Remove the unused TRL dev Docker image

[albertvillanova]

This PR removes the TRL dev Docker image, which has had no consumer for almost a year.

Motivation

The Docker images were introduced in #1215 as pre-baked GPU environments for the slow tests workflow, which was created one day later in #1223 and consumed them as a container: matrix. The dev image is the descendant of huggingface/trl-source-gpu, named "Latest TRL + HF ecosystem from source", whose only purpose was to give the slow tests an environment built from the main branch.

#3931 replaced the two old images with huggingface/trl and huggingface/trl:dev. Three days later, #4085 switched the slow tests to pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel with uv pip install ".[dev]", dropping both images.

Since then, huggingface/trl:dev has had no consumer:

• No workflow references it.
• It is not mentioned anywhere in the documentation, unlike huggingface/trl, which the Jobs guide documents as --image huggingface/trl.

It is still rebuilt and pushed on every single push to main, which is about 150 builds per month of a 9.6 GB image that nothing pulls.

Removing it is also a prerequisite for building the released image on release rather than on every push to main: the dev image tracks the main branch, so it is the only reason the workflow needs to run per commit.

Changes

• Remove the dev image Dockerfile
• Remove the dev image build job from the Docker build workflow

Before merging

Does anyone rely on huggingface/trl:dev? It has never been documented, so I assume not, but I would rather confirm than break an undocumented workflow. Keeping this as a draft until then.

CC: @huggingface/trl

---

Note

Low Risk
CI-only cleanup with no application or runtime code changes; only removes an undocumented image that nothing in the repo references.

Overview
Stops building and publishing huggingface/trl:dev by deleting the trl-dev workflow job and removing docker/trl-dev/Dockerfile.

The main docker-build workflow now only builds the release huggingface/trl image (version-tagged from PyPI plus latest), with Slack notification unchanged for that job.

^{Reviewed by Cursor Bugbot for commit 26b5a89. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[qgallouedec]

We can't really know. it's a public image, someone could be pulling it without ever referencing it in the repo, and docker hub pull counts are per-repo not per-tag (AFAIK) so they won't answer it either.

fwiw I've used it a few times: working on trl-jobs, and just running hf jobs when I wanted main rather than the last release. So at least one consumer 😅

It's not really costing us anything (same tag gets overwritten, layers are shared), so i'd rather keep it. if the goal is to stop the workflow running per commit (which again I think is cheap enough), build it nightly on a cron instead of on every push to main.