# huggingface/transformers v5.18.0: Release 5.18.0 — retained record

> Retained github record from github:huggingface/transformers.

Canonical: https://highsignal.app/data/records/36f9052b079fc4de

[← sources](https://highsignal.app/data)


source: [github](https://highsignal.app/data/github) (github:huggingface/transformers)  published 2026-09-30 16:46 ingested 2026-10-01 02:35

[original source ↗](https://github.com/huggingface/transformers/releases/tag/v5.18.0)

## extract

## New Model additions


### Nemotron 3 Diarization

<img width="1680" height="900" alt="image" src="https://github.com/user-attachments/assets/fe735cb3-9e5b-43ad-8f60-9dec8425aec7" />

Nemotron 3 Diarization is an open-weight streaming speaker diarization model designed to determine "who spoke when" in real-world audio. It supports both streaming and offline inference, handles up to eight speakers, and orders speaker outputs by each speaker's first arrival in the input audio.

The model uses the Arrival-Order Speaker Cache (AOSC) [1](https://huggingface.co/papers/2507.18446) and FIFO queue introduced for Streaming Sortformer [1](https://huggingface.co/papers/2507.18446), [2](https://huggingface.co/papers/2409.06656). A single checkpoint supports configurable latency profiles, from an 80 ms input buffer to a 30.4 s offline-style buffer, and configurable output frame resolution in multiples of 10 ms. With chunked inference, the maximum audio duration is not limited.

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/nemotron3_diarization)
* Add Nemotron3Diarization (#49056) by @eustlb in [#49056](https://github.com/huggingface/transformers/pull/49056)

### NemotronH Omni

NemotronH Omni is a multimodal reasoning model from NVIDIA that pairs the [NemotronH](https://huggingface.co/docs/transformers/main/en/model_doc/nemotron_h) hybrid
Mamba-Transformer language model with a [RADIO](https://huggingface.co/docs/transformers/main/en/model_doc/radio) vision encoder and an optional Parakeet-based sound encoder.
Image (and video) patches are projected through a RADIO tower and a pixel-shuffle MLP into the language model's
embedding space at the `<image>` / `<video>` context-token positions; audio clips are projected in the same way at
`<audio>` positions. The result is a single autoregressive model that reasons jointly over text, images, video and
sound.

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/nemotron_h_omni)
* Add support for Nemotron Omni (#46509) by @meatybobby in [#46509](https://github.com/huggingface/transformers/pull/46509)

### HyperCLOVAX Vision V2

HyperCLOVAX Vision V2 is a multimodal vision-language model developed by NAVER. It combines the [HyperClovaX](https://huggingface.co/docs/transformers/main/en/model_doc/hyperclovax) language model backbone with a [Qwen2.5-VL](https://huggingface.co/docs/transformers/main/en/model_doc/qwen2_5_vl) vision encoder. The model supports text, image, and video inputs and is capable of chain-of-thought reasoning via built-in thinking tokens (`<think>...</think>`).

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/hyperclovax_vision_v2)
* add HyperClovaX Vision (#44314) by @jp1924 in [#44314](https://github.com/huggingface/transformers/pull/44314)

### GTE

GTE was proposed in [mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval](https://huggingface.co/papers/2407.19669) by Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li and Min Zhang.

GTE is a BERT-style bidirectional encoder that replaces absolute position embeddings with RoPE, uses a gated MLP, and applies layer normalization after each residual connection. The same architecture backs Alibaba's `gte-*-v1.5`, `gte-multilingual-*` and `gte-en-mlm-*` checkpoints as well as Snowflake's `snowflake-arctic-embed-m-v2.0`.

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/gte)
* model: Add GTE to Transformers (#48416) by @harshaljanjani in [#48416](https://github.com/huggingface/transformers/pull/48416)


## Breaking changes

* 🚨 [ROCm] gpt-oss: route FA3 to aiter-flash-attn, generate ROCm fixtures (#46837) by @Abdennacer-Badaoui
* 🚨 Speed up detr image processing (#48066) by @guarin
* 🚨 Remap indexers layer_type (#48974) by @Cyrilvallez
* 🚨 [vLLM] Fix video token counting for Transformers backend video inputs (Part 1) (#48894) by @harshaljanjani
* 🚨 🚨Bring some dinos to modern standards (#46266) by @molbap
* :rotating_light: [`Kernels`] Bump version (#48714) by @vasqu


