ArthurZ's picture
ArthurZ HF Staff
PR dots use GitHub Primer state colours (merged/open/draft/closed) + draft detection
a5c6a96 verified
Raw History Blame Contribute Delete
365 kB
// Generated by build2.py — do not hand-edit. // Spans are REAL: first commit of the earliest PR -> last commit / merge of the latest. // Source board: https://miro.com/app/board/uXjVL_vDn1k=/ const AXIS = ["2023-12-14", "2026-09-01"]; const RELEASES = [{"n": "v4.42.0", "d": "2024-06-27", "big": false}, {"n": "v4.43.0", "d": "2024-07-23", "big": false}, {"n": "v4.44.0", "d": "2024-08-06", "big": false}, {"n": "v4.45.0", "d": "2024-09-25", "big": false}, {"n": "v4.46.0", "d": "2024-10-24", "big": false}, {"n": "v4.47.0", "d": "2024-12-05", "big": false}, {"n": "v4.48.0", "d": "2025-01-10", "big": false}, {"n": "v4.49.0", "d": "2025-02-17", "big": false}, {"n": "v4.50.0", "d": "2025-03-21", "big": false}, {"n": "v4.51.0", "d": "2025-04-05", "big": false}, {"n": "v4.52.0", "d": "2025-05-20", "big": false}, {"n": "v4.53.0", "d": "2025-06-26", "big": false}, {"n": "v4.54.0", "d": "2025-07-25", "big": false}, {"n": "v4.55.0", "d": "2025-08-05", "big": false}, {"n": "v4.56.0", "d": "2025-08-29", "big": false}, {"n": "v4.57.0", "d": "2025-10-03", "big": false}, {"n": "v5.0.0rc0", "d": "2025-12-01", "big": true}, {"n": "v5.0.0rc1", "d": "2025-12-11", "big": false}, {"n": "v5.0.0rc2", "d": "2026-01-07", "big": false}, {"n": "v5.0.0rc3", "d": "2026-01-14", "big": false}, {"n": "v5.0.0", "d": "2026-01-26", "big": true}, {"n": "v5.1.0", "d": "2026-02-05", "big": false}, {"n": "v5.2.0", "d": "2026-02-16", "big": false}, {"n": "v5.3.0", "d": "2026-03-04", "big": false}, {"n": "v5.4.0", "d": "2026-03-27", "big": false}, {"n": "v5.5.0", "d": "2026-04-02", "big": false}, {"n": "v5.6.0", "d": "2026-04-22", "big": false}, {"n": "v5.7.0", "d": "2026-04-28", "big": false}, {"n": "v5.8.0", "d": "2026-05-05", "big": false}, {"n": "v5.9.0", "d": "2026-05-20", "big": false}, {"n": "v5.10.0", "d": "2026-06-03", "big": false}, {"n": "v5.11.0", "d": "2026-06-10", "big": false}, {"n": "v5.12.0", "d": "2026-06-12", "big": false}, {"n": "v5.13.0", "d": "2026-07-03", "big": false}, {"n": "v5.14.0", "d": "2026-07-15", "big": false}, {"n": "v5.15.0", "d": "2026-08-10", "big": false}, {"n": "v5.16.0", "d": "2026-08-26", "big": false}]; const MODELS = [{"m": "rt_detr", "d": "2024-06-22", "pr": 29077, "by": "Sangbum Daniel Choi", "t": "New model support RTDETR"}, {"m": "instructblipvideo", "d": "2024-06-25", "pr": 30182, "by": "Raushan Turganbay", "t": "Add video modality for InstrucBLIP"}, {"m": "llava_next_video", "d": "2024-06-26", "pr": 31252, "by": "Raushan Turganbay", "t": "Add LLaVa NeXT Video"}, {"m": "gemma2", "d": "2024-06-27", "pr": 31659, "by": "Arthur", "t": "Add gemma 2"}, {"m": "zoedepth", "d": "2024-07-08", "pr": 30136, "by": "NielsRogge", "t": "Add ZoeDepth"}, {"m": "hiera", "d": "2024-07-12", "pr": 30356, "by": "Naman Garg", "t": "Adding hiera"}, {"m": "chameleon", "d": "2024-07-17", "pr": 31534, "by": "Raushan Turganbay", "t": "Chameleon: add model"}, {"m": "nemotron", "d": "2024-08-06", "pr": 31699, "by": "Ao Tang", "t": "Add Nemotron HF Support"}, {"m": "mamba2", "d": "2024-08-06", "pr": 32080, "by": "Pablo Montalvo", "t": "Add codestral mamba2"}, {"m": "qwen2_audio", "d": "2024-08-08", "pr": 32137, "by": "Yunfei Chu", "t": "Add Qwen2-Audio"}, {"m": "falcon_mamba", "d": "2024-08-12", "pr": 32615, "by": "Younes Belkada", "t": "Add new model"}, {"m": "dac", "d": "2024-08-19", "pr": 31494, "by": "Kamil Akesbi", "t": "Add Descript-Audio-Codec model"}, {"m": "qwen2_vl", "d": "2024-08-26", "pr": 32318, "by": "Shijie", "t": "support qwen2-vl"}, {"m": "granite", "d": "2024-08-27", "pr": 31502, "by": "Mayank Mishra", "t": "Granite language models"}, {"m": "olmoe", "d": "2024-09-03", "pr": 32406, "by": "Niklas Muennighoff", "t": "Add OLMoE"}, {"m": "llava_onevision", "d": "2024-09-05", "pr": 32673, "by": "Raushan Turganbay", "t": "Llava Onevision: add model"}, {"m": "pixtral", "d": "2024-09-14", "pr": 33449, "by": "Arthur", "t": "Add support for Pixtral"}, {"m": "mimi", "d": "2024-09-18", "pr": 33565, "by": "Yoach Lacombe", "t": "Codec integration"}, {"m": "granitemoe", "d": "2024-09-20", "pr": 33207, "by": "Mayank Mishra", "t": "Granitemoe"}, {"m": "omdet_turbo", "d": "2024-09-25", "pr": 31843, "by": "Yoni Gozlan", "t": "Add OmDet-Turbo"}, {"m": "mllama", "d": "2024-09-25", "pr": 33703, "by": "Arthur", "t": "Add MLLama"}, {"m": "idefics3", "d": "2024-09-25", "pr": 32473, "by": "Andrés Marafioti", "t": "Add Idefics 3!"}, {"m": "phimoe", "d": "2024-10-04", "pr": 33363, "by": "Amit Garg", "t": "PhiMoE"}, {"m": "zamba", "d": "2024-10-04", "pr": 30950, "by": "pglorio", "t": "Add Zamba"}, {"m": "myt5", "d": "2024-10-06", "pr": 31286, "by": "TomLim", "t": "[WIP] Add Tokenizer for MyT5 Model"}, {"m": "moshi", "d": "2024-10-16", "pr": 33624, "by": "Yoach Lacombe", "t": "Moshi integration"}, {"m": "glm", "d": "2024-10-18", "pr": 33823, "by": "Cyril Vallez", "t": "add Glm"}, {"m": "olmo_1124", "d": "2024-11-18", "pr": 34551, "by": "Shane A", "t": "Add OLMo November 2024"}, {"m": "ijepa", "d": "2024-12-05", "pr": 33125, "by": "João Marcelo", "t": "Add I-JEPA"}, {"m": "aria", "d": "2024-12-06", "pr": 34157, "by": "Aymeric Roucher", "t": "Add Aria"}, {"m": "timm_wrapper", "d": "2024-12-11", "pr": 34564, "by": "Pavel Iakubovskii", "t": "Add TimmWrapper"}, {"m": "cohere2", "d": "2024-12-13", "pr": 35224, "by": "alexrs-cohere", "t": "Add Cohere2 model"}, {"m": "colpali", "d": "2024-12-17", "pr": 33736, "by": "Tony Wu", "t": "Add ColPali to 🤗 transformers"}, {"m": "bamba", "d": "2024-12-19", "pr": 34982, "by": "Yu Chin Fabian Lim", "t": "Add the Bamba Model"}, {"m": "modernbert", "d": "2024-12-19", "pr": 35158, "by": "Benjamin Warner", "t": "Add ModernBERT to Transformers"}, {"m": "dinov2_with_registers", "d": "2024-12-24", "pr": 35348, "by": "NielsRogge", "t": "Add DINOv2 with registers"}, {"m": "diffllama", "d": "2025-01-07", "pr": 34083, "by": "松本和真", "t": "Add diffllama"}, {"m": "textnet", "d": "2025-01-08", "pr": 34979, "by": "Jade Choghari", "t": "Add TextNet"}, {"m": "vitpose", "d": "2025-01-08", "pr": 30530, "by": "NielsRogge", "t": "Add ViTPose"}, {"m": "vitpose_backbone", "d": "2025-01-08", "pr": 30530, "by": "NielsRogge", "t": "Add ViTPose"}, {"m": "moonshine", "d": "2025-01-10", "pr": 34784, "by": "eustlb", "t": "Add Moonshine"}, {"m": "emu3", "d": "2025-01-10", "pr": 33770, "by": "Raushan Turganbay", "t": "[WIP] Emu3: add model"}, {"m": "helium", "d": "2025-01-13", "pr": 35669, "by": "Arthur", "t": "Add-helium"}, {"m": "superglue", "d": "2025-01-20", "pr": 29886, "by": "StevenBucaille", "t": "Add SuperGlue model"}, {"m": "qwen2_5_vl", "d": "2025-01-23", "pr": 35569, "by": "ShuaiBai623", "t": "add qwen2.5vl"}, {"m": "zamba2", "d": "2025-01-27", "pr": 34517, "by": "pglorio", "t": "Add Zamba2"}, {"m": "got_ocr2", "d": "2025-01-31", "pr": 34721, "by": "Yoni Gozlan", "t": "Add GOT-OCR 2.0 to Transformers"}, {"m": "dab_detr", "d": "2025-02-04", "pr": 30803, "by": "David", "t": "Add DAB-DETR for object detection"}, {"m": "rt_detr_v2", "d": "2025-02-06", "pr": 34773, "by": "Jade Choghari", "t": "Adding RT-DETRv2 for object detection"}, {"m": "depth_pro", "d": "2025-02-10", "pr": 34583, "by": "Armaghan Shakir", "t": "Add Apple's Depth-Pro for depth estimation"}, {"m": "granitemoeshared", "d": "2025-02-14", "pr": 35894, "by": "Mayank Mishra", "t": "add shared experts for upcoming Granite 4.0 language models"}, {"m": "smolvlm", "d": "2025-02-20", "pr": 36126, "by": "Orr Zohar", "t": "SmolVLM2"}, {"m": "siglip2", "d": "2025-02-21", "pr": 36323, "by": "Pavel Iakubovskii", "t": "Add SigLIP 2"}, {"m": "aya_vision", "d": "2025-03-04", "pr": 36521, "by": "Arthur", "t": "Add aya"}, {"m": "gemma3", "d": "2025-03-12", "pr": 36658, "by": "Ryan Mullins", "t": "Gemma3"}, {"m": "mistral3", "d": "2025-03-18", "pr": 36790, "by": "Cyril Vallez", "t": "Add Mistral3"}, {"m": "shieldgemma2", "d": "2025-03-20", "pr": 36678, "by": "Ryan Mullins", "t": "Shieldgemma2"}, {"m": "prompt_depth_anything", "d": "2025-03-21", "pr": 35401, "by": "Haotong LIN", "t": "Add Prompt Depth Anything Model"}, {"m": "phi4_multimodal", "d": "2025-03-25", "pr": 36939, "by": "Cyril Vallez", "t": "Add Phi4 multimodal"}, {"m": "deepseek_v3", "d": "2025-03-28", "pr": 35926, "by": "Minho Ryu", "t": "[WIP] add deepseek-v3"}, {"m": "qwen3", "d": "2025-03-31", "pr": 36878, "by": "Bo Zheng", "t": "Adding Qwen3 and Qwen3MoE"}, {"m": "qwen3_moe", "d": "2025-03-31", "pr": 36878, "by": "Bo Zheng", "t": "Adding Qwen3 and Qwen3MoE"}, {"m": "llama4", "d": "2025-04-05", "pr": 37307, "by": "Arthur", "t": "Add llama4"}, {"m": "glm4", "d": "2025-04-09", "pr": 37388, "by": "Arthur", "t": "Add glm4"}, {"m": "granite_speech", "d": "2025-04-11", "pr": 36801, "by": "Alex Brooks", "t": "Add Granite Speech Support"}, {"m": "qwen2_5_omni", "d": "2025-04-14", "pr": 36752, "by": "BakerBunker", "t": "Add Qwen2.5-Omni"}, {"m": "mlcd", "d": "2025-04-15", "pr": 36182, "by": "Huajie Tan", "t": "Add MLCD model"}, {"m": "timesfm", "d": "2025-04-16", "pr": 34082, "by": "Jinan Zhou", "t": "Add TimesFM Time Series Forecasting Model"}, {"m": "janus", "d": "2025-04-17", "pr": 36053, "by": "Yaswanth Gali", "t": "Add Janus model"}, {"m": "internvl", "d": "2025-04-18", "pr": 35968, "by": "Yoni Gozlan", "t": "Add InternVL (2.5 MPO)"}, {"m": "bitnet", "d": "2025-04-28", "pr": 37742, "by": "Mohamed Mekkouri", "t": "Add Bitnet model"}, {"m": "sam_hq", "d": "2025-04-28", "pr": 35147, "by": "sushmanth reddy", "t": "Samhq model addition"}, {"m": "d_fine", "d": "2025-04-29", "pr": 36261, "by": "Vladislav Bronzov", "t": "Add D-FINE Model into Transformers"}, {"m": "hgnet_v2", "d": "2025-04-29", "pr": 36261, "by": "Vladislav Bronzov", "t": "Add D-FINE Model into Transformers"}, {"m": "granitemoehybrid", "d": "2025-05-05", "pr": 37658, "by": "Sukriti Sharma", "t": "Add GraniteMoeHybrid support for 4.0"}, {"m": "csm", "d": "2025-05-07", "pr": 36719, "by": "eustlb", "t": "Add CSM model"}, {"m": "falcon_h1", "d": "2025-05-21", "pr": 38249, "by": "Younes Belkada", "t": "[MODEL] Add Falcon H1"}, {"m": "colqwen2", "d": "2025-06-02", "pr": 35778, "by": "Tony Wu", "t": "Add ColQwen2 to 🤗 transformers"}, {"m": "minimax", "d": "2025-06-04", "pr": 35831, "by": "Armaghan Shakir", "t": "Add support for MiniMax's MiniMax-Text-01"}, {"m": "vjepa2", "d": "2025-06-11", "pr": 38746, "by": "Pavel Iakubovskii", "t": "Add V-JEPA 2"}, {"m": "lightglue", "d": "2025-06-17", "pr": 31718, "by": "StevenBucaille", "t": "Add LightGlue model"}, {"m": "arcee", "d": "2025-06-24", "pr": 38621, "by": "Crystalcareai", "t": "Add Arcee model support"}, {"m": "stt", "d": "2025-06-24", "pr": 38909, "by": "eustlb", "t": "Add kyutai stt"}, {"m": "glm4v", "d": "2025-06-25", "pr": 38431, "by": "Yuxuan Zhang", "t": "GLM-4.1V Model support"}, {"m": "t5gemma", "d": "2025-06-25", "pr": 38332, "by": "Biao Zhang", "t": "Encoder-Decoder Gemma"}, {"m": "dots1", "d": "2025-06-25", "pr": 38143, "by": "redmoe-moutain", "t": "[Model] add dots1"}, {"m": "smollm3", "d": "2025-06-25", "pr": 38755, "by": "Anton Lozhkov", "t": "Add SmolLM3"}, {"m": "dia", "d": "2025-06-26", "pr": 38405, "by": "Jaeyong Sung", "t": "Add Dia model"}, {"m": "gemma3n", "d": "2025-06-26", "pr": 39059, "by": "Ryan Mullins", "t": "Gemma 3n"}, {"m": "eomt", "d": "2025-06-27", "pr": 37610, "by": "Yaswanth Gali", "t": "✨ Add EoMT Model || 🚨 Fix Mask2Former loss calculation"}, {"m": "doge", "d": "2025-07-08", "pr": 35891, "by": "Jingze Shi", "t": "Add Doge model"}, {"m": "aimv2", "d": "2025-07-08", "pr": 36625, "by": "Yaswanth Gali", "t": "Add Aimv2 model"}, {"m": "deepseek_v2", "d": "2025-07-09", "pr": 36400, "by": "Vladislav Bronzov", "t": "Add DeepSeek V2 Model into Transformers"}, {"m": "lfm2", "d": "2025-07-10", "pr": 39340, "by": "Paul Pak", "t": "LFM2"}, {"m": "perception_lm", "d": "2025-07-11", "pr": 37878, "by": "Shuming Hu", "t": "PerceptionLM"}, {"m": "modernbert_decoder", "d": "2025-07-15", "pr": 38967, "by": "Orion Weller", "t": "Add ModernBERT Decoder Models - ModernBERT, but trained with CLM!"}, {"m": "voxtral", "d": "2025-07-18", "pr": 39429, "by": "eustlb", "t": "Add voxtral"}, {"m": "glm4_moe", "d": "2025-07-21", "pr": 39393, "by": "Yuxuan Zhang", "t": "GLM-4 Update"}, {"m": "ernie4_5", "d": "2025-07-21", "pr": 39228, "by": "Anton Vlasjuk", "t": "[`Ernie 4.5`] Add ernie text models"}, {"m": "ernie4_5_moe", "d": "2025-07-21", "pr": 39228, "by": "Anton Vlasjuk", "t": "[`Ernie 4.5`] Add ernie text models"}, {"m": "efficientloftr", "d": "2025-07-22", "pr": 36355, "by": "StevenBucaille", "t": "Add EfficientLoFTR model"}, {"m": "deepseek_vl", "d": "2025-07-25", "pr": 36248, "by": "Armaghan Shakir", "t": "Add support for DeepseekAI's DeepseekVL"}, {"m": "deepseek_vl_hybrid", "d": "2025-07-25", "pr": 36248, "by": "Armaghan Shakir", "t": "Add support for DeepseekAI's DeepseekVL"}, {"m": "xlstm", "d": "2025-07-25", "pr": 39665, "by": "Cyril Vallez", "t": "Add xlstm model"}, {"m": "evolla", "d": "2025-07-26", "pr": 36232, "by": "Xibin Bayes Zhou", "t": "Add evolla rebase main"}, {"m": "exaone4", "d": "2025-07-26", "pr": 39129, "by": "lgai-exaone", "t": "Add EXAONE 4.0 model"}, {"m": "cohere2_vision", "d": "2025-07-31", "pr": 39810, "by": "Raushan Turganbay", "t": "[Model] Cohere2 Vision"}, {"m": "mm_grounding_dino", "d": "2025-08-01", "pr": 37925, "by": "rziga", "t": "Add MM Grounding DINO"}, {"m": "gpt_oss", "d": "2025-08-05", "pr": 39923, "by": "Arthur", "t": "Add GPT OSS model from OpenAI"}, {"m": "glm4v_moe", "d": "2025-08-08", "pr": 39805, "by": "Yuxuan Zhang", "t": "GLM-4.5V Model Support"}, {"m": "sam2", "d": "2025-08-14", "pr": 32317, "by": "Sangbum Daniel Choi", "t": "Add Segment Anything 2 (SAM2)"}, {"m": "sam2_video", "d": "2025-08-14", "pr": 32317, "by": "Sangbum Daniel Choi", "t": "Add Segment Anything 2 (SAM2)"}, {"m": "dinov3_convnext", "d": "2025-08-14", "pr": 40167, "by": "Pavel Iakubovskii", "t": "Fix docs typo"}, {"m": "dinov3_vit", "d": "2025-08-14", "pr": 40167, "by": "Pavel Iakubovskii", "t": "Fix docs typo"}, {"m": "xcodec", "d": "2025-08-15", "pr": 38248, "by": "Manal ML", "t": "Add X-Codec model"}, {"m": "ovis2", "d": "2025-08-18", "pr": 37088, "by": "Eon Kim", "t": "Add Ovis2 model and processor implementation"}, {"m": "kosmos2_5", "d": "2025-08-19", "pr": 31711, "by": "tic-top", "t": "Add Kosmos-2.5"}, {"m": "metaclip_2", "d": "2025-08-20", "pr": 39826, "by": "NielsRogge", "t": "Add MetaCLIP 2"}, {"m": "florence2", "d": "2025-08-20", "pr": 38188, "by": "Duc-Viet Hoang", "t": "Add support for Florence-2"}, {"m": "hunyuan_v1_dense", "d": "2025-08-22", "pr": 39606, "by": "yjc9696", "t": "HunYuan opensource"}, {"m": "hunyuan_v1_moe", "d": "2025-08-22", "pr": 39606, "by": "yjc9696", "t": "HunYuan opensource"}, {"m": "seed_oss", "d": "2025-08-22", "pr": 40272, "by": "Fazzie", "t": "Addiing ByteDance Seed Seed-OSS"}, {"m": "apertus", "d": "2025-08-28", "pr": 39381, "by": "EduardDurech", "t": "Add Apertus"}, {"m": "qwen3_next", "d": "2025-09-10", "pr": 40771, "by": "Bo Zheng", "t": "Adding Support for Qwen3-Next"}, {"m": "ministral", "d": "2025-09-11", "pr": 40247, "by": "Manuel de Prada Corral", "t": "⚠️ 🔴 Add ministral model"}, {"m": "vaultgemma", "d": "2025-09-12", "pr": 40851, "by": "Ryan Mullins", "t": "add: differential privacy research model"}, {"m": "qwen3_vl", "d": "2025-09-15", "pr": 40795, "by": "JJJYmmm", "t": "Adding Support for Qwen3-VL Series"}, {"m": "qwen3_vl_moe", "d": "2025-09-15", "pr": 40795, "by": "JJJYmmm", "t": "Adding Support for Qwen3-VL Series"}, {"m": "olmo3", "d": "2025-09-16", "pr": 40778, "by": "Shane A", "t": "Add Olmo3 model"}, {"m": "longcat_flash", "d": "2025-09-17", "pr": 40730, "by": "Pablo Montalvo", "t": "Add LongCat-Flash"}, {"m": "flex_olmo", "d": "2025-09-18", "pr": 40921, "by": "Shane A", "t": "Add FlexOlmo model"}, {"m": "lfm2_vl", "d": "2025-09-18", "pr": 40624, "by": "Raushan Turganbay", "t": "Add new model LFM2-VL"}, {"m": "blt", "d": "2025-09-19", "pr": 38579, "by": "Ita Zaporozhets", "t": "blt wip"}, {"m": "qwen3_omni_moe", "d": "2025-09-21", "pr": 41025, "by": "BakerBunker", "t": "Adding support for Qwen3Omni"}, {"m": "parakeet", "d": "2025-09-25", "pr": 39062, "by": "Nithin Rao", "t": "Add Parakeet"}, {"m": "edgetam", "d": "2025-09-29", "pr": 39800, "by": "Yoni Gozlan", "t": "Add EdgeTAM"}, {"m": "edgetam_video", "d": "2025-09-29", "pr": 39800, "by": "Yoni Gozlan", "t": "Add EdgeTAM"}, {"m": "lfm2_moe", "d": "2025-10-07", "pr": 41401, "by": "Paul Pak", "t": "[Model] Lfm2Moe"}, {"m": "cwm", "d": "2025-10-09", "pr": 41199, "by": "Jacob Kahn", "t": "Add Code World Model (CWM)"}, {"m": "video_llama_3", "d": "2025-10-13", "pr": 40499, "by": "Kehan Li", "t": "[model] Add VideoLLaMA3 implementation"}, {"m": "audioflamingo3", "d": "2025-11-12", "pr": 40290, "by": "Lasha Koroshinadze", "t": "[models] Add AudioFlamingo3 integration"}, {"m": "glm46v", "d": "2025-11-15", "pr": 42122, "by": "Yuxuan Zhang", "t": "GLM-V update with new processor"}, {"m": "sam3", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "sam3_tracker", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "sam3_tracker_video", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "sam3_video", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "nanochat", "d": "2025-11-27", "pr": 41634, "by": "burtenshaw", "t": "[MODEL] Nanochat implementation"}, {"m": "afmoe", "d": "2025-11-29", "pr": 42168, "by": "Raghav Ravishankar", "t": "Add afmoe model"}, {"m": "t5gemma2", "d": "2025-12-01", "pr": 41834, "by": "Biao Zhang", "t": "T5gemma2"}, {"m": "ministral3", "d": "2025-12-01", "pr": 42498, "by": "Patrick von Platen", "t": "[Ministral 3] Add ministral 3"}, {"m": "fast_vlm", "d": "2025-12-02", "pr": 41112, "by": "Kamila Luczaj (Woo-chai)", "t": "Add FastVLM"}, {"m": "lasr", "d": "2025-12-05", "pr": 42648, "by": "eustlb", "t": "Lasr model"}, {"m": "paddleocr_vl", "d": "2025-12-11", "pr": 42178, "by": "zhang-prog", "t": "[Model] Add PaddleOCR-VL Model Support"}, {"m": "pe_audio", "d": "2025-12-16", "pr": 42905, "by": "eustlb", "t": " Sam: Perception Encoder Audiovisual"}, {"m": "pe_audio_video", "d": "2025-12-16", "pr": 42905, "by": "eustlb", "t": " Sam: Perception Encoder Audiovisual"}, {"m": "pe_video", "d": "2025-12-16", "pr": 42905, "by": "eustlb", "t": " Sam: Perception Encoder Audiovisual"}, {"m": "jais2", "d": "2025-12-17", "pr": 42684, "by": "sarathc-cerebras", "t": "adds jais2 model support"}, {"m": "pixio", "d": "2025-12-17", "pr": 42795, "by": "Lihe Yang", "t": "Add Pixio pre-trained models"}, {"m": "ernie4_5_vl_moe", "d": "2025-12-19", "pr": 39585, "by": "Anton Vlasjuk", "t": "[`Ernie 4.5`] Ernie VL models"}, {"m": "glmasr", "d": "2025-12-24", "pr": 42875, "by": "Yuxuan Zhang", "t": "GLM-ASR Support"}, {"m": "minimax_m2", "d": "2026-01-09", "pr": 42028, "by": "Roger Young", "t": "Add support for MiniMax-M2"}, {"m": "lw_detr", "d": "2026-01-12", "pr": 40991, "by": "StevenBucaille", "t": "Add LWDetr model"}, {"m": "glm4_moe_lite", "d": "2026-01-13", "pr": 43031, "by": "Yuxuan Zhang", "t": "[GLM-4.7] GLM-Lite Supoort"}, {"m": "glm_image", "d": "2026-01-13", "pr": 43100, "by": "Yuxuan Zhang", "t": "[GLM-Image] AR Model Support for GLM-Image"}, {"m": "lighton_ocr", "d": "2026-01-14", "pr": 41621, "by": "Aubertin Baptiste", "t": "Add LightOnOCR model implementation"}, {"m": "solar_open", "d": "2026-01-22", "pr": 43244, "by": "Inseo Song", "t": "Add Solar-Open Model"}, {"m": "glm_ocr", "d": "2026-01-27", "pr": 43391, "by": "Yuxuan Zhang", "t": "[GLM-OCR] GLM-OCR Support"}, {"m": "youtu", "d": "2026-01-28", "pr": 43166, "by": "Lu junru", "t": "Add Youtu-LLM model"}, {"m": "pp_doclayout_v3", "d": "2026-01-29", "pr": 43098, "by": "zhang-prog", "t": "[Model] Add PP-DocLayoutV3 Model Support"}, {"m": "eomt_dinov3", "d": "2026-02-02", "pr": 41212, "by": "NielsRogge", "t": "Add EoMT with DINOv3 backbone"}, {"m": "exaone_moe", "d": "2026-02-04", "pr": 43080, "by": "Junwon Hwang", "t": "Add EXAONE-MoE implementations"}, {"m": "moonshine_streaming", "d": "2026-02-04", "pr": 43702, "by": "eustlb", "t": "Add moonshine streaming"}, {"m": "vibevoice_acoustic_tokenizer", "d": "2026-02-06", "pr": 43400, "by": "Eric Bezzam", "t": "Add VibeVoice Acoustic Tokenizer"}, {"m": "qwen3_5", "d": "2026-02-09", "pr": 43830, "by": "Bo Zheng", "t": "Adding Support for Qwen3.5"}, {"m": "qwen3_5_moe", "d": "2026-02-09", "pr": 43830, "by": "Bo Zheng", "t": "Adding Support for Qwen3.5"}, {"m": "glm_moe_dsa", "d": "2026-02-09", "pr": 43858, "by": "Cyril Vallez", "t": "Add GlmMoeDsa"}, {"m": "voxtral_realtime", "d": "2026-02-16", "pr": 43769, "by": "eustlb", "t": "Add Voxtral Realtime"}, {"m": "higgs_audio_v2", "d": "2026-02-19", "pr": 40294, "by": "Shuai Zheng", "t": "Add Higgs Audio V2 Model"}, {"m": "higgs_audio_v2_tokenizer", "d": "2026-02-19", "pr": 40294, "by": "Shuai Zheng", "t": "Add Higgs Audio V2 Model"}, {"m": "colmodernvbert", "d": "2026-02-23", "pr": 42504, "by": "Paul Teiletche", "t": "Add ModernVBERT models"}, {"m": "modernvbert", "d": "2026-02-23", "pr": 42504, "by": "Paul Teiletche", "t": "Add ModernVBERT models"}, {"m": "olmo_hybrid", "d": "2026-02-26", "pr": 43358, "by": "Yanhong Li", "t": "Add OLMo Hybrid model"}, {"m": "pp_doclayout_v2", "d": "2026-02-27", "pr": 43018, "by": "zhang-prog", "t": "[Model] Add PP-DocLayoutV2 Model Support"}, {"m": "timesfm2_5", "d": "2026-02-27", "pr": 41763, "by": "Kashif Rasul", "t": "Timesfm 2.5"}, {"m": "vibevoice_asr", "d": "2026-03-02", "pr": 43625, "by": "Eric Bezzam", "t": "Add VibeVoice ASR"}, {"m": "nemotron_h", "d": "2026-03-03", "pr": 44390, "by": "Li Ding", "t": "add support for nemotron_3"}, {"m": "eurobert", "d": "2026-03-04", "pr": 39455, "by": "Arthur", "t": "Add eurobert"}, {"m": "chmv2", "d": "2026-03-11", "pr": 44595, "by": "Yoni Gozlan", "t": "Add CHMv2"}, {"m": "pp_ocrv5_server_det", "d": "2026-03-13", "pr": 43274, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_server_det Model Support"}, {"m": "pp_lcnet", "d": "2026-03-13", "pr": 43247, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_mobile_det Model Support"}, {"m": "pp_lcnet_v3", "d": "2026-03-13", "pr": 43247, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_mobile_det Model Support"}, {"m": "pp_ocrv5_mobile_det", "d": "2026-03-13", "pr": 43247, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_mobile_det Model Support"}, {"m": "pi0", "d": "2026-03-16", "pr": 44160, "by": "Pablo Montalvo", "t": "Add model lerobot PI0 to transformers"}, {"m": "mistral4", "d": "2026-03-16", "pr": 44760, "by": "Julien Denize", "t": "Add Mistral 4"}, {"m": "jina_embeddings_v3", "d": "2026-03-19", "pr": 44251, "by": "Sai-Suraj-27", "t": "Add `Jina-Embeddings-V3` Model"}, {"m": "pp_ocrv5_mobile_rec", "d": "2026-03-19", "pr": 44808, "by": "zhang-prog", "t": "[Model] Add PP-OCRv5_server_rec and PP-OCRv5_mobile_rec models Support"}, {"m": "pp_ocrv5_server_rec", "d": "2026-03-19", "pr": 44808, "by": "zhang-prog", "t": "[Model] Add PP-OCRv5_server_rec and PP-OCRv5_mobile_rec models Support"}, {"m": "pp_chart2table", "d": "2026-03-20", "pr": 43767, "by": "XingweiDeng", "t": "[Model] Add PP-Chart2Table Model Support"}, {"m": "slanext", "d": "2026-03-21", "pr": 43707, "by": "Liu Jiaxuan", "t": "[Model] Add SLANeXt Model Support"}, {"m": "uvdoc", "d": "2026-03-21", "pr": 43385, "by": "XingweiDeng", "t": "[Model] Add UVDoc Model Support"}, {"m": "videomt", "d": "2026-03-25", "pr": 44285, "by": "NielsRogge", "t": "Add VidEoMT"}, {"m": "cohere_asr", "d": "2026-03-26", "pr": 45023, "by": "eustlb", "t": "Add cohere asr"}, {"m": "musicflamingo", "d": "2026-03-30", "pr": 43538, "by": "Lasha Koroshinadze", "t": "Add Music Flamingo"}, {"m": "nomic_bert", "d": "2026-04-02", "pr": 43067, "by": "Sonny Cooper", "t": "Internalise the NomicBERT model"}, {"m": "gemma4", "d": "2026-04-02", "pr": 45192, "by": "Ryan Mullins", "t": "casually dropping the most capable open weights on the planet"}, {"m": "sam3_lite_text", "d": "2026-04-13", "pr": 44320, "by": "NielsRogge", "t": "Add SAM3-LiteText"}, {"m": "qianfan_ocr", "d": "2026-04-17", "pr": 45280, "by": "bairongz", "t": "add Qianfan-OCR model definition"}, {"m": "slanet", "d": "2026-04-22", "pr": 45532, "by": "zhang-prog", "t": "[Model] Add SLANet Model Support"}, {"m": "hy_v3", "d": "2026-04-22", "pr": 45574, "by": "Anton Vlasjuk", "t": "Fix typos"}, {"m": "openai_privacy_filter", "d": "2026-04-22", "pr": 45580, "by": "Anton Vlasjuk", "t": "[`Privacy Filter`] Add model"}, {"m": "deimv2", "d": "2026-04-27", "pr": 44339, "by": "Harshal Janjani", "t": "model: Add DEIMv2 to Transformers"}, {"m": "laguna", "d": "2026-04-28", "pr": 45673, "by": "Joe Rowell", "t": "Laguna XS.2 implementation"}, {"m": "minicpmv4_6", "d": "2026-04-28", "pr": 45689, "by": "Anton Vlasjuk", "t": "fixing more typos"}, {"m": "granite_speech_plus", "d": "2026-04-29", "pr": 45695, "by": "Zvi Kons", "t": "Support for a new Granite-Speech-Plus model"}, {"m": "pp_formulanet", "d": "2026-04-30", "pr": 45626, "by": "zhang-prog", "t": "[Model] Add PP-FormulaNet Model Support"}, {"m": "deepseek_v4", "d": "2026-05-02", "pr": 45643, "by": "Arthur", "t": "Add DeepSeek V4"}, {"m": "exaone4_5", "d": "2026-05-04", "pr": 45471, "by": "Junwon Hwang", "t": "Add EXAONE 4.5 implementations"}, {"m": "gemma4_assistant", "d": "2026-05-05", "pr": 45788, "by": "Sindhu Raghuram Panyam", "t": "First model"}, {"m": "granite4_vision", "d": "2026-05-05", "pr": 45597, "by": "artem-spector", "t": "Add Granite 4.1 Vision (granite4_vision)"}, {"m": "rf_detr", "d": "2026-05-07", "pr": 36895, "by": "StevenBucaille", "t": "Add RF-DETR"}, {"m": "hyperclovax", "d": "2026-05-08", "pr": 44956, "by": "bigshanedogg", "t": "Add HyperCLOVAX SEED Think 14B"}, {"m": "hrm_text", "d": "2026-05-18", "pr": 46025, "by": "yifei wu", "t": "Add hrm text"}, {"m": "pp_lcnet_v4", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_medium_det", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_small_det", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_small_rec", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_tiny_rec", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "cohere2_moe", "d": "2026-05-20", "pr": 46115, "by": "Cyril Vallez", "t": "Add new cohere2_moe model"}, {"m": "glmga", "d": "2026-05-26", "pr": 46184, "by": "Yuxuan Zhang", "t": "[GLM-4.6V] Update with GLM-GA Processor"}, {"m": "mellum", "d": "2026-05-28", "pr": 46112, "by": "Madeesh Kannan", "t": "feat: Add support for JetBrains' `Mellum` v2 code generation model"}, {"m": "deepseek_ocr2", "d": "2026-06-02", "pr": 45075, "by": "Eon Kim", "t": "Add Deepseek-OCR-2 model"}, {"m": "sapiens2", "d": "2026-06-03", "pr": 45919, "by": "guarin", "t": "Add Sapiens2 Model"}, {"m": "gemma4_unified", "d": "2026-06-03", "pr": 46385, "by": "Douglas Reid", "t": "who needs encoders?"}, {"m": "gemma4_unified_assistant", "d": "2026-06-03", "pr": 46385, "by": "Douglas Reid", "t": "who needs encoders?"