Pin layer_types so the tiny Gemma3 and Olmo3 models cover both attention types

[albertvillanova]

This PR pins layer_types in the tiny Gemma3 and Olmo3 generation scripts so that each model has one layer of each attention type.

Fix #6961.

Motivation

Both tiny models have 2 layers, and their sliding-window periods (6 for Gemma3, 4 for Olmo3) make the derived pattern yield two sliding layers and no full-attention layer. Two consequences:

• The global RoPE path is never built or executed, so _compute_linear_scaling_rope_parameters (Gemma3) and _compute_yarn_parameters (Olmo3) are dead code in our fixtures, even though both configs still carry the reference's global rope settings.
• transformers logs Unrecognized keys in rope_parameters for 'rope_type'='default': {'full_attention', 'sliding_attention'} on every config load, up to 104 times per CI job, because it classifies the nested per-layer rope dict against the layer types the model actually instantiates. Reported upstream as Nested per-layer rope_parameters is misparsed when layer_types omits one of its keys transformers#48392.

Solution

Pin layer_types explicitly, as the Gemma4 and DiffusionGemma scripts already do, and as the Nemotron 3 scripts do for block types ("one of each block type").

With 2 layers the reference ratios (5:1 for Gemma3, 3:1 for Olmo3) cannot be represented at all, so the choice is not between a faithful and an unfaithful pattern, but between covering one attention type and covering both.

The Hub models have been regenerated. Verified in this CI run: the log line drops from 104 to 0 in every job, no other transformers log line changes, and the test totals are unchanged (2369 passed, 149 skipped, 7 xfailed).

Changes

• Pin layer_types to ["sliding_attention", "full_attention"] in the tiny Gemma3 generation script, replacing the previous None that derived the pattern from num_hidden_layers
• Pin the same layer_types in the tiny Olmo3 generation script

---

Note

Low Risk
Only affects tiny test fixture generation scripts and regenerated Hub weights; no runtime TRL or production model behavior changes.

Overview
Pins layer_types to ["sliding_attention", "full_attention"] in the tiny Gemma3 and Olmo3 Hub generation scripts so each 2-layer fixture includes one sliding and one full-attention block.

With only two layers, the auto-derived pattern from production ratios would use sliding attention everywhere, so global RoPE paths never ran in tests and transformers spammed unrecognized rope_parameters warnings on config load. This matches how other tiny generators (e.g. Gemma4, DiffusionGemma) already pin one of each attention type.

^{Reviewed by Cursor Bugbot for commit f229067. 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.

[albertvillanova]

I created the corresponding PRs on the Hub for both tiny models:

• https://huggingface.co/trl-internal-testing/tiny-Gemma3ForConditionalGeneration/discussions/11
• https://huggingface.co/trl-internal-testing/tiny-Olmo3ForCausalLM/discussions/2

And I tested them by pinning both revisions through MODEL_REVISIONS in tests/conftest.py: the log message is gone, 104 occurrences down to 0, with the test totals unchanged (2369 passed, 149 skipped, 7 xfailed).

https://github.com/huggingface/trl/actions/runs/33168783898/job/98840532231

Both Hub PRs need to be merged once approved: this PR's own CI still reads the tiny models from their main revision, so the message will keep showing here until then.

[albertvillanova]

Merging this PR as a simple fix of the CI and as already empirically verified by GH run: https://github.com/huggingface/trl/actions/runs/33168783898/job/98840532231