## Bugfixes and improvements

* fix incorrect hub tokenizer class (#48641) by @itazap
* [CI] Deduplicate Nvidia/AMD CI reply comments and add headers (#48655) by @ydshieh
* Update dev (#48654) by @vasqu
* [tests] Fix NougatModelIntegrationTest: pin artifact revision and update golden values (#48638) by @ydshieh
* [fix] Fix GlmOcr integration tests: wrong token IDs and image token decode bug (#48650) by @ydshieh
* [GLM 5.3 Flash] Preserve original names when saving text checkpoints (#48676) by @Dovis01
* Register activation kernel layers on XPU (#47858) by @jiqing-feng
* Mirror the registered pytree flatten when building automatic dynamic shapes (#48578) by @IlyasMoutawwakil
* Fix `AutoImageProcessor` requiring torchvision when only Pillow is installed (#48616) by @blipbyte
* Assert cached decode matches recomputing without a cache (#48289) by @IlyasMoutawwakil
* [higgs_audio_v2] Fix: use config.num_codebooks in audio labels tensor (Simple Fix!) (#48560) by @b-re-w
* [`MoE`] Fix eager EP (#48653) by @vasqu
* [MusicgenMelody] Fix conditioning silently dropped at generation step 0 (#48679) by @ydshieh
* Fix deepspeed ci (#48640) by @SunMarc
* Synthetic test assets (#48589) by @tarekziade
* Skip the expert-parallel sentinel masking when expert parallelism is off (#48201) by @qgallouedec
* [Fix] yolos offload issue (#48688) by @molbap
* [serge] Fix 2 integration tests for model `minimax` failing with `output_mismatch` (tensor values differ (2)) (#48515) by @sergereview[bot]
* [serge] Fix 2 integration tests for model `mistral` failing with `other` (other (2)) (#48429) by @sergereview[bot]
* Fix Qwen2.5-VL temporal RoPE for fractional video intervals (#48669) by @yeyeyeping
* Fix slow integration tests on XPU (#48611) by @jiqing-feng
* QA: Added a `MemoryCleanupMixin` class for tests (#48681) by @tarekziade
* [serge] Fix 1 integration test for model `flex_olmo` failing with `other` (other (1)) (#48668) by @sergereview[bot]
* Fix D-FINE / RT-DETR main loss being computed over the denoising queries (#48528) by @stefan-it
* [CI] Replace hardcoded username allowlists with author_association check in workflow triggers (#48712) by @ydshieh
* Add flash_attention_4 in attn_implementation AutoModel docstring (#48684) by @3manifold
* Keep the expert-parallel sentinel slots out of the router gradient (#48689) by @qgallouedec
* [DeepseekV3] OOM cascade root-cause investigation (generator ref leak in conversion_mapping) (#48720) by @ydshieh
* Use huggingface_hub httpx export (#48685) by @Wauplin
* unifying device_mesh init to enable PP + TP inference (#48155) by @3outeille
* Enable FSDP2 + expert parallelism via a 2-D (fsdp, tp) device mesh (#48516) by @qgallouedec
* Switch daily CI to torch 2.14 — update expected outputs (#48750) by @ydshieh
* fix videomae load error (#48675) by @sywangyi
* Fix silently random-initializing `RTDetrModel`/`SEWDForCTC` loads (wrong `base_model_prefix`) (#48744) by @<NOT FOUND>
* Fix Glm4vMoeIntegrationTest: offload_folder + MemoryCleanupMixin (#48776) by @ydshieh
* [InternVL] Normalize num_patches before np.cumsum (#48469) by @lorenzozanee
* Kernel api doc, part 2 (#46889) by @michaelbenayoun
* Fix hidden state selection on Gemma4 assistant's first prefill step (#48704) by @glistening
* Fix inkling embedding norm (#48786) by @Cyrilvallez
* Fix TextToAudioPipeline crash for tokenizer-only models (#48505) by @jiqing-feng
* Cast pixel values to the patch embedding dtype in DeepSeek-OCR-2 (#48632) by @jiqing-feng
* add kernel mapping entry for RMSNormGated, KDA, Conv1D on XPU (#48702) by @kaixuanliu
* Bump `peft` version requirement (#48716) by @shniubobo
* [serge] Fix 12 integration tests for model `edgetam` failing with `import_or_config` (other (12)) (#48322) by @sergereview[bot]
* Restore legacy tensor-parallel initialization for compatibility (#48797) by @3outeille
* Fix kosmos flaky test (#48800) by @IlyasMoutawwakil