}, {"m": "cosmos3_omni", "d": "2026-06-05", "pr": 46146, "by": "MaciejBalaNV", "t": "Added cosmos3 model"}, {"m": "diffusion_gemma", "d": "2026-06-10", "pr": 46540, "by": "Joao Gante", "t": "GPU go brr"}, {"m": "deepseek_v32", "d": "2026-06-11", "pr": 41251, "by": "Arthur", "t": "Add deepseek 3.2 exp"}, {"m": "minimax_m3_vl", "d": "2026-06-12", "pr": 46600, "by": "Arthur", "t": "Add minimax m3vl"}, {"m": "nemotron_asr_streaming", "d": "2026-06-19", "pr": 46332, "by": "eustlb", "t": "Add Nemotron ASR Streaming"}, {"m": "videoprism", "d": "2026-06-20", "pr": 39895, "by": "MAHIR DAIYAN", "t": "Add Videoprism"}, {"m": "minicpm3", "d": "2026-06-23", "pr": 41116, "by": "Minho Ryu", "t": "Add MiniCPM3"}, {"m": "xcodec2", "d": "2026-06-25", "pr": 44178, "by": "Eric Bezzam", "t": "Add xcodec2 model"}, {"m": "qwen3_asr", "d": "2026-06-26", "pr": 43838, "by": "Muhammed", "t": "Qwen3 ASR and Forced Aligner"}, {"m": "nemotron3_5_asr", "d": "2026-06-27", "pr": 46565, "by": "eustlb", "t": "Add Nemotron 3.5 ASR Streaming"}, {"m": "radio", "d": "2026-07-01", "pr": 46425, "by": "meatybobby", "t": "Add support for RADIO models"}, {"m": "zaya", "d": "2026-07-01", "pr": 45862, "by": "JJJYmmm", "t": "[new model] Add Zyphra/ZAYA1-8B"}, {"m": "mimo_v2_flash", "d": "2026-07-01", "pr": 45144, "by": "casinca", "t": "Add Xiaomi MiMo-V2"}, {"m": "hunyuan_vl", "d": "2026-07-03", "pr": 46417, "by": "Mi Jiazhi", "t": "Add HunYuan VL model"}, {"m": "kimi_k25", "d": "2026-07-03", "pr": 45630, "by": "Raushan Turganbay", "t": "Add new model: Kimi2-6"}, {"m": "tipsv2", "d": "2026-07-07", "pr": 46347, "by": "Zixuan Zhu", "t": "Add TIPSv2"}, {"m": "tipsv2_dpt", "d": "2026-07-07", "pr": 46347, "by": "Zixuan Zhu", "t": "Add TIPSv2"}, {"m": "inkling", "d": "2026-07-15", "pr": 47347, "by": "Pablo Montalvo", "t": "Add Inkling model"}, {"m": "cosmos3_edge", "d": "2026-07-16", "pr": 47181, "by": "Atharva Joshi", "t": "Add Cosmos3 Edge model support"}, {"m": "axk1", "d": "2026-07-24", "pr": 46867, "by": "Minsang Kim", "t": "add_axk1"}, {"m": "axk2", "d": "2026-07-25", "pr": 47528, "by": "Anton Vlasjuk", "t": "Add AXK2 from SKT"}, {"m": "granite_swa", "d": "2026-07-29", "pr": 47179, "by": "Davis Wertheimer", "t": "Add Granite-swa and Granitemoe-swa model support"}, {"m": "granitemoe_swa", "d": "2026-07-29", "pr": 47179, "by": "Davis Wertheimer", "t": "Add Granite-swa and Granitemoe-swa model support"}, {"m": "muse_glimmer", "d": "2026-08-10", "pr": 47867, "by": "Cyril Vallez", "t": "Add MuseGlimmer"}, {"m": "muse_glimmer_assistant", "d": "2026-08-10", "pr": 47867, "by": "Cyril Vallez", "t": "Add MuseGlimmer"}, {"m": "cohere_compass", "d": "2026-08-10", "pr": 47878, "by": "calpt", "t": "Add CohereCompass modeling"}, {"m": "step3p7", "d": "2026-08-16", "pr": 46658, "by": "Ita Zaporozhets", "t": "[new model] step 3.7"}, {"m": "esmc", "d": "2026-08-20", "pr": 46419, "by": "Matt", "t": "Port ESMC and ESMFold2 to Transformers"}, {"m": "esmfold2", "d": "2026-08-20", "pr": 46419, "by": "Matt", "t": "Port ESMC and ESMFold2 to Transformers"}, {"m": "granite_speech5", "d": "2026-08-25", "pr": 48288, "by": "eustlb", "t": "gs"}, {"m": "qwen4_exp", "d": "2026-08-26", "pr": 48337, "by": "Cyril Vallez", "t": "Add Qwen4Exp model"}, {"m": "glm5_next", "d": "2026-08-26", "pr": 48342, "by": "Shijin Zhang", "t": "[Glm 5.3 Flash] GLM 5.3 Flash Support"}, {"m": "vibevoice", "d": "2026-08-27", "pr": 40546, "by": "Zhiliang Peng", "t": "Implement VibeVoice"}]; const MODEL_ROW = {"n": 267, "span": ["2024-06-22", "2026-08-27"], "owners": [["eustlb", 14], ["Arthur", 12], ["zhang-prog", 12], ["Yoni Gozlan", 10]], "authors": 126}; const TRACKS = [ { "name": "Deprecation → v5", "hue": 258, "blurb": "Cut the library to one framework, then ship the major version.", "items": [ { "t": "Remove TensorFlow, Flax & JAX support", "tag": "3 frameworks -> 1", "s": "done", "prs": [ { "n": 38758, "d": "2025-06-11", "t": "Deprecate TF + JAX", "st": "MERGED", "by": "Rocketknight1", "a": 27, "r": 50030, "f": 121, "c": 3, "b": "The time has finally come 🔫 🥃" }, { "n": 38760, "d": "2025-06-11", "t": "Prepare for TF+Jax deprecation", "st": "MERGED", "by": "Rocketknight1", "a": 4, "r": 37, "f": 2, "c": 2, "b": "This PR removes a lot of TF + Jax testing in repo-consistency and .circleci. In a follow-up PR, I'll delete a lot of the test files, but this needs to be in a separate PR because repo-consistency checks run against the target branch, not the PR branch, and so they will fail if we remove other tests " }, { "n": 40372, "d": "2025-08-25", "t": "[docs] flax/jax purge", "st": "MERGED", "by": "gante", "a": 16, "r": 183, "f": 7, "c": 2, "b": "This should be the last PR removing JAX/FLAX from our docs 🤗 After this PR, the only JAX/FLAX mentions in our docs are either historical (e.g. \"(written in JAX)\" as a note regarding the model origin) or useful for model addition information (e.g. tips to debug models originally written in jax). Meth" }, { "n": 40654, "d": "2025-09-03", "t": "Remove TF/Flax examples", "st": "MERGED", "by": "Rocketknight1", "a": 1, "r": 21961, "f": 74, "c": 3, "b": "As the title says, the frameworks have been deprecated, so we should probably drop the examples as well! One thing to note: check_copies.py had a check_full_copies function that checked when files should be exact copies of each other. This was only used for TF/Flax examples. I've removed the FULL_CO" }, { "n": 40760, "d": "2025-09-18", "t": "🚨🚨🚨 Fully remove Tensorflow and Jax support library-wide", "st": "MERGED", "by": "Cyrilvallez", "a": 3792, "r": 181269, "f": 854, "c": 35, "b": "Apart from obvious tf/jax support, I believe the following should be the only potential breaking changes to torch-only code: pipelines do not take framework argument anymore onnx config methods do not take framework argument anymore It may break current torch code if users do framework=\"pt\" explicit" }, { "n": 41058, "d": "2025-09-23", "t": "Remove mention of TensorFlow/Flax/JAX from English documentation", "st": "MERGED", "by": "cyyever", "a": 3, "r": 4, "f": 2, "c": 1, "b": "There is a remaining one. Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's th" } ], "d": "Deprecated then fully removed TF/Flax/JAX code paths, examples, and docs library-wide, collapsing transformers to a single PyTorch backend.", "note": null, "links": [], "src": "speed.json", "span": [ "2025-06-11", "2025-09-23" ], "roll": { "owners": [ [ "Rocketknight1", 3 ], [ "Cyrilvallez", 1 ], [ "cyyever", 1 ], [ "gante", 1 ] ], "n": 6, "add": 3843, "del": 253484, "files": 1060, "commits": 46 } }, { "t": "Trainer & distributed API modernization", "tag": "trainer API decluttered", "s": "wip", "prs": [ { "n": 39395, "d": "2025-07-21", "t": "🚨🚨🚨 [Trainer] Enable `average_tokens_across_devices` by default in `TrainingArguments`", "st": "MERGED", "by": "Krish0909", "a": 2, "r": 2, "f": 1, "c": 8, "b": "Fixes #39392 This change improves loss calculation correctness for multi-GPU training by enabling proper token averaging across devices by default. What does this PR do? Changes the default value of average_tokens_across_devices from False to True in TrainingArguments. This ensures more accurate los" }, { "n": 41166, "d": "2025-09-29", "t": "[v5] Remove `model_parallel` deprecated feature ", "st": "MERGED", "by": "SunMarc", "a": 61, "r": 1301, "f": 68, "c": 6, "b": "This PR removes deprecates code related to model_parallel feature in the modeling as well as in Trainer. This was a feature that was added only for a few models like t5 and gpt2 but eventually we switched to using device_map and tp now. For the models, we removes is_parallelizable and model_parallel" }, { "n": 41131, "d": "2025-09-30", "t": "[v5] More Training Args cleaning ", "st": "MERGED", "by": "SunMarc", "a": 2, "r": 18, "f": 3, "c": 2, "b": "This PR removes EvaluationStrategy as it is deprecated" }, { "n": 44989, "d": "2026-04-13", "t": "🚨 Distributed training API", "st": "DRAFT", "by": "3outeille", "a": 542, "r": 0, "f": 6, "c": 8, "b": "Distributed Training API Goal # torchrun --nproc_per_node=4 train_fsdp_tp.py import os import torch from torch.utils.data import DataLoader from datasets import load_dataset from transformers import AutoModelForCausalLM, AutoTokenizer from transformers.distributed import DistributedConfig from trans" }, { "n": 45640, "d": "2026-05-20", "t": "🚨🚨🚨 [Trainer] Default to FSDP2, simplify API around fsdp + fsdp_config", "st": "MERGED", "by": "SunMarc", "a": 235, "r": 211, "f": 4, "c": 9, "b": "? This PR defaults to FSDP version to 2. We cleanup the docstring to only show FSDPv2 related args and I cleaned a bit the codebase to separate FSDPv1 arg and FSDPv2 args logic. I've also added a bunch of deprecation messages for FSDPv1 + we will deprecate passing string in fsdp args. This shouldn't" }, { "n": 46917, "d": "2026-07-22", "t": "Remove deprecated training args and `is_fast` property", "st": "MERGED", "by": "cyyever", "a": 1, "r": 54, "f": 3, "c": 1, "b": "Removes overdue deprecations: Deprecated TrainingArguments fields warmup_ratio and logging_dir (plus their __post_init__ migration logic and the warmup_ratio arg in set_lr_scheduler). Docs updated to use warmup_steps (which accepts a float ratio). Deprecated is_fast property from image processing ba" } ], "d": "Removed deprecated TrainingArguments fields and model_parallel, defaulted average_tokens_across_devices, and moved Trainer to default FSDP2 with a simplified distributed config.", "note": null, "links": [], "src": "speed.json", "span": [ "2025-07-14", "2026-07-22" ], "roll": { "owners": [ [ "SunMarc", 3 ], [ "Krish0909", 1 ], [ "3outeille", 1 ], [ "cyyever", 1 ] ], "n": 6, "add": 843, "del": 1586, "files": 85, "commits": 34 } }, { "t": "[v5] breaking-change cleanup series", "tag": "one clean v4->v5 cut", "s": "wip", "prs": [ { "n": 40997, "d": "2025-09-19", "t": "🚨 [v5] remove deprecated entry point", "st": "MERGED", "by": "gante", "a": 99, "r": 111, "f": 30, "c": 3, "b": "Removes the transformers-cli entrypoint, which is deprecated in favor of transformers" }, { "n": 40998, "d": "2025-09-19", "t": "🚨 [v5] remove generate output retrocompatibility aliases", "st": "MERGED", "by": "gante", "a": 0, "r": 66, "f": 3, "c": 2, "b": "See title :)" }, { "n": 41017, "d": "2025-09-24", "t": "🚨 [V5] Remove deprecated training arguments ", "st": "MERGED", "by": "cyyever", "a": 12, "r": 234, "f": 5, "c": 4, "b": "Remove deprecated training arguments from TrainingArguments" }, { "n": 41166, "d": "2025-09-29", "t": "[v5] Remove `model_parallel` deprecated feature ", "st": "MERGED", "by": "SunMarc", "a": 61, "r": 1301, "f": 68, "c": 6, "b": "This PR removes deprecates code related to model_parallel feature in the modeling as well as in Trainer. This was a feature that was added only for a few models like t5 and gpt2 but eventually we switched to using device_map and tp now. For the models, we removes is_parallelizable and model_parallel" }, { "n": 41123, "d": "2025-09-30", "t": "[v5] Remove deprecated prediction loop", "st": "MERGED", "by": "SunMarc", "a": 6, "r": 530, "f": 11, "c": 5, "b": "This PR removes the legacy prediction loop that was deprecated 4 years ago. I think that it should be fine to deprecate it even without any warning for v5 which it would have been better to have. I checked files on github and no one really seemed to use the legacy one. Also SequentialDistributedSamp" }, { "n": 41127, "d": "2025-09-30", "t": "[v5] Remove train kwargs", "st": "MERGED", "by": "SunMarc", "a": 0, "r": 12, "f": 1, "c": 5, "b": "This PR removes train kwargs. There were used for deprecated arguments but since we are removing them, we can also safely remove the kwargs." }, { "n": 41128, "d": "2025-09-30", "t": "[v5] Remove tokenizer from Trainer", "st": "MERGED", "by": "SunMarc", "a": 0, "r": 16, "f": 2, "c": 5, "b": "This PR removes tokenizers from Trainer as it will be deprecated for v5. Since we removed it from the init, I think it should also make sense to remove the setter and the property method for tokenizer" }, { "n": 41131, "d": "2025-09-30", "t": "[v5] More Training Args cleaning ", "st": "MERGED", "by": "SunMarc", "a": 2, "r": 18, "f": 3, "c": 2, "b": "This PR removes EvaluationStrategy as it is deprecated" }, { "n": 41161, "d": "2025-09-30", "t": "[v5] Remove old sagemaker api support ", "st": "MERGED", "by": "SunMarc", "a": 21, "r": 80, "f": 1, "c": 3, "b": "This PR cleans a bit the sagemaker code. We remove code related to the old sagemaker API (<1.10) as it is 5 years old. In our setup file, the recommended version is >2.0 also." }, { "n": 41107, "d": "2025-10-01", "t": "🚨 [v5] Remove SinkCache", "st": "MERGED", "by": "cyyever", "a": 0, "r": 25, "f": 4, "c": 1, "b": "It could be removed in V5." }, { "n": 41214, "d": "2025-10-01", "t": "[V5] Remove deprecated transformers.onnx", "st": "MERGED", "by": "cyyever", "a": 1, "r": 1306, "f": 9, "c": 4, "b": "Remove the source files, doc and the dependencies." }, { "n": 41122, "d": "2025-10-02", "t": "🚨 [V5] Remove deprecated `resume_download`", "st": "MERGED", "by": "cyyever", "a": 0, "r": 120, "f": 22, "c": 5, "b": "resume_download is a placeholder..." }, { "n": 41223, "d": "2025-10-02", "t": "🚨 [v5] remove deprecated `generate` classes (constraints and beam scorers)", "st": "MERGED", "by": "gante", "a": 0, "r": 1696, "f": 6, "c": 1, "b": "The decoding methods using these classes were recently moved to the hub, and these classes were deprecated. They were scheduled to be removed in minor versions that won't exist, so they get tagged as v5 removals :) (cc @manueldeprada )" }, { "n": 41406, "d": "2025-10-07", "t": "🚨 [v5] Remove deprecated cache classes", "st": "OPEN", "by": "gante", "a": 4, "r": 198, "f": 6, "c": 1, "b": "🚨🚨🚨 DO NOT MERGE 🚨🚨🚨 Depends on #41889 Merging this PR will break from transformers import Trainer on the latest PEFT release (0.17.1), which in turn breaks our make fixup when PEFT is installed. PEFT main already sorted this issue (here), but it is yet unreleased as of 7-oct-2025. cc @BenjaminBossa" }, { "n": 41417, "d": "2025-10-08", "t": "🚨 [v5] Prune `prune_heads`", "st": "MERGED", "by": "gante", "a": 352, "r": 3692, "f": 421, "c": 4, "b": "Spiritual successor to #41076 -- removes prune_heads and related code/tests/docs. prune_heads was a feature that we stopped including in models a while ago. For instance, the test_head_pruning_save_load_from_config_init test passes in 28 model classes, but is skipped in 444." }, { "n": 41287, "d": "2025-10-09", "t": "[v5] remove load_in_4bit and load_in_8bit", "st": "MERGED", "by": "SunMarc", "a": 349, "r": 245, "f": 59, "c": 9, "b": "This PR removes load_in_4bit and load_in_8bit deprecated arguments for v5. This is something that I didn't want to do in a minor release as this might be breaking for a lot of users." }, { "n": 41314, "d": "2025-10-09", "t": "🚨 [v5] Rendundant code in nested configs", "st": "MERGED", "by": "zucchini-nlp", "a": 707, "r": 1046, "f": 143, "c": 10, "b": "This PR deletes from_xxx_config and similar methods from everywhere and instead ensures that the config init can accept subconfigs of all formats (dict or object) Along the way, I found that not all configs call super init in the end and thus they can't set recursively attn implementation on subconf" }, { "n": 41402, "d": "2025-10-09", "t": "[v5] rm `utils/tf_ops/`", "st": "MERGED", "by": "gante", "a": 0, "r": 245, "f": 1, "c": 1, "b": "TF-related folder that flew under the radar :) (seems unused in the lib, latest commit was 4 years ago)" }, { "n": 41448, "d": "2025-10-09", "t": "🚨 [v5] Rename left traces of `past_key_value` in BERT-like models", "st": "MERGED", "by": "zucchini-nlp", "a": 599, "r": 599, "f": 71, "c": 1, "b": "As per title, finalizes the work from #41425 and renames past_key_value as past_key_values in all files without deprecation cycle It was a find and replace, so tests might fail 😅" }, { "n": 41505, "d": "2025-10-13", "t": "🚨 [v5] `generate` delegates default cache initialization to the model", "st": "MERGED", "by": "gante", "a": 114, "r": 50, "f": 31, "c": 10, "b": "See PR title. Now that all traces of legacy caches were removed, we can trust the model to initialize its own cache! This means we no longer need to set cache_implementation=\"xxx\" defaults in new models, assuming the model's forward pass defaults to the right cache class. Also fixes related bugs, un" }, { "n": 41474, "d": "2025-10-16", "t": "🚨 [v5] Toggle the serialization format in processors", "st": "MERGED", "by": "zucchini-nlp", "a": 137, "r": 227, "f": 23, "c": 11, "b": "As per title, in v5 we will start saving in the new nested format. The loading is still supported though for old-format configs Also, we don't need an optional_attributes field that does not do what it is called. Since it is only chat template, let's just manually pop and set it. For audio tokenizer" }, { "n": 41607, "d": "2025-10-21", "t": "[v5] Delete `videos` from image processing classes ", "st": "MERGED", "by": "zucchini-nlp", "a": 93, "r": 1026, "f": 19, "c": 6, "b": "As per title, it was deprecated for v5" }, { "n": 41700, "d": "2025-10-21", "t": "[v5] Remove deprecated tranformers.onnx", "st": "MERGED", "by": "echarlaix", "a": 93, "r": 5309, "f": 91, "c": 8, "b": "remove deprecated transformers.onnx for v5 remove all XxxOnnxConfig from concerned models config files Also following #41214" }, { "n": 41648, "d": "2025-10-22", "t": "[v5] Delete legacy chat template saving", "st": "MERGED", "by": "zucchini-nlp", "a": 77, "r": 147, "f": 8, "c": 7, "b": "As per title" }, { "n": 41626, "d": "2025-10-31", "t": "[v5] Return a BatchEncoding dict from apply_chat_template by default", "st": "MERGED", "by": "Rocketknight1", "a": 57, "r": 214, "f": 13, "c": 11, "b": "Tokenizers return a BatchEncoding dict by default, but apply_chat_template doesn't. This is just an accident of how I wrote it originally, which we were stuck with for backward compatibility reasons. Ideally, I think apply_chat_template should return exactly the same format as tokenizers, since it a" }, { "n": 41996, "d": "2025-11-04", "t": "[v5] Deprecate Text2Text and related pipelines", "st": "MERGED", "by": "Rocketknight1", "a": 44, "r": 6, "f": 3, "c": 4, "b": "Text2Text is mostly used only by older models, and via the Summarization/Translation pipelines. However, all of these tasks are now handled by general-purpose LLMs that provide much higher translation/summarization quality than older single-purpose models. This PR retains the pipelines, but deprecat" }, { "n": 41633, "d": "2025-11-07", "t": "[v5] 🚨Refactor subprocessors handling in processors", "st": "MERGED", "by": "yonigozlan", "a": 368, "r": 839, "f": 149, "c": 29, "b": "Refactor the handling of subprocessors in processors. Main change is that we deduce the subprocessors from the init signature instead of having to manually add \"subprocessor_class\" attributes. This means we can remove all attributes attribute in processors, along with all \"subprocessor\"_class attrib" }, { "n": 41681, "d": "2025-11-14", "t": "🚨 Delete deprecations with end-cycle in v4.xx and v5.0", "st": "MERGED", "by": "zucchini-nlp", "a": 121, "r": 2807, "f": 99, "c": 19, "b": "Deletes deprecation for the v5 release. cc @yonigozlan and @eustlb to verify that we can really delete them, since it is mostly deprecated stuff from processors For core maintainers: The max-size and pad-pixels args can't be deleted due to many models on the hub still using it. Instead I decided to " }, { "n": 41889, "d": "2025-11-27", "t": "🚨 [v5][PEFT] Bump min version requirement of PEFT to 0.18.0", "st": "MERGED", "by": "BenjaminBossan", "a": 36, "r": 169, "f": 6, "c": 12, "b": "PEFT is an optional dependency of transformers with a min version of 0.5.0. However, starting with transformers v5, older PEFT versions will not work anymore (see #41406). The minimum PEFT version will be 0.18.0. This PR updates the PEFT integration to require PEFT 0.18.0. This allows us to eliminat" }, { "n": 42567, "d": "2025-12-04", "t": "[V5] Return a BatchEncoding dict from apply_chat_template by default again", "st": "MERGED", "by": "Rocketknight1", "a": 18, "r": 136, "f": 7, "c": 11, "b": "This is basically PR #41626 again! Some of it got clobbered in the tokenizer refactor, but it's just as good the second time." }, { "n": 43325, "d": "2026-02-24", "t": "🚨 More V5 pipeline cleanup", "st": "MERGED", "by": "Rocketknight1", "a": 140, "r": 2679, "f": 116, "c": 43, "b": "More V5 pipeline cleanup, followup to #43256 and #43306: (changelist updated Jan 26th) question-answering and visual-question-answering removed image-to-image removed Updated the default text-generation and image-text-to-text models Removed some references to image-to-text, summarization, translatio" } ], "d": "Sept 2025-onward streak of 🚨[v5] PRs deleting deprecated entry points, generate() aliases, legacy cache tuple format, ONNX export, SageMaker support, prune_heads, resume_download, load_in_4bit/8bit, and more.", "note": "36 total [v5]-tagged PRs found in title search (33 merged, 1 still open #41406, 2 closed/abandoned #41294 #41405); list above is representative, not exhaustive.", "links": [], "src": "speed.json", "span": [ "2025-09-19", "2026-02-24" ], "roll": { "owners": [ [ "gante", 7 ], [ "SunMarc", 7 ], [ "zucchini-nlp", 6 ], [ "cyyever", 4 ] ], "n": 31, "add": 3511, "del": 25149, "files": 1431, "commits": 243 } }, { "t": "Remove slow tokenizers/processors, fast by default", "tag": "no more slow codepath", "s": "done", "prs": [ { "n": 41174, "d": "2025-10-01", "t": "🚨 [v5] Delete feature extractors used for vision", "st": "MERGED", "by": "zucchini-nlp", "a": 19, "r": 1367, "f": 65, "c": 7, "b": "As per title, let's clean up for v5 These were supposed to be deleted anyway and we had the warning logged for a long time, even before I joined. Feature Extractors are now reserved for audio models only" }, { "n": 41321, "d": "2025-10-07", "t": "[v5] Delete left traces of feature extractor", "st": "MERGED", "by": "zucchini-nlp", "a": 1, "r": 522, "f": 26, "c": 1, "b": "After rebasing #40793 and getting new changes on TypedDict, I found out that some traces from feature extractor were still left. Thanks for type validation we can catch inconsistencies easily :) This PR goes over the deprecated properties as well and deletes them. Also it deletes current_processor w" }, { "n": 40936, "d": "2025-11-27", "t": "rm slow tokenizers", "st": "MERGED", "by": "itazap", "a": 16908, "r": 55397, "f": 443, "c": 277, "b": "Tokenization Just as we moved towards a single backend library for model definition, we want Tokenizer to be a lot more intuitive. With v5, you can now initialize an empty LlamaTokenizer and train it directly on your new task! Defining a new tokenizer object should be as simple as this: from transfo" }, { "n": 41388, "d": "2026-01-21", "t": "🚨Default to fast image processors for all models", "st": "MERGED", "by": "yonigozlan", "a": 404, "r": 259, "f": 64, "c": 81, "b": "Following the trial testing with Qwen_VL image processors, this extends defaulting to fast image processors even for checkpoints saved with a slow one to all models. Also made sure that all processors use AutoImageProcessor to instantiate their image_processor_class. On that point, defining default " } ], "d": "Deleted the pure-Python slow tokenizer path and legacy vision feature extractors, and flipped every model to fast (Rust/vectorized) tokenizers and image processors by default.", "note": null, "links": [], "src": "speed.json", "span": [ "2025-09-26", "2026-01-21" ], "roll": { "owners": [ [ "zucchini-nlp", 2 ], [ "itazap", 1 ], [ "yonigozlan", 1 ] ], "n": 4, "add": 17332, "del": 57545, "files": 598, "commits": 366 } }, { "t": "v5.0 release cut & migration guide", "tag": "v4->v5 migration guide", "s": "done", "prs": [ { "n": 41681, "d": "2025-11-14", "t": "🚨 Delete deprecations with end-cycle in v4.xx and v5.0", "st": "MERGED", "by": "zucchini-nlp", "a": 121, "r": 2807, "f": 99, "c": 19, "b": "Deletes deprecation for the v5 release. cc @yonigozlan and @eustlb to verify that we can really delete them, since it is mostly deprecated stuff from processors For core maintainers: The max-size and pad-pixels args can't be deleted due to many models on the hub still using it. Instead I decided to " }, { "n": 42399, "d": "2025-11-25", "t": "Add more items to migration guide", "st": "MERGED", "by": "Wauplin", "a": 33, "r": 4, "f": 1, "c": 1, "b": "PR on top of #42386 cc @LysandreJik I have added parts I could help with. Let me know if you want me to change anything :) Feel free to merge anytime." }, { "n": 42386, "d": "2025-11-26", "t": "Initial migration guide", "st": "MERGED", "by": "LysandreJik", "a": 274, "r": 0, "f": 1, "c": 4, "b": "WIP (related to #40822)" }, { "n": 42470, "d": "2025-11-28", "t": "Update migration guide - generation config", "st": "MERGED", "by": "zucchini-nlp", "a": 7, "r": 7, "f": 2, "c": 5, "b": "cc @vasqu , as discussed in TRL's issue thread (huggingface/trl#4595) Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Git" }, { "n": 42477, "d": "2025-12-01", "t": "Migration guide: tokenization", "st": "MERGED", "by": "LysandreJik", "a": 120, "r": 81, "f": 1, "c": 3, "b": "" }, { "n": 42510, "d": "2025-12-01", "t": "Small migration guide update", "st": "MERGED", "by": "ArthurZucker", "a": 70, "r": 0, "f": 1, "c": 3, "b": "Update" }, { "n": 42526, "d": "2025-12-01", "t": "CLI Migration guide", "st": "MERGED", "by": "LysandreJik", "a": 8, "r": 3, "f": 1, "c": 2, "b": "Light updates" }, { "n": 42532, "d": "2025-12-01", "t": "Migration guide typo", "st": "MERGED", "by": "LysandreJik", "a": 2, "r": 1, "f": 1, "c": 1, "b": "" }, { "n": 45556, "d": "2026-04-28", "t": "Add image processors refactor to v5 migration guide", "st": "MERGED", "by": "yonigozlan", "a": 96, "r": 38, "f": 1, "c": 3, "b": "As discussed internally @vasqu Cc @stevhliu" } ], "d": "Closed out end-of-cycle deprecations, then wrote and iterated the official v4->v5 migration guide covering generation config, tokenization, CLI and image processors.", "note": "Per git tags: v5.0.0rc0 tagged 2025-12-01, rc1-rc3 through 2026-01-14, v5.0.0 released 2026-01-26; repo now at v5.16.0.dev0.", "links": [], "src": "speed.json", "span": [ "2025-10-15", "2026-04-28" ], "roll": { "owners": [ [ "LysandreJik", 4 ], [ "zucchini-nlp", 2 ], [ "Wauplin", 1 ], [ "ArthurZucker", 1 ] ], "n": 9, "add": 731, "del": 2941, "files": 108, "commits": 41 } }, { "t": "Pipeline API v5 cleanup", "tag": "pipelines aligned with v5", "s": "wip", "prs": [ { "n": 41996, "d": "2025-11-04", "t": "[v5] Deprecate Text2Text and related pipelines", "st": "MERGED", "by": "Rocketknight1", "a": 44, "r": 6, "f": 3, "c": 4, "b": "Text2Text is mostly used only by older models, and via the Summarization/Translation pipelines. However, all of these tasks are now handled by general-purpose LLMs that provide much higher translation/summarization quality than older single-purpose models. This PR retains the pipelines, but deprecat" }, { "n": 42359, "d": "2025-11-26", "t": "🚨 Image-text pipeline expects correctly formatted chat", "st": "MERGED", "by": "zucchini-nlp", "a": 30, "r": 60, "f": 2, "c": 2, "b": "As per title, we can break for v5 and enforce a correct usage of chat templates. I will request review after adjusting tests" }, { "n": 42430, "d": "2025-12-04", "t": "🚨 Clean up `image-text-to-text` pipeline", "st": "OPEN", "by": "ebezzam", "a": 127, "r": 178, "f": 4, "c": 37, "b": "This PR standardizes / cleans up image-to-text-text pipeline to use the chat template logic from the base pipeline class. TODO: add function to convert from OpenAI format to HF-format to address this failing test after #42359 removed add_images_to_messages (see here)" }, { "n": 43325, "d": "2026-02-24", "t": "🚨 More V5 pipeline cleanup", "st": "MERGED", "by": "Rocketknight1", "a": 140, "r": 2679, "f": 116, "c": 43, "b": "More V5 pipeline cleanup, followup to #43256 and #43306: (changelist updated Jan 26th) question-answering and visual-question-answering removed image-to-image removed Updated the default text-generation and image-text-to-text models Removed some references to image-to-text, summarization, translatio" } ], "d": "Deprecated redundant pipelines (Text2Text) and cleaned up chat-formatting and image-text-to-text pipeline behavior to match v5 conventions.", "note": null, "links": [], "src": "speed.json", "span": [ "2025-11-04", "2026-02-24" ], "roll": { "owners": [ [ "Rocketknight1", 2 ], [ "ebezzam", 1 ], [ "zucchini-nlp", 1 ] ], "n": 4, "add": 341, "del": 2923, "files": 125, "commits": 86 } } ] }, { "name": "Speed & serving", "hue": 8, "blurb": "Make transformers fast enough to serve, not just to prototype.", "items": [ { "t": "Speculative & assisted decoding", "tag": "draft-model speedups", "s": "wip", "prs": [ { "n": 27979, "d": "2023-12-19", "t": "Generate: speculative decoding", "st": "MERGED", "by": "gante", "a": 90, "r": 18, "f": 1, "c": 8, "b": "Useful context: In a recent PR (#27750), the candidate generation in assisted generation got abstracted, so we can host new candidate generation techniques (such as #27722). This PR: Reworks assisted candidate generation to call .generate(), instead of having its own custom generation loop. For most" }, { "n": 33258, "d": "2024-09-11", "t": "Dynamic number of speculative tokens in order to accelerate speculative decoding", "st": "MERGED", "by": "jmamou", "a": 57, "r": 0, "f": 6, "c": 21, "b": "This PR adds support to dynamic number of speculative tokens in order to accelerate speculative decoding. It is an unsupervised version of the dynamic speculation lookahead from Dynamic Speculation Lookahead Accelerates Speculative Decoding of Large Language Models. We add the argument assistant_con" }, { "n": 33657, "d": "2024-09-25", "t": "🚨🚨 Setting default behavior of assisted decoding", "st": "MERGED", "by": "jmamou", "a": 24, "r": 8, "f": 4, "c": 8, "b": "Following #33258 (comment) Setting default behavior of assisted decoding: the number of speculative tokens per assistant model iteration is dynamically determined according to the confidence in the assistant model prediction. By default, num_assistant_tokens=20 num_assistant_tokens_schedule=\"constan" }, { "n": 35029, "d": "2025-02-26", "t": "Universal Speculative Decoding `CandidateGenerator`", "st": "MERGED", "by": "keyboardAnt", "a": 638, "r": 46, "f": 3, "c": 112, "b": "📄 ICML oral (top %1): Accelerating LLM Inference with Lossless Speculative Decoding Algorithms for Heterogeneous Vocabularies, https://arxiv.org/abs/2502.05202 - Nadav Timor, Jonathan Mamou, Daniel Korat, Moshe Berchansky, Oren Pereg, Gaurav Jain, Moshe Wasserblat, David Harel This PR is a collabora" }, { "n": 45979, "d": "2026-07-09", "t": "[Generation] Add static ensemble verification for lossy speculative decoding", "st": "MERGED", "by": "kasakh", "a": 239, "r": 3, "f": 5, "c": 5, "b": "Adds support for static ensemble verification in speculative decoding (assisted generation), a training-free method that increases draft token acceptance rates by relaxing the verification distribution. Fixes #45865 Motivation Standard speculative decoding rejects many plausible tokens because it st" } ], "d": "Built out generate()'s assisted/speculative decoding path: dynamic speculative-token count, corrected default behavior, and a Universal Speculative Decoding candidate generator for mismatched tokenizers.", "note": null, "links": [], "src": "speed.json", "span": [ "2023-12-14", "2026-07-09" ], "roll": { "owners": [ [ "jmamou", 2 ], [ "gante", 1 ], [ "keyboardAnt", 1 ], [ "kasakh", 1 ] ], "n": 5, "add": 1048, "del": 75, "files": 19, "commits": 154 } }, { "t": "torch.compile coverage & speedups", "tag": "3-5x faster compiled decode", "s": "done", "prs": [ { "n": 29114, "d": "2024-02-21", "t": "`torch.compile` compatibility with `generate` + static cache", "st": "MERGED", "by": "fxmarty", "a": 76, "r": 51, "f": 3, "c": 16, "b": "This PR adds the support of torch.compile when using generate + static cache." }, { "n": 30560, "d": "2024-05-20", "t": "Add support for torch.compile dynamic shapes", "st": "MERGED", "by": "warner-benjamin", "a": 190, "r": 60, "f": 27, "c": 4, "b": "This PR adds support for compiling models with dynamic shapes dynamic=True to almost all models with SDPAttention implementations which currently do not support dynamic shapes. #30442 added support for Llama, Gemma, OLMo, & Cohere. The only model not modified is DBRX, which needs the changes from bo" }, { "n": 30919, "d": "2024-05-23", "t": "Finish adding support for torch.compile dynamic shapes", "st": "MERGED", "by": "warner-benjamin", "a": 16, "r": 6, "f": 3, "c": 1, "b": "Following #30560, this PR updates new models (Jamba & JetMoe) and DBRX to support compiling with dynamic shapes. DBRX appears to support the FA2 SDPA kernel now, although the model is too large for me to test. cc @ArthurZucker and @fxmarty" }, { "n": 32227, "d": "2024-07-31", "t": ">3-5x faster torch.compile forward compilation for autoregressive decoder models", "st": "MERGED", "by": "fxmarty", "a": 1602, "r": 424, "f": 22, "c": 19, "b": "As detailed in #29114 and #29165, the current implementation of GenerationMixin.generate passes 2D integer tensors of variable length as attention_mask to the models forward, that are transformed into 4D tensors of static shape in the models forward for decoder models. This is an issue when using to" }, { "n": 34247, "d": "2024-11-22", "t": "Auto compile when static cache", "st": "MERGED", "by": "ArthurZucker", "a": 16, "r": 2, "f": 1, "c": 12, "b": "Add automatic compile for static cache. This can be tested with: from transformers import AutoModelForCausalLM, AutoTokenizer import torch import time device = \"cuda\" model = AutoModelForCausalLM.from_pretrained(\"meta-llama/Meta-Llama-3-8B\", torch_dtype=torch.float16).to(device) tokenizer = AutoToke" }, { "n": 39505, "d": "2025-07-23", "t": "Rename `supports_static_cache` to `can_compile_fullgraph`", "st": "MERGED", "by": "zucchini-nlp", "a": 143, "r": 148, "f": 130, "c": 5, "b": "As per title, let's be specific about how the attribute is used. I will see if we can entirely get rid of the flag, maybe there's another way to infer which models support compile or not. At least from what I know, all MoE and non-transformer arch models don't support it" }, { "n": 40137, "d": "2025-08-19", "t": "🚨🚨 Switch default compilation to fullgraph=False", "st": "MERGED", "by": "Cyrilvallez", "a": 38, "r": 101, "f": 16, "c": 14, "b": "As we sometimes talked offline. Graph breaks are in general not a problem, and requiring fullgraph=True is too restrictive! Would also allow us to use torch.compiler.disable where needed sometimes, to skip bigger regions which are problematic and not worth it" } ], "d": "Made generate() compile-compatible with static cache, fixed dynamic-shape recompilation triggers, and auto-enabled torch.compile whenever a static cache is used.", "note": "PR #32227 title states '>3-5x faster torch.compile forward compilation for autoregressive decoder models' (merged 2024-07-31).", "links": [], "src": "speed.json", "span": [ "2024-02-19", "2025-08-19" ], "roll": { "owners": [ [ "fxmarty", 2 ], [ "warner-benjamin", 2 ], [ "ArthurZucker", 1 ], [ "zucchini-nlp", 1 ] ], "n": 7, "add": 2081, "del": 792, "files": 202, "commits": 71 } }, { "t": "Fast image processors architecture", "tag": "vectorized image preprocessing", "s": "wip", "prs": [ { "n": 28847, "d": "2024-06-11", "t": "Fast image processor", "st": "MERGED", "by": "amyeroberts", "a": 1643, "r": 811, "f": 64, "c": 40, "b": "Adds a ViTImageProcessorFast class, which uses a torchvision backend to speed up image processor import requests from PIL import Image from transformers import AutoImageProcessor # Load the image image = Image.open(requests.get(\"http://images.cocodataset.org/val2017/000000039769.jpg\", stream=True).r" }, { "n": 35069, "d": "2025-02-04", "t": "Refactoring of ImageProcessorFast", "st": "MERGED", "by": "yonigozlan", "a": 4072, "r": 2269, "f": 66, "c": 59, "b": "This PR introduces a significant refactoring of how fast image processors can be implemented in Transformers. Motivation The primary goal of this refactoring is to simplify the process of creating fast image processors when a slow image processor already exists. Unlike the current BaseImageProcessor" }, { "n": 43514, "d": "2026-03-19", "t": "🚨🚨 Refactor Image Processors to support different backends", "st": "MERGED", "by": "yonigozlan", "a": 33586, "r": 57296, "f": 675, "c": 59, "b": "Image Processor Backend Refactor Summary Replaces the dual-file BaseImageProcessor (slow/PIL) + BaseImageProcessorFast (fast/torchvision) design with a unified backend architecture. The image_processing_utils_fast module is removed; all logic lives in image_processing_utils and image_processing_back" }, { "n": 48066, "d": "2026-08-24", "t": "🚨 Speed up detr image processing", "st": "OPEN", "by": "guarin", "a": 626, "r": 562, "f": 11, "c": 8, "b": "Speeds up preprocessing on GPU for all detr models by using group_images_by_shape. With batch size 32 and ragged inputs between 512px to 2048px I got 27-38% speed up on an H100: model baseline step ms speedup-branch step ms improvement rtdetrv2 20.84 15.23 27% faster rf-detr 25.85 16.37 37% faster d" } ], "d": "Introduced batched, tensor-native fast image processors (replacing per-image PIL loops), then refactored them to support multiple backends (torch/numpy).", "note": "Open PR #48066 measured '27-38% speed up on an H100' for DETR-family GPU preprocessing (batch 32, 512-2048px inputs).", "links": [], "src": "speed.json", "span": [ "2024-06-10", "2026-08-24" ], "roll": { "owners": [ [ "yonigozlan", 2 ], [ "amyeroberts", 1 ], [ "guarin", 1 ] ], "n": 4, "add": 39927, "del": 60938, "files": 816, "commits": 166 } }, { "t": "Flex Attention & attention sinks", "tag": "flexible, compiled masking", "s": "wip", "prs": [ { "n": 34896, "d": "2024-12-04", "t": "[`GPTNeoX`] Flex Attention + Refactor", "st": "MERGED", "by": "vasqu", "a": 371, "r": 249, "f": 6, "c": 20, "b": "Adds flex attention and the refactor according to #34809 However, I discovered several issues in the current version of gemma2 (#34282): It seems like that flex attention needs a transpose afterwards like sdpa Loading flex attn with from pretrained didn't work and hence, current tests use another at" }, { "n": 37285, "d": "2025-04-07", "t": "Fixing flex attention for torch=2.6.0", "st": "MERGED", "by": "salmanmohammadi", "a": 11, "r": 2, "f": 1, "c": 6, "b": "This PR fixes this issue originally raised in pytorch core pytorch/pytorch#146260 for flex attention on 2.6.0, by setting mode=\"max-autotune-no-cudagraphs\". Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read" }, { "n": 37381, "d": "2025-04-14", "t": "Fix mask handling for flex attention in llama/gemma2/mistral/qwen2", "st": "MERGED", "by": "flukeskywalker", "a": 423, "r": 177, "f": 59, "c": 4, "b": "Issue In gemma2 and mistral modelling code, where the _update_causal_mask() methods were not handling the case of the atn_implementation being the recently added \"flex_attention\". Main Fix For Mistral and Gemma2: Added logic for converting attention mask passed as as a tensor to a BlockMask required" }, { "n": 39194, "d": "2025-07-04", "t": "Add packed tensor format support for flex/sdpa/eager through the mask!", "st": "MERGED", "by": "Cyrilvallez", "a": 303, "r": 9, "f": 65, "c": 16, "b": "As per the title. import torch from transformers import AutoModelForCausalLM from transformers.masking_utils import create_causal_mask model = AutoModelForCausalLM.from_pretrained(\"meta-llama/Llama-3.2-1B\", torch_dtype=torch.float16) batch_size = 1 sequence_length = 10 cache_position = torch.arange(" }, { "n": 41083, "d": "2025-09-29", "t": "Fix attention sink implementation in flex attention", "st": "MERGED", "by": "SamuelBarryCS", "a": 25, "r": 6, "f": 2, "c": 13, "b": "What Fixes #41026 where @jonny-so correctly identified that attention sinks were being incorrectly applied within the [score_mod] function of flex attention The previous implementation attempted to apply attention sinks by manipulating pre-softmax scores within score_mod, which is incorrect because " }, { "n": 41249, "d": "2025-10-02", "t": "[`Flex Attn`] Fix lse x attention sinks logic ", "st": "MERGED", "by": "vasqu", "a": 3, "r": 3, "f": 1, "c": 1, "b": "We always return lse except for CPU #40352 (it simply isn't supported in torch yet) Let's rather raise a proper error for this than encountering a verbose torch exception" }, { "n": 44684, "d": "2026-03-18", "t": "update flex attention to use `return_aux` instead of `return_lse` when torch verison >= 2.9", "st": "MERGED", "by": "ntenenz", "a": 41, "r": 5, "f": 1, "c": 4, "b": "… What does this PR do? In torch versions >= 2.9.0, it requests the lse from flex_attenetion using AuxRequest instead of the deprecated return_lse, which triggers a warning and can break tracing. Fixes #44683 Before submitting This PR fixes a typo or improves the docs (you can dismiss the other chec" } ], "d": "Added torch's FlexAttention backend (compiled block-sparse masking) with packed-tensor support, plus attention-sink support for sink-based models such as gpt-oss.", "note": null, "links": [], "src": "speed.json", "span": [ "2024-11-23", "2026-03-18" ], "roll": { "owners": [ [ "vasqu", 2 ], [ "salmanmohammadi", 1 ], [ "flukeskywalker", 1 ], [ "Cyrilvallez", 1 ] ], "n": 7, "add": 1177, "del": 451, "files": 135, "commits": 64 } }, { "t": "Unified attention interface across models", "tag": "one attention API", "s": "wip", "prs": [ { "n": 35235, "d": "2024-12-18", "t": "🚨All attention refactor🚨", "st": "MERGED", "by": "ArthurZucker", "a": 5635, "r": 9778, "f": 107, "c": 99, "b": "Todo in this PR: Cohere Chameleon DBRX Gemma Gemma2 GLM (modular donc rien à faire je crois) gpt_neoX et GPT2 Granite Jamba JetMoe Mimi Mistral Mixtral Mllama Moshi Nemotron OPT Phi Ph3 PhiMoe Qwen2 qwen2Moe qwen2VL SableML StartCoder2 -> Modular normalement oK Idefics1,2,3 Olmo Olmo2 Siglip Whisper" }, { "n": 35342, "d": "2024-12-20", "t": "Cleaner attention interfaces", "st": "MERGED", "by": "Cyrilvallez", "a": 30, "r": 12, "f": 4, "c": 5, "b": "Improve the attention interfaces with comments/dtype fix" }, { "n": 36889, "d": "2025-03-26", "t": "Allow easy registration of custom attention functions", "st": "MERGED", "by": "Cyrilvallez", "a": 171, "r": 11, "f": 6, "c": 20, "b": "Basic design idea: from transformers import AttentionInterface def my_new_attention(): ... AttentionInterface.register(\"my_new_attention\", my_new_attention)" }, { "n": 37866, "d": "2025-05-22", "t": "🚨🚨[core] Completely rewrite the masking logic for all attentions", "st": "MERGED", "by": "Cyrilvallez", "a": 2984, "r": 6808, "f": 129, "c": 193, "b": "As per the title. The goal is to properly separate masking logic from modeling code itself, to continue our objective of simplifying the library. Code is much simpler to understand Much more general: always work for all lengths, and all attention implementations, e.g.: flex attention now works with " }, { "n": 39423, "d": "2025-07-18", "t": "🚨🚨 Fix and simplify attention implementation dispatch and subconfigs handling", "st": "MERGED", "by": "Cyrilvallez", "a": 472, "r": 323, "f": 33, "c": 46, "b": "A few issues: the checks and attention switch should be performed recursively for all submodels, not only the \"top-most\" model flash can dispatch even with cpu/disk offloading, as the module is put back on device prior to computation some models with the old class attention interface should not be a" }, { "n": 46738, "d": "2026-06-30", "t": "🚨 Modeling changes for export, compile, and hybrid-attention standardization", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 2353, "r": 3720, "f": 152, "c": 48, "b": "Modeling changes for export, compile, and hybrid-attention standardization A broad sweep of modeling-side changes that make a large set of models cleanly exportable (ONNX / torch.export / ExecuTorch), fullgraph-compilable, and consistent in how they declare their layer types and build masks/caches. " }, { "n": 47014, "d": "2026-07-30", "t": "🚨 Enable SDPA (and other attention backends) for T5 and propagate to the T5 family", "st": "MERGED", "by": "jiqing-feng", "a": 1259, "r": 1364, "f": 20, "c": 32, "b": "What this PR does Refactors the T5 attention stack to route through ALL_ATTENTION_FUNCTIONS, so T5 can dispatch to sdpa / eager instead of the old eager-only path, and propagates the change to the copied-from family. Now _supports_sdpa = True / _supports_attention_backend = True and running test_eag" }, { "n": 46196, "d": "2026-08-20", "t": "🚨[wav2vec2] Support attn_implementation=sdpa dispatch", "st": "MERGED", "by": "YangKai0616", "a": 1767, "r": 3073, "f": 33, "c": 53, "b": "The wav2vec2_conformer/bert and SeamlessM4T/v2 models fail during the initialization phase when using from_pretrained(..., attn_implementation=\"sdpa\"). This PR enables standard attn_implementation=\"sdpa\" dispatch to these models." }, { "n": 48333, "d": "2026-08-27", "t": "Use the shared attention interface in encoder-decoder attention blocks", "st": "OPEN", "by": "ArthurZucker", "a": 1016, "r": 993, "f": 14, "c": 3, "b": "Routes 15 bart-style *Attention forwards through ALL_ATTENTION_FUNCTIONS, enabling sdpa/flash. No weight renames; modules read hyperparameters from config. Deltas fp32-rounding only (<=8.4e-07). ProphetNet attentions were pre-softmax logits, now probabilities. FSMT split out to #48357; Autoformer an" }, { "n": 48357, "d": "2026-08-27", "t": "Modernize FSMT: shared attention interface and remove legacy fairseq code", "st": "OPEN", "by": "ArthurZucker", "a": 536, "r": 792, "f": 2, "c": 1, "b": "Reshapes FSMT after BART: attention via ALL_ATTENTION_FUNCTIONS, masks built in the model, batch-first hidden states, new decorators, sdpa/flash support. Drops ~250 lines of fairseq helpers; fixes forward(input_ids, labels=...). No weight renames; logits match within 3.3e-5 on wmt19-en-de." } ], "d": "Replaced per-model eager/SDPA/FlashAttention forks with one ALL_ATTENTION_FUNCTIONS registry and config.attn_implementation dispatch, now propagated to T5, Wav2Vec2, FSMT and encoder-decoder families.", "note": null, "links": [], "src": "speed.json", "span": [ "2024-12-16", "2026-08-27" ], "roll": { "owners": [ [ "Cyrilvallez", 4 ], [ "ArthurZucker", 3 ], [ "IlyasMoutawwakil", 1 ], [ "jiqing-feng", 1 ] ], "n": 10, "add": 16223, "del": 26874, "files": 500, "commits": 500 } }, { "t": "transformers serve & chat CLI", "tag": "OpenAI-compatible server", "s": "wip", "prs": [ { "n": 35714, "d": "2025-01-22", "t": "[Chat] Add Chat from TRL 🐈 ", "st": "MERGED", "by": "gante", "a": 695, "r": 103, "f": 7, "c": 7, "b": "Adds a chat command, allowing users to explore text models with a minimal effort. Except for the transformers-specific command boilerplate, this chat is a copy of the chat in TRL. The purpose of this PR is to establish a baseline from something that already works well. In subsequent PRs, a few QoL u" }, { "n": 38443, "d": "2025-06-30", "t": "Split `transformers chat` and `transformers serve` ", "st": "MERGED", "by": "LysandreJik", "a": 915, "r": 310, "f": 9, "c": 25, "b": "This PR splits the transformers chat frontend from its backend (now moved to transformers serve). We take this opportunity to have transformers serve act as a server with an OpenAI-compatible API, via http, using SSE to stream tokens. You can take it for a spin either by spawning a server: transform" }, { "n": 39155, "d": "2025-07-16", "t": "Responses API in `transformers serve`", "st": "MERGED", "by": "LysandreJik", "a": 948, "r": 391, "f": 8, "c": 10, "b": "This PR offers an initial scaffolding of transformers serve so that it has both /v1/chat/completions and /v1/responses endpoints. So far, it only implements a streaming version of both endpoints. To implement the responses endpoint, I'm basing myself off of the \"Streaming\" tab in the OpenAPI API ref" }, { "n": 39454, "d": "2025-07-23", "t": "Transformers serve VLM", "st": "MERGED", "by": "LysandreJik", "a": 268, "r": 75, "f": 2, "c": 8, "b": "Add VLM support for transformers serve. In case you're interested in playing around with it, Open WebUI is a good tool to work with. Launching transformers serve with VLM support: cd <PATH_TO_TRANSFORMERS> git checkout transformers-serve-vlm transformers serve --enable_cors To install Open WebUI: pi" }, { "n": 41446, "d": "2025-10-15", "t": "Enable non-streaming mode in `transformers serve`", "st": "MERGED", "by": "LysandreJik", "a": 211, "r": 73, "f": 4, "c": 3, "b": "Needs this to be merged first: #41444 Tests and docs need to be added before undraft" }, { "n": 44558, "d": "2026-04-22", "t": "Add /v1/completions endpoint (OpenAI legacy completions API) to `transformers serve`", "st": "MERGED", "by": "rain-1", "a": 567, "r": 1, "f": 5, "c": 1, "b": "Adds support for the legacy text completions endpoint, which accepts a freeform text prompt (no chat template) and returns generated text in choices[].text. Supports both streaming and non-streaming modes, suffix for fill-in-the-middle insertion, and proper finish_reason detection. What does this PR" } ], "d": "Built a local OpenAI-compatible inference server (`transformers serve`) split out from the `transformers chat` CLI, adding a Responses API, VLM support, and a /v1/completions endpoint.", "note": null, "links": [], "src": "speed.json", "span": [ "2025-01-15", "2026-04-22" ], "roll": { "owners": [ [ "LysandreJik", 4 ], [ "gante", 1 ], [ "rain-1", 1 ] ], "n": 6, "add": 3604, "del": 953, "files": 35, "commits": 54 } }, { "t": "Cache architecture refactor", "tag": "one cache, per-layer design", "s": "done", "prs": [ { "n": 39106, "d": "2025-07-22", "t": "[cache refactor] Move all the caching logic to a per-layer approach", "st": "MERGED", "by": "manueldeprada", "a": 2505, "r": 2167, "f": 64, "c": 44, "b": "This PR completes Part 1 of the cache refactor tracked in #38077. Summary: Now Cache is structured in a list of layers. Ports all existing cache types (Static, Dynamic, Offloaded, Quantized, Hybrid, etc.) to use layer composition Backward compatibility, tests passing. Modification logic (e.g., reset" }, { "n": 39590, "d": "2025-07-23", "t": "Fix DynamicCache and simplify Cache classes a bit", "st": "MERGED", "by": "Cyrilvallez", "a": 147, "r": 184, "f": 6, "c": 8, "b": "As per the title. DynamicCache was broken because self.layers needs to exist if we want to append to it -> test_multi_gpu_data_parallel_forward was failing cc @manueldeprada for viz" }, { "n": 39797, "d": "2025-08-08", "t": "[core] Refactor the Cache logic to make it simpler and more general", "st": "MERGED", "by": "Cyrilvallez", "a": 779, "r": 1449, "f": 48, "c": 51, "b": "Big simplifications everywhere, but most notably: all caches are initialized lazily -> no more issues of devices with device_map, which would lead to breaking the Static dynamo addresses due to device movement + no issue of dimensions with TP + much simpler to prepare for generate (all properties ar" }, { "n": 40420, "d": "2025-08-28", "t": "Use the config for DynamicCache initialization in all modelings", "st": "MERGED", "by": "Cyrilvallez", "a": 195, "r": 181, "f": 131, "c": 4, "b": "As per the title" }, { "n": 47094, "d": "2026-07-07", "t": "[sdpa] Allow prefill to use FA kernel with StaticCache", "st": "MERGED", "by": "Cyrilvallez", "a": 43, "r": 66, "f": 2, "c": 6, "b": "As per the title. In general, we have no reason to constrain the prefill to have a mask with sdpa, as most usual situations will not need it. It's more efficient to let it dispatch to the FA kernel if we can. It also has the following benefits: much faster, with quadratic benefits as the prefill siz" } ], "d": "Rebuilt Cache/DynamicCache/StaticCache around a per-layer design driven by model config, simplifying cache classes and letting SDPA prefill use FlashAttention kernels with StaticCache.", "note": null, "links": [], "src": "speed.json", "span": [ "2025-06-29", "2026-07-07" ], "roll": { "owners": [ [ "Cyrilvallez", 4 ], [ "manueldeprada", 1 ] ], "n": 5, "add": 3669, "del": 4047, "files": 251, "commits": 113 } }, { "t": "TGI backend", "tag": "shipped outside this repo", "s": "done", "prs": [], "note": "Landed in huggingface/text-generation-inference — no transformers PR to link.", "links": [ [ "text-generation-inference", "https://github.com/huggingface/text-generation-inference" ] ], "d": null, "src": "curated", "span": null, "roll": { "owners": [], "n": 0, "add": 0, "del": 0, "files": 0, "commits": 0 } } ] }, { "name": "Continuous batching", "hue": 14, "blurb": "Serving-grade batching inside the library — the big 2026 push.", "items": [ { "t": "CB core engine, scheduler & API", "tag": "request scheduler + generate_batch API", "s": "done", "prs": [ { "n": 38085, "d": "2025-05-22", "t": "Add CB", "st": "MERGED", "by": "ArthurZucker", "a": 3467, "r": 11, "f": 21, "c": 110, "b": "Snippet: import time import datasets import torch from transformers import AutoModelForCausalLM, AutoTokenizer from transformers.generation import GenerationConfig torch.set_float32_matmul_precision(\"high\") model_id = \"meta-llama/Llama-3.2-3b-Instruct\" model = AutoModelForCausalLM.from_pretrained( m" }, { "n": 39457, "d": "2025-07-22", "t": "[Paged-Attention] Handle continuous batching for repetition penalty", "st": "MERGED", "by": "kashif", "a": 80, "r": 5, "f": 3, "c": 9, "b": "We handle the repetition penalty for continuous batching Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or " }, { "n": 40426, "d": "2025-08-26", "t": "Continuous batching refactor", "st": "MERGED", "by": "remi-or", "a": 2174, "r": 1547, "f": 10, "c": 27, "b": "This PR refactors the files related to continuous batching and changes a few parts of the logic: the way the max_batch_tokens and num_blocks is computed; the static inputs of the CB are now sliced to avoid passing an arbitrary amount of padding in the forward the example file for CB has been reworke" }, { "n": 40692, "d": "2025-09-08", "t": "[Continous Batching] fix do_Sample=True in continuous batching", "st": "MERGED", "by": "kashif", "a": 87, "r": 7, "f": 4, "c": 6, "b": "This pull request introduces improvements to sampling support in continuous batching generation, enhances tensor shape handling for logit processors, and expands test coverage for batch generation with sampling. The main changes are grouped below: Sampling and Generation Behavior: Enabled sampling i" }, { "n": 40828, "d": "2025-09-12", "t": "Fixes for continuous batching", "st": "MERGED", "by": "remi-or", "a": 474, "r": 199, "f": 11, "c": 22, "b": "Some architectures like llama alter the attention mask if it is not a tensor, which was not compatible with the way CB created and handled the attention mask. Now, arguments like attention_mask, cumulative_seqlens_k and max_seqlen_k are tensors or ints unless the model is hybrid, in which case they " }, { "n": 41655, "d": "2025-10-16", "t": "Switch to CB if cache_implementation == paged", "st": "MERGED", "by": "remi-or", "a": 75, "r": 3, "f": 2, "c": 7, "b": "This PR fixes one of the issues raised in #41501 This PR adds paged as a static cache implementation and switches from generate to generate_batch if it's the chosen implementation. It also raises errors if an argument for generate is passed to generate_batch but is not supported, and warn the user t" }, { "n": 42699, "d": "2025-12-11", "t": "Better continuous batching tests", "st": "MERGED", "by": "remi-or", "a": 206, "r": 298, "f": 2, "c": 10, "b": "Since a lot of features have been added to continuous batching, this PR aims o refactor the associated tests so we can catch new failures. The structure of the tests used to be one