Who contributed to Hugging Face on August 31, 2026?

7 developers shipped this update, including sergereview[bot], sergereview, albertvillanova, dependabot, alexanderar, sywangyi, and tarekziade.

What were the notable Hugging Face updates?

[serge] Fix 4 integration tests for model `generation` failing with `output_mismatch` (list output differs (4)) (#48133), Pin layer_types so the tiny Cohere2 model covers both attention types, and Pin layer_types so the tiny Gemma3 and Olmo3 models cover both attention types.

Aug 31, 2026

$ the-wire · showcase

Transformers tightens integration tests, TRL pins hybrid attention coverage

By RepoJournal · Filed 06:59 UTC on August 31, 2026 · About Hugging Face

7 people shipped this

albertvillanova @albertvillanova 3 cited

sergereview[bot] @sergereview[bot] 1 cited

alexanderar @alexanderar 1 cited

tarekziade @tarekziade 1 cited

sergereview @sergereview

dependabot @dependabot

sywangyi @sywangyi

Hugging Face's transformers repo fixed four integration tests, TRL pinned layer types so tiny hybrid models actually cover both attention paths, and diffusers shipped a major new LTX-2.5 pipeline.

The transformers repo fixed four integration tests in `tests/generation/test_utils.py` that were failing with `output_mismatch` (list output differs (4)) [1]. The fix ensures the generation tests pass as expected. In TRL, three PRs pin `layer_types` for tiny Cohere2, Gemma3, and Olmo3 models so their 2-layer architectures cover both full-attention and sliding-window layers [2][3]. Previously, the sliding-window period derived patterns that yielded only sliding layers, leaving the global RoPE path unbuilt and untested. For Cohere2, RoPE applies only on sliding layers, making the missed path more consequential. The PRs fix issues #6961 and #6962. Separately, TRL made its ruff lint rule set explicit by using `select` instead of `extend-select`, because "Ruff 0.16 raised its default rule selection from 59 to 413 rules" [4]. This prevents TRL from silently inheriting rule families like `SIM`, `S`, `TRY`, and `PL*` it never opted into. Diffusers added the `LTX2DFRPipeline` for LTX-2.5, a Diffusion Fidelity Rendering pipeline ported from Lightricks [5], which generates keyframe slots that relax temporal compression so video can be conditioned on genuinely new frames. Finally, transformers-ci states the modular layout as fact in its agent instructions, after a session spent 15 of 46 turns (a third of its budget at ~40k input tokens per turn) on a six-way OR grep to discover it [6].

Action items

→ Pull the transformers generation integration test fix (ccba41e) into your repo if you track transformers huggingface/transformers [plan]
→ Review TRL PRs #6962, #6963, #6964 for hybrid model coverage and lint explicitness huggingface/trl [plan]
→ Monitor the LTX-2.5 DFR pipeline rollout if you use diffusers for video generation huggingface/diffusers [monitor]

References

[1] [serge] Fix 4 integration tests for model `generation` failing with `output_mismatch` (list output differs (4)) (#48133) ↗ huggingface/transformers
[2] Pin layer_types so the tiny Cohere2 model covers both attention types ↗ huggingface/trl
[3] Pin layer_types so the tiny Gemma3 and Olmo3 models cover both attention types ↗ huggingface/trl
[4] Use ruff `select` instead of `extend-select` to keep the rule set explicit ↗ huggingface/trl
[5] Add LTX-2.5 DFR pipeline (keyframe slots, spatial detailing, tiled temporal rounds) ↗ huggingface/diffusers
[6] State the modular layout as fact instead of making the agent grep for it ↗ huggingface/transformers-ci