* Add integration tests for MuseGlimmerAssistantModel (#48796) by @ydshieh
* Fix `AutoModel.from_pretrained` not restoring `modules_to_save` weights (#48595) by @shniubobo
* [docs] mrope and axial rope (#48717) by @stevhliu
* Fix decoder only path for older bert variants (#48785) by @Cyrilvallez
* [Fix] Clean-up ternaries in the DeepSeek family (#48447) by @remi-or
* [utils] Add MUSA support for Flash Attention 2 (#48612) by @XiaomingFun233
* [generate] Drop attention mask early without padding (#48814) by @Cyrilvallez
* Pin utf-8 in test_can_init_all_missing_weights source read (Windows non-UTF-8 locale fix) (#48819) by @dltsum
* Relax static-cache tolerance in test_generate_with_static_cache (1e-5 → 5e-5) (#48815) by @ydshieh
* Detect nested rope_parameters without relying on layer_types (#48798) by @hmellor
* Support for MoE in the GGUF integration (#48529) by @SunMarc
* [docs] gguf (#46357) by @stevhliu
* [apply_chat_template] pass sampling_rate to call (#48794) by @eustlb
* [CI] Add link checker (#48160) by @stevhliu
* Qwen3.8 GGUF (#48660) by @SunMarc
* docs: fix broken #combining-with-fsdp2 anchor in expert_parallelism.md (#48854) by @ydshieh
* [GPTNeoXJapanese] Fix RoPE ignoring partial_rotary_factor (#48652) by @blipbyte
* QA: deactivate rule 41 (#48852) by @tarekziade
* Fix `reset` on the dynamic cache layers (#48809) by @jiqing-feng
* [generate] Make all methods and logits processors agnostic to lm_head output size (#48846) by @Cyrilvallez
* Fix kosmos (#48853) by @Cyrilvallez
* Add test for causal only variant of some encoder-decoder models (#48760) by @nandan2003
* Fix ESMFold2 ligand iPLDDT weighting: mol_type non-polymer code is 3, not 4 (#48831) by @faustomilletari
* Fix flaky RfDetr test_save_load (#48869) by @ydshieh
* fix(ci): harden GitHub Actions workflows (#48160) (#48834) by @hf-security-analysis[bot]
* Fix Glm4MoeIntegrationTest: split into 3 tests, switch to GLM-4.5-Air (#48820) by @ydshieh
* Fix `StaticCache` for Mllama and enable `torch.compile` (#48141) by @jiqing-feng
* Fix Minimax M2 partial_rotary_factor by remapping legacy rotary_dim (#48486) by @dajiaohuang
* Fix Pixtral image processor do_pad option (#48872) by @jzakrzew
* [CB] Add pause mechanism (#48462) by @remi-or
* Fix DFlash sampled candidates losing the batch dimension (#48281) by @VaggelisGian
* Fix sliding window cache when assistant model calls generate (#48280) by @VaggelisGian
* Fix GLM rope_parameters dict shared mutation across test instances (#48895) by @ydshieh
* [`feat`] Allow untying `hidden_states[-1]` from `last_hidden_state` via the model config (#48087) by @tomaarsen
* Fix Idefics2 padding when the first example has no image (#48753) by @Arnavsharma2
* Don't fail model loading when the accelerator can't report free memory (#48136) by @studioego
* [Improvement] Rework the docstring and comments of mHC (#48888) by @remi-or
* [serge] Fix 3 integration tests for model `deepseek_vl` failing with `other` (other (3)) (#48536) by @sergereview[bot]
* Add auto_docstring task model overrides (#47815) by @guarin
* Make LongcatFlashConfig self-consistent without building the model (#48899) by @hmellor
* [`DSA`] Only save latents on dsa with indexer as well (#48876) by @vasqu
* [Chat Parsing] Coerce oneOf tool arguments (#48719) by @yonigozlan
* Always materialize the causal mask in Doge so sdpa stays causal (#48821) by @blipbyte
* Fix docstring argument names that don't match signatures (#48474) by @IshaanPotle
* Mark test_generate_with_static_cache as flaky for olmo and bigbird_pegasus (#48856) by @ydshieh
* Clean up some MoE models' integration tests (#48833) by @ydshieh
* Fix T5 tied weights order for lm_head (#48238) by @vaibhavmashal
* Size the device_map buffer from the largest leaf module (#47211) by @dhruv7477
* Fix compile cache tests (#48923) by @Cyrilvallez