test per model / attention implementation. Now we have one test backend, that makes sure the generation with and withou" }, { "n": 42839, "d": "2025-12-15", "t": "[CB] Easy optimizations for continuous batching", "st": "MERGED", "by": "remi-or", "a": 40, "r": 24, "f": 3, "c": 10, "b": "Summary This PR adds a few optimizations for continuous batching: removes non-needed torch.cuda.synchronize add sorting of the inputs to maximize prefix caching hits sampling is done on the GPU removed an extranuous axis from the output_ids Performance Attention Version Generated tokens Duration (s)" }, { "n": 42942, "d": "2025-12-18", "t": "Fix result retrieval starvation and terminate request-scoped iteration on completion", "st": "OPEN", "by": "pythongiant", "a": 45, "r": 23, "f": 1, "c": 1, "b": "This PR fixes two correctness issues in the continuous batching result consumption logic. First, it improves ContinuousBatchingManager.get_result to avoid starvation and incorrect timeout behavior when multiple requests share a single output queue. Instead of immediately re-queuing mismatched result" }, { "n": 42921, "d": "2026-01-06", "t": "[CB] Support the `num_return_sequences` argument", "st": "MERGED", "by": "remi-or", "a": 333, "r": 39, "f": 9, "c": 14, "b": "Summary This PR adds the options to fork requests during continuous batching, which duplicates the request and uses as much as possible the existing cache. This is then leveraged to make the num_return_sequences argument available in CB. This PR enables parallel decoding, which will be useful for RL" }, { "n": 43237, "d": "2026-01-14", "t": "[CB] Refactor: extract common scheduling logic into Scheduler base class", "st": "MERGED", "by": "dstekanov", "a": 74, "r": 94, "f": 1, "c": 6, "b": "This PR refactors the FIFOScheduler and PrefillFirstScheduler classes to extract common scheduling logic into the base Scheduler class (fixed TODO: further common-ize the two classes) Local verification Verified basic continuous batching utilities locally on CPU (macOS) by running existing tests: py" }, { "n": 43569, "d": "2026-02-02", "t": "[CB] Refactor logic for inputs and outputs outside of the main API", "st": "MERGED", "by": "remi-or", "a": 509, "r": 416, "f": 4, "c": 7, "b": "This PR refactors all of the logic linked to the inputs and outputs of continuous batching into a new class ContinuousBatchingIOs itself in a new file. This will be quite useful when we introduce async scheduling, because there will be two sets of host-side IOs and one set of device-side IOs. To avo" }, { "n": 43626, "d": "2026-02-03", "t": "[CB] Keep order of incoming requests", "st": "MERGED", "by": "remi-or", "a": 40, "r": 33, "f": 2, "c": 5, "b": "The test tests/generation/test_paged_attention.py::TestBatchGeneration::test_generate_batch_consistency is failing because CB shuffles incoming request to maximize prefix caching. This is not a problem unless the user expects outputs to have the same order as the inputs when using generate_batch, wh" }, { "n": 43960, "d": "2026-02-23", "t": "[CB] [Major] Asynchronous batching", "st": "MERGED", "by": "remi-or", "a": 1194, "r": 666, "f": 11, "c": 39, "b": "Summary This PR adds the asynchronous batching feature to continuous batching (CB). Asynchronous batching, through the use of more VRAM and CUDA streams and events, greatly reduces the CPU overhead of preparing and updating batches by hiding it with GPU compute. In practice, the GPU can run uninterr" }, { "n": 44227, "d": "2026-03-03", "t": "[CB] Small fixes", "st": "MERGED", "by": "remi-or", "a": 21, "r": 13, "f": 4, "c": 2, "b": "This PR adds a logging message when infering the behavior of use async and fixes an error when evicting a graph from the graph buffer." }, { "n": 44434, "d": "2026-03-13", "t": "[CB] Add dedicated config", "st": "MERGED", "by": "remi-or", "a": 413, "r": 303, "f": 9, "c": 22, "b": "This PR adds a dedicated config for continuous batching, which is starting to have a lot parameters. This will give the user a clear view of what is possible and make adding new parameters easier. No breaking changes through account_for_cb_deprecated_arguments which is a function that ensures the ol" }, { "n": 44673, "d": "2026-03-13", "t": "[CB] [Bug] Fix crashes when running without cuda", "st": "MERGED", "by": "remi-or", "a": 73, "r": 15, "f": 3, "c": 4, "b": "This PR fixes a bug in continuous batching where non-CUDA devices cannot use the feature because some CUDA-exclusive objects are always instantiated. It also adds a test to make sure this will not break again in the future." }, { "n": 44835, "d": "2026-03-23", "t": "[CB] Add an option to return logprobs", "st": "MERGED", "by": "remi-or", "a": 187, "r": 83, "f": 6, "c": 14, "b": "Summary This PR adds the return_logprobs flag to the continuous batching, enabling the user to retrieve the log probabilites of the tokens generated. Tests Added a test to compare with regular generate and it passes. All tests pass. Performance No noticeable impact." }, { "n": 44924, "d": "2026-03-23", "t": "Continuous batching thread safety", "st": "MERGED", "by": "Qubitium", "a": 9, "r": 1, "f": 1, "c": 20, "b": "Fix two nogil threading bugs (reproduced on 3.14) : Continus Batching crashes with torch graph errors with 2 threads on 2 separate model instances (same model path, but two distinct instances). Cause is missing capture_error_mode=\"thread_local\" so the graph capture operate safely in thread_local mod" }, { "n": 44435, "d": "2026-03-26", "t": "[CB] Persistent manager", "st": "MERGED", "by": "remi-or", "a": 138, "r": 54, "f": 6, "c": 9, "b": "This PR adds the option to have a ContinuousBatchingManager not be destroyed after generation is over. This allows the user to re-use the manager without requiring him to know any other entry point for CB apart from generate_batch or the context manager. If we want to have sleep function at some poi" }, { "n": 45112, "d": "2026-03-31", "t": "[CB] Add warmup feature", "st": "MERGED", "by": "remi-or", "a": 233, "r": 46, "f": 5, "c": 16, "b": "This PR adds a warmup phase before generation starts, turned on by default. It allows for better diagnostics and a more representative user experience than without warmup, where the cost of wamup is payed during the first request rather than in a phase before generation starts. Warmup is skipped if " }, { "n": 45026, "d": "2026-04-03", "t": "[CB] Add per-request logits processors", "st": "MERGED", "by": "remi-or", "a": 714, "r": 202, "f": 11, "c": 2, "b": "Summary This PR adds per-request logits processors and overalls the way CB handles logits processors. It introduces batched logits processing with per-request parameters for continuous batching, enabling each request in a batch to use different sampling parameters (temperature, top_k, top_p). This i" }, { "n": 45179, "d": "2026-04-03", "t": "[CB] Tweaks to update and minor fixes", "st": "MERGED", "by": "remi-or", "a": 327, "r": 288, "f": 6, "c": 13, "b": "Summary This PR ads minor changes to cache.update, updates the memory handler with all new features and refactors a few parts of the code to make it more readable. Cache indexing: Replace fancy indexing (cache[idx, :, :]) with explicit torch.index_select / index_copy_, which have cleaner behavior un" }, { "n": 45530, "d": "2026-04-23", "t": "[CB] Changes for long generation", "st": "MERGED", "by": "remi-or", "a": 189, "r": 110, "f": 8, "c": 17, "b": "Summary This PR fixes some issues related to memory to make long generation (16K+) easier. Fix KV dedup for decode batches (scheduler.py): Decode-only batches don't consume the read_indices cache budget, so don't reject them on that basis. Also gate the decode fast path on max_blocks_per_request > 0" }, { "n": 45582, "d": "2026-04-24", "t": "generate: drop stale num_return_sequences warning on continuous batching path", "st": "MERGED", "by": "joaquinhuigomez", "a": 2, "r": 6, "f": 1, "c": 4, "b": "The continuous-batching branch in generate warned that num_return_sequences was unsupported alongside num_beams, but generate_batch() already honors generation_config.num_return_sequences when expanding requests. The warning fires for any run that explicitly sets num_return_sequences even though the" }, { "n": 45653, "d": "2026-05-01", "t": "[CB] Better overall script and decode bucketting", "st": "MERGED", "by": "remi-or", "a": 401, "r": 180, "f": 5, "c": 12, "b": "Summary This PR updates the continuous_batching_overall script to better reflect the way users use continuous batching (ie. using simpler or the default config) on realistic tasks (GSM8K or RL rollouts of various lengths). It also changes the way decode-only batches are bucketed: instead of using li" }, { "n": 45654, "d": "2026-05-06", "t": "[CB] Refactor any model-related code in a separate class", "st": "MERGED", "by": "remi-or", "a": 655, "r": 568, "f": 9, "c": 10, "b": "Summary This PR simplifies the continuous batching code by refactoring away any model-related code in the ModelRunner class. This is because the ContinuousBatchingProcessor class is starting to get too big and this will help keep things organized. Draft until #45653 is merged. Performance label samp" }, { "n": 45984, "d": "2026-05-19", "t": "[CB] Remove OpenTelemetry", "st": "MERGED", "by": "remi-or", "a": 1, "r": 1753, "f": 19, "c": 3, "b": "This PR remove the code related to OpenTelemetry in continuous batching because: it has not been updated since the rework of CB started (~September 2025) it is not used by serve it is not used in CB If telemetry is needed, it will be added back, after manager rework is done and TP is finalized, both" }, { "n": 46251, "d": "2026-06-04", "t": "[CB] Fix crashes when fork is not possible", "st": "MERGED", "by": "remi-or", "a": 30, "r": 2, "f": 2, "c": 4, "b": "Currently there is a crash if a user tries to fork a request but there isn't enoguh blocks to do that. This PR fixes this: before forking, we count the maximum number of forks possible according to the current number of free blocks. We limit the number of forks using this number, and the non-forked " }, { "n": 46434, "d": "2026-06-09", "t": "[CB] Add a \"max_requests_per_batch\" parameter", "st": "MERGED", "by": "remi-or", "a": 259, "r": 154, "f": 11, "c": 23, "b": "Summary This PR adds a max_requests_per_batch to the continuous batching config, which can be manually set by the user or auto-inferred from the number of requests passed. Introducing this parameters create a clear difference between: the tensors that should be sized by the number of tokens in a bat" }, { "n": 46593, "d": "2026-06-13", "t": "[CB] Fix seqlens and use TypedDict", "st": "MERGED", "by": "remi-or", "a": 68, "r": 61, "f": 3, "c": 9, "b": "Summary This PR fixes a race condition where the cumulative_seqlens tensors would be padded on the device during their transfer. It seems the bug did not happen often in a non-distributed setting, but it appeared when running the continuous batching overall script in DP 8. It is now fixed: the paddi" }, { "n": 46660, "d": "2026-06-15", "t": "[CB] Slice logits inside the model", "st": "MERGED", "by": "remi-or", "a": 18, "r": 12, "f": 1, "c": 3, "b": "Summary This PR leverages an awesome kwarg in transformers: logits_to_keep. Simply, this drop the tokens for which we won't use the next token predictions before we enter the lm_head. Reduces the memory footprints of large prefill batches, increase their throughput, makes everyone happy. Performance" }, { "n": 46670, "d": "2026-06-16", "t": "[Continuous Batching] Snapshot generation outputs without mutating request state", "st": "MERGED", "by": "Incheonkirin", "a": 58, "r": 7, "f": 2, "c": 1, "b": "This fixes RequestState.to_generation_output() in continuous batching so that building a GenerationOutput does not mutate the active request state or return live aliases for growing output buffers. Streaming continuous batching can call to_generation_output() after every generated token. On current " }, { "n": 46765, "d": "2026-06-24", "t": "[CB] Fix issues with FA read / writes", "st": "MERGED", "by": "remi-or", "a": 117, "r": 21, "f": 6, "c": 8, "b": "Summary This PR fixes an issue where the padding tokens from flash attention would write NaN in the cache at the same place where they read their cache from. This should not be a problem, expect NaNs screw up the softmax of real tokens for which the QK tensor is computed in the same block as the pad" }, { "n": 46712, "d": "2026-06-26", "t": "[CB] Changes to increase max_batch_tokens", "st": "MERGED", "by": "remi-or", "a": 392, "r": 272, "f": 10, "c": 3, "b": "Summary This PR aims at increasing the performances of continuous batching, especially for prefill batches. To that order, it changes the following things: rework the cache size estimator by adding a default max_batch_tokens of 8192 and applying a VRAM-based bound. In effect, this means that the pre" }, { "n": 47038, "d": "2026-07-07", "t": "Keep continuous batching config separate from GenerationConfig", "st": "CLOSED", "by": "VectorPeak", "a": 44, "r": 6, "f": 3, "c": 9, "b": "Addresses #47039 This PR keeps ContinuousBatchingConfig separate from GenerationConfig serialization. Continuous batching configuration is runtime/hardware-specific, while GenerationConfig is model-generation-specific, so the two configs should live side by side rather than nested inside generation_" }, { "n": 47330, "d": "2026-07-16", "t": "[CB] Automatically switch attention implementation to flash", "st": "MERGED", "by": "remi-or", "a": 102, "r": 19, "f": 3, "c": 10, "b": "This PR change the behavior of CB when it comes to attn_implementation. Before, we only ensured that the attention was paged, which is necessary for CB. With this PR, it also checks if flash is available (version 3 then 2, either through the package or through kernels) and if it is, it automatically" }, { "n": 47291, "d": "2026-07-24", "t": "Deprecate CB config in gen configuration", "st": "MERGED", "by": "remi-or", "a": 19, "r": 0, "f": 2, "c": 3, "b": "This PR deprecates passing a ContinuousBatchingConfig through the GenerationConfig which is a different object: the generation config is model-centered while the CB config is hardware-centered. Closes #47039 ." }, { "n": 47341, "d": "2026-08-18", "t": "[Fix] Small FA-related test failures in CB", "st": "MERGED", "by": "remi-or", "a": 34, "r": 13, "f": 2, "c": 4, "b": "This PR fixes two flash-attention related issues in CB: block table is not supported for sliding window attention, yet the order of checks makes it so that a sliding attention model can turn on the block table fast path if it uses FA2. This was previously not an issue because FA2 disable the block t" }, { "n": 48334, "d": "2026-08-27", "t": "[CB] Fail faster", "st": "MERGED", "by": "remi-or", "a": 62, "r": 25, "f": 5, "c": 7, "b": "This PR fixes #48303 and makes error handling more redundant in CB: the fatal_error attribute lived awkwardly next to background_thread_status, an object meant to keep track of the bg thread status (shockingly). Since fatal_error is there for consumers to check if the thread is still running, and if" }, { "n": 48370, "d": "2026-08-28", "t": "[CB] Fix wrong device scoping", "st": "MERGED", "by": "remi-or", "a": 41, "r": 0, "f": 2, "c": 3, "b": "This PR fixes errors that could happen when CB is on multiple devices and some thread are wrongly scoped: they think they are on device 0 at creation when really they might be on device n. This leads to error during capture like #48312 or wrong VRAM measurements." } ], "d": "Founding CB PR (#38085) through the async scheduler, dedicated config, per-request logits processors/logprobs, warmup, request ordering and decode bucketing.", "note": "1 correctness bugfix (#42942) still open", "links": [], "src": "cb.json", "span": [ "2025-04-15", "2026-08-28" ], "roll": { "owners": [ [ "remi-or", 32 ], [ "kashif", 2 ], [ "ArthurZucker", 1 ], [ "dstekanov", 1 ] ], "n": 41, "add": 13405, "del": 7579, "files": 229, "commits": 508 } }, { "t": "CB in transformers serve", "tag": "serving-grade batching in the OpenAI-compatible server", "s": "done", "prs": [ { "n": 39149, "d": "2025-07-03", "t": "Fix continuous batching in `transformers serve`", "st": "MERGED", "by": "LysandreJik", "a": 38, "r": 28, "f": 1, "c": 6, "b": "" }, { "n": 40479, "d": "2025-09-02", "t": "fix: continuous batching in `transformers serve`", "st": "MERGED", "by": "McPatate", "a": 136, "r": 58, "f": 4, "c": 7, "b": "Fixing continuous batching in transformers serve. added a --continuous_batching cmd line flag to enable, open to change this! max_new_tokens can sometimes be None, set a default so it doesn't break CB which expects it to be set can't repro my previous error, removed the added code and added a test t" }, { "n": 41624, "d": "2025-10-16", "t": "Fix serving continuous batching", "st": "MERGED", "by": "SunMarc", "a": 25, "r": 22, "f": 5, "c": 6, "b": "Serving is broken with continuous batching to due recent PRs. This PR fixes it" }, { "n": 43594, "d": "2026-01-30", "t": "[CB] [Serve] Fix broken serve tests", "st": "MERGED", "by": "remi-or", "a": 22, "r": 13, "f": 2, "c": 6, "b": "This PR fixes two issues related to CB and serve: when the max_length argument as set to None when creating a request, CB had an error because it always expected max_length to be an int. This is now an expected behavior, and we have a hidden variable to avoid checking if the variable is None the CB " }, { "n": 45057, "d": "2026-03-31", "t": "[serving] Fix continuous batching JSON response serialization", "st": "MERGED", "by": "NathanHB", "a": 318, "r": 37, "f": 3, "c": 10, "b": "Change model_dump_json() to model_dump() to avoid double JSON encoding. When using continuous batching with stream=false, the response was being double-encoded as a string instead of returning a proper JSON object. Added a UV script to run GPQA using transformer-serve: hf jobs uv run examples/pytorc" } ], "d": "Wires continuous batching into transformers serve: request handling, JSON response streaming/serialization, and broken serve-test fixes.", "note": null, "links": [], "src": "cb.json", "span": [ "2025-07-01", "2026-03-31" ], "roll": { "owners": [ [ "LysandreJik", 1 ], [ "McPatate", 1 ], [ "SunMarc", 1 ], [ "remi-or", 1 ] ], "n": 5, "add": 539, "del": 158, "files": 15, "commits": 35 } }, { "t": "CB docs & dev UX", "tag": "documented CB", "s": "done", "prs": [ { "n": 40758, "d": "2025-09-08", "t": "docs: add continuous batching to serving", "st": "MERGED", "by": "McPatate", "a": 36, "r": 2, "f": 1, "c": 3, "b": "Adding a description of continuous batching in serving docs" }, { "n": 41847, "d": "2025-11-03", "t": "docs: add continuous batching page", "st": "MERGED", "by": "McPatate", "a": 204, "r": 0, "f": 5, "c": 8, "b": "" }, { "n": 42329, "d": "2025-12-05", "t": "[docs] Attention backends + continuous batching", "st": "MERGED", "by": "stevhliu", "a": 242, "r": 212, "f": 7, "c": 7, "b": "Refactors the optimization docs, starting with the attention backend and continuous batching docs (more updates to come in subsequent PRs!). reorg the toctree to create a clearer inference optimization section for inference (adds the AttentionInterface, continuous batching, and kernels guides) renam" }, { "n": 44896, "d": "2026-03-30", "t": "[docs] continuous batching", "st": "MERGED", "by": "stevhliu", "a": 354, "r": 81, "f": 4, "c": 2, "b": "updates the continuous batching docs new page for the API reference adds sections for new features like CUDA graphs, async batching, prefix caching, logprobs (depending on when its merged) clearer example of generation with varying loads using continuous_batching_context_manager new page explaining " }, { "n": 46334, "d": "2026-06-02", "t": "[docs] xpu continuous batching", "st": "MERGED", "by": "stevhliu", "a": 8, "r": 1, "f": 1, "c": 1, "b": "adds docs for #46028 (enabling cb decode fast path on XPU)" }, { "n": 46019, "d": "2026-06-08", "t": "[docs] tp for continuous batching", "st": "MERGED", "by": "stevhliu", "a": 64, "r": 1, "f": 1, "c": 2, "b": "adds docs for #45821" }, { "n": 46925, "d": "2026-07-07", "t": "[docs] continuous batching (offloading behavior, max batch tokens, block size minimum)", "st": "MERGED", "by": "stevhliu", "a": 34, "r": 7, "f": 2, "c": 3, "b": "Adds docs for several recent continuous batching PRs (#46765, #46712, #46587): New section for max_batch_tokens, the trade off and when to raise or lower it New section explaining block_size Reframe soft reset as the fallback path for offloading Describe new offloading behavior (demand-drive, newest" }, { "n": 48381, "d": "2026-08-28", "t": "[docs] Pass ContinuousBatchingConfig and sliding window models ", "st": "MERGED", "by": "stevhliu", "a": 7, "r": 2, "f": 1, "c": 1, "b": "Lets users know to pass ContinuousBatchingConfig to the continuous_batching_config arg instead of GenerationConfig and mentions that the decode fast path is disabled for sliding window models." } ], "d": "Doc pages for continuous batching: quick start, attention backends, tensor parallelism, XPU, and tuning offloading/max_batch_tokens/block_size.", "note": null, "links": [], "src": "cb.json", "span": [ "2025-09-08", "2026-08-28" ], "roll": { "owners": [ [ "stevhliu", 6 ], [ "McPatate", 2 ] ], "n": 8, "add": 949, "del": 306, "files": 22, "commits": 27 } }, { "t": "Paged KV cache & memory management", "tag": "paged attention + prefix caching + CPU offload", "s": "wip", "prs": [ { "n": 40688, "d": "2025-09-09", "t": "Support sliding window in CB", "st": "MERGED", "by": "remi-or", "a": 1018, "r": 429, "f": 11, "c": 40, "b": "This PR introduces support for sliding window attention in CB, using an allocator that can support both full attention and sliding window attention. Code works with almost similar performances (665 -> 655 tok/s in the example) but I am putting the PR in draft status until: Code is tested on Nvidia w" }, { "n": 41078, "d": "2025-09-26", "t": "Fix flash-attn for paged_attention when no kernels", "st": "MERGED", "by": "remi-or", "a": 16, "r": 3, "f": 2, "c": 7, "b": "Currently there is an issue when using paged_attention without the kernels package and with flash_attn installed: the flash_attn_varlen_func is no longer referenced when reaching its call. For some reason, by assigning it to a value, this can be avoided. The issue was observed on A100 and Mi325. In " }, { "n": 41228, "d": "2025-09-30", "t": "Fix sliding window attn mask", "st": "MERGED", "by": "remi-or", "a": 95, "r": 3, "f": 2, "c": 5, "b": "It was reporedt in #41184 that the sliding window attention mask has errors in CB. This PR fixes this and adds a test for the build_attention_mask function, so it does not happen again." }, { "n": 41370, "d": "2025-10-06", "t": "[`CB`] Refactors the way we access paged", "st": "MERGED", "by": "ArthurZucker", "a": 33, "r": 43, "f": 8, "c": 6, "b": "The user decides on the function, but CB handles itself which \"interface\" wrapper to use. This should make stuff easier" }, { "n": 42094, "d": "2025-11-17", "t": "Add prefix sharing to continuous batching", "st": "MERGED", "by": "remi-or", "a": 570, "r": 231, "f": 7, "c": 28, "b": "This PR adds a prefix sharing mechanism to the continuous batching API like the one present in VLLM. It only activates if the model is a full-attention model, as is the case in VLLM. The mechanism has two main components: block hashing: each block in the cache, once it is filled up, is given a hash " }, { "n": 42547, "d": "2025-12-08", "t": "Fixed paged|FA2 kernel loading logic and UT.", "st": "MERGED", "by": "YangKai0616", "a": 16, "r": 11, "f": 3, "c": 16, "b": "1. [General] When running the test case tests/generation/test_continuous_batching.py::ContinuousBatchingTest::test_continuous_batching_parity_qwen_flash, it reports error as follows: flash_attn_func = getattr(kernel, \"flash_attn_func\", None) flash_attn_varlen_func = getattr(kernel, \"flash_attn_varle" }, { "n": 42877, "d": "2025-12-18", "t": "[CB] Allow block sharing in hybrid models", "st": "MERGED", "by": "remi-or", "a": 168, "r": 94, "f": 6, "c": 5, "b": "Summary This PR increases the granularity of KV cache sharing. Previously, sharing was only enabled for full-attention only models. With this PR, sharing is enabled for any model that has full attention layers. It can still be disabled with a flag. This PR paves the way for parallel decoding, by mak" }, { "n": 43150, "d": "2026-01-13", "t": "[CB] Soft-reset offloading", "st": "MERGED", "by": "remi-or", "a": 383, "r": 108, "f": 8, "c": 20, "b": "Summary This PR adds a basic form of offloading to the continuous batching API. Currently, when we cannot schedule any request because the KV cache is full, we throw an error, which is less than optimal. This PR fixes this by adding a mechanism where instead of throwing an error, we de-schedule the " }, { "n": 42869, "d": "2026-01-29", "t": "enable cpu paged cache", "st": "MERGED", "by": "jiqing-feng", "a": 57, "r": 29, "f": 5, "c": 50, "b": "CPU can also use paged cache with eager or sdpa: python continuous_batching_simple.py --attn sdpa Without this change, the previous command error would be like: Error in generation loop: unsupported operand type(s) for -: 'NoneType' and 'int' Traceback (most recent call last): File \"/home/jiqing/tra" }, { "n": 44379, "d": "2026-03-09", "t": "[CB] Add paged_attention kernel", "st": "MERGED", "by": "remi-or", "a": 579, "r": 235, "f": 14, "c": 6, "b": "Summary ! This PR is in draft, waiting for #44227 to be merged This PR adds support for the flash_attention_with_kvcache kernel in continuoys batching. This is very efficient for decode-only batches, hence especially useful for long generations. Right now, this needs to be explicitly turned on by th" }, { "n": 45184, "d": "2026-04-27", "t": "[CB] [Major] Add CPU request offloading", "st": "MERGED", "by": "remi-or", "a": 447, "r": 60, "f": 7, "c": 14, "b": "Summary This PR adds CPU offloading to continuous batching. It's in raft until perf and test status are reported. When the GPU KV cache is full and a request must be evicted, we check if there is enough VRAM to copy the request's KV cache to a pre-allocated pinned CPU buffer instead of discarding th" }, { "n": 45733, "d": "2026-05-07", "t": "[CB] Fixes for SDPA and CPU offloading", "st": "CLOSED", "by": "remi-or", "a": 0, "r": 0, "f": 0, "c": 3, "b": "This PR fixes two issues: SDPA attn used the wrong total_seqlen_k which led to crashes CPU offloading did not copy the cache correctly when blocks were non-contiguous (fancy indexing) This PR fixes both issues." }, { "n": 46587, "d": "2026-06-17", "t": "[CB] Fix offloading", "st": "MERGED", "by": "remi-or", "a": 260, "r": 128, "f": 6, "c": 5, "b": "Summary This PR changes the offloading mechanism so that it can offload multiple requests at the same time. Before, offloading triggered when we could not schedule a batch: we offloaded ONE request, and waited for the next forward to schedule the next batch. This lead to an arbitrary number of reque" }, { "n": 47911, "d": "2026-08-12", "t": "Update sliding window indexing in continuous batching", "st": "OPEN", "by": "SindhuRaghuram97", "a": 19, "r": 4, "f": 5, "c": 2, "b": "Fixes the retrieval of historical tokens in sliding window layers during paged attention decoding. The Bug: In SlidingAttentionCacheAllocator, the starting index for reading from the rolling buffer was pointing to stale tokens from the beginning of the buffer (instead of skipping them) once generati" }, { "n": 47809, "d": "2026-08-25", "t": "[CB] [Major] Upgrade the cache to support different attention types", "st": "DRAFT", "by": "remi-or", "a": 2693, "r": 1852, "f": 25, "c": 35, "b": "Before: full attention, sliding attention After: any attention type. Probably. Draft as not ready yet. But tests pass and perf look similar. A new cache type bundled in would be a nice addition, to prove this is actually usefull" } ], "d": "Adds the paged_attention kernel, vLLM-style prefix/block sharing, sliding-window & hybrid-model cache support, and CPU offloading of KV blocks/requests.", "note": "#47809 (major cache rework for mixed attention types) still open", "links": [], "src": "cb.json", "span": [ "2025-09-09", "2026-08-25" ], "roll": { "owners": [ [ "remi-or", 11 ], [ "SindhuRaghuram97", 1 ], [ "jiqing-feng", 1 ], [ "YangKai0616", 1 ] ], "n": 15, "add": 6354, "del": 3230, "files": 109, "commits": 242 } }, { "t": "CB performance: CUDA graphs, torch.compile & benchmarking", "tag": "cuda graphs + compile + benchmark_v2 suite", "s": "done", "prs": [ { "n": 41421, "d": "2025-10-13", "t": "Restore cuda graphs to continuous batching", "st": "MERGED", "by": "remi-or", "a": 380, "r": 233, "f": 8, "c": 18, "b": "This PR restores cuda graphs in continuous batching. The main changes associated with this are: the logic of how to generate tokens have been moved to the CB processor, which also handles the cuda graphs the generation step automatically slices the tensors to remove all padding unless cuda graphs ar" }, { "n": 41408, "d": "2025-10-14", "t": "Benchmark overhaul", "st": "MERGED", "by": "remi-or", "a": 1007, "r": 1864, "f": 13, "c": 25, "b": "This PR overhauls the benchmarking suite that is included in transformers. The benchmarking suite is now based around three main components: BenchmarkingConfig is a dataclass-like object which contains everything needed to reproduce a benchmark on the same machine: input length, generation length, w" }, { "n": 41662, "d": "2025-10-16", "t": "Small changes to benchmarking script", "st": "MERGED", "by": "remi-or", "a": 88, "r": 77, "f": 4, "c": 2, "b": "This PR: adds throughput to the final pretty print at the end of the benchmark runs adds the information of the output shape to the decoded text of the output (usefull to double check if throughput is right) changes the behavior of the run_benchmarks.py script so that it will only run 3 configs unle" }, { "n": 41848, "d": "2025-11-03", "t": "More data in benchmarking", "st": "MERGED", "by": "remi-or", "a": 30, "r": 31, "f": 3, "c": 4, "b": "This PR changes the benchmarking workflow to gather more data per benchmark. cc. @McPatate I only changed one workflow file, not sure I did not miss others. It also adds a check on the benchmark config to disable FA when it is not installed to avoid an exception." }, { "n": 42008, "d": "2025-11-04", "t": "Reduce the number of benchmark in the CI", "st": "MERGED", "by": "remi-or", "a": 88, "r": 120, "f": 3, "c": 1, "b": "In recent changes, we increased the number of benchmarks being run in the benchmarking CI. This has led to a relatively long CI runtime ~1hr on A10 runners. This might be wasteful. This PR changes how the run_benchmark.py generates the benchmark configs, favoring a benchmarking \"level\" system instea" }, { "n": 41916, "d": "2025-11-07", "t": "feat(ci): add continuous batching to benchmarks", "st": "MERGED", "by": "McPatate", "a": 190, "r": 107, "f": 10, "c": 15, "b": "" }, { "n": 42357, "d": "2025-11-24", "t": "Make benchmarking lighter: clean-up result files and remove non-needed arguments", "st": "MERGED", "by": "remi-or", "a": 97, "r": 122, "f": 5, "c": 9, "b": "This PR simplifies the benchmarking suite by: removing the timestamps for each token generation from the result file: we now only keep the time to first token and the (mean) inter-token latency removes the SDPA backend argument which is both un-used in practice and introduces a lot of edge cases sta" }, { "n": 42339, "d": "2025-11-24", "t": "fix(benchmarks): correct sdpa_backend inconsistency and attn_implementation for continuous batching", "st": "MERGED", "by": "engmohamedsalah", "a": 2, "r": 2, "f": 1, "c": 1, "b": "Summary This PR fixes two bugs in BenchmarkConfig reported in issue #42211: sdpa_backend inconsistency (line 105): The warning message states \"sdpa_backend must be None\" but the code was setting it to \"math\". Changed to None to match the warning message. Invalid attn_implementation (line 243): Chang" }, { "n": 42408, "d": "2025-11-27", "t": "Benchmark simplification", "st": "MERGED", "by": "remi-or", "a": 178, "r": 157, "f": 6, "c": 12, "b": "This PR does two things: in the benchmark_v2 suite, merge the two time_generate function (with and without CB) into one rename the the attributes of RequestState so they make more sense This will make the code more readable and concise." }, { "n": 42516, "d": "2025-12-03", "t": "Add torch compile to CB", "st": "MERGED", "by": "remi-or", "a": 309, "r": 188, "f": 6, "c": 11, "b": "This PR adds the option to wrap the GPU-side of continuous batching with torch.compile It also enable the benchmarking of CB with a compile config, and simplifies how benchmarking handles everything related to compile by using a transformers CompileConfig object" }, { "n": 43521, "d": "2026-01-28", "t": "[CB] Minor perf improvements and ty compatibility", "st": "MERGED", "by": "remi-or", "a": 228, "r": 169, "f": 8, "c": 21, "b": "Summary This PR introduces minor changes to the continuous batching feature: Performance optimize the get_seqlens_k function to avoid looping over all cache managers logits indexing is now done inside the cuda graph, as there was no apparent reason it should be done outside fuse the get_read/write_i" }, { "n": 44578, "d": "2026-03-19", "t": "[CB] Better parametrization for compile", "st": "MERGED", "by": "remi-or", "a": 449, "r": 121, "f": 11, "c": 27, "b": "Summary This PR adds three attributes to the compile config, to have granularity over how varlen (handles mixed prefil and decode batches) and decode (only decode batches) are compiled. We want to have this kind of granularity because varlen has 2 dynamic axis (number of query tokens and KV tokens) " }, { "n": 45911, "d": "2026-05-13", "t": "[CB] Hide activation footprint by using the CUDA graph pool", "st": "MERGED", "by": "remi-or", "a": 33, "r": 5, "f": 2, "c": 3, "b": "After this PR, activations created when warming up for a new CUDA graph catpure will be allocated inside the existing CUDA graph pool, as long as torch version is >= 2.4 . This is because the Mempool object of pytorch was inroduced in 2.5 . This ensures there is no OOM when warming for a new CUDA gr" }, { "n": 46532, "d": "2026-06-10", "t": "(Draft) [CB] Change padding from interval to pow2", "st": "CLOSED", "by": "remi-or", "a": 131, "r": 117, "f": 11, "c": 11, "b": "Draft until we have a full picture of the performance improvements." }, { "n": 46490, "d": "2026-06-11", "t": "[CB] [Minor] Add data-parallel to overall script", "st": "MERGED", "by": "remi-or", "a": 82, "r": 39, "f": 3, "c": 8, "b": "This PR adds a data-parallel mode to the continuous_batching_overall.py script that we use to benchmark the performance of CB across a range of use cases. Until now, the script did not allow to leverage all the GPUs on a node because it only supported TP: if a user wanted to run 16 benchmarks on a 8" }, { "n": 46533, "d": "2026-06-12", "t": "[CB] [Minor] Add parameter to tune default compile level", "st": "MERGED", "by": "remi-or", "a": 47, "r": 23, "f": 5, "c": 5, "b": "Summary Currently, continuous batching exposes the parameter use_default_compile_configs to the users so that they may enable torch.compile easily. But that parameter enables compile with max-autotune-no-cudagraphs and dynamic=False for best performances. This makes for a long warmup, which is not g" } ], "d": "Restores CUDA graph capture for CB, adds torch.compile support with tunable compile levels, and builds/iterates the benchmark_v2 suite used to measure CB throughput.", "note": null, "links": [], "src": "cb.json", "span": [ "2025-10-13", "2026-06-12" ], "roll": { "owners": [ [ "remi-or", 14 ], [ "engmohamedsalah", 1 ], [ "McPatate", 1 ] ], "n": 16, "add": 3339, "del": 3375, "files": 99, "commits": 173 } }, { "t": "CB hardware backend support (XPU / MPS / CPU)", "tag": "cross-platform paged decode", "s": "wip", "prs": [ { "n": 41830, "d": "2025-11-10", "t": "fix continuous batching issues, extend ut cases to xpu", "st": "MERGED", "by": "yao-matrix", "a": 61, "r": 21, "f": 3, "c": 15, "b": "@SunMarc , pls help review, thx very much." }, { "n": 46028, "d": "2026-05-27", "t": "Support FA2 flash_attn_with_kvcache for XPU continuous batching", "st": "MERGED", "by": "YangKai0616", "a": 15, "r": 10, "f": 1, "c": 6, "b": "XPU now supports flash_attn_with_kvcache in kernels-community/flash-attn2. This PR enables the relevant path in CB. transformers pipeline test results: | label | samples | avg_in | max_new | time (s) | tokens | tok/s | mem (GB) | |------------------------|-----------|----------|-----------|---------" }, { "n": 46792, "d": "2026-06-26", "t": "Enable CB paged decode fast path for MPS hub kernels + sensible block_size default", "st": "OPEN", "by": "ArthurZucker", "a": 52, "r": 7, "f": 4, "c": 3, "b": "Summary Three changes to make continuous batching (CB) work well on Apple Silicon (MPS) with the metal-flash-sdpa kernel: 1. Enable paged decode fast path for hub kernels (initialization.py) ensure_decode_fast_path_is_available previously only allowed the paged decode fast path (block tables + flash" }, { "n": 47159, "d": "2026-07-08", "t": "Extend continuous batching memory prediction test to XPU", "st": "MERGED", "by": "sywangyi", "a": 17, "r": 16, "f": 1, "c": 2, "b": "Intel XPU: @IlyasMoutawwakil pytest -q tests/generation/test_continuous_batching.py::TestMemoryHandlerPrediction 5 passed" }, { "n": 47185, "d": "2026-07-15", "t": "Resolve continuous batching XPU availability checks at runtime", "st": "MERGED", "by": "kaixuanliu", "a": 37, "r": 30, "f": 1, "c": 5, "b": "This PR updates continuous batching XPU availability checks to resolve through transformers.utils at runtime instead of importing is_torch_xpu_available directly into continuous batching modules. This makes the behavior consistent with other runtime accelerator checks such as torch.cuda.is_available" }, { "n": 47163, "d": "2026-07-23", "t": "add paged attention tests support for XPU", "st": "MERGED", "by": "kaixuanliu", "a": 16, "r": 2, "f": 1, "c": 2, "b": "" }, { "n": 48024, "d": "2026-08-28", "t": "[Continuous Batching] Gate FlashAttention kernel fallbacks on CUDA capability >= 8.0", "st": "OPEN", "by": "Ultron09", "a": 29, "r": 0, "f": 2, "c": 1, "b": "Fixes #47926 CC: @albertvillanova @remi-or @IlyasMoutawwakil Problem In continuous batching, switch_to_cb_friendly_attn automatically switches attention from sdpa to flash_attention_3 (or flash_attention_2) by checking is_flash_attn_3_available(kernels_fallback_ok=True). Because is_flash_attn_3_avai" }, { "n": 48365, "d": "2026-08-31", "t": "This PR generalizes continuous batching graph support from CUDA-only …", "st": "OPEN", "by": "sywangyi", "a": 177, "r": 68, "f": 8, "c": 8, "b": "…graphs to accelerator graphs, with support for both CUDA and XPU graph capture. The main changes are: Add use_accelerator_graph to ContinuousBatchingConfig while keeping use_cuda_graph and cuda_graph_booleans as backward-compatible aliases. Rename the continuous batching graph decision path from CU" } ], "d": "Extends CB's paged-decode fast path and test coverage beyond CUDA: Intel XPU flash-attn/availability checks, Apple MPS decode path, generalizing CUDA-graph support off CUDA-only.", "note": null, "links": [], "src": "cb.json", "span": [ "2025-10-23", "2026-08-31" ], "roll": { "owners": [ [ "sywangyi", 2 ], [ "kaixuanliu", 2 ], [ "yao-matrix", 1 ], [ "YangKai0616", 1 ] ], "n": 8, "add": 404, "del": 154, "files": 21, "commits": 42 } }, { "t": "CB multimodal support", "tag": "multimodal continuous batching (wip)", "s": "wip", "prs": [ { "n": 44436, "d": "2026-03-09", "t": "Fix continuous batching for multimodal models", "st": "MERGED", "by": "jw9603", "a": 8, "r": 0, "f": 1, "c": 5, "b": "Fixes #44423 continuous_batching_chat_completion was missing input preprocessing and tokenize=True in apply_chat_template, causing 'str' object has no attribute 'to' for multimodal models. Added the same get_model_modality + get_processor_inputs_from_inbound_messages preprocessing already used in ge" }, { "n": 46376, "d": "2026-07-16", "t": "[CB] [Major] Add multimodal support to continous batching", "st": "OPEN", "by": "remi-or", "a": 1100, "r": 184, "f": 10, "c": 14, "b": "This PR adds support for image and text models to continuous batching. Tests run, will add image and audio perf numbers later. This is more a beta versions, we are missing a bunch of features: CUDA graphs for encoder, multiple modalities at once, etc. But better start from here and build little by l" } ], "d": "Flagship #46376 (open) adds multimodal model support to continuous batching; #44436 fixed a chat-completion preprocessing crash for multimodal CB requests.", "note": null, "links": [], "src": "cb.json", "span": [ "2026-03-04", "2026-07-16" ], "roll": { "owners": [ [ "remi-or", 1 ], [ "jw9603", 1 ] ], "n": 2, "add": 1108, "del": 184, "files": 11, "commits": 19 } }, { "t": "Tensor-parallel / distributed CB", "tag": "multi-GPU CB via tensor parallelism", "s": "done", "prs": [ { "n": 45821, "d": "2026-05-18", "t": "[CB] [Major] Add tensor paralellism", "st": "MERGED", "by": "remi-or", "a": 807, "r": 183, "f": 9, "c": 30, "b": "This PR adds support for TP in continuous batching. The major changes required to do this were: add inter-process communication for the requests states add per TP group seeding add hints to prevent NCCL graph mixing change hash function to avoid python hash which is salted depending on the process I" }, { "n": 46205, "d": "2026-05-29", "t": "Guard DeviceMesh import in continuous batching", "st": "MERGED", "by": "danyalahmed1995", "a": 13, "r": 3, "f": 1, "c": 3, "b": "Closes #46170. What Does This PR Do ? Basically this if torch.distributed.is_available(): from torch.distributed.device_mesh import DeviceMesh else: DeviceMesh = object Self explanatory I suppose. Tested: local import check for torch.distributed.device_mesh PYTHONPATH=src python -c \"from transformer" }, { "n": 46070, "d": "2026-06-02", "t": "[CB] [Major] Rework manager to have clearer control flow + handle TP", "st": "MERGED", "by": "remi-or", "a": 463, "r": 286, "f": 6, "c": 7, "b": "Summary This PR reworks the continuous batching manager in order to make its control flow clearer and separate the background and main thread. Notably: the methods inside the ContinuousBatchingManager class are now clearly categrorized the status of the background thread now lives in a dedicated obj" } ], "d": "Adds tensor-parallel support to continuous batching and reworks the batching manager for clearer TP-aware control flow.", "note": null, "links": [], "src": "cb.json", "span": [ "2026-05-14", "2026-06-02" ], "roll": { "owners": [ [ "remi-or", 2 ], [ "danyalahmed1995", 1 ] ], "n": 3, "add": 1283, "del": 472, "files": 16, "commits": 40 } }, { "t": "CB + training integration (colocated weight sharing)", "tag": "RL-style train+generate on one weight copy", "s": "wip", "prs": [ { "n": 48301, "d": "2026-08-25", "t": "Fail fast when the continuous batching background thread has died", "st": "CLOSED", "by": "qgallouedec", "a": 27, "r": 6, "f": 1, "c": 3, "b": "Fixes the secondary problem reported in #48298 (silent timeouts after the background thread dies). Part of #48294. Change: add_request raises immediately when fatal_error is set, and get_result raises once the remaining results are drained, both chaining from self.fatal_error so the root cause is in" }, { "n": 48299, "d": "2026-08-26", "t": "Resolve continuous batching config against the text config for composite models", "st": "MERGED", "by": "qgallouedec", "a": 3, "r": 3, "f": 1, "c": 3, "b": "Fixes #48298. Part of #48294. Change: the three call sites in ContinuousBatchingManager that read decoder attributes (resolve_continuous_batching_config, PagedAttentionCache, ContinuousBatchProcessor) pass self.model.config.get_text_config() instead of self.model.config. get_text_config() returns se" }, { "n": 48302, "d": "2026-08-26", "t": "Create the continuous batching CPU group with local synchronization", "st": "MERGED", "by": "qgallouedec", "a": 9, "r": 2, "f": 1, "c": 4, "b": "Part of #48294. DistributedHelper creates the TP CPU (gloo) group with a plain dist.new_group, which requires every rank of the default group to participate in the call. That holds when the TP group spans the whole world (the current CB TP setup), but not when the TP group is a subset: with one CB m" }, { "n": 48304, "d": "2026-08-27", "t": "Wait for the first request in the async continuous batching bootstrap", "st": "MERGED", "by": "qgallouedec", "a": 33, "r": 27, "f": 1, "c": 9, "b": "Fixes #48303. Part of #48294. Change: the async-batching bootstrap waits for the first request with the same stop conditions as the main loop (hard stop, or flush with nothing pending), instead of raising Failed to bootstrap the first batch when the queue is still empty at background-thread start. V" }, { "n": 48314, "d": "2026-08-27", "t": "Quiet continuous batching at default verbosity", "st": "MERGED", "by": "qgallouedec", "a": 6, "r": 6, "f": 1, "c": 4, "b": "Fixes #48313. Part of #48294. Changes: The status messages (attention auto-switch, warmup start/finish, progress-bar note) move from warning to info. The Solving n requests progress bar becomes opt-in (progress_bar=False by default): like generate, generate_batch prints nothing at default verbosity." }, { "n": 48382, "d": "2026-08-27", "t": "Let a thread that shares the model take a turn on it under continuous…", "st": "DRAFT", "by": "qgallouedec", "a": 57, "r": 8, "f": 2, "c": 2, "b": "Adds ContinuousBatchingManager.training_turn(), so a thread that shares the model with the generation loop can use it without deadlocking under TP. The engine decodes in a background thread. A thread that keeps using the same model, typically a trainer updating it in place, issues its own collective" }, { "n": 48297, "d": "2026-08-28", "t": "Raise when a paged attention forward is called with no cache", "st": "MERGED", "by": "qgallouedec", "a": 46, "r": 20, "f": 3, "c": 6, "b": "Fixes #48296. paged|sdpa and paged|eager treat a missing cache as \"no KV cache to update\" and carry on. They are written for the packed inputs and the 4D mask that continuous batching prepares, so on a standard forward they attend bidirectionally: no error, no warning, loss 1.99 vs 3.09 on Qwen3-0.6" }, { "n": 48383, "d": "2026-08-28", "t": "Give each tensor parallel model in a process its own device mesh", "st": "OPEN", "by": "qgallouedec", "a": 31, "r": 3, "f": 2, "c": 3, "b": "Two tensor parallel models in the same process silently share a communicator, which deadlocks as soon as they are used at the same time. initialize_tensor_parallelism built its mesh with init_device_mesh(device_type, (tp_size,)). That mesh has no dimension names, and when tp_size == world_size torch" }, { "n": 48294, "d": "2026-08-31", "t": "[Umbrella] Continuous batching fixes for training-while-generating on one weight copy", "st": "DRAFT", "by": "qgallouedec", "a": 1242, "r": 147, "f": 19, "c": 67, "b": "Draft / umbrella PR: this branch is the sum of six independent fixes, each also opened as its own PR (the mergeable units). It exists so the whole set can be run and reviewed together. Context We explored running a training loop and continuous batching generation simultaneously on ONE copy of the we" } ], "d": "New umbrella effort (opened 2026-08-25) letting a training loop and CB share one model weight copy/process group: thread turn-taking, per-model device mesh, CPU sync group.", "note": "Brand new; umbrella PR and 2 sub-PRs still open", "links": [], "src": "cb.json", "span": [ "2026-08-25", "2026-08-31" ], "roll": { "owners": [ [ "qgallouedec", 9 ] ], "n": 9, "add": 1454, "del": 222, "files": 31, "commits": 101 } } ] }, { "name": "vLLM & SGLang", "hue": 265, "blurb": "Be the model definition that inference engines read from.", "items": [ { "t": "Transformers backend in vLLM/SGLang", "tag": "backend runs natively inside vLLM/SGLang", "s": "done", "prs": [ { "n": 36091, "d": "2025-02-12", "t": "Add pipeline parallel plan to `PretrainedConfig` and `PreTrainedModel`", "st": "MERGED", "by": "hmellor", "a": 188, "r": 0, "f": 50, "c": 25, "b": "Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. Di" }, { "n": 36889, "d": "2025-03-26", "t": "Allow easy registration of custom attention functions", "st": "MERGED", "by": "Cyrilvallez", "a": 171, "r": 11, "f": 6, "c": 20, "b": "Basic design idea: from transformers import AttentionInterface def my_new_attention(): ... AttentionInterface.register(\"my_new_attention\", my_new_attention)" }, { "n": 36934, "d": "2025-05-02", "t": "Aligning modling code for GPT2 to work with vLLM (fallback)", "st": "MERGED", "by": "ariG23498", "a": 13, "r": 0, "f": 3, "c": 23, "b": "This PR changes the modeling code for GPT2 to support working on vLLM using the transformers fallback backend. The changes are as follows: Introduction of kwargs which is used to propagate information about attention_indices in vLLM Adding a base_model_tp_plan, which is currently empty. This is a re" }, { "n": 37866, "d": "2025-05-22", "t": "🚨🚨[core] Completely rewrite the masking logic for all attentions", "st": "MERGED", "by": "Cyrilvallez", "a": 2984, "r": 6808, "f": 129, "c": 193, "b": "As per the title. The goal is to properly separate masking logic from modeling code itself, to continue our objective of simplifying the library. Code is much simpler to understand Much more general: always work for all lengths, and all attention implementations, e.g.: flex attention now works with " }, { "n": 44696, "d": "2026-03-18", "t": "Fix `supports_{tp/pp}_plan`", "st": "MERGED", "by": "hmellor", "a": 19, "r": 13, "f": 1, "c": 4, "b": "In configs, base_model_pp_plan and base_model_tp_plan default to None In models, _pp_plan and _tp_plan look like they default to None based on the class variables, but will actually always be a dict because of post_init. This means that supports_pp_plan and supports_tp_plan always return True. This " }, { "n": 44699, "d": "2026-03-20", "t": "Fix several based models' pipeline parallel support", "st": "MERGED", "by": "hmellor", "a": 219, "r": 148, "f": 52, "c": 6, "b": "These models have base_model_pp_plans but currently do not work because the base model's forward pass depends on all the layers being Qwen2VLDecoderLayer. i.e. if one of the layers is removed/replaced with Identity, decoder_layer.attention_type will fail with an attribute error. This PR fixes the is" }, { "n": 47198, "d": "2026-07-14", "t": ":rotating_light: Fix GPTBigCode and GPTNeoX for the Transformers modelling backend for vLLM", "st": "MERGED", "by": "hmellor", "a": 18, "r": 14, "f": 4, "c": 9, "b": "Adds a remapping from embed_out to lm_head for GPTNeoX Adds _supports_attention_backend = True to GPTBigCode because it does" }, { "n": 47547, "d": "2026-07-29", "t": "Fix CUDA Graph breaking host to device copy from scalar tensor allocation", "st": "MERGED", "by": "hmellor", "a": 249, "r": 243, "f": 130, "c": 4, "b": "Fixed a CUDA-graph-capture crash in several models by replacing torch.tensor(self.config.<x>_token_id, device=inputs_embeds.device) (a host→device copy that graph capture forbids) with torch.full((), self.config.