## huggingface/transformers — v5.18.0…v5.19.0

_60 commits._

### Features
- **[`distributed`] Add expert-parallel token dispatch, default for Qwen3 MoE (#48865)** (d28b793)
- **[`distributed`] Add dense and expert device mesh (#48857)** (f4ae762)
- **Allow registering new processing backends (#48984)** (4c9422f)
- **Revert "[CI] ssh-runner: add optional cache_type input to switch between bucket and EFS runners (#49159)" (#49270)** (3693f8d)
- **Add Trainer.loss_is_scaled_for_ga to declare whether compute_loss already scales for gradient accumulation (#49240)** (094f710)
- **Support per-layer cache configuration (#48178)** (35924ec)
- **[`GTE`] Add memory mixin (#49234)** (a005fc8)
- **[CB] Make CB more device agnostic and add support for XPU (#49156)** (0bb187f)
- **🚨 Attention 🚨 Make regular attention support CB (#49101)** (5f0b641)

### Fixes
- **Fix DataCollatorForLanguageModeling ignoring seed=0 (#49349)** (b4526a7)
- **Fix explicit kernelization modes for evaluation models (#49339)** (5a27d2d)
- **fix(cohere_compass): add video to input_modalities (#49320) (#49323)** (23c21ac)
- **Fix WatermarkDetector repeated-ngram counting (#49315)** (3758eba)
- **Fix minicpm 4.6V video batching and CI (#49259)** (0cf6d17)
- **Fix the `labels` docstring of CLIPSegForImageSegmentation (#49001)** (080c288)
- **Fix run_glue.py label ids becoming strings with pandas>=3 (#49326)** (fda3605)
- **[Fix] Remove old test file with two ancient tests (#49280)** (f5ab856)
- **Fix chunked prefill with multi-axis position ids (Qwen3-VL) (#49319)** (2c8526d)
- **Fix multi-image feature ordering during generation batch expansion (#49276)** (9d68278)
- **Fix quantized cache (#48700)** (492f240)
- **Why override when you can fix in `Mixin` (#49105)** (3d207f8)
- **Fix encoder repetition penalty source rows after batch expansion (#49262)** (1050257)
- **Fix pp chart2table inference (#49225)** (9d8c3a2)
- **fix precomputed flash kwargs clashing (#47928)** (000a29f)
- **[ESM/Persimmon/NllbMoe/ESMFold] use safetensors repos to fix Xet FUSE SIGBUS/IOError on bucket CI runners (#49194)** (361bfb1)
- **fix the slack API call (#49220)** (2624191)

### Backend
- **v5.19.0** (c4c4605)
- **Smthn smthn (#49364)** (92cd495)
- **Revert "[CI] Temporarily disable AMD scheduled CI trigger" (#49271) (#49362)** (a7e8ff2)
- **ALM testx fIxing... (#49355)** (571c5cc)
- **GPT-OSS: read the clamped SwiGLU's alpha and limit from the config (#49346)** (f1d0d9f)
- **[`distributed`] Default MoE `ep_plans` to token dispatch (#49160)** (9921402)
- **[`distributed`]: adapt Trainer to work Expert Parallel (#48873)** (8625785)
- **Move CI to Python 3.11, with the version set in `.python-version` (#49311)** (422bd07)
- **Bump doc-builder main docs workflow pin (#49350)** (03011bd)
- **[`distributed`] decoupled tp_plan and ep_plan (#48859)** (2033cf0)
- **Bump doc-builder workflow pin (#49342)** (475e93a)
- **remove colwise_gather_output for lm_head for DeepseekV4  (#49314)** (8073e6d)
- **🚨 Return router logits from every MoE model that computes them (#48920)** (c8b25ab)
- **[CI] Temporarily disable AMD scheduled CI trigger (#49271)** (263c5eb)
- **Remove the unused pandas entry from setup.py deps (#49329)** (c3cd9d0)
- **Do RoPE with mul instead of matmul (#49265)** (c633787)
- **Deprecated pipelines should redirect; removed pipelines should raise (#49308)** (56d3afc)
- **Enable model test on upcoming `torch_tpu` backend (#49207)** (26c0a7f)
- **Preserve masked EOS scores in exponential decay length penalty (#49260)** (1aac958)
- **[CI] Avoid fetching all artifacts when only specific ones are needed (#49282)** (4692303)
- **[CI] check_bad_commit: use EFS cache to avoid Xet FUSE OOM (exit 137) (#49273)** (02d8fb9)
- **🚨 Select the OWLv2 image query by objectness (#49200)** (42adb29)
- **Don't forward text-level flash-attention kwargs to multimodal encoders (#49227)** (35dff09)
- **Reapply modular examples (#49230)** (99168b2)
- **Remove some stuff that does not exist (#49232)** (fe4107c)
- **Remove deprecation cycle on mask and cache (#49231)** (f9ecae3)
- **Remove rotaries deprecation cycle (#49228)** (0a896aa)
- **[docs] TP for PEFT adapters (#49060)** (8c8b92b)
- **correct final metrics due to resuming from checkpoint (#49208)** (7fb5bcd)
- **🚨 Attention 🚨 Deprecate "paged|" prefix for SDPA and flash  (#49112)** (9bda615)
- **🚨 [CB] 🚨 Fuse update for index and block table path (#49088)** (1014cee)
- **[Refactor] Make some flash-attention utils more readable (#49071)** (e7cceb9)
- **🚨 [CB] 🚨 Little fixes before removing "paged" (#49069)** (2112ec4)
- **Dev update (#49212)** (d6c1e71)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/huggingface/transformers?utm_source=github-action)._