* Update arXiv citation in NeoMME model doc (#48926) by @tonywu71
* Standardise Aria's MoE onto the experts interface (#48907) by @hmellor
* Fix possessive typo in Whisper long-form warning (#48877) by @erikdw
* Modular conversion small fix (#48934) by @zucchini-nlp
* fix(vibevoice-asr): use integer ceiling division for audio token count (#48864) by @ege-arhan
* Improve lazy import error messages (#48602) by @karatarassul4-max
* Fix tests due to dropping attn mask (#48903) by @SunMarc
* docs: fix docstring parameters that do not match signatures (#48904) by @simpleqt
* DOC: Improve documentation for remap_legacy_layer_types function (#48651) by @mathewOracle
* Align special tokens on the text config (#48847) by @qgallouedec
* docs: fix dead doc links in i18n READMEs and the xlnet docstring (#48905) by @simpleqt
* Fix AXK2 integration test: update CUDA expected text and rename class (#48941) by @ydshieh
* Resolve the Hub revision once per load instead of passing a private _commit_hash around (#47611) by @Wauplin
* Keep image processor backends in sync on keys and dtypes (#48739) by @yupengtang
* Fix infeasible cost matrix errors in hugarian matcher losses (#47730) by @guarin
* Fix more stuff (#48979) by @zucchini-nlp
* Skip an unnecessary image copy in the torchvision image normalization path (#48897) by @jzakrzew
* Better attn default for gguf (#48935) by @SunMarc
* Add pose estimation keypoint preprocessing to Sapiens2ImageProcessor (#47199) by @Sainava
* Fix image processor class-level size mutation and min_pixels handling (#48916) by @Gracy769
* Contract the mamba2 chunk scan with einsum instead of broadcast-then-sum (#48978) by @tarekziade
* Fix typos in DeepseekV4 comments (#48953) by @Janiarafath
* Reduce peak memory in examples_torch CI job (OOM fix) (#48983) by @ydshieh
* Fix vibevoice TTS batched audio index (#48902) by @ebezzam
* [CB] [Major] Upgrade the cache to support different attention types (#47809) by @remi-or
* Fix SwitchTransformers Top1 router: raw logits, expert capacity accounting, and router losses (#48421) by @yurekami
* Fix Qwen3OmniMoeIntegrationTest OOM (#48987) by @ydshieh
* Let `prefix_allowed_tokens_fn` override model `-inf` and raise an exception on unsatisfiable generation constraints. (#48927) by @ksh108405
* [generate] Simplify candidate generators by removing required `update_candidate_strategy` (#48982) by @Cyrilvallez
* [generate] Always correctly restrict assisted decoding with max length/eos token  (#48981) by @Cyrilvallez
* QA: applied ruff rule PLW1514 (#48990) by @tarekziade
* Update tokenizer gguf support  (#48656) by @SunMarc
* [vLLM] Fix video token counting for Transformers backend video inputs (Part 2) (#48900) by @harshaljanjani
* Update ggml kernels path  (#48991) by @SunMarc
* Enable compressed-tensors FP8 kernels on MPS (torch >= 2.15) (#48985) by @Isalia20
* Fix mps autocast handling in rotary embeddings (#49006) by @Isalia20
* Fix static cache per layer head shapes (#48619) by @dacorvo
* Summarization examples: download NLTK punkt_tab, not punkt (#49014) by @davanstrien
* [docs] Fix legacy hf CLI references (transformers) (#48988) by @Wauplin
* Fix RecurrentGemma compiled generation with StaticCache (#48961) by @sywangyi
* updated GraniteMoeHybrid expectations (CPU) (#49016) by @tarekziade
* OpenVINO HF Exporter (#47003) by @IlyasMoutawwakil
* fix noisy comments (#49013) by @tarekziade
* Fix assisted decoding for VLM due to dropping attn mask  (#49019) by @SunMarc
* Deprecate min-max pixels (#49021) by @zucchini-nlp
* Keep special token ids the tokenizer does not define (#48708) by @albertvillanova
* Added more usage of MemoryCleanupMixin (#49011) by @tarekziade
* Shorten noisy OpenVINO SDPA comment (#49033) by @tarekziade
* PEFT x Dtensor-based TP integration (#48485) by @michaelbenayoun
* fix(zamba):  add use_associative_scan config flag to avoid torch.compile slowdown (#48331) by @msnliu
* Scope GITHUB_TOKEN permissions per job (#49046) by @hf-security-analysis[bot]
* Fix mask creation not being skipped under `torch.compile` (#48975) by @jiqing-feng
* [Chat] Loading GGUF models served with the Chat CLI (#49031) by @ariG23498
* Pin GitHub Actions to commit SHAs (#49049) by @hf-security-analysis[bot]
* Only seed numpy in BigBird block-sparse attention during training (#49037) by @jiqing-feng
* QA: Fix llama4 leak (#49042) by @tarekziade
* Fix startup failures: drop permissions reusable-workflow callers cannot grant (#49055) by @paulinebm
* Fix startup failures: drop pull-requests: read from check_failed_tests.yml (#49057) by @paulinebm
* Register the remaining mamba-ssm kernel layers on XPU (#49035) by @jiqing-feng
* Add workflow for building XPU CI Docker images (#49039) by @regisss
* Fix assistant masks for processors (#48793) by @Rocketknight1
* Add MPS maintainer (#49052) by @Rocketknight1
* [docs] Loading behavior (#49023) by @stevhliu
* [Nemotron3Diarization] nit: hub pr merged to main (#49065) by @eustlb
* fix(ci): harden GitHub Actions workflows (#49057) (#49059) by @hf-security-analysis[bot]
* Honor `config.output_router_logits` in the MoE VLM wrappers (#48885) by @qgallouedec
* Fix XPU Docker image build (#49073) by @regisss
* Fix assisted eos token condition (#49075) by @Cyrilvallez
* [serge] Fix OOM in Moshi integration tests with MemoryCleanupMixin (#48839) by @sergereview[bot]
* Fix odd head_dim validation for RoPE configurations (#48524) by @somuai
* Keep already-decoded array arguments unchanged (#49062) by @yonigozlan
* Fix NaN in Parakeet eager attention with padded batches (#49070) by @ArthurZucker
* Bump huggingface_hub upper bound to <3.0 (transformers) (#49083) by @Wauplin
* Keep `attention_mask` as `None` in OPT's causal mask creation (#49002) by @jiqing-feng
* Add `Trainer.end` (#48875) by @qgallouedec
* Add image processing tester init (#48829) by @guarin
* [Executorch] Add MLX recipe (#48910) by @metascroy
* [Zamba] Fix associative scan breaking ONNX export and OOM in integration test (#49092) by @ydshieh
* Fix RGB early-return skipping PNG tRNS compositing (#49005) by @cs-fisha
* Use namespaced dataset ids in docs and PyTorch examples (#49017) by @davanstrien
* Fix MaskFormerSwin attention mask dtype to follow hidden states (#49032) by @kaixuanliu
* [NemotronH-Omni] Fix device mismatch in test tensor creation (#49102) by @ydshieh
* [serge] Fix 2 integration tests for model `cvt` failing with `output_mismatch` (tensor values differ (2)) (#49076) by @sergereview[bot]
* [serge] Fix 2 integration tests for model `hy_v3` failing with `output_mismatch` (tensor values differ (2)) (#49068) by @sergereview[bot]
* [serge] Fix 2 integration tests for model `pvt_v2` failing with `other` (#49067) by @sergereview[bot]
* Keep the eos ids the config declares (#49082) by @albertvillanova
* Use grouped_mm on TPU devices under torch.compile (#49097) by @salkan0
* [serge] Fix 2 integration tests for model `jamba` failing with `other` (other (2)) (#49044) by @sergereview[bot]
* Remap the legacy Gemma 1 hidden_act in the config post-init (#49084) by @PCfVW
* Fix exporters import on torch < 2.9 (is_contiguous_or_false) (#49124) by @ydshieh
* [gemma3n] fix audio test fixture: use hf_hub_download instead of load_dataset (#49128) by @ydshieh
* qwen3 models map to wrong tokenizer class on the hub (#49116) by @itazap
* Restore the Unicode whitespace set in the GPT-SW3 tokenizer (#48912) by @David-Wu1119
* [AMD] Fix some integration tests (#49153) by @Abdennacer-Badaoui
* [Gemma] Fix test_model_7b_fp16_static_cache expected value for cuda 8 after #49084 (#49133) by @ydshieh
* Fix UMT5 decoder self-attention not being causal (#49135) by @the-cross-art
* Skip flash tests that fall back to a hub kernel when kernels is missing (#49129) by @Abdennacer-Badaoui
* Auto generate model inits (#478

## retained text (truncated)



```
## New Model additions


### Nemotron 3 Diarization

<img width="1680" height="900" alt="image" src="https://github.com/user-attachments/assets/fe735cb3-9e5b-43ad-8f60-9dec8425aec7" />

Nemotron 3 Diarization is an open-weight streaming speaker diarization model designed to determine "who spoke when" in real-world audio. It supports both streaming and offline inference, handles up to eight speakers, and orders speaker outputs by each speaker's first arrival in the input audio.

The model uses the Arrival-Order Speaker Cache (AOSC) [1](https://huggingface.co/papers/2507.18446) and FIFO queue introduced for Streaming Sortformer [1](https://huggingface.co/papers/2507.18446), [2](https://huggingface.co/papers/2409.06656). A single checkpoint supports configurable latency profiles, from an 80 ms input buffer to a 30.4 s offline-style buffer, and configurable output frame resolution in multiples of 10 ms. With chunked inference, the maximum audio duration is not limited.

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/nemotron3_diarization)
* Add Nemotron3Diarization (#49056) by @eustlb in [#49056](https://github.com/huggingface/transformers/pull/49056)

### NemotronH Omni

NemotronH Omni is a multimodal reasoning model from NVIDIA that pairs the [NemotronH](https://huggingface.co/docs/transformers/main/en/model_doc/nemotron_h) hybrid
Mamba-Transformer language model with a [RADIO](https://huggingface.co/docs/transformers/main/en/model_doc/radio) vision encoder and an optional Parakeet-based sound encoder.
Image (and video) patches are projected through a RADIO tower and a pixel-shuffle MLP into the language model's
embedding space at the `<image>` / `<video>` context-token positions; audio clips are projected in the same way at
`<audio>` positions. The result is a single autoregressive model that reasons jointly over text, images, video and
sound.

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/nemotron_h_omni)
* Add support for Nemotron Omni (#46509) by @meatybobby in [#46509](https://github.com/huggingface/transformers/pull/46509)

### HyperCLOVAX Vision V2

HyperCLOVAX Vision V2 is a multimodal vision-language model developed by NAVER. It combines the [HyperClovaX](https://huggingface.co/docs/transformers/main/en/model_doc/hyperclovax) language model backbone with a [Qwen2.5-VL](https://huggingface.co/docs/transformers/main/en/model_doc/qwen2_5_vl) vision encoder. The model supports text, image, and video inputs and is capable of chain-of-thought reasoning via built-in thinking tokens (`<think>...</think>`).

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/hyperclovax_vision_v2)
* add HyperClovaX Vision (#44314) by @jp1924 in [#44314](https://github.com/huggingface/transformers/pull/44314)

### GTE

GTE was proposed in [mGTE: Generalized Long-Context Text Representation and Reranking Models for Multilingual Text Retrieval](https://huggingface.co/papers/2407.19669) by Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li and Min Zhang.

GTE is a BERT-style bidirectional encoder that replaces absolute position embeddings with RoPE, uses a gated MLP, and applies layer normalization after each residual connection. The same architecture backs Alibaba's `gte-*-v1.5`, `gte-multilingual-*` and `gte-en-mlm-*` checkpoints as well as Snowflake's `snowflake-arctic-embed-m-v2.0`.

**Links:** [Documentation](https://huggingface.co/docs/transformers/main/en/model_doc/gte)
* model: Add GTE to Transformers (#48416) by @harshaljanjani in [#48416](https://github.com/huggingface/transformers/pull/48416)


## Breaking changes

* 🚨 [ROCm] gpt-oss: route FA3 to aiter-flash-attn, generate ROCm fixtures (#46837) by @Abdennacer-Badaoui
* 🚨 Speed up detr image processing (#48066) by @guarin
* 🚨 Remap indexers layer_type (#48974) by @Cyrilvallez
* 🚨 [vLLM] Fix video token counting for Transformers backend video inputs (Part 1) (#48894) by @harshaljanjani