<x>_token_id, dtype=torch.long, device=inputs_embeds.device), which fills the scalar dir" } ], "d": "Backend support letting vLLM/SGLang run transformers models directly: pipeline-parallel plan, custom-attention registration, masking rewrite, per-model/CUDA-graph fixes (GPT2, GPTBigCode, GPTNeoX).", "note": null, "links": [], "src": "vllm.json", "span": [ "2025-02-07", "2026-07-29" ], "roll": { "owners": [ [ "hmellor", 5 ], [ "Cyrilvallez", 2 ], [ "ariG23498", 1 ] ], "n": 8, "add": 3861, "del": 7237, "files": 375, "commits": 284 } }, { "t": "Processor & config standardization for vLLM", "tag": "stable processor & config API for vLLM", "s": "done", "prs": [ { "n": 37915, "d": "2025-05-27", "t": "[transformers x vLLM] standardize processors", "st": "MERGED", "by": "zucchini-nlp", "a": 1167, "r": 79, "f": 28, "c": 19, "b": "Part of #37780. The design was tested on different model types: Oke, I verified that the inference works for all models, unless I forgot about some new ones. Here is the list I tested. A few models (blip2, gotOcr, gemma3) won't be supported in the first release. Gemma3 is already planned after we me" }, { "n": 39583, "d": "2025-07-24", "t": "Update recent processors for vLLM backend", "st": "MERGED", "by": "zucchini-nlp", "a": 79, "r": 9, "f": 6, "c": 2, "b": "As peer title, models needed tiny changes here and there" }, { "n": 44255, "d": "2026-03-04", "t": "[vllm + v5 fix] handle TokenizersBackend fallback properly for v5", "st": "MERGED", "by": "itazap", "a": 404, "r": 205, "f": 24, "c": 48, "b": "What this PR does Given the different issues that were noticed by @hmellor on vLLM, we wanted to make sure we did not end up with crazy breaks. We ran a full test suite (code can be found in #44298) and the results showed 22 model converters that had potential issues. We test what would be the AutoT" }, { "n": 44447, "d": "2026-03-05", "t": "[vLLM] Fix backward compatibility with hardcoded subprocessors classes in processors", "st": "MERGED", "by": "yonigozlan", "a": 8, "r": 0, "f": 1, "c": 3, "b": "Fixes MiniCPM-o-2_6 related tests failures in vLLM, and improve backward compatibility with remote code in general. Cc @hmellor @zucchini-nlp" }, { "n": 45058, "d": "2026-03-27", "t": "Allow advanced users to override `model_type` in `AutoConfig.from_pretrained`", "st": "MERGED", "by": "hmellor", "a": 9, "r": 0, "f": 1, "c": 1, "b": "Some checkpoints, such as https://huggingface.co/omni-research/Tarsier2-Recap-7b, have the wrong model_type in their config.json. This PR allows advanced users (vLLM) to pass model_type into AutoConfig.from_pretrained via kwargs to override what's in the config.json." }, { "n": 45326, "d": "2026-04-21", "t": "feat[vLLM × v5]: Add vLLM compatibility for audio models", "st": "MERGED", "by": "harshaljanjani", "a": 19, "r": 0, "f": 11, "c": 4, "b": "→ This PR introduces compat fixes across several audio models to ensure they can be loaded and used by a companion vLLM PR. These changes are deliberate and are blocking this vLLM PR which adds audio backend compatibility to vLLM. Once this PR is merged, the other PR will be marked ready for review!" }, { "n": 46258, "d": "2026-05-29", "t": "Processor compatibility with vLLM ", "st": "MERGED", "by": "zucchini-nlp", "a": 6, "r": 3, "f": 1, "c": 1, "b": "As per title, recent refactor broken vLLM. Can be checked with vllm serve Qwen/Qwen3-VL-8B-Instruct --gpu-memory-utilization 0.7 --max-model-len 8192. It fails on main and passed on PR branch Plan is to update all models first, and check what can we do with vllm after that. Personally, I want us to " }, { "n": 46472, "d": "2026-06-17", "t": "feat[vLLM x v5]: Expose max_source_positions on VibeVoiceAsrConfig", "st": "MERGED", "by": "harshaljanjani", "a": 37, "r": 0, "f": 3, "c": 9, "b": "→ Fixes vllm-project/vllm#39330 (comment) → Exposes max_source_positions on VibeVoiceAsrConfig so profiling can resolve the model's max audio token budget. → Companion to the vLLM Transformers audio backend PR. vLLM's get_max_audio_tokens looks for one of max_source_positions, max_position_embedding" }, { "n": 47384, "d": "2026-07-23", "t": "Use new `per_layer_config` for Gemma 4 so that heterogeneous attention config is explicit", "st": "MERGED", "by": "hmellor", "a": 214, "r": 191, "f": 21, "c": 17, "b": "Gemma 4 contains a different head_dim and num_key_value_heads for the full attention and sliding attention layers. Currently, which layers use which is implicit during model construction. This PR: leverages the heterogeneous config support added in #45333 to represent this at the config level delete" }, { "n": 47483, "d": "2026-07-23", "t": "feat[vLLM x v5]: Make audio optional and support multiple audios in VibeVoice ASR processor", "st": "MERGED", "by": "harshaljanjani", "a": 118, "r": 24, "f": 2, "c": 1, "b": "→ Unblocks the vLLM Transformers audio backend PR by making VibeVoiceAsrProcessor compatible with test_processing_correctness. → Makes audio optional (audio: AudioInput | None = None) and relaxes the processor from a strict 1:1 text-audio mapping to support multiple audios per prompt, matching other" }, { "n": 47614, "d": "2026-08-03", "t": "feat[vLLM]: Support text replacement offsets in the remaining old-format processors", "st": "MERGED", "by": "harshaljanjani", "a": 169, "r": 170, "f": 4, "c": 5, "b": "→ Change alluding to the discussion in vllm-project/vllm#50107 currently blocking the vLLM PR. → vllm-project/vllm#50107 (comment) → vllm-project/vllm#50107 (comment) cc: @zucchini-nlp @eustlb Code Agent Policy I confirm that this is not a pure code agent PR. Before submitting This PR fixes a typo o" } ], "d": "Standardized processors, tokenizer backends, and per-layer/model_type config so vLLM/SGLang's backend can call them directly, including v5-era audio-processor and subprocessor back-compat fixes.", "note": null, "links": [], "src": "vllm.json", "span": [ "2025-05-01", "2026-08-03" ], "roll": { "owners": [ [ "harshaljanjani", 4 ], [ "zucchini-nlp", 3 ], [ "hmellor", 2 ], [ "yonigozlan", 1 ] ], "n": 11, "add": 2230, "del": 681, "files": 102, "commits": 110 } }, { "t": "Weight-conversion API (conversion_mapping)", "tag": "vLLM reuses HF's weight-conversion API", "s": "done", "prs": [ { "n": 41580, "d": "2025-11-13", "t": "Refactor weight loading", "st": "MERGED", "by": "ArthurZucker", "a": 8733, "r": 7189, "f": 586, "c": 387, "b": "CORE REFACTORING, loading, converting, logging More helpful debugging report when loading weights If you just want to fuse qkv: It can. You just need to make sure you change the model code and pouf! WeightConverter( [\"self_attn.q_proj\", \"self_attn.k_proj\", \"self_attn.v_proj\"], \"self_attn.qkv_proj\", " }, { "n": 43071, "d": "2026-01-29", "t": "Separate renaming from merging in Mixtral conversion mapping", "st": "CLOSED", "by": "hmellor", "a": 4, "r": 4, "f": 1, "c": 3, "b": "This allows the Transformers modeling backend for vLLM to load Mixtral with Transformers v5 because it currently (once vllm-project/vllm#31545 is merged) only supports WeightRenaming but not WeightConverter" }, { "n": 43621, "d": "2026-01-30", "t": "[vllm compat] Separate renaming from conversion ops", "st": "MERGED", "by": "Cyrilvallez", "a": 13, "r": 12, "f": 1, "c": 6, "b": "As per the title. Following specs described by @hmellor for vllm compatibility, no logic change should be required. Only the following small changes to mappings should be enough" }, { "n": 47460, "d": "2026-07-23", "t": "Isolate MLA KV expansion to make it easier to bypass", "st": "MERGED", "by": "hmellor", "a": 218, "r": 116, "f": 16, "c": 10, "b": "Dedicated MLA kernels often consume the compressed kv latents directly. While there are no such kernels in Transformers today, vLLM does include them. Therefore, in order for the Transformers backend for vLLM to leverage these kernels, we need a way to get the compressed kv latents directly into the" } ], "d": "Created conversion_mapping.py (get_model_conversion_mapping), the weight-loading API vLLM's backend imports, then split renaming from tensor-merging so MoE models like Mixtral load under v5.", "note": "hmellor's #43071 (closed) first proposed splitting renaming from merging for Mixtral; Cyrilvallez's #43621 generalized it and merged.", "links": [], "src": "vllm.json", "span": [ "2025-11-03", "2026-07-23" ], "roll": { "owners": [ [ "hmellor", 2 ], [ "ArthurZucker", 1 ], [ "Cyrilvallez", 1 ] ], "n": 4, "add": 8968, "del": 7321, "files": 604, "commits": 406 } }, { "t": "RoPE config API standardization for v5", "tag": "stable rope config for vLLM's parser", "s": "done", "prs": [ { "n": 42019, "d": "2025-11-05", "t": "Add backward compatibility helpers to ease v5 RoPE config transition", "st": "CLOSED", "by": "hmellor", "a": 43, "r": 0, "f": 1, "c": 2, "b": "Overrides the following 3 methods if rope_parameters is found to be nested based on layer_types: __getitem__ - Checks all subdicts and returns the value as long as it is the same in all subdicts __contains__ - Checks that all subdicts contain the key and return True if they all do __setitem__ - Sets" }, { "n": 42390, "d": "2025-11-25", "t": "Add backward compatibility for `rope_theta` in the same way as `rope_scaling`", "st": "CLOSED", "by": "hmellor", "a": 8, "r": 0, "f": 1, "c": 1, "b": "Custom models may still be using rope_scaling and rope_theta. If they are, we should also add a property for rope_theta for full backward compatibility." }, { "n": 42421, "d": "2025-11-26", "t": "Remove `rope_theta` attribute map for `ModernBert`", "st": "CLOSED", "by": "hmellor", "a": 0, "r": 1, "f": 1, "c": 1, "b": "This attribute being present via attribute_map causes problems for vLLM's backward compatibility. We have the following code snippet for backward compatibility with custom models which were written for Transformers v4 https://github.com/vllm-project/vllm/blob/d9d342d214b8c13f71215318a6d9252cc4a5ca47" }, { "n": 42517, "d": "2025-12-02", "t": "Add backward compatibility for methods which have been moved to `RotaryEmbeddingConfigMixin`", "st": "MERGED", "by": "hmellor", "a": 19, "r": 1, "f": 1, "c": 5, "b": "This function was moved in #42255. Many custom code models import this function, which is arguably good practise as it shows they are leveraging our first party utilities. This PR reinstates the function by calling the methods which now live in RotaryEmbeddingConfigMixin so that these custom code mo" } ], "d": "Backward-compatibility shims (rope_theta property, RotaryEmbeddingConfigMixin) keeping vLLM's own config parser working through the rope_scaling to rope_parameters (v5) migration.", "note": "#42019/#42390/#42421 (hmellor) were closed unmerged; the same vLLM back-compat need was met via #42517 and related merged RoPE PRs.", "links": [], "src": "vllm.json", "span": [ "2025-11-04", "2025-12-02" ], "roll": { "owners": [ [ "hmellor", 4 ] ], "n": 4, "add": 70, "del": 2, "files": 4, "commits": 9 } }, { "t": "Output-capturing API for MoE routing", "tag": "stable hook to capture MoE routing", "s": "done", "prs": [ { "n": 43765, "d": "2026-02-06", "t": "[core] Faster and thread-safe `check_model_inputs` implementation", "st": "MERGED", "by": "Cyrilvallez", "a": 728, "r": 565, "f": 121, "c": 17, "b": "The problem Currently, check_model_inputs needs to iterate on all modules and monkey-patch all needed submodule's forward on-the-fly, before restoring them afterwards. This brings 2 big issues: It's NOT thread-safe (see also issue #42673): due to the monkey patching, 2 threads running the model's fo" }, { "n": 44925, "d": "2026-04-14", "t": "[Doc] MoE routing capture and replay recipe ", "st": "MERGED", "by": "kashif", "a": 96, "r": 0, "f": 1, "c": 9, "b": "This PR adds a first-class MoE routing capture/replay API for Qwen2Moe and introduces shared MoE routing helpers for reuse by other MoE model families. It adds: a structured MoERouting payload in modeling outputs output_moe_routing=True to return exact selected experts from the real forward path moe" } ], "d": "A faster, thread-safe check_model_inputs implementation added output_capturing.py, the module vLLM imports; documented via a MoE router-decision capture-and-replay recipe.", "note": null, "links": [], "src": "vllm.json", "span": [ "2026-02-06", "2026-04-14" ], "roll": { "owners": [ [ "Cyrilvallez", 1 ], [ "kashif", 1 ] ], "n": 2, "add": 824, "del": 565, "files": 122, "commits": 26 } }, { "t": "vLLM CI infrastructure", "tag": "catches vLLM breakage before merge", "s": "done", "prs": [ { "n": 45139, "d": "2026-04-08", "t": "Fix vllm cis", "st": "MERGED", "by": "ArthurZucker", "a": 757, "r": 976, "f": 99, "c": 29, "b": "More fixes" }, { "n": 46383, "d": "2026-06-08", "t": "Add vLLM smoke test to CI", "st": "MERGED", "by": "hmellor", "a": 57, "r": 0, "f": 1, "c": 2, "b": "vLLM used to run forward compatibility tests with Transformers main but the signal was in the wrong place so it was deleted in vllm-project/vllm#43785. This PR adds this same signal back on the Transformers side so that the Transformers team can see how their changes affect one of the biggest depend" }, { "n": 47864, "d": "2026-08-10", "t": "Update vLLM CI workflow", "st": "MERGED", "by": "ydshieh", "a": 232, "r": 57, "f": 2, "c": 51, "b": "Add a nightly CI workflow (.github/workflows/vllm-ci-caller.yml) that tests transformers compatibility with vLLM nightly CPU builds. What it does Installs the latest vLLM nightly CPU wheel + transformers from source Runs vLLM's multimodal processing test suite across 4 parallel shards Runs test_mapp" }, { "n": 47934, "d": "2026-08-12", "t": "CI: add vllm-test-init and vllm-test-transformers jobs on dedicated runners", "st": "MERGED", "by": "ydshieh", "a": 297, "r": 16, "f": 1, "c": 31, "b": "Summary Restructures the vLLM CI workflow to fix persistent OOM failures and adds dedicated matrix jobs for better isolation and parallelism. New jobs vllm-test-init: runs test_can_initialize_small_subset across 4 shards on aws-m8i-8xl-cache; mirrors vLLM's Buildkite CI. Gemma3nForCausalLM is desele" } ], "d": "Dedicated CI: a vLLM smoke test plus vllm-test-init/vllm-test-transformers jobs on dedicated runners, so PRs that break the vLLM transformers backend are caught before merge.", "note": "#47864/#47934 authored by ydshieh; hmellor contributed the smoke test (#46383) and a CI fix (#45139).", "links": [], "src": "vllm.json", "span": [ "2026-03-31", "2026-08-12" ], "roll": { "owners": [ [ "ydshieh", 2 ], [ "hmellor", 1 ], [ "ArthurZucker", 1 ] ], "n": 4, "add": 1343, "del": 1049, "files": 103, "commits": 113 } } ] }, { "name": "Standardization & modular", "hue": 34, "blurb": "Cement transformers as the source of truth, then let others read from it.", "items": [ { "t": "Modular converter tooling maturation", "tag": "reliable modular->modeling codegen", "s": "wip", "prs": [ { "n": 33737, "d": "2024-09-26", "t": "Fix modular model converter unable to generate Processor classes", "st": "MERGED", "by": "tonywu71", "a": 1, "r": 1, "f": 1, "c": 1, "b": "Modular 🤗 transformers is awesome, but a typo in the code prevents the user to generate a processing_xxx.py file from modular_xxx.py. This PR solves this problem. Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contr" }, { "n": 33991, "d": "2024-10-08", "t": "Improve modular converter", "st": "MERGED", "by": "Cyrilvallez", "a": 299, "r": 71, "f": 11, "c": 11, "b": "This improves the modular file converter. support for top-level functions newly defined in the modular_xxx.py code and used in a class inheriting from another model class without the function dependency (and recursively adding all other top-level functions that may be needed for that function) e.g. " }, { "n": 34051, "d": "2024-10-10", "t": "Small Fix to modular converter", "st": "MERGED", "by": "MekkCyber", "a": 2, "r": 1, "f": 1, "c": 3, "b": "This update addresses a small issue in the modular converter. The relative path used in AUTO_GENERATED_MESSAGE is generated using a regex that assumes all models reside in the src/transformers directory. This approach fails when handling examples located in the examples/ folder. The fix is simple : " }, { "n": 34266, "d": "2024-10-29", "t": "Fixes for Modular Converter on Windows", "st": "MERGED", "by": "hlky", "a": 5, "r": 5, "f": 1, "c": 5, "b": "This PR fixes some issues with Modular Converter on Windows. Relative path regex. rf\"(src{os.sep}transformers{os.sep}.*|examples{os.sep}.*)\" results in (src\\transformers\\.*|examples\\.*), due to \\t and \\. this regex fails, we replace \\ on os.path.abspath instead, and use (src/transformers/.*|examples" }, { "n": 35642, "d": "2025-01-21", "t": "modular_model_converter bugfix on assignments", "st": "MERGED", "by": "nikosanto13", "a": 113, "r": 98, "f": 12, "c": 3, "b": "This PR improves the logic of modular_model_converter.py script in order to keep the assignments from modular files. Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was th" }, { "n": 35795, "d": "2025-01-21", "t": "Deterministic sorting in modular converter when adding new functions", "st": "MERGED", "by": "Cyrilvallez", "a": 6, "r": 0, "f": 1, "c": 1, "b": "Fix non-deterministic sorting when importing from different file which appeared in #35692. This can cause the CI checker to error depending on the system the conversion was performed on. This fixes it." }, { "n": 44833, "d": "2026-03-18", "t": "Fix and re-run modular converter on examples", "st": "MERGED", "by": "Cyrilvallez", "a": 299, "r": 590, "f": 19, "c": 3, "b": "Fix the examples after the config change (#41250), and re-run conversion as in general modelings changed quite a bit in the lib." }, { "n": 45528, "d": "2026-04-20", "t": "qa: re-run modular converter when the script itself is modified", "st": "MERGED", "by": "tarekziade", "a": 103, "r": 6, "f": 2, "c": 1, "b": "When changing the checker itself, we want to ignore the guaranteed_no_diff shortcut since changing the checker might impact models not in the PR" }, { "n": 45046, "d": "2026-04-27", "t": "chore: speedup modular converter (~30%)", "st": "MERGED", "by": "tarekziade", "a": 62, "r": 26, "f": 1, "c": 7, "b": "First optimization Adding a module-source cache in utils/modular_model_converter.py. When the converter needs the same imported modeling file more than once, it now reuses the already-read source and parsed LibCST tree, invalidated by the file’s mtime, instead of reopening and reparsing the file. Se" } ], "d": "Ongoing hardening of modular_model_converter.py: Windows fixes, deterministic function ordering, Processor class generation, ~30% speedup, auto re-run when script/examples change.", "note": null, "links": [], "src": "standard.json", "span": [ "2024-09-26", "2026-04-27" ], "roll": { "owners": [ [ "Cyrilvallez", 3 ], [ "tarekziade", 2 ], [ "tonywu71", 1 ], [ "MekkCyber", 1 ] ], "n": 9, "add": 890, "del": 798, "files": 49, "commits": 35 } }, { "t": "Attention interface unification", "tag": "single ALL_ATTENTION_FUNCTIONS path", "s": "wip", "prs": [ { "n": 35235, "d": "2024-12-18", "t": "🚨All attention refactor🚨", "st": "MERGED", "by": "ArthurZucker", "a": 5635, "r": 9778, "f": 107, "c": 99, "b": "Todo in this PR: Cohere Chameleon DBRX Gemma Gemma2 GLM (modular donc rien à faire je crois) gpt_neoX et GPT2 Granite Jamba JetMoe Mimi Mistral Mixtral Mllama Moshi Nemotron OPT Phi Ph3 PhiMoe Qwen2 qwen2Moe qwen2VL SableML StartCoder2 -> Modular normalement oK Idefics1,2,3 Olmo Olmo2 Siglip Whisper" }, { "n": 35342, "d": "2024-12-20", "t": "Cleaner attention interfaces", "st": "MERGED", "by": "Cyrilvallez", "a": 30, "r": 12, "f": 4, "c": 5, "b": "Improve the attention interfaces with comments/dtype fix" }, { "n": 37866, "d": "2025-05-22", "t": "🚨🚨[core] Completely rewrite the masking logic for all attentions", "st": "MERGED", "by": "Cyrilvallez", "a": 2984, "r": 6808, "f": 129, "c": 193, "b": "As per the title. The goal is to properly separate masking logic from modeling code itself, to continue our objective of simplifying the library. Code is much simpler to understand Much more general: always work for all lengths, and all attention implementations, e.g.: flex attention now works with " }, { "n": 42848, "d": "2026-02-09", "t": ":rotating_light: [`Attn`] New attn mask interface everywhere", "st": "MERGED", "by": "vasqu", "a": 1482, "r": 7100, "f": 117, "c": 61, "b": "As per title ~" }, { "n": 46738, "d": "2026-06-30", "t": "🚨 Modeling changes for export, compile, and hybrid-attention standardization", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 2353, "r": 3720, "f": 152, "c": 48, "b": "Modeling changes for export, compile, and hybrid-attention standardization A broad sweep of modeling-side changes that make a large set of models cleanly exportable (ONNX / torch.export / ExecuTorch), fullgraph-compilable, and consistent in how they declare their layer types and build masks/caches. " }, { "n": 47451, "d": "2026-07-22", "t": "Move `value` padding into the attention interfaces that need it", "st": "MERGED", "by": "hmellor", "a": 26, "r": 87, "f": 14, "c": 4, "b": "MLA attention forwards currently use is_flash_attention_requested(self.config) as a stand-in for \"the selected attention kernel requires equal query/value head dimensions\". If true, it then pads value_states up to qk_head_dim before attention and slicing the output back afterwards: if is_flash_atten" }, { "n": 47014, "d": "2026-07-30", "t": "🚨 Enable SDPA (and other attention backends) for T5 and propagate to the T5 family", "st": "MERGED", "by": "jiqing-feng", "a": 1259, "r": 1364, "f": 20, "c": 32, "b": "What this PR does Refactors the T5 attention stack to route through ALL_ATTENTION_FUNCTIONS, so T5 can dispatch to sdpa / eager instead of the old eager-only path, and propagates the change to the copied-from family. Now _supports_sdpa = True / _supports_attention_backend = True and running test_eag" }, { "n": 46196, "d": "2026-08-20", "t": "🚨[wav2vec2] Support attn_implementation=sdpa dispatch", "st": "MERGED", "by": "YangKai0616", "a": 1767, "r": 3073, "f": 33, "c": 53, "b": "The wav2vec2_conformer/bert and SeamlessM4T/v2 models fail during the initialization phase when using from_pretrained(..., attn_implementation=\"sdpa\"). This PR enables standard attn_implementation=\"sdpa\" dispatch to these models." }, { "n": 48333, "d": "2026-08-27", "t": "Use the shared attention interface in encoder-decoder attention blocks", "st": "OPEN", "by": "ArthurZucker", "a": 1016, "r": 993, "f": 14, "c": 3, "b": "Routes 15 bart-style *Attention forwards through ALL_ATTENTION_FUNCTIONS, enabling sdpa/flash. No weight renames; modules read hyperparameters from config. Deltas fp32-rounding only (<=8.4e-07). ProphetNet attentions were pre-softmax logits, now probabilities. FSMT split out to #48357; Autoformer an" }, { "n": 48357, "d": "2026-08-27", "t": "Modernize FSMT: shared attention interface and remove legacy fairseq code", "st": "OPEN", "by": "ArthurZucker", "a": 536, "r": 792, "f": 2, "c": 1, "b": "Reshapes FSMT after BART: attention via ALL_ATTENTION_FUNCTIONS, masks built in the model, batch-first hidden states, new decorators, sdpa/flash support. Drops ~250 lines of fairseq helpers; fixes forward(input_ids, labels=...). No weight renames; logits match within 3.3e-5 on wmt19-en-de." } ], "d": "Multi-year effort routing every model's attention/masking through a shared interface; core rewrite landed, but encoder-decoder and legacy models (FSMT, Bart-family) are still being migrated.", "note": null, "links": [], "src": "standard.json", "span": [ "2024-12-16", "2026-08-27" ], "roll": { "owners": [ [ "ArthurZucker", 3 ], [ "Cyrilvallez", 2 ], [ "vasqu", 1 ], [ "IlyasMoutawwakil", 1 ] ], "n": 10, "add": 17088, "del": 33727, "files": 592, "commits": 499 } }, { "t": "Community integration & runtime docs", "tag": "docs for external inference runtimes", "s": "wip", "prs": [ { "n": 39550, "d": "2025-07-22", "t": "[docs] Create page on inference servers with transformers backend", "st": "MERGED", "by": "zucchini-nlp", "a": 260, "r": 49, "f": 3, "c": 31, "b": "As per title, I added the basic info about existing inference engines so feel free to add more examples/tips etc. This PR creates a space where we can host docs on all third-party servers and we can submit PRs in vLLM/SGLang/TGI pointing to this page" }, { "n": 42932, "d": "2026-01-07", "t": "[docs] inference engines", "st": "MERGED", "by": "stevhliu", "a": 248, "r": 208, "f": 5, "c": 3, "b": "starts off a new section dedicated to our integrations with inference engines" }, { "n": 44304, "d": "2026-03-03", "t": "docs: Add NeMo Automodel community integration docs", "st": "MERGED", "by": "adil-a", "a": 151, "r": 0, "f": 3, "c": 12, "b": "Adds NeMo Automodel under the community integrations tab. Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Plea" }, { "n": 44682, "d": "2026-03-13", "t": "transformers serve + llamacpp", "st": "OPEN", "by": "SunMarc", "a": 301, "r": 73, "f": 2, "c": 1, "b": "Llama cpp integration in transformers serve. Minor changes to add llama.cpp integration Mostly changes on serve to fix latency for streaming and non streaming" }, { "n": 46754, "d": "2026-06-24", "t": "Documentation for the kernel API", "st": "MERGED", "by": "michaelbenayoun", "a": 204, "r": 0, "f": 4, "c": 5, "b": "This PR provides the documentation for the big API changes and extension provided by #46339" }, { "n": 46889, "d": "2026-06-25", "t": "Kernel api doc, part 2", "st": "OPEN", "by": "michaelbenayoun", "a": 29, "r": 9, "f": 1, "c": 3, "b": "Addresses comments from @vasqu in #46754." } ], "d": "Docs on transformers as a backend for external inference engines/servers (NeMo Automodel, llama.cpp serve, Hub kernels); coverage is uneven vs. mlx/candle/sglang/tensorrt-llm.", "note": "No dedicated merged/open PRs found specifically for mlx, candle, sglang, tensorrt-llm, executorch, axolotl, or nanotron integration docs despite targeted searches.", "links": [], "src": "standard.json", "span": [ "2025-07-21", "2026-06-25" ], "roll": { "owners": [ [ "michaelbenayoun", 2 ], [ "zucchini-nlp", 1 ], [ "stevhliu", 1 ], [ "adil-a", 1 ] ], "n": 6, "add": 1193, "del": 339, "files": 18, "commits": 55 } }, { "t": "Output-collection standardization", "tag": "decorator-based output tracing", "s": "wip", "prs": [ { "n": 40342, "d": "2025-09-05", "t": "🚨 Allow `check_model_inputs` in core VLMs", "st": "MERGED", "by": "zucchini-nlp", "a": 845, "r": 2494, "f": 43, "c": 30, "b": "Unblocks #39722 so we can use check_model_inputs in VLMs copied from llava. Otehrwise #39722 would need to re-define model classes to change forward pass and delete output_xxx kwargs" }, { "n": 40994, "d": "2025-10-06", "t": "Check model inputs - hidden states", "st": "MERGED", "by": "zucchini-nlp", "a": 393, "r": 367, "f": 173, "c": 9, "b": "Fixes #40964 In most vision models the output.hidden_states are the hiddens right after encoder blocks, i.e. before the last layernorm. Therefore for these models output.hidden_states != output.last_hidden_state Currently check_model_inputs assumes that last hidden state is the correct one to return" }, { "n": 40811, "d": "2025-10-06", "t": "Fixes in check_model_inputs, GPTBigCodeModel and ImageGPTModel", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 34, "r": 27, "f": 3, "c": 9, "b": "Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. Di" }, { "n": 43765, "d": "2026-02-06", "t": "[core] Faster and thread-safe `check_model_inputs` implementation", "st": "MERGED", "by": "Cyrilvallez", "a": 728, "r": 565, "f": 121, "c": 17, "b": "The problem Currently, check_model_inputs needs to iterate on all modules and monkey-patch all needed submodule's forward on-the-fly, before restoring them afterwards. This brings 2 big issues: It's NOT thread-safe (see also issue #42673): due to the monkey patching, 2 threads running the model's fo" }, { "n": 43862, "d": "2026-02-10", "t": "Separate `check_model_inputs` into `capture_outputs` and `merge_with_config_defaults` + ensure correctness", "st": "MERGED", "by": "Cyrilvallez", "a": 1642, "r": 996, "f": 261, "c": 20, "b": "As per the title. check_model_inputs is becoming very complex and doing more than what it should. Let's separate into 2 clear decorators: capture_outputs: everything related to capture outputs merge_with_config_defaults: everything related to handling of some kwargs Also, add a test to ensure that c" }, { "n": 43996, "d": "2026-02-14", "t": "Refactor FNet and CVT output tracing", "st": "OPEN", "by": "beelapranay", "a": 44, "r": 134, "f": 2, "c": 1, "b": "Refactors FNet and CVT output tracing to use the standardized decorators where appropriate. FNet now uses @capture_outputs with _can_record_outputs to collect hidden states. CVT keeps manual hidden-state collection to preserve vision-model semantics. Wrapper heads use @can_return_tuple to simplify r" }, { "n": 44015, "d": "2026-02-15", "t": "Refactor GPT2-based models to standardized output collection interface", "st": "OPEN", "by": "akashadsare", "a": 95, "r": 231, "f": 3, "c": 1, "b": "This PR migrates GPT-2 and its derivatives (GPTBigCode, Decision Transformer) to the new standardized output collection interface using the @capture_outputs and @can_return_tuple decorators. By leveraging the _can_record_outputs metadata in the model classes, we remove the need for manual loops and " }, { "n": 44024, "d": "2026-02-17", "t": "Focalnet standardized outputs", "st": "OPEN", "by": "mmahjoub5", "a": 15, "r": 60, "f": 1, "c": 2, "b": "This PR refactors the FocalNet implementation to adopt the new @capture_outputs and @can_return_tuple decorators for standardized output collection, as part of #43979. The goal is to align FocalNet with the new output capturing infrastructure and remove legacy return_dict / output_hidden_states hand" }, { "n": 44066, "d": "2026-02-18", "t": "Refactor GPT-J to use standardized output tracing (#43979)", "st": "OPEN", "by": "jayavelubalaji-ai", "a": 35, "r": 107, "f": 2, "c": 3, "b": "Migrate GPT-J from manual boilerplate output collection to the new decorator-based output tracing system: Add _can_record_outputs to GPTJPreTrainedModel Add @capture_outputs and @merge_with_config_defaults to GPTJModel.forward Add @can_return_tuple to GPTJForCausalLM, GPTJForSequenceClassification, " }, { "n": 44114, "d": "2026-02-18", "t": "Migrate wav2vec2, wav2vec2_conformer, and wav2vec2_bert to standardized output collection decorators", "st": "OPEN", "by": "23atharvaS", "a": 716, "r": 1237, "f": 19, "c": 12, "b": "Summary This PR migrates the wav2vec2 family to the standardized output-capturing interface (@capture_outputs + @can_return_tuple) and includes follow-up compatibility fixes required to make full CI green. What changed Core migration (wav2vec2, wav2vec2_conformer, wav2vec2_bert) Added _can_record_ou" }, { "n": 44161, "d": "2026-02-19", "t": "Refactor LongT5 to use @capture_outputs and @can_return_tuple decorators for unified output handling (Fixes #43979)", "st": "OPEN", "by": "cogniera", "a": 379, "r": 170, "f": 1, "c": 1, "b": "This PR refactors the LongT5 model to use the @capture_outputs and @can_return_tuple decorators for standardized output handling across the model stack. The refactor removes manual handling of: output_attentions output_hidden_states return_dict manual tuple construction explicit attention/hidden sta" }, { "n": 43590, "d": "2026-03-12", "t": "Remove many output_attentions and other traced outputs on 100+ models ", "st": "MERGED", "by": "molbap", "a": 5117, "r": 12955, "f": 234, "c": 191, "b": "In model additions, we often see old standards not using check_model_inputs, can_return_tuple and it's often a first review comment/something that can slip through. Doing a wide scan to try to remove all occurrences systematically. Background Every model used to manually resolve output_attentions, o" }, { "n": 44947, "d": "2026-03-26", "t": "Add doc page for capturing outputs", "st": "MERGED", "by": "zucchini-nlp", "a": 123, "r": 2, "f": 4, "c": 10, "b": "The doc was generated by Claude. I deleted unnecessary repetitions and fixed a few moments to be more precise. We don't really need to merge it now so if you think the text is too LLM, feel free to take this as an initial template and close the PR" } ], "d": "Decorator-based output tracing replaces manual output_attentions/hidden_states plumbing; core landed, ~20 more per-model migrations open (GPT2, GPT-J, RegNet, wav2vec2, etc.).", "note": null, "links": [], "src": "standard.json", "span": [ "2025-08-21", "2026-03-26" ], "roll": { "owners": [ [ "zucchini-nlp", 3 ], [ "Cyrilvallez", 2 ], [ "IlyasMoutawwakil", 1 ], [ "molbap", 1 ] ], "n": 13, "add": 10166, "del": 19345, "files": 867, "commits": 306 } }, { "t": "Modular v2 / transformers format draft", "tag": "next-gen modular authoring format", "s": "open", "prs": [ { "n": 41289, "d": "2025-10-03", "t": "Add modular detector", "st": "MERGED", "by": "molbap", "a": 677, "r": 1, "f": 3, "c": 7, "b": "Adds a new util to parse a new modeling_* file, whether or not it is from transformers, and lists the closest possible matches from existing models in the library using embeddings from Qwen3. Why? Because most models should be able to be defined as modulars from each other. It's not exactly the case" }, { "n": 43743, "d": "2026-08-20", "t": "Modular playground", "st": "OPEN", "by": "itazap", "a": 3988, "r": 729, "f": 9, "c": 32, "b": "Modular Detector Updates: Hub / index: FAISS index. The new index is at: itazap/transformers_code_embeddings_v3 (old: hf-internal-testing/transformers_code_embeddings) Model: Qwen/Qwen3-Embedding-4B, now using last-token pooling instead of mean pooling Sanitization: type hints, dtype save/cast boile" }, { "n": 48271, "d": "2026-08-26", "t": "draft modular v2 + transformers format", "st": "DRAFT", "by": "ArthurZucker", "a": 15323, "r": 3245, "f": 175, "c": 21, "b": "" } ], "d": "Embedding-based modular detector shipped to find parent classes; #48271 drafts a modular v2 'transformers format' and a modular playground UI, both still unmerged.", "note": "Core direction (#48271) is an open draft, not yet reviewed/merged; scope may still change.", "links": [], "src": "standard.json", "span": [ "2025-10-02", "2026-08-26" ], "roll": { "owners": [ [ "molbap", 1 ], [ "itazap", 1 ], [ "ArthurZucker", 1 ] ], "n": 3, "add": 19988, "del": 3975, "files": 187, "commits": 60 } }, { "t": "TokenizersBackend rollout", "tag": "unified fast-tokenizer backend", "s": "wip", "prs": [ { "n": 42563, "d": "2025-12-09", "t": "Refactor-tokenization-more", "st": "MERGED", "by": "ArthurZucker", "a": 1274, "r": 2854, "f": 84, "c": 124, "b": "Update TokenizersBackend to simplify converting from sentencepiece. Add the model attribute to guide the conversion. Adds the convert_from_spm for special cases when the spm is not straight forward. Simplify init for TokenizersBackend to support back passing vocab file / merges file path." }, { "n": 42894, "d": "2026-01-07", "t": "use `TokenizersBackend`", "st": "MERGED", "by": "ArthurZucker", "a": 249, "r": 243, "f": 29, "c": 61, "b": "Fixes a bunch of issues (including #42874) With this PR, we now kinda enforce that if there is a specific python path to be used (meaning there is a XXXXTokenizer class that has python code whatever the backend) the tokenizer's saved tokenizer_class needs to match the class mapped in tokenization_au" }, { "n": 44294, "d": "2026-02-26", "t": "Fix: use `TokenizersBackend` for Olmo3 to preserve custom `pre_tokenizer`", "st": "MERGED", "by": "mario-sanz", "a": 1, "r": 1, "f": 1, "c": 1, "b": "Fixes a bug where GPT2Tokenizer.