development.

**Solution**
- Removes `docs/methodology/energy/03-proposal.md`, a plan built on powerd and
  the planned meters, and its link from the docs index.
- Removes the powerd, Shelly and KM003C passages and the links to the proposal
  from 01-landscape.md and 02-device-capabilities.md. The probe results stay.
**Problem**
The docs did not say which adapter the MacBook Neo runs on, and the internal
energy proposal said that every fleet MacBook runs on a 140 W unit.

**Solution**
- device-conditions.md: laptops run on the adapter they ship with, which can
  supply less than the laptop draws. The Neo ships with a 20 W adapter and can
  draw up to 33 W, so under load it can also draw from its battery.
**Problem**
device-conditions.md named two coolers per phone, one on the screen and one on
the back, with the Neveika 001 as the standard. Every phone now carries one
Black Shark Live Streaming Cooler on its back; the Neveika coolers are retired
so that no cooler touches the screen. The thermal studies also called their
October Neveika setup "the fleet setup".

**Solution**
- device-conditions.md: states the one-cooler rig and why no cooler sits on the
  screen (phantom touches). Keeps the cooler sweep as data, notes that it has no
  row for the rig's cooler, and updates the fleet table.
- Rewrites four settings rationales that assumed a cooler against the panel
  (home-screen icons, brightness on Android and iOS, accidental touch
  protection).
- Galaxy S26 Ultra, Pixel 11 Pro and iPhone 18 Pro studies: point to the
  current rig in their measurement-rig sections, and date the Neveika setup as
  the October 2026 fleet setup.
**Problem**
`a_peak_reads_ten_times_as_often_as_an_observation` required one peak read per
50 ms of wall clock. On GitHub's hosted macOS runners, each 10 ms tick waits on
file reads and stretched to about 70 ms, so the test failed twice on this PR
(10 reads in 705 ms, 14 in 764 ms). The code under test was correct.

**Solution**
- Asserts that the peak sampler reads more often than the observation sampler
  that runs beside it. A busy machine slows both, and a peak sampler on the
  100 ms tick would still fail.
- Keeps the observation's upper bound unchanged.

**Testing**
- `cargo nextest run -p pipette-memprobe`: the test passes 20 of 20 times idle
  and 10 of 10 with every core saturated (`yes` x 28 on 14 cores).
- `cargo fmt --all --check`, `cargo clippy -p pipette-memprobe --all-targets`.
@m3m0ryh0l3
m3m0ryh0l3 merged commit 8715581 into main Nov 5, 2026
57 of 59 checks passed
@m3m0ryh0l3
m3m0ryh0l3 deleted the kyle/internal-pipette-clients-sync branch November 5, 2026 15:48
@github-actions github-actions Bot mentioned this pull request Nov 9, 2026
m3m0ryh0l3 added a commit that referenced this pull request Nov 9, 2026
**Problem**
This public repository lagged 19 commits behind the internal one, from
49d8c48b (the last sync, #27) to be6ca299.

**Solution**
- Brings the tree up to date with internal `main` at `be6ca299`:
  - an attempt limit, `attempt_timeout_seconds` (300 s default), that stops
    a cell's attempt and reports a `failure_code`
  - `--eval-shard` to run one group of an eval cell into a part file
  - the CLI worker turns Android doze off at start
  - iOS carries the public PrismML llama.cpp `prism-b10828-a153697` beside
    `prism-b10777` (new submodule `vendor/llamacpp-prismml-b10828`)
  - the `pipette-ctl` MCP server is removed
  - the Core AI sidecar moves to `pipette-artifacts` and installs in the
    runtime store
  - the internal docs fixes that followed the #27 review
- Pins `pipette-plan-types` to public `pipette-mgmt` at `b74d06c5`, which
  carries the `FailureCode` type the attempt limit needs.
- `pipette-ios-proxy` takes internal's crate-level
  `RUNTIME_BUILD_FLAG_PREFIX` in place of its public copy.
- Keeps the public scrubs: no `lfm3-audio` or `asr` llama.cpp variants, no
  app-token step or private pins in CI, no TensorRT Edge-LLM docs, no
  private repository links, lab host names, release tags or power-lab
  details.
- Points `docs/client-parity.md` at the #27 sync commit `8715581`, and
  corrects claims that a check against the code found: `pipette-artifacts`
  builds the Core AI sidecar, `--benchmark-flags` has five single-setting
  flags, the iOS 300 s default covers the throughput cells, and the
  attempt-clock diagram names no TensorRT engine build.

**Testing**
- `cargo nextest run --workspace --locked --no-fail-fast`: 6,271 passed,
  12 skipped
- `cargo fmt --all --check` and
  `cargo clippy --workspace --all-targets --locked -- -D warnings` are clean
- `./ci/lint.sh` passes every check, shellcheck included
- Not run locally: the Android Gradle build, the iOS app build and the
  Windows-only `gpu-keepalive` job
m3m0ryh0l3 added a commit that referenced this pull request Nov 9, 2026
**Problem**
This public repository lagged 19 commits behind the internal one, from
`49d8c48b` (the #27 sync) to `be6ca299`.

**Solution**
- Adds `attempt_timeout_seconds` (300 s default), `--eval-shard`, doze
off at Android worker start, and the PrismML `prism-b10828` llama.cpp
pin on iOS
- Removes the `pipette-ctl` MCP server and moves the Core AI sidecar
build into `pipette-artifacts`
- Pins `pipette-plan-types` to public `pipette-mgmt` at `b74d06c5`,
which carries `FailureCode`
- Keeps the public scrubs: no private llama.cpp variants, private CI
steps, TensorRT Edge-LLM docs, private repository links, lab host names
or release tags
- Brings in the internal docs fixes after the #27 review, and corrects
four claims that a check against the code found

**Testing**
- `cargo nextest run --workspace --locked --no-fail-fast`: 6,271 passed,
12 skipped
- `cargo fmt --all --check`, `cargo clippy --workspace --all-targets
--locked -- -D warnings` and `./ci/lint.sh` are clean
- Not run locally: the Android Gradle build, the iOS app build and the
Windows-only `gpu-keepalive` job
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant