main
Better and more extensive tests for RoPE (#46912)
Refactor image loading in tests to use load_test_image helper (#47218)
:rotating_light: Validate config attributes (#41250)
processor tests: use tiny Hub repos to reduce CI memory (#47115)
🚨 fix + tests dense & MoE TP all reduce (decoder only) (#43722)
🚨 Remove torch.fx support (#41683)
Fix Aria checkpoint key conversion mapping (#47151)
Use common floats_list method for feature extractor tests. (#46956)
[AudioFlamingo3] Fix bfloat16 dtype mismatch in audio encoder positional embedding (#47258)
[Mistral] Add native tekken tokenizer support to AutoTokenizer (#47507)
[PoC] HF exporters (#41992)
[Fix] Fix multi-head latent attention (MLA) (#47761)
Fix ci style (#47818)
tests: reduce processor test memory usage and use tiny test assets (#47168)
🚨 [`Kernels`] Refactor all linear attn models & native kernels fallback (#47630)
Fix path traversal when saving Bark voice preset embeddings (#46237)
Enforce min length in some generate tests (#44401)
rm slow tokenizers (#40936)
Return logits from semantic segmentation post-process (#46163)
Fix models for which we don't have a dedicated tokenizer class, and the listed one is incorrect (#45936)
Delete irrelevant test that sometimes fails (#42515)
Enable FURB rules in ruff (#41395)
Fix left-padding token selection in `BioGptForSequenceClassification` (#46782)
🚨🚨 Refactor Image Processors to support different backends (#43514)
🚨 More V5 pipeline cleanup (#43325)
Gemma4 resizing per layer inputs (#45324)
🚨 Enable SDPA (and other attention backends) for T5 and propagate to the T5 family (#47014)
:rotating_light: [`Attn`] New attn mask interface everywhere (#42848)
BLT: vectorize patch length processing (#47385)
Fix CANINE fp16 dtype mismatch in attention mask causing ONNX export failure (#47083)
Switch decorator order (#46853)
Fix model parallel issue for altclip model and ChineseClip model (#45487)
Refactor CLIP-like models (#44431)
Revert "Fix CodeLlama tokenizer dropping leading whitespace on decode (#47488)" (#47861)
Fix failing `CodeGenModelTests` (#42730)
Fix tests: 'Cohere2MoeModel' object has no attribute 'hf_device_map' (#46337)
Fix CohereASR training-loss double-shift (same as Moonshine fix #46784) (#46895)
Add CohereCompass modeling (#47878)
Fall back to flat kwarg when modality dict is passed without it (#46195)
tests: reduce processor test memory usage by using tiny Hub checkpoints (#47213)
Fix some failed test cases related with XPU Expectations (#47173)
Fix Cosmos 3 Edge Patch packing order (#47399)
add XPU Expectation for cosmos3_omni tests (#46880)
Use Python 3.9 syntax in tests (#37343)
Harmonize `input_embeds` to `inputs_embeds` everywhere (#43916)
Fix failing `salesforce-ctrl`, `xlm` & `gpt-neo` model generation tests (#43180)
fix: compute auxiliary losses when denoising is disabled in D-FINE (#45601)
[fix] update integration test values (#47146)
Fix DBRX MoE hidden size and expert GLU transposes (#47671)
Fix missing post_processor in DebertaV2Tokenizer causing no special t… (#44570)
Better processing tests (#46374)
model: Add DEIMv2 to Transformers (#44339)
🚨 Refactor ViT to updated standards (#41693)
[generate] Never use `cache_position` anymore in generation (#44816)
docs(tests): fix typos in test comments (#47859)
Make Transformers more torch-exportable and dynamo-friendly (#42317)
Strict export cleanup (#44293)
fix(flaky): enforce manual seed to reduce flakiness (#43794)
[causallm tester] automate pipeline mappings + bloom tests (#41318)
Dynamic weight conversion is recursive (#44300)
SAM2 Video support fp16 (#43268)
fix bug embedding_size mismatch with hidden_size in electra model test (#44657)
Single config attribute for weight tying (#42815)
fix(flaky): Different approach to make sure loss exists (#43804)
feat(pipelines): accept numpy arrays and tensors in ImageClassificationPipeline (#39607) (#46573)
Refactor all linear attention models to latest best standards for convolution (#47452)
:rotating_light: Processors update the rest (#46556)
Acc fix in xpu (#47500)
cancel deterministic for XPU in gemma4 tests (#47790)
First model (#45788)
who needs encoders? (#46385)
Update torch minimum version to 2.4 (#41307)
Fix MoE decode optimization for experts living in a submodel (#47107)
Improve `has_similar_generate_outputs` assertions (#44166)
[GLM-4.6V] Update with GLM-GA Processor (#46184)
Fix GPT2 attention scaling ignored in SDPA/FlashAttention (#44397)
Fix failing `GPTNeoModelLanguageGenerationTest` (#44515)
[`Kernels`] Refactor function handling (#46883)
[`Tokenizers`] Move gpt sw3 tokenizer out (#45404)
Fix GraniteConfig type hints to accept int for multiplier fields (#45019)
[`CI`] Fix some failures introduced by myself :grimacing: (#46751)
🚨 [ALM] Add base model without head (#45534)
Fix failing tests for granite_swa and granitemoe_swa (#47723)
🚨 Modeling changes for export, compile, and hybrid-attention standardization (#46738)
Use `HF_TOKEN` directly and remove `require_read_token` (#43233)
Add xcodec2 model (#44178)
fix(hrm_text): Add XPU Expectations for tests (#46214)
🚨 [v5] Refactor RoPE for layer types (#39847)
Fix Hunyuan-VL PIL image resize parity with reference preprocessing (#47233)
hy_v3: add XPU expectations (#45858)
hyperclovax: add XPU Expectations for CI test (#45926)
Remove many output_attentions and other traced outputs on 100+ models (#43590)
skip invalid test cases for inkling tests (#47493)
[XPU] Add flash_attn2 support for XPU (#41956)
Fix flash attention with StaticCache (#46914)
Remove remaining vestiges of the TranslationPipeline (#43869)
Fix failing `MarianIntegrationTests` (#44519)
[vllm + v5 fix] handle TokenizersBackend fallback properly for v5 (#44255)
Simplify and standardize processor tests (#41773)
Fix failing tests for mimo_v2_flash (#47284)
[Fix] Remove old automatic cross attn pattern from output recorders (#47117)
[cache] Remove all deprecated classes (#43168)
[style] Rework ruff rules and update all files (#43144)
[generate] Completely stop relying on `cache_position` to prepare inputs (#44130)
[CI] Temporarily skip Mistral4 tests as they almost all fail (#44825)
skip `MLukeTokenizerTest` temporarily (#42520)
[CI] AMD docker: bump to ROCm 7.2.2 / PyTorch 2.10 + prebuilt FA wheel (#45913)
Add ModernVBERT models (#42504)
Add MuseGlimmer (#47867)
[Qwen3ASR] Add hotword parsing, and fix language parsing and training. (#47111)
fix(models, testing): Fix Llama4 vision rotary meta tensor initialization and MyT5 get_tokenizer signature (#44581)
[MODEL] Nanochat implementation (#41634)
[NemotronAsrStreaming] fix pipeline (#46870)
Add Inkling model (#47347)
Refactor-tokenization-more (#42563)
Add ForSequenceClassification heads for the OLMo family (#45551)
Align OlmoHybrid to use a native cache in generate (#47604)
Add `base_model_tp_plan` to `OlmoeConfig` (#44668)
[`OAI Privacy Filter`] Add integration test (#45725)
Fix some processors (#47608)
fix(testing): Fix PaliGemma 2 and PaddleOCR-VL test failures on main (#44765)
Some new models added stuff that was already removed (#43179)
fix: prevent accelerate from splitting vision encoder by setting _no_… (#43047)
fix(models): Fix Perceiver interpolate_pos_encoding interpolating to the source size (#44899)
[v5] remove load_in_4bit and load_in_8bit (#41287)
Fix failing `PhiIntegrationTests` (#43214)
fix(testing): Fix Parakeet, Evolla, Pi0, and Phi-3 test failures on main CI (#45004)
Fix `PhimoeIntegrationTest` (#46539)
Add Pixio pre-trained models (#42795)
Updated the image cache for Paddle models according to the latest API (#45562)
[Model] Add PP-FormulaNet Model Support (#45626)
Fix more model tester missing `parent` issue (#44685)
[Model] Add PP-OCRv6 Models Support (#45838)
PP-OCRv6: update documentation and slow tests (#46576)
Avoid `Image.open` failure (#44645)
Fix loss alignment and Trainer token counting for encoder decoder models (#46903)
Add support for batched Qwen2.5/3-Omni audio generation (#47186)
Honor a concrete dtype in AutoModel for composite checkpoints (#46514)
Use tiny Hub checkpoint in Qwen3ASR processor test (#47833)
extend tests/models/qwen3_next/test_modeling_qwen3_next.py::Qwen3Next… (#47184)
Exportable kimi (#47096)
Add support for RADIO models (#46425)
Untangle config inheritance (#41541)
[RecurrentGemma] Support attn_implementation dispatch (#46320)
Fix post processing RF-DETR (#46041)
Allow position_ids_start=2 on DataCollatorWithFlattening for RoBERTa etc. (#47525)
Fix RT-DETR indexing error when num_feature_levels exceeds backbone o… (#46833)
Fix half-precision torch.compile crash in DETR-family sine position embeddings (#47238)
🚨Fix memory leaks caused by lru decorators in vision models (#45922)
[core] Faster and thread-safe `check_model_inputs` implementation (#43765)
[`Sam`] Fixup training flags (#43567)
Add native masked MSE loss for Sapiens2ForPoseEstimation (#46764)
Fix failing `SmolLM3IntegrationTest` (#45048)
Add XPU support to the tests for solar_open (#43579)
🚨 Generation config defaults are now `None` (#42702)
Fix a regression in encoder-decoder generation cache initialization (#46111)
🚨 Delete duplicate code in backbone utils (#43323)
Fix TimesFM 2.5 window_size AttributeError (#47363)
Fix/video classification pipeline video processor (#46256)
fix(utils): Resolve backbone utils test regressions (#45594)
fix(models): Fix dtype mismatch in SwitchTransformers and TimmWrapperModel (#45074)
Fix model tests for tipsv2 (#47356)
tipsv2_dpt: fix failing tests for XPU (#47292)
XPU now supports backward for the FA2 fixed path (#43905)
feat[vLLM x v5]: Make audio optional and support multiple audios in VibeVoice ASR processor (#47483)
[Weight Converter] More fine-grained mappings on classes, scoping for every transforms (including weight converter) (#45661)
PR CI with torch 2.13 (#47767)
Fix: accept scalar audio formats in Voxtral transcription requests (#47045)
fix(models): Resolve regressions in Wav2Vec2PhonemeCTCTokenizer (wav2vec2-lv-60-espeak-cv-ft) (#45199)
[v5] 🚨Refactor subprocessors handling in processors (#41633)
CPU can incur a slow path on non-contiguous magnitudes (#47351)
Fix `x_clip`: 8 failed test cases (#45394)
Fix Xcodec2 attention to be non-causal. (#46963)
Fix recurrent-layer padding mask being skipped on continued forwards (chunked prefill, cache continuation) (#47087)
Fix failing tests for zaya (#47268)
Move test model folders (#17034)