__init__ unconditionally overwrites the pre_tokenizer with a generic ByteLevel, discarding the custom Sequence pre_tokenizer defined in the model's tokenizer.json. This causes incorrect tokenization of consecutive newlines for Olmo3 models (e.g., \\n\\nhello is split in" }, { "n": 44255, "d": "2026-03-04", "t": "[vllm + v5 fix] handle TokenizersBackend fallback properly for v5", "st": "MERGED", "by": "itazap", "a": 404, "r": 205, "f": 24, "c": 48, "b": "What this PR does Given the different issues that were noticed by @hmellor on vLLM, we wanted to make sure we did not end up with crazy breaks. We ran a full test suite (code can be found in #44298) and the results showed 22 model converters that had potential issues. We test what would be the AutoT" }, { "n": 45262, "d": "2026-04-06", "t": "doc: fix TokenizersBackend.convert_to_native_format docstring", "st": "MERGED", "by": "lowzhao", "a": 1, "r": 1, "f": 1, "c": 1, "b": "Fix docstring spelling mistake TokenizersBackend.convert_to_native_format. @classmethod def convert_to_native_format(cls, trust_remote_code=False, **kwargs): \"\"\"s <---- additional s Likely caused by mistyped ctrl/cmd + s command. Before submitting This PR fixes a typo or improves the docs (you can d" }, { "n": 45813, "d": "2026-05-18", "t": "fix: route Granite models to TokenizersBackend to preserve tokenizer.json pre-tokenizer", "st": "MERGED", "by": "kndtran", "a": 18, "r": 4, "f": 2, "c": 4, "b": "Change TOKENIZER_MAPPING_NAMES for Granite model types from \"GPT2Tokenizer\" to \"TokenizersBackend\" so that AutoTokenizer loads tokenizer.json faithfully instead of routing through GPT2Tokenizer.__init__ which hardcodes a wrong pre-tokenizer. Fixes #45812 Code Agent Policy The Transformers repo is cu" }, { "n": 46091, "d": "2026-06-19", "t": "Update tokenizer mappings to use TokenizersBackend for additional models", "st": "MERGED", "by": "itazap", "a": 84, "r": 24, "f": 3, "c": 17, "b": "see the description / report here: #45936 this PR is just the auto changes from the PR above. it's for models that don't have their own Tokenizer class so we don't have test_tokenization_*.py for these models. fixes: #45920, #46710, #45488, #46489" }, { "n": 47302, "d": "2026-07-15", "t": "[docs] TokenizersBackend fallback", "st": "MERGED", "by": "stevhliu", "a": 34, "r": 34, "f": 1, "c": 1, "b": "Adds docs to describe the routing behavior behind how a checkpoint falls back to TokenizersBackend when there is no dedicated tokenizer class, an incorrect tokenizer class is stored on the Hub, or overriding a specific checkpoint (see #45936)." }, { "n": 48211, "d": "2026-08-26", "t": "Fix AutoTokenizer returning TokenizersBackend for DeepSeek-R1-Distill-Qwen models", "st": "MERGED", "by": "ydshieh", "a": 1, "r": 1, "f": 1, "c": 1, "b": "What MODEL_IDS_TO_TOKENIZERS_BACKEND contained \"deepseek-ai/deepseek-r1-distill-*\" — a wildcard that is too broad. Since _config_name_or_path is lowercased before fnmatch matching, it also matches DeepSeek-R1-Distill-Qwen-1.5B (and all other Qwen-based distills), routing them to TokenizersBackend in" }, { "n": 47017, "d": "2026-08-31", "t": "Route byte-level llama tokenizers to TokenizersBackend", "st": "OPEN", "by": "subin9", "a": 121, "r": 0, "f": 2, "c": 4, "b": "Model type \"llama\" spans both SentencePiece (Llama-1/2) and byte-level (Llama-3 / tiktoken) tokenizers under one Hub tokenizer_class (LlamaTokenizerFast). In v5, LlamaTokenizer.init unconditionally installs a Metaspace pre-tokenizer/decoder, which silently drops spaces for byte-level repos (see #454" } ], "d": "New TokenizersBackend preserves tokenizer.json pre_tokenizer/decoder fidelity; rolled out model-by-model (Olmo3, Granite, DeepSeek-R1-Distill-Qwen); byte-level Llama routing still open.", "note": null, "links": [], "src": "standard.json", "span": [ "2025-12-03", "2026-08-31" ], "roll": { "owners": [ [ "ArthurZucker", 2 ], [ "itazap", 2 ], [ "mario-sanz", 1 ], [ "lowzhao", 1 ] ], "n": 10, "add": 2187, "del": 3367, "files": 148, "commits": 262 } }, { "t": "mlinter model-structure lint + CI review bot", "tag": "automated model-structure enforcement", "s": "wip", "prs": [ { "n": 43687, "d": "2026-02-03", "t": "Added S110 - try-except-pass rule", "st": "MERGED", "by": "tarekziade", "a": 164, "r": 150, "f": 18, "c": 31, "b": "adds Bandit's S110 that makes sure we don't have a dry Except fixes all occurrences mark a couple of spots where we could tighten the Exception catch all I focused on making changes under src/transformers keep the same behavior" }, { "n": 44174, "d": "2026-03-12", "t": "Expand model-structure lint rules with a fast AST-based, ruff-like framework", "st": "MERGED", "by": "tarekziade", "a": 1367, "r": 24, "f": 4, "c": 12, "b": "This draft expands utils/check_modeling_structure.py into a rule-driven linter for model code, with new checks and tests, while keeping runtime very low. Key features: The checker is intentionally AST-only (no heavy imports/execution), so it stays fast (<3 seconds against all models) It's built in a" }, { "n": 44939, "d": "2026-03-24", "t": "refactor: mlinter as its own package", "st": "MERGED", "by": "tarekziade", "a": 2038, "r": 1446, "f": 25, "c": 8, "b": "Refactored and cleaned up model linter separated package one rule per module refactored legacy checks into their own rules simplified pattern, duplication removal" }, { "n": 45720, "d": "2026-04-30", "t": "chore(mlinter): implement part of rule TRF003", "st": "OPEN", "by": "tarekziade", "a": 225, "r": 848, "f": 22, "c": 2, "b": "Enables rule TRF003 and implement it on 10 models" }, { "n": 47172, "d": "2026-07-09", "t": "Switch mlinter to 0.1.2", "st": "MERGED", "by": "tarekziade", "a": 165, "r": 3, "f": 4, "c": 4, "b": "runs on current mlinter 0.1.2 models that fail to pass the linter TRF014 (trust_remote_code) │ auto, legitimate loader exemption we can keep that TRF016 (dead do_* flags) │ kimi_k25, minimax_m3_vl TRF018 (super()._init_weights) | cosmos3_omni, diffusion_gemma, eomt_dinov3, exaone4_5, gemma4, higgs_a" }, { "n": 47819, "d": "2026-08-06", "t": "Post mlinter findings as inline PR review comments", "st": "MERGED", "by": "ydshieh", "a": 198, "r": 6, "f": 2, "c": 2, "b": "What this does Adds a post-mlinter-review job to pr-ci-post-dashboard-link.yml that: Downloads the mlinter-findings artifact produced by the check_code_quality job Posts the findings as a single PR review with inline comments anchored to the diff The review includes a summary table (rule → count) an" }, { "n": 47830, "d": "2026-08-07", "t": "Use content hash for mlinter review deduplication", "st": "MERGED", "by": "ydshieh", "a": 68, "r": 8, "f": 1, "c": 3, "b": "Summary Improves deduplication in post_mlinter_review.py: instead of paginating all PR reviews to find a marker (O(n) API calls), store mlinter state in the PR description itself (O(1)). Before: scan all reviews for <!-- mlinter-review -->, skip if found — always skips after the first post regardles" }, { "n": 47866, "d": "2026-08-10", "t": "Enable new trf rules", "st": "OPEN", "by": "zucchini-nlp", "a": 0, "r": 10, "f": 11, "c": 1, "b": "Fixes errors found by huggingface/transformers-mlinter#24. Some will requires a lot of time to fix (i mean adding missing test files) so I hardcoded them in allowlist The rule shouldn't allow any new files to skip tests, at least the bare minimum functionality needs to be tested with dummy inputs" }, { "n": 47918, "d": "2026-08-18", "t": "Moving mlinter to 0.1.4", "st": "MERGED", "by": "tarekziade", "a": 2249, "r": 131, "f": 54, "c": 13, "b": "Bumps transformers-mlinter to 0.1.4 to pick up the new rules (TRF020 to TRF057), and resolves everything those rules flag. The base pinned transformers-mlinter==0.1.2, whose utils/rules.toml carried only TRF001–TRF019. This bump adds 37 new rules in the TRF020–TRF057 range, and every finding they re" }, { "n": 48117, "d": "2026-08-19", "t": "Enable mlinter findings artifact for inline PR reviews", "st": "MERGED", "by": "ydshieh", "a": 1, "r": 1, "f": 1, "c": 1, "b": "Summary Last piece of the mlinter inline review pipeline, verified in #48109. Adds --output-json mlinter-findings.json to the mlinter invocation so the check_code_quality job uploads a mlinter-findings artifact. The post-mlinter-review job (already on main) then downloads it and posts inline review " }, { "n": 48259, "d": "2026-08-28", "t": "Bump transformers-mlinter to 0.1.5 and clear the new findings", "st": "MERGED", "by": "tarekziade", "a": 872, "r": 536, "f": 23, "c": 13, "b": "setup.py points transformers-mlinter at main ahead of the 0.1.5 release. The new rules surface 30 findings; this branch takes that to 0. make typing and make check-repo pass. ⚠️ Before merge: pin setup.py back to ==0.1.5 once released (make fix-repo propagates to dependency_versions_table.py). Findi" }, { "n": 48398, "d": "2026-08-28", "t": "Enforce TRF057 everywhere: drop its allowlist and cutoff date", "st": "OPEN", "by": "tarekziade", "a": 3198, "r": 2158, "f": 325, "c": 1, "b": "The fix was done mechanicallyi by AI see below: TRF057 requires @auto_docstring on public model, config, output and processor classes and on their public methods. It was running with six allowlisted models and a cutoff_date, so it only ever fired on new code. This clears the backlog and removes both" } ], "d": "AST-based linter (TRF0xx rules) for modeling/config/modular files, split into its own pip package (transformers-mlinter) and wired into CI to post inline PR review comments.", "note": "TRF057 allowlist/cutoff removal (#48398) still open, so full enforcement isn't final yet.", "links": [], "src": "standard.json", "span": [ "2026-02-02", "2026-08-28" ], "roll": { "owners": [ [ "tarekziade", 8 ], [ "ydshieh", 3 ], [ "zucchini-nlp", 1 ] ], "n": 12, "add": 10545, "del": 5321, "files": 490, "commits": 91 } }, { "t": "Transpilation: transformers → MLX", "tag": "conversion skill, 51★", "s": "done", "prs": [], "d": "Agent Skill converting transformers LLMs to mlx-lm — MoEs, MLA, hybrid SSM+attention, GatedDeltaNet, with per-layer logit comparison.", "links": [ [ "huggingface/transformers-to-mlx", "https://github.com/huggingface/transformers-to-mlx" ], [ "MLX docs", "https://huggingface.co/docs/transformers/community_integrations/mlx" ] ], "note": null, "src": "curated", "span": [ "2026-04-15", "2026-06-16" ], "roll": { "owners": [], "n": 0, "add": 0, "del": 0, "files": 0, "commits": 0 } }, { "t": "Visual modular — the Atlas", "tag": "SVG diagram for every model", "s": "wip", "prs": [], "d": "Generated architecture gallery: a full view per model plus a diff view of what each modular file overrides vs its parent.", "links": [ [ "Atlas: model-architectures", "https://huggingface.co/spaces/transformers-community/model-architectures" ] ], "note": null, "src": "curated", "span": [ "2026-06-02", "2026-07-01" ], "roll": { "owners": [], "n": 0, "add": 0, "del": 0, "files": 0, "commits": 0 } } ] }, { "name": "API & weight loading", "hue": 122, "blurb": "One way to do things, and other libraries can build on it.", "items": [ { "t": "Pipelines & generation API standardization", "tag": "v5 decoding/pipeline API cleanup", "s": "done", "prs": [ { "n": 33899, "d": "2024-10-08", "t": "Image pipelines spec compliance", "st": "MERGED", "by": "Rocketknight1", "a": 153, "r": 44, "f": 13, "c": 15, "b": "This PR checks for Hub spec compliance for a lot of image pipelines that have very similar inputs. In most cases, no changes are required except deprecating the timeout argument, and renaming the images input to inputs." }, { "n": 40428, "d": "2025-08-26", "t": "🚨 Remove Contrastive Search decoding strategy", "st": "MERGED", "by": "manueldeprada", "a": 107, "r": 1109, "f": 26, "c": 11, "b": "Removes Contrastive Search generation strategy from the codebase. Directs users to the transformers-community/contrastive-search repository. It has been a warning for a few releases, but now trust_remote_code=True is required to run contrastive search." }, { "n": 40518, "d": "2025-09-01", "t": "🚨 Remove Constrained Beam Search decoding strategy", "st": "MERGED", "by": "manueldeprada", "a": 516, "r": 1234, "f": 19, "c": 25, "b": "Removes Constrained Beam Search generation strategy from the codebase. Directs users to the transformers-community/constrained-beam-search repository. It has been a warning for a few releases, but now trust_remote_code=True is required to run constrained beam search. Depends on #40480 and #40495" }, { "n": 40495, "d": "2025-09-01", "t": "🚨 Remove Group Beam Search decoding strategy", "st": "MERGED", "by": "manueldeprada", "a": 63, "r": 1432, "f": 32, "c": 6, "b": "Removes Group Beam Search generation strategy from the codebase. Directs users to the transformers-community/group-beam-search repository. It has been a warning for a few releases, but now trust_remote_code=True is required to run group beam search. Depends on #40480" }, { "n": 40928, "d": "2025-09-24", "t": "🚨Refactor: Update text2text generation pipelines to use max_new_tokens…", "st": "MERGED", "by": "lilin-1", "a": 12, "r": 14, "f": 1, "c": 7, "b": "name: Pull Request about: Create a pull request to contribute to 🤗 Transformers title: \"Refactor: Update text2text generation pipelines to use max_new_tokens and resolve max_length warning\" labels: bug, summarization assignees: '' Related Issue Closes #40768 Summary This PR fixes an incorrect warnin" }, { "n": 43325, "d": "2026-02-24", "t": "🚨 More V5 pipeline cleanup", "st": "MERGED", "by": "Rocketknight1", "a": 140, "r": 2679, "f": 116, "c": 43, "b": "More V5 pipeline cleanup, followup to #43256 and #43306: (changelist updated Jan 26th) question-answering and visual-question-answering removed image-to-image removed Updated the default text-generation and image-text-to-text models Removed some references to image-to-text, summarization, translatio" } ], "d": "V5 cleanup removed legacy decoding strategies (contrastive search, constrained/group beam search) and standardized text2text pipelines on max_new_tokens; image pipelines made Hub-spec compliant.", "note": null, "links": [], "src": "standard.json", "span": [ "2024-10-02", "2026-02-24" ], "roll": { "owners": [ [ "manueldeprada", 3 ], [ "Rocketknight1", 2 ], [ "lilin-1", 1 ] ], "n": 6, "add": 991, "del": 6512, "files": 207, "commits": 107 } }, { "t": "auto_docstring decorator system", "tag": "auto-generated argument docs", "s": "wip", "prs": [ { "n": 38242, "d": "2025-06-23", "t": "Add support for auto_docstring with model outputs", "st": "MERGED", "by": "yonigozlan", "a": 6972, "r": 8993, "f": 177, "c": 21, "b": "As the title says. However not sure if worth it, as there are a lot of different descriptions currently for the same args. The descriptions that are different from the one in ModelOutputArgs are left in the ModelOutput docstring, which is why the there's not that much more deletions than additions i" }, { "n": 39439, "d": "2025-07-18", "t": "Improve @auto_docstring doc and rename `args_doc.py` to `auto_docstring.py`", "st": "MERGED", "by": "yonigozlan", "a": 180, "r": 30, "f": 4, "c": 6, "b": "As the title says :). Cc @Cyrilvallez @stevhliu" }, { "n": 39433, "d": "2025-07-18", "t": "Fix placeholders replacement logic in auto_docstring", "st": "MERGED", "by": "yonigozlan", "a": 18, "r": 22, "f": 1, "c": 2, "b": "The placeholders logic in auto_docstring seems to be broken currently, so that placeholders like {image_processor_class} are not replaced properly. This PR simplifies greatly the logic and fixes the issue" }, { "n": 39564, "d": "2025-07-25", "t": "Fix auto_docstring crashing when dependencies are missing", "st": "MERGED", "by": "yonigozlan", "a": 9, "r": 5, "f": 1, "c": 2, "b": "As the title says, simply use a try except for now until I come up with a better solution" }, { "n": 42512, "d": "2025-12-01", "t": "Handle decorator with optional arguments better", "st": "MERGED", "by": "hmellor", "a": 195, "r": 193, "f": 157, "c": 1, "b": "In #40994 this decorator was changed so that, without arguments, it had to be used like @check_model_inputs() def ... instead of @check_model_inputs def ... This PR fixes that so that custom models which use this decorator can continue to work." }, { "n": 43748, "d": "2026-02-25", "t": "[auto_docstring] Improve typing parsing and add tests", "st": "MERGED", "by": "yonigozlan", "a": 1372, "r": 227, "f": 9, "c": 26, "b": "Improve types parsing to handle \"|\" syntax and other edge cases, add much needed tests Cc @Cyrilvallez ;)" }, { "n": 45056, "d": "2026-03-27", "t": "[`auto_docstring`] needs to be only run on __doc__ ", "st": "OPEN", "by": "ArthurZucker", "a": 331, "r": 28, "f": 4, "c": 3, "b": "This is mega long due I wanted to check benches. Its not super super huge but a win is a win" }, { "n": 45130, "d": "2026-04-15", "t": "[docs] @auto_docstring decorator", "st": "MERGED", "by": "stevhliu", "a": 122, "r": 154, "f": 2, "c": 5, "b": "updates the @auto_docstring docs as part of the process of improving the model addition docs updates title to be more precise as it can mean documenting a model in a model.md file as well flat hierarchy before → each component-type example has their own section now so its easier to find and referenc" }, { "n": 47737, "d": "2026-08-07", "t": "Add new args in auto-docstring", "st": "MERGED", "by": "zucchini-nlp", "a": 489, "r": 1339, "f": 58, "c": 16, "b": "As per title, add image/video grid thw in auto docs and deletes it from modeling code. Also noticed that labels can be deleted as it is in ModelArgs already. And, allows video processors be auto-doc'ced IMPORTANT: From this PR on we will be explicitly unpacking all kwargs in the docstring if we can." }, { "n": 47815, "d": "2026-08-26", "t": "Add auto_docstring task model overrides", "st": "OPEN", "by": "guarin", "a": 157, "r": 2096, "f": 292, "c": 5, "b": "Adds task specific overrides of ModelArgs for @auto_docstring decorator: class ModelForArgs: \"\"\"Task-specific overrides for `ModelArgs`.\"\"\" class ForSequenceClassification: labels = { \"description\": \"\"\" Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., " } ], "d": "@auto_docstring generates standard argument docs for shared model/config APIs, replacing hand-written docstrings; typing parsing, task-specific overrides, and edge-case robustness still being refined.", "note": null, "links": [], "src": "standard.json", "span": [ "2025-05-20", "2026-08-26" ], "roll": { "owners": [ [ "yonigozlan", 5 ], [ "hmellor", 1 ], [ "stevhliu", 1 ], [ "zucchini-nlp", 1 ] ], "n": 10, "add": 9845, "del": 13087, "files": 705, "commits": 87 } }, { "t": "Cache refactor: native mamba/hybrid + cache_position removal", "tag": "unified first-class Cache API", "s": "done", "prs": [ { "n": 38635, "d": "2025-07-16", "t": "[cache] make all classes cache compatible finally", "st": "MERGED", "by": "zucchini-nlp", "a": 5715, "r": 6839, "f": 268, "c": 65, "b": "As per title, and let's get rid of _supports_cache/_supports_quantized_cache flags. We will assume all models from now on support cache and initialize a DynamicCache (model-specific cache in case of mamba) by default For the static cache, we can't yet assume all models support it because even if the" }, { "n": 41505, "d": "2025-10-13", "t": "🚨 [v5] `generate` delegates default cache initialization to the model", "st": "MERGED", "by": "gante", "a": 114, "r": 50, "f": 31, "c": 10, "b": "See PR title. Now that all traces of legacy caches were removed, we can trust the model to initialize its own cache! This means we no longer need to set cache_implementation=\"xxx\" defaults in new models, assuming the model's forward pass defaults to the right cache class. Also fixes related bugs, un" }, { "n": 43679, "d": "2026-02-04", "t": "🚨 Generation cache preparation", "st": "MERGED", "by": "zucchini-nlp", "a": 151, "r": 112, "f": 12, "c": 11, "b": "I also want to see if linear cache thing can be squeezed in this PR. If it requires big diffs, I'll split into two Fixes #43673 Sidenote: kinda breaking but in a good way. Prev models init their cache in model.forward without passing config. That means sliding windows weren't always respected (e.g. " }, { "n": 44181, "d": "2026-03-04", "t": "[core] 🚨 Completely remove cache positions", "st": "MERGED", "by": "Cyrilvallez", "a": 898, "r": 2698, "f": 169, "c": 51, "b": "As per the title! Follow up of #44130 and #44226. Finally remove the cache_position everywhere (not ALL models, but all the recent/most important models that use modular). This means we still create them and pass them around in generate, but they are ignored by the models. I will gradually remove th" }, { "n": 44816, "d": "2026-03-19", "t": "[generate] Never use `cache_position` anymore in generation", "st": "MERGED", "by": "Cyrilvallez", "a": 98, "r": 375, "f": 57, "c": 14, "b": "As per the title. This is the last of many PR to remove the cache_position. At this point, all the models were already updated to not use them, and they are fully ignored in all the modelings. So this removes their creation and usage in generate, so they are not passed as kwarg anywhere anymore. Thi" }, { "n": 44950, "d": "2026-03-31", "t": "🚨 [Cache] Native mamba & hybrid cache", "st": "MERGED", "by": "Cyrilvallez", "a": 1346, "r": 4113, "f": 64, "c": 59, "b": "As per the title. This PR finally makes mamba layer caches first class citizen, and adds native support for them. It supports the following layers combinations: all mamba layers alternating attention layer/mamba layer layers that are BOTH mamba and attention (zamba models) For this, it adds the 2 fo" } ], "d": "Mamba/hybrid layer caches made first-class citizens (#44950, supports all-mamba and alternating layer patterns); cache_position fully removed from generate(), simplifying the Cache API.", "note": null, "links": [], "src": "standard.json", "span": [ "2025-06-04", "2026-03-31" ], "roll": { "owners": [ [ "Cyrilvallez", 3 ], [ "zucchini-nlp", 2 ], [ "gante", 1 ] ], "n": 6, "add": 8322, "del": 14187, "files": 601, "commits": 210 } }, { "t": "PretrainedConfig standardization", "tag": "typed/validated config attributes", "s": "wip", "prs": [ { "n": 41250, "d": "2026-03-16", "t": ":rotating_light: Validate config attributes", "st": "MERGED", "by": "zucchini-nlp", "a": 22671, "r": 39076, "f": 891, "c": 114, "b": "As per title. Continues from #40793 and supersedes #36534 NOTE: config classes can't accept positional args anymore! I don't think anyone would use pos args anyway but marring the PR as breaking NoteHigh Risk Refactors PreTrainedConfig and many model config classes to @dataclass + huggingface_hub @s" }, { "n": 44953, "d": "2026-03-24", "t": "Config kwargs", "st": "MERGED", "by": "zucchini-nlp", "a": 861, "r": 824, "f": 536, "c": 9, "b": "Decouples kwargs manipulation from hub's strict decorator, and ensures that all subclasses of a PreTrainedConfig accept any kwargs which is what we supported prev. Not all remote code has @strict or has an __init__ defined and therefore can fail to accept arbitrary kwargs. For example, this should b" }, { "n": 47613, "d": "2026-07-28", "t": "Don't warn about class defaults when diffing a config", "st": "OPEN", "by": "guptaishaan", "a": 35, "r": 3, "f": 2, "c": 1, "b": "Fixes #47612 to_diff_dict instantiates self.