…831)

The `/template` Ray-bump automation had no per-template depset recipe,
so the Ray 2.56.0 batch behaved inconsistently. This teaches the skill
(and the cursor preflight) how depsets interact with a bump. **Docs +
preflight only — no CI or lockfile changes** (those are #834 / #835 /
#833).

## Why
- **No per-template depset recipe** — `bump-ray-version.md` only pointed
at the whole-repo *batch* `upgrade-dependencies.md`; nothing covered
no-lock templates or how to read the repo-global `check-depsets` gate.
- **#819 was a workflow drop** — the cleanest 2.56.0 PR stuck because
the agent opened a draft and never dispatched `/test-template`.
- **`.cursor` was never the block** — `raydepsets` bundles `uv` and
recompiled fine in the cursor env; but preflight never smoke-tested the
toolchain, so a real breakage would only surface mid-bump.

## Changes
- `bump-ray-version.md` — per-template depset recipe (no-lock vs
has-lock), batch-vs-per-template preamble,
`check-depsets`-is-repo-global note, explicit **Done criteria**.
- `upgrade-dependencies.md` / `dependencies.md` — label the human/batch
path; document the gate's collateral behavior.
- `.cursor/preflight.sh` — depset-toolchain smoke test (fail at
preflight, not mid-bump).

## Testing
`pre-commit` on changed files passed. Docs/preflight only — no runtime
impact; validated end-to-end by retriggering `template-updater` against
the fixed skill.

---------

Signed-off-by: Aydin Abiar <aydin@anyscale.com>

…875)

## Summary
Bump `ray-tune-train-integration` to Ray 2.56.0.

## Changes
- `BUILD.yaml`: `cluster_env.image_uri` →
`anyscale/ray:2.56.0-py312-cu129`.
- `dependencies/template.depsets.yaml`: repoint the templates `expand`
entry from `ray2551_py312_cu128` → `ray2560_py312_cu128`. The
`ray2560_py312_cu128` bundle and base `ray_2.56.0_img_py312.lock`
already landed via #833, so only the expand pointer flipped (Depset-bump
SHORTCUT).
- `templates/ray-tune-train-integration/python_depset.lock`: recompiled
via `./update_deps.sh --name
ray_tune_train_integration_depset_2.56.0_3.12`; byte-identical to the
previous lock (requirements don`t pull Ray transitively;
`--unsafe-package ray` in append_flags).

## Tests / validation
- **Local:** `pre-commit run --all-files` — passed.
- **CI:** all GH-Actions checks green (`pre-commit`, `build`,
`validate-build-yaml`, `check-depsets`). Buildkite `template-test` build
#478 passed (~15m58s) —
https://buildkite.com/anyscale/template-test/builds/478.

## Publish
Automation tokens `BUILDKITE_API_TOKEN` lacks the `write_builds` scope,
so this run cannot trigger the `tmpl-publish` build itself. After merge,
a reviewer must trigger it manually per
`.claude/skills/template/references/publish-to-backend.md`:
- Pipeline: https://buildkite.com/anyscale/tmpl-publish
- `input-tmpl-name` fields: `tmpl-name=ray-tune-train-integration`,
`tmpl-branch=main`, `tmpl-commit=HEAD`.

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Aydin Abiar <Aydin-ab@users.noreply.github.com>

…887)

## Summary
Bump llm_batch_inference_vision to Ray 2.56.0.

## Changes
- `BUILD.yaml`: `cluster_env.image_uri` →
`anyscale/ray-llm:2.56.0-py312-cu130` (was `2.55.1-py311-cu128`). Ray
2.56.0 `anyscale/ray-llm` only ships `py312-cu130`, so this also moves
the Python/CUDA pair (tag verified on Docker Hub).
- `templates/llm_batch_inference_vision/job.yaml`: `image_uri` →
`anyscale/ray-llm:2.56.0-py312-cu130`.
- `templates/llm_batch_inference_vision/README.md` + `README.ipynb`:
in-doc `image_uri` example → `anyscale/ray-llm:2.56.0-py312-cu130`.
- `dependencies/template.depsets.yaml`: repoint this template's `expand`
entry `build_arg_sets` `ray2551_py311_cu128` → `ray2560_py312_cu130`;
PyTorch `--index` → `cu130`. The `ray2560_py312_cu130` bundle +
`rayllm_2.56.0_py312_cu130.lock` base already existed from #833's
follow-up, so only this expand pointer + index flipped.
- `templates/llm_batch_inference_vision/requirements.txt`: refresh the
comment to reference `anyscale/ray-llm:2.56.0` and pin
`huggingface-hub==1.13.0` to match the new base image (was `0.36.2`).
`numpy==1.26.4`, `pandas==2.3.3`, `pyarrow==19.0.1`, and
`datasets==3.6.0` kept unchanged — the 2.56.0 ray-llm image still ships
numpy 1.x + pyarrow 19 + scipy compiled against numpy 1.x (verified from
the CI package diff), so keeping the numpy-1 stack is what actually lets
the layered lock reinstall as a no-op.
- `templates/llm_batch_inference_vision/batch_inference_vision.py` +
`batch_inference_vision_scaled.py` + `README.ipynb`/`README.md`: migrate
`ray.data.llm.build_llm_processor` → `build_processor`. Ray 2.56.0
removed the deprecated `build_llm_processor` alias in
ray-project/ray#63569.
- Regenerated `templates/llm_batch_inference_vision/python_depset.lock`
via `./update_deps.sh --name
llm_batch_inference_vision_depset_2.56.0_3.12_cu130`.

## Fix iterations
Two failure→fix cycles (both agent-fixable, not infra):
1. **Build #490** — `ImportError: A module that was compiled using NumPy
1.x cannot be run in NumPy 2.2.6` from `scipy.io._sparsetools` on first
`import datasets`. The Ray 2.56.0 rayllm deplock declares
`numpy==2.2.6`, but the actual `anyscale/ray-llm:2.56.0-py312-cu130`
image still ships numpy 1.26.4 + pyarrow 19.0.1 (with scipy compiled
against numpy 1.x). Fix: pin the layered lock to the runtime versions
(`numpy==1.26.4`, `pyarrow==19.0.1`), matching the classic "un-pinned
floats" gotcha in `references/dependencies.md`.
2. **Build #494** — `ImportError: cannot import name
'build_llm_processor' from 'ray.data.llm'`. Ray 2.56.0 removed the
deprecated alias (ray-project/ray#63569). Fix: rename to
`build_processor` in the two `.py` files and in the notebook +
regenerated README.

## Tests / validation
- **Local:** `pre-commit run --all-files` — passed.
- **CI:** Buildkite [`template-test` build
#497](https://buildkite.com/anyscale/template-test/builds/497) — passed
(~12m 40s).

## Publish
`BUILDKITE_API_TOKEN` provided to this Cloud Agent lacks the
`write_builds` scope, so this run cannot programmatically `create_build`
on `tmpl-publish` (see
`.claude/skills/template/references/publish-to-backend.md`). After
merge, a reviewer should trigger a fresh `tmpl-publish` build
(`org_slug=anyscale`, `pipeline_slug=tmpl-publish`, `branch=master`,
`message=llm_batch_inference_vision`) and drive the manual gates dev →
staging → production per that reference. Pipeline:
https://buildkite.com/anyscale/tmpl-publish.

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Aydin Abiar <Aydin-ab@users.noreply.github.com>