Add LFM2.5-VL support and testing

[qgallouedec]

Released 2026-08-12: LiquidAI/LFM2.5-VL-3B — 3.1B on-device VLM (SigLIP2 vision tower + LFM2 conv/attention text backbone), open weights.

• Tiny model script lfm2_vl_for_conditional_generation.py (Lfm2VlForConditionalGeneration), SmolVLM-style sub-config scale-down; pinned to transformers==5.0.0 (first release with lfm2_vl, and the TokenizersBackend floor the LFM2.5 tokenizer needs). Config diff vs the reference is 11 entries, all scale-downs.
• Vendored lfm2_2_5_vl.jinja and aliased it onto the existing LFM2.5 branch of add_response_schema (same <|tool_call_start|>[...]<|tool_call_end|> format, so it reuses lfm2_2_5_template). No training variant: the template already ships {% generation %} markers.
• Tests: VLM param lists in test_sft_trainer.py (3), test_dpo_trainer.py, test_kto_trainer.py, plus the response-schema-VLM, tool-calling and parse-response lists in test_chat_template_utils.py.
• Tiny model pushed: trl-internal-testing/tiny-Lfm2VlForConditionalGeneration-2.5.

---

Note

Low Risk
Mostly new template, tiny-model generator, and test parametrization; no changes to core training logic beyond extending existing LFM2.5 response-schema branching.

Overview
Adds LFM2.5-VL (LiquidAI vision-language) to TRL’s supported model families and test matrix.

A new lfm2_2_5_vl.jinja chat template is vendored (multimodal <image> blocks, same tool/thinking/{% generation %} behavior as LFM2.5 text). add_response_schema treats it like LFM2.5 and reuses lfm2_2_5_template for response parsing. Docs list LFM2.5-VL among recognized families.

A scripts/generate_tiny_models/.../lfm2_vl_for_conditional_generation.py script builds trl-internal-testing/tiny-Lfm2VlForConditionalGeneration-2.5 (transformers ≥ 5.0). That tiny model is wired into VLM parametrized tests in SFT, DPO, KTO, and chat template suites (response schema, tool calling, parse-response with thinking field).

^{Reviewed by Cursor Bugbot for commit c65247c. 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.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3e1a3bc755

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Handle escaped quotes before reusing the LFM parser

When an LFM2.5-VL tool-call string contains an apostrophe, such as O'Reilly, the new template renders it as 'O\'Reilly', but the reused _LFM2_2_5_ARG_VALUE parser only recognizes single-quoted strings with '[^']*'. It therefore truncates the value at the escaped apostrophe, so parse_response cannot reliably round-trip these tool calls; the response template needs an escape-aware value pattern and decoding for the VL format.

Useful? React with 👍 / 👎.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 3e1a3bc. Configure here.}