* 🚨 🚨Bring some dinos to modern standards (#46266) by @molbap
* :rotating_light: [`Kernels`] Bump version (#48714) by @vasqu


## Bugfixes and improvements

* fix incorrect hub tokenizer class (#48641) by @itazap
* [CI] Deduplicate Nvidia/AMD CI reply comments and add headers (#48655) by @ydshieh
* Update dev (#48654) by @vasqu
* [tests] Fix NougatModelIntegrationTest: pin artifact revision and update golden values (#48638) by @ydshieh
* [fix] Fix GlmOcr integration tests: wrong token IDs and image token decode bug (#48650) by @ydshieh
* [GLM 5.3 Flash] Preserve original names when saving text checkpoints (#48676) by @Dovis01
* Register activation kernel layers on XPU (#47858) by @jiqing-feng
* Mirror the registered pytree flatten when building automatic dynamic shapes (#48578) by @IlyasMoutawwakil
* Fix `AutoImageProcessor` requiring torchvision when only Pillow is installed (#48616) by @blipbyte
* Assert cached decode matches recomputing without a cache (#48289) by @IlyasMoutawwakil
* [higgs_audio_v2] Fix: use config.num_codebooks in audio labels tensor (Simple Fix!) (#48560) by @b-re-w
* [`MoE`] Fix eager EP (#48653) by @vasqu
* [MusicgenMelody] Fix conditioning silently dropped at generation step 0 (#48679) by @ydshieh
* Fix deepspeed ci (#48640) by @SunMarc
* Synthetic test assets (#48589) by @tarekziade
* Skip the expert-parallel sentinel masking when expert parallelism is off (#48201) by @qgallouedec
* [Fix] yolos offload issue (#48688) by @molbap
* [serge] Fix 2 integration tests for model `minimax` failing with `output_mismatch` (tensor values differ (2)) (#48515) by @sergereview[bot]
* [serge] Fix 2 integration tests for model `mistral` failing with `other` (other (2)) (#48429) by @sergereview[bot]
* Fix Qwen2.5-VL temporal RoPE for fractional video intervals (#48669) by @yeyeyeping
* Fix slow integration tests on XPU (#48611) by @jiqing-feng
* QA: Added a `MemoryCleanupMixin` class for tests (#48681) by @tarekziade
* [serge] Fix 1 integration test for model `flex_olmo` failing with `other` (other (1)) (#48668) by @sergereview[bot]
* Fix D-FINE / RT-DETR main loss being computed over the denoising queries (#48528) by @stefan-it
* [CI] Replace hardcoded username allowlists with author_association check in workflow triggers (#48712) by @ydshieh
* Add flash_attention_4 in attn_implementation AutoModel docstring (#48684) by @3manifold
* Keep the expert-parallel sentinel slots out of the router gradient (#48689) by @qgallouedec
* [DeepseekV3] OOM cascade root-cause investigation (generator ref leak in conversion_mapping) (#48720) by @ydshieh
* Use huggingface_hub httpx export (#48685) by @Wauplin
* unifying device_mesh init to enable PP + TP inference (#48155) by @3outeille
* Enable FSDP2 + expert parallelism via a 2-D (fsdp, tp) device mesh (#48516) by @qgallouedec
* Switch daily CI to torch 2.14 — update expected outputs (#48750) by @ydshieh
* fix videomae load error (#48675) by @sywangyi
* Fix silently random-initializing `RTDetrModel`/`SEWDForCTC` loads (wrong `base_model_prefix`) (#48744) by @<NOT FOUND>
* Fix Glm4vMoeIntegrationTest: offload_folder + MemoryCleanupMixin (#48776) by @ydshieh
* [InternVL] Normalize num_patches before np.cumsum (#48469) by @lorenzozanee
* Kernel api doc, part 2 (#46889) by @michaelbenayoun
* Fix hidden state selection on Gemma4 assistant's first prefill step (#48704) by @glistening
* Fix inkling embedding norm (#48786) by @Cyrilvallez
* Fix TextToAudioPipeline crash for tokenizer-only models (#48505) by @jiqing-feng
* Cast pixel values to the patch embedding dtype in DeepSeek-OCR-2 (#48632) by @jiqing-feng
* add kernel mapping entry for RMSNormGated, KDA, Conv1D on XPU (#48702) by @kaixuanliu
* Bump `peft` version requirement (#48716) by @shniubobo
* [serge] Fix 12 integration tests for model `edgetam` failing with `import_or_config` (other (12)) (#48322) by @sergereview[bot]
