Proposing to remove SSDTrainer #7596

Proposing to remove trl.experimental.ssd (SSDTrainer, SSDConfig), added in #5505 (2026-04-14), directly into trl.experimental. Reporting the inputs discussed in #7182.

Usage.

• 4 trainings in the last 30 days as per our telemetry.
• 2 models carry both the ssd and trl tags, none since May: https://huggingface.co/models?other=ssd,trl&sort=created
• Nobody has spoken up for it in [RFC] Criteria for removing code from trl.experimental #7182.

External issues and pull requests.

• None from outside the maintainer team in the 6 months since it landed.

Subsumed.

• Yes. The paper's recipe is two steps: sample from the model at a training-time temperature and truncation, then standard SFT on the samples. Generating with vLLM and training with SFTTrainer covers it. What goes away is the sampling inside the training loop.

Cost.

• 842 lines of trainer and config, 1,255 with test and doc page, plus 420 lines of example scripts in examples/ssd_codegen/.
• It carries its own copy of the generation path (vLLM and transformers, weight sync), so every generation sweep has to touch it: Sync the model config pad token id with the tokenizer in the experimental trainers #7128, Write the aligned token ids to the text config #7315, Seed before creating the model in experimental trainers #7401, End completions on every eos id the model declares in the experimental trainers #7506 in the last month.

Downstream consumers.

• None. GitHub code search finds three zero-star personal repos referencing it, everything else is vendored copies of the trl tree and docs mirrors.

Owner.

• @kashif landed it, no feature work since.

Age.

• 6 months in trl.experimental, shipped in every release from v1.2.0 to v1.15.0 with only upkeep commits since it landed. Long enough to have been found.

This issue aims to collect feedback from the community, if any. If none:

• remove the module, test, doc page and examples/ssd_codegen/, and
• keep the paper index entry pointing to the last release that shipped it.