__class__() to diff against, and @strict runs the class validators on that throwaway object. So validate_token_ids warns about the class defaults, not about the config the user actually loaded. Siglip2TextConfig defaults to vocab_size=32000 with CLIP's bos" }, { "n": 48282, "d": "2026-08-25", "t": "Infer config class default fields without init a dummy instance", "st": "OPEN", "by": "zucchini-nlp", "a": 185, "r": 39, "f": 14, "c": 3, "b": "Potentially fixes #47612 and the like. As discussed in slack we dont want to rely on init a dummy class for default, at the same time we don't really want to save every field in config probably which makes json files huge to inspect. Do not merge yet! Though ideas are always welcome" }, { "n": 48360, "d": "2026-08-27", "t": "Standardise head_dim onto the config and drop the guards that stood in for it", "st": "OPEN", "by": "ArthurZucker", "a": 29, "r": 582, "f": 90, "c": 1, "b": "head_dim is now defaulted once in the base config instead of derived in every model, so the 100 per-model divisibility guards it stood in for are gone. Reformer maps its attention_head_size. Interacts with #46211 — see the commit message." } ], "d": "Config attribute validation and default-kwargs tracking tighten PretrainedConfig; open work covers diff-vs-defaults warnings, standardizing head_dim, and inferring defaults without instantiation.", "note": null, "links": [], "src": "standard.json", "span": [ "2025-10-01", "2026-08-27" ], "roll": { "owners": [ [ "zucchini-nlp", 3 ], [ "guptaishaan", 1 ], [ "ArthurZucker", 1 ] ], "n": 5, "add": 23781, "del": 40524, "files": 1533, "commits": 128 } }, { "t": "Dynamic weight loader / WeightConverter", "tag": "declarative checkpoint conversion", "s": "wip", "prs": [ { "n": 41580, "d": "2025-11-13", "t": "Refactor weight loading", "st": "MERGED", "by": "ArthurZucker", "a": 8733, "r": 7189, "f": 586, "c": 387, "b": "CORE REFACTORING, loading, converting, logging More helpful debugging report when loading weights If you just want to fuse qkv: It can. You just need to make sure you change the model code and pouf! WeightConverter( [\"self_attn.q_proj\", \"self_attn.k_proj\", \"self_attn.v_proj\"], \"self_attn.qkv_proj\", " }, { "n": 42246, "d": "2025-11-18", "t": "[loading] Fix device when source and target are different", "st": "MERGED", "by": "Cyrilvallez", "a": 19, "r": 32, "f": 2, "c": 4, "b": "The device_map specifies the target keys when loading. The PR updates the loading accordingly, otherwise we have issues when using a device_map with any model using a _conversion_mapping (the VLMs) for example, where source and targets are different. Currently, the only thing saving us is the fact t" }, { "n": 42254, "d": "2025-11-20", "t": "Update conversion mapping to separate renaming from converting", "st": "MERGED", "by": "ArthurZucker", "a": 525, "r": 329, "f": 5, "c": 34, "b": "Update the conversion: to support torchao" }, { "n": 42242, "d": "2025-11-24", "t": "[loading] Re-add and improve disk offloading support", "st": "MERGED", "by": "Cyrilvallez", "a": 186, "r": 110, "f": 8, "c": 32, "b": "This PR re-add support for weight offloading to disk using the device_map, which was temporarily dropped in #41580 to simplify the PR (because weights need to be resaved correctly after performing custom Ops during dynamic loading). It further improve the offloading mechanism as everything is now of" }, { "n": 42479, "d": "2025-11-28", "t": "Fix minor bugs in ConversionOps and WeightConverter", "st": "MERGED", "by": "SunMarc", "a": 13, "r": 12, "f": 3, "c": 10, "b": "This fixes a couple of minor details in ConversionOps and WeightConverter. We still need to pass full_layer_name as we won't be able to get the module otherwise. Also dim is not always defined, so the repr was failing. cc @Cyrilvallez for vis" }, { "n": 43621, "d": "2026-01-30", "t": "[vllm compat] Separate renaming from conversion ops", "st": "MERGED", "by": "Cyrilvallez", "a": 13, "r": 12, "f": 1, "c": 6, "b": "As per the title. Following specs described by @hmellor for vllm compatibility, no logic change should be required. Only the following small changes to mappings should be enough" }, { "n": 43926, "d": "2026-02-16", "t": "[Deespeed] fix WeightConverter.convert() use", "st": "MERGED", "by": "kashif", "a": 24, "r": 105, "f": 2, "c": 6, "b": "Fixes #43257 Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. Did y" }, { "n": 44396, "d": "2026-03-03", "t": "[ Dynamic weight loader] fix remote code when format matches", "st": "MERGED", "by": "ArthurZucker", "a": 3, "r": 16, "f": 2, "c": 8, "b": "Some renaming should just never be applied when the weight format already matches. (this is actually regardless of remote code). This allows us to remove 1 test added in 8c1b5d3 ! FYI @vasqu we discussed offline with @Cyrilvallez and this makes sense mostly for \"errors\" or special cases were a model" }, { "n": 44627, "d": "2026-03-17", "t": "Move VLM conversions to the main mapping", "st": "MERGED", "by": "zucchini-nlp", "a": 141, "r": 367, "f": 78, "c": 20, "b": "The diff in revert mapping is needed, otherwise we get failures in a few models, see https://app.circleci.com/pipelines/github/huggingface/transformers/167425/workflows/fa96efe5-f810-408e-bafd-de03b7e881aa/jobs/2208432/tests The literal dot is a nice feat to have in general, we actually mean a liter" }, { "n": 44974, "d": "2026-04-14", "t": "Refactor core_model_loading to support FSDP shard-on-read loading", "st": "OPEN", "by": "3outeille", "a": 428, "r": 82, "f": 5, "c": 3, "b": "TODO: Saving seems to take a bit of time tho. Need investigation Need to check if it works in 1D (FSDP or TP)and 2D (FSDP + TP). Running the script from #44996 (env_pr-44974-fsdp-core-model-loading) ➜ pr-44974-fsdp-core-model-loading git:(pr-44974-fsdp-core-model-loading) ✗ torchrun --nproc_per_node" }, { "n": 45448, "d": "2026-04-20", "t": "[loading] Clean way to add/remove full parts in checkpoint names", "st": "MERGED", "by": "Cyrilvallez", "a": 432, "r": 124, "f": 5, "c": 37, "b": "As per the title. The issue The problem is that transforms that want to remove a full part of a model name (such as a prefix, e.g. the model. start) are non bijective in general, i.e. we completely lose the information when they are dropped. So adding them back later when saving is impossible withou" }, { "n": 45782, "d": "2026-05-06", "t": "Fix WeightConverter regex incorrectly matching shared_experts as experts in DeepSeek V4", "st": "MERGED", "by": "silencelamb", "a": 5, "r": 5, "f": 1, "c": 1, "b": "What this fixes The WeightConverter patterns for DeepSeek V4 routed experts use the bare substring experts. in source and target patterns. Since shared_experts contains the substring experts, the regex experts.down_proj also matches model.layers.X.mlp.shared_experts.down_proj.weight, and experts.gat" }, { "n": 45661, "d": "2026-05-07", "t": "[Weight Converter] More fine-grained mappings on classes, scoping for every transforms (including weight converter)", "st": "MERGED", "by": "yonigozlan", "a": 843, "r": 339, "f": 27, "c": 20, "b": "This PR aims to solve three different issues with the existing conversion mapping system. The registry only accepted model_type strings. This made it impossible to have different conversions for e.g. DetrForSegmentation vs DetrForObjectDetection both map to \"detr\", so the segmentation-specific mask-" }, { "n": 44794, "d": "2026-05-14", "t": "Refacto GGUF weight conversion", "st": "OPEN", "by": "ArthurZucker", "a": 1741, "r": 610, "f": 13, "c": 16, "b": "Refactor GGUF loading to use our weight-conversion pipeline + swap the slow NumPy dequant for a pure-PyTorch port that also runs on MPS / CUDA (vendored from diffusers / existing places) 1. Use the weight-conversion framework _GGUF_ARCH_CONVERTERS is now a flat per-model_type list of WeightRenaming " }, { "n": 47611, "d": "2026-08-27", "t": "Resolve the Hub revision once per load instead of passing a private _commit_hash around", "st": "DRAFT", "by": "Wauplin", "a": 585, "r": 245, "f": 34, "c": 1, "b": "Follow-up on huggingface_hub v1.26.0, which shipped resolve_revision / ResolvedRevision. Same change as vllm-project/vllm#49990 (in production since) and huggingface/diffusers#14340. Alternative to #47583, same goal, opposite approach. ImportantBlocked on huggingface/huggingface_hub#4767, which make" }, { "n": 48373, "d": "2026-08-27", "t": "handle transpose during model loading with TP (Qwen3-VL-MoE)", "st": "OPEN", "by": "3outeille", "a": 127, "r": 8, "f": 8, "c": 15, "b": "Explanation Specifying a tp_plan for qwen3-vl-moe is not enough to fix #48308. A fix is as well needed during the loading process. The loading order is: checkpoint weight -> shard according to TP plan -> WeightConverter -> load local weight The tp_plan for qwen3_vl_moe is gate_up_proj: packed_colwis" } ], "d": "core_model_loading.py + WeightConverter/ConversionOps replace ad hoc state_dict renaming with declarative mappings; separates renaming from conversion, adds disk offload and TP-aware transpose.", "note": "GGUF conversion rewrite (#44794) and FSDP shard-on-read support (#44974) still open.", "links": [], "src": "standard.json", "span": [ "2025-11-03", "2026-08-27" ], "roll": { "owners": [ [ "ArthurZucker", 4 ], [ "Cyrilvallez", 4 ], [ "3outeille", 2 ], [ "SunMarc", 1 ] ], "n": 16, "add": 13818, "del": 9585, "files": 780, "commits": 600 } }, { "t": "Model patching API", "tag": "declarative library-wide monkey-patching", "s": "done", "prs": [ { "n": 43917, "d": "2026-03-02", "t": "Model patching API", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 1157, "r": 3, "f": 5, "c": 27, "b": "Integrates a library wide monkey patching API to fix #43284 and allow things like restructuring the expert weights and fusing the qkv projections. A concrete and complete example: Details from typing import Unpack import torch import torch.nn as nn from transformers import AutoModelForCausalLM, Conc" }, { "n": 43376, "d": "2026-03-19", "t": "Fix `KeyError` when patching mistral regex", "st": "MERGED", "by": "LeonardoEmili", "a": 7, "r": 1, "f": 2, "c": 6, "b": "Seems like the same fix_mistral_regex is provided multiple times, replacing get with pop to avoid running into KeyError fixes the issue. File \"/mambaforge/envs/cf/lib/python3.10/site-packages/transformers/models/auto/processing_auto.py\", line 400, in from_pretrained return PROCESSOR_MAPPING[type(con" }, { "n": 43603, "d": "2026-04-13", "t": "Fix unintended Hub metadata calls from _patch_mistral_regex", "st": "MERGED", "by": "vaibhav-research", "a": 53, "r": 2, "f": 3, "c": 4, "b": "TokenizersBackend._patch_mistral_regex() is a Mistral-specific tokenizer patch, but the current implementation may call huggingface_hub.model_info() during detection. That triggers an HTTP request to /api/models/<repo_id> and can occur even for non-Mistral repos in environments where outbound networ" }, { "n": 45359, "d": "2026-04-13", "t": "Fix Kimi-K2.5 tokenizer regression and _patch_mistral_regex AttributeError", "st": "MERGED", "by": "ArthurZucker", "a": 3, "r": 4, "f": 2, "c": 1, "b": "Fixes #45356 Summary Remove kimi_k25 from MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS: its remote TikTokenTokenizer is the only correct backend — the model has no tokenizer.json, and its added_tokens_decoder has non-sequential IDs (gaps + [UNK]/[PAD] at 163838/163839) that TokenizersBackend.add_tokens" }, { "n": 47798, "d": "2026-08-05", "t": "Fix patching in some models", "st": "MERGED", "by": "zucchini-nlp", "a": 31, "r": 51, "f": 5, "c": 2, "b": "CI reported failing slow tests" } ], "d": "Library-wide patching API (#43917) lets users restructure expert weights or fuse QKV projections declaratively instead of hand-patching modeling code; follow-up regex/tokenizer patch bugs fixed.", "note": null, "links": [], "src": "standard.json", "span": [ "2026-01-20", "2026-08-05" ], "roll": { "owners": [ [ "IlyasMoutawwakil", 1 ], [ "zucchini-nlp", 1 ], [ "LeonardoEmili", 1 ], [ "vaibhav-research", 1 ] ], "n": 5, "add": 1251, "del": 61, "files": 17, "commits": 40 } } ] }, { "name": "Training & the 5 parallelisms", "hue": 48, "blurb": "PP · TP · EP · DP · CP · FSDP behind one native API.", "items": [ { "t": "Parallelism interface foundation", "tag": "ParallelInterface + device_mesh base", "s": "done", "prs": [ { "n": 34800, "d": "2024-11-19", "t": "protect tensor parallel usage", "st": "MERGED", "by": "ArthurZucker", "a": 11, "r": 5, "f": 2, "c": 1, "b": "Fixes #34795" }, { "n": 37877, "d": "2025-05-20", "t": "parallelism goes brrr", "st": "MERGED", "by": "nmntz", "a": 1515, "r": 136, "f": 6, "c": 59, "b": "Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. Di" }, { "n": 38262, "d": "2025-05-21", "t": "Protect ParallelInterface", "st": "MERGED", "by": "ArthurZucker", "a": 22, "r": 17, "f": 3, "c": 2, "b": "" } ], "d": "Introduced the pluggable ParallelInterface abstraction ('parallelism goes brrr') plus device_mesh plumbing and usage protections that TP/FSDP/EP/CP all build on.", "note": null, "links": [], "src": "training.json", "span": [ "2024-11-19", "2025-05-21" ], "roll": { "owners": [ [ "ArthurZucker", 2 ], [ "nmntz", 1 ] ], "n": 3, "add": 1548, "del": 158, "files": 11, "commits": 62 } }, { "t": "TP-aware from_pretrained / rank-local shard loading", "tag": "rank-local sharded checkpoint loading", "s": "done", "prs": [ { "n": 36335, "d": "2025-02-26", "t": "Update form pretrained to make TP a first class citizen", "st": "MERGED", "by": "ArthurZucker", "a": 442, "r": 340, "f": 36, "c": 79, "b": "Match keys with loaded safe without opening the file makes sure that even when the checkpoint is sharded, we are only load our rank's slice of the tensor init nccl by default for \"most\" use cases. make device_map=\"auto\" equivalent to tp_plan=\"auto\" when nccl is initialized Updates after #36380 becau" }, { "n": 41445, "d": "2025-10-13", "t": "[`from_pretrained`] Small refactor `from_pretrained`: move around unrelated stuff", "st": "MERGED", "by": "ArthurZucker", "a": 627, "r": 548, "f": 21, "c": 41, "b": "As we are moving towards having a dynamic weight loader, we need to have a simple from_pretrained call. This should help" } ], "d": "Made TP a first-class citizen of from_pretrained so each rank loads only its slice of a sharded safetensors checkpoint instead of the full state dict, cutting load time.", "note": null, "links": [], "src": "training.json", "span": [ "2025-02-21", "2025-10-13" ], "roll": { "owners": [ [ "ArthurZucker", 2 ] ], "n": 2, "add": 1069, "del": 888, "files": 57, "commits": 120 } }, { "t": "FP8 / quantization-aware training", "tag": "fp8/mxfp4 backward + QAT", "s": "wip", "prs": [ { "n": 40180, "d": "2025-08-14", "t": "Enable native mxfp4 training support for GPT-OSS models", "st": "OPEN", "by": "akacmazz", "a": 921, "r": 28, "f": 4, "c": 1, "b": "Enable native mxfp4 training support for GPT-OSS models Fixes #40170 What does this PR do? This PR adds the foundational infrastructure to enable native mxfp4 training for GPT-OSS models, addressing the issue where users currently must convert 4-bit weights to bfloat16 for training (using 4X more me" }, { "n": 41419, "d": "2025-10-07", "t": "First QAT for Finegrained FP8", "st": "OPEN", "by": "MekkCyber", "a": 121, "r": 5, "f": 6, "c": 4, "b": "Adds QAT support for finegrained FP8, simply calling prepare_model should prepare the model for qat: quantization_config = FineGrainedFP8Config( weight_block_size=(128, 128), ) model = AutoModelForMaskedLM.from_pretrained(model_name, device_map=\"auto\", dtype=torch.bfloat16) quantizer = FineGrainedFP" }, { "n": 41897, "d": "2025-11-04", "t": "[FPQuant] MXFP8 and MXFP4 backwards support", "st": "MERGED", "by": "BlackSamorez", "a": 20, "r": 5, "f": 5, "c": 6, "b": "This PR adds MXFP4 and MXFP8 backwards support in combination with MXFP4 forward, allowing for lightning-fast QAT on Blackwell GPUs. It's blocked by QuTLASS v0.2.0 and fp_quant v0.3.0 release we're hoping to release within a day or two. Before submitting This PR fixes a typo or improves the docs (yo" } ], "d": "Added MXFP8/MXFP4 backward kernels enabling fp8 training (merged); a first quantization-aware-training PR for finegrained FP8 and native MXFP4 training for GPT-OSS remain open.", "note": null, "links": [], "src": "training.json", "span": [ "2025-08-14", "2025-11-04" ], "roll": { "owners": [ [ "BlackSamorez", 1 ], [ "MekkCyber", 1 ], [ "akacmazz", 1 ] ], "n": 3, "add": 1062, "del": 38, "files": 15, "commits": 11 } }, { "t": "Context Parallel (CP)", "tag": "context parallelism in Trainer", "s": "wip", "prs": [ { "n": 40205, "d": "2025-08-26", "t": "[Trainer] accelerate contextparallel support in trainer", "st": "MERGED", "by": "kashif", "a": 200, "r": 51, "f": 2, "c": 33, "b": "Add support for context parallel in the Trainer" }, { "n": 42496, "d": "2025-11-29", "t": "feat: allow CP with trainer", "st": "OPEN", "by": "kmehant", "a": 68, "r": 2, "f": 1, "c": 1, "b": "does batch padding to cp_size * 2 with shift_label calculation (we can omit this if you think this has to be outside the train loop) adds cp accelerate context loss calculation for cp and cp + dp cases Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if t" }, { "n": 48167, "d": "2026-08-21", "t": "Apply context parallelism to the evaluation path", "st": "MERGED", "by": "qgallouedec", "a": 4, "r": 2, "f": 1, "c": 2, "b": "With context parallelism enabled, eval_loss is reported as cp_size times its true value. Training loss is correct; only evaluation is wrong. Same everything, varying only cp_size: Run it once per cp_size (no config file needed): FSDP=\"--use_fsdp --fsdp_version 2 --fsdp_auto_wrap_policy TRANSFORMER_B" }, { "n": 48442, "d": "2026-08-31", "t": "Add `supports_context_parallel` to `PreTrainedModel`", "st": "OPEN", "by": "qgallouedec", "a": 24, "r": 0, "f": 1, "c": 2, "b": "@SunMarc suggested in huggingface/accelerate#4177 that this check belongs here rather than in accelerate It reads layer_types and sliding_window, which are transformers concepts, and there is already _supports_sdpa / _supports_flex_attn for this kind of capability flag. CP can only express full caus" } ], "d": "Added context-parallel support via Accelerate in Trainer (#40205) and to the evaluation path (#48167); open PRs add a supports_context_parallel model flag and broader CP+Trainer support.", "note": null, "links": [], "src": "training.json", "span": [ "2025-08-15", "2026-08-31" ], "roll": { "owners": [ [ "qgallouedec", 2 ], [ "kashif", 1 ], [ "kmehant", 1 ] ], "n": 4, "add": 296, "del": 55, "files": 5, "commits": 38 } }, { "t": "Training telemetry (Trackio)", "tag": "Trackio default training telemetry", "s": "done", "prs": [ { "n": 40950, "d": "2025-10-02", "t": "Deprecate Trackio environment variables and deploy to Spaces by default", "st": "MERGED", "by": "qgallouedec", "a": 52, "r": 15, "f": 2, "c": 8, "b": "This PR Deploys Trackio by default (instead of keeping local). See internal discussion here https://huggingface.slack.com/archives/C08SW1X12C8/p1758056496519289?thread_ts=1757706409.065919&cid=C08SW1X12C8 and Drops environment variables in favour of explicit args. See gradio-app/trackio#212 (review)" }, { "n": 41375, "d": "2025-10-09", "t": "`report_to` default changed to \"none\" + cleaning deprecated env var", "st": "MERGED", "by": "SunMarc", "a": 40, "r": 115, "f": 5, "c": 3, "b": "This PR changes the default for reporting to \"none\" as we move to v5. Moreover, we clean some deprecations messages around some env env + we deprecate logging_dir that was solely used for tensorboard. Instead the user needs to pass a global variable." }, { "n": 43371, "d": "2026-02-09", "t": "Remove GPU tracking from TrackioCallback and remove env var support", "st": "MERGED", "by": "qgallouedec", "a": 4, "r": 46, "f": 1, "c": 10, "b": "Trackio now natively supports GPU tracking, see https://github.com/gradio-app/trackio/releases/tag/trackio%400.14.0, so I suggest that we remove it from this callback." }, { "n": 45329, "d": "2026-04-13", "t": "Update `trackio` integration to use Buckets and \"freeze\" Space after training", "st": "MERGED", "by": "abidlabs", "a": 152, "r": 57, "f": 2, "c": 48, "b": "Updates TrackioCallback and TrainingArguments for the latest version of Trackio using HF Buckets as the backend, and control over creating a static Space for the Trackio dashboard during or at the end of training. These are now the TrainingArguments relevant for trackio: trackio_space_id — Only affe" } ], "d": "Made Trackio the default training telemetry integration (auto-deployed to Spaces, report_to default changed), then iterated on GPU-tracking removal and Buckets-based storage.", "note": null, "links": [], "src": "training.json", "span": [ "2025-09-17", "2026-04-13" ], "roll": { "owners": [ [ "qgallouedec", 2 ], [ "SunMarc", 1 ], [ "abidlabs", 1 ] ], "n": 4, "add": 248, "del": 233, "files": 10, "commits": 69 } }, { "t": "Core weight-loading refactor", "tag": "rewrote from_pretrained weight loading", "s": "done", "prs": [ { "n": 41445, "d": "2025-10-13", "t": "[`from_pretrained`] Small refactor `from_pretrained`: move around unrelated stuff", "st": "MERGED", "by": "ArthurZucker", "a": 627, "r": 548, "f": 21, "c": 41, "b": "As we are moving towards having a dynamic weight loader, we need to have a simple from_pretrained call. This should help" }, { "n": 41580, "d": "2025-11-13", "t": "Refactor weight loading", "st": "MERGED", "by": "ArthurZucker", "a": 8733, "r": 7189, "f": 586, "c": 387, "b": "CORE REFACTORING, loading, converting, logging More helpful debugging report when loading weights If you just want to fuse qkv: It can. You just need to make sure you change the model code and pouf! WeightConverter( [\"self_attn.q_proj\", \"self_attn.k_proj\", \"self_attn.v_proj\"], \"self_attn.qkv_proj\", " }, { "n": 44396, "d": "2026-03-03", "t": "[ Dynamic weight loader] fix remote code when format matches", "st": "MERGED", "by": "ArthurZucker", "a": 3, "r": 16, "f": 2, "c": 8, "b": "Some renaming should just never be applied when the weight format already matches. (this is actually regardless of remote code). This allows us to remove 1 test added in 8c1b5d3 ! FYI @vasqu we discussed offline with @Cyrilvallez and this makes sense mostly for \"errors\" or special cases were a model" } ], "d": "Refactored the core weight loading/conversion pipeline (ArthurZucker), the shared foundation the TP/FSDP sharded-loading paths above sit on.", "note": null, "links": [], "src": "training.json", "span": [ "2025-10-07", "2026-03-03" ], "roll": { "owners": [ [ "ArthurZucker", 3 ] ], "n": 3, "add": 9363, "del": 7753, "files": 609, "commits": 436 } }, { "t": "Sequence/Data Parallel (Ulysses/ALST)", "tag": "Ulysses sequence-parallel integration", "s": "wip", "prs": [ { "n": 41832, "d": "2025-11-21", "t": "HF Trainer: ALST/Ulysses sequence parallelism integration via HF Accelerate", "st": "MERGED", "by": "stas00", "a": 453, "r": 21, "f": 5, "c": 55, "b": "Integrates HF Accelerate's support for ALST/Ulysses sequence parallelism huggingface/accelerate#3817 into HF Trainer TODO: docs - no idea where? the FSDP/CP is not documented, or any parallelism for that matter. tests need to merge huggingface/accelerate#3817 need to wait for a new HF Accelerate rel" }, { "n": 46775, "d": "2026-06-24", "t": "New SP support in Trainer", "st": "OPEN", "by": "SunMarc", "a": 118, "r": 4, "f": 8, "c": 4, "b": "This PR adds compatibility with the new native SP support from accelerate. With this, you can perform long context training with FSDPv2, DeepSpeed or even DDP. Requires huggingface/accelerate#4084 cc @qgallouedec for viz" } ], "d": "Integrated DeepSpeed Ulysses/ALST sequence parallelism into HF Trainer via Accelerate (merged); a broader 'New SP support in Trainer' PR from SunMarc remains open.", "note": null, "links": [], "src": "training.json", "span": [ "2025-10-23", "2026-06-24" ], "roll": { "owners": [ [ "stas00", 1 ], [ "SunMarc", 1 ] ], "n": 2, "add": 571, "del": 25, "files": 13, "commits": 59 } }, { "t": "Native sharding via safetensors", "tag": "safetensors-only sharded checkpoints", "s": "done", "prs": [ { "n": 42734, "d": "2025-12-09", "t": "🚨🚨 [saving] Default to 50GB shards, and remove non-safe serialization", "st": "MERGED", "by": "Cyrilvallez", "a": 404, "r": 683, "f": 81, "c": 14, "b": "As per the title. This PR fixes both feature requests #42555 and #42556. Note that increasing the shard size was a decision that was only taken after very careful considerations, and the following benchmarks: Benchmarking scripts First, the following script was used to create model checkpoints of se" }, { "n": 42904, "d": "2025-12-18", "t": "Document new default shard size + dropped unsafe serialization", "st": "MERGED", "by": "Wauplin", "a": 14, "r": 0, "f": 1, "c": 4, "b": "Follow-up PR after #42734 from @Cyrilvallez (related to #42555 and #42556). Let's document the changes introduced in #42734 in the MIGRATION_GUIDE_V5.md." } ], "d": "Raised the default save shard size to 50GB and removed the non-safe (pickle) serialization path, making safetensors the sole native sharding format for saved checkpoints.", "note": "Chased the maintainer's memory of a personal draft PR: searched all 584 ArthurZucker PRs (titles+bodies) plus huggingface/safetensors repo for shard/safetensors/sharding/DCP terms - found none under his authorship. #42734 (Cyril Vallez, merged) is the real, substantive PR matching the description; treat the 'own draft PR' recollection as not corroborated.", "links": [], "src": "training.json", "span": [ "2025-12-09", "2025-12-18" ], "roll": { "owners": [ [ "Cyrilvallez", 1 ], [ "Wauplin", 1 ] ], "n": 2, "add": 418, "del": 683, "files": 82, "commits": 18 } }, { "t": "Unified DistributedConfig API", "tag": "single config for TP/FSDP/EP/CP", "s": "open", "prs": [ { "n": 42765, "d": "2026-02-17", "t": "Add distributed training CI", "st": "DRAFT", "by": "3outeille", "a": 618, "r": 29, "f": 8, "c": 17, "b": "it needs several PR: Add FSDP v2 (with dtensor for now) natively to Transformers (not rely on transformer anymore) => needs to compare correctness through DDP Check save/load of TP/FSDP/FSDP+TP distributed training ci use gloo backend no subprocess check if it bitwise convergent in terms of loss & g" }, { "n": 44989, "d": "2026-04-13", "t": "🚨 Distributed training API", "st": "DRAFT", "by": "3outeille", "a": 542, "r": 0, "f": 6, "c": 8, "b": "Distributed Training API Goal # torchrun --nproc_per_node=4 train_fsdp_tp.py import os import torch from torch.utils.data import DataLoader from datasets import load_dataset from transformers import AutoModelForCausalLM, AutoTokenizer from transformers.distributed import DistributedConfig from trans" } ], "d": "Open draft proposing one DistributedConfig entry point unifying TP/FSDP/EP/CP setup, plus a dedicated distributed-training CI; both still draft/open.", "note": null, "links": [], "src": "training.json", "span": [ "2025-12-10", "2026-04-13" ], "roll": { "owners": [ [ "3outeille", 2 ] ], "n": 2, "add": 1160, "del": 29, "files": 14, "commits": 25 } }, { "t": "Expert Parallel (EP) training correctness", "tag": "EP training fixes (grads, Trainer, FSDP2)", "s": "wip", "prs": [ { "n": 43722, "d": "2026-03-04", "t": "🚨 fix + tests dense & MoE TP all reduce (decoder only)", "st": "MERGED", "by": "3outeille", "a": 1381, "r": 796, "f": 77, "c": 188, "b": "Let's make sure it works for decoder only first (We skip VLM + Encoder-decoder for now) Introduction, forward, backward, generation (with convert mapping triggering) test agains TP vs non-TP baseline from transformers import AutoModelForCausalLM, AutoTokenizer import torch import os from torch.distr" }, { "n": 45621, "d": "2026-05-04", "t": "Better Grouped GEMM + EP