* Restore legacy tensor-parallel initialization for compatibility (#48797) by @3outeille
* Fix kosmos flaky test (#48800) by @IlyasMoutawwakil
* Add integration tests for MuseGlimmerAssistantModel (#48796) by @ydshieh
* Fix `AutoModel.from_pretrained` not restoring `modules_to_save` weights (#48595) by @shniubobo
* [docs] mrope and axial rope (#48717) by @stevhliu
* Fix decoder only path for older bert variants (#48785) by @Cyrilvallez
* [Fix] Clean-up ternaries in the DeepSeek family (#48447) by @remi-or
* [utils] Add MUSA support for Flash Attention 2 (#48612) by @XiaomingFun233
* [generate] Drop attention mask early without padding (#48814) by @Cyrilvallez
* Pin utf-8 in test_can_init_all_missing_weights source read (Windows non-UTF-8 locale fix) (#48819) by @dltsum
* Relax static-cache tolerance in test_generate_with_static_cache (1e-5 → 5e-5) (#48815) by @ydshieh
* Detect nested rope_parameters without relying on layer_types (#48798) by @hmellor
* Support for MoE in the GGUF integration (#48529) by @SunMarc
* [docs] gguf (#46357) by @stevhliu
* [apply_chat_template] pass sampling_rate to call (#48794) by @eustlb
* [CI] Add link checker (#48160) by @stevhliu
* Qwen3.8 GGUF (#48660) by @SunMarc
* docs: fix broken #combining-with-fsdp2 anchor in expert_parallelism.md (#48854) by @ydshieh
* [GPTNeoXJapanese] Fix RoPE ignoring partial_rotary_factor (#48652) by @blipbyte
* QA: deactivate rule 41 (#48852) by @tarekziade
* Fix `reset` on the dynamic cache layers (#48809) by @jiqing-feng
* [generate] Make all methods and logits processors agnostic to lm_head output size (#48846) by @Cyrilvallez
* Fix kosmos (#48853) by @Cyrilvallez
* Add test for causal only variant of some encoder-decoder models (#48760) by @nandan2003
* Fix ESMFold2 ligand iPLDDT weighting: mol_type non-polymer code is 3, not 4 (#48831) by @faustomilletari
* Fix flaky RfDetr test_save_load (#48869) by @ydshieh
* fix(ci): harden GitHub Actions workflows (#48160) (#48834) by @hf-security-analysis[bot]
* Fix Glm4MoeIntegrationTest: split into 3 tests, switch to GLM-4.5-Air (#48820) by @ydshieh
* Fix `StaticCache` for Mllama and enable `torch.compile` (#48141) by @jiqing-feng
* Fix Minimax M2 partial_rotary_factor by remapping legacy rotary_dim (#48486) by @dajiaohuang
* Fix Pixtral image processor do_pad option (#48872) by @jzakrzew
* [CB] Add pause mechanism (#48462) by @remi-or
* Fix DFlash sampled candidates losing the batch dimension (#48281) by @VaggelisGian
* Fix sliding window cache when assistant model calls generate (#48280) by @VaggelisGian
* Fix GLM rope_parameters dict shared mutation across test instances (#48895) by @ydshieh
* [`feat`] Allow untying `hidden_states[-1]` from `last_hidden_state` via the model config (#48087) by @tomaarsen
* Fix Idefics2 padding when the first example has no image (#48753) by @Arnavsharma2
* Don't fail model loading when the accelerator can't report free memory (#48136) by @studioego
* [Improvement] Rework the docstring and comments of mHC (#48888) by @remi-or
* [serge] Fix 3 integration tests for model `deepseek_vl` failing with `other` (other (3)) (#48536) by @sergereview[bot]
* Add auto_docstring task model overrides (#47815) by @guarin
* Make LongcatFlashConfig self-consistent without building the model (#48899) by @hmellor
* [`DSA`] Only save latents on dsa with indexer as well (#48876) by @vasqu
* [Chat Parsing] Coerce oneOf tool arguments (#48719) by @yonigozlan
* Always materialize the causal mask in Doge so sdpa stays causal (#48821) by @blipbyte
* Fix docstring argument names that don't match signatures (#48474) by @IshaanPotle
* Mark test_generate_with_static_cache as flaky for olmo and bigbird_pegasus (#48856) by @ydshieh
* Clean up some MoE models' integration tests (#48833) by @ydshieh
* Fix T5 tied weights order for lm_head (#48238) by @vaibhavmashal
* Size the 
```



document fetched 2026-10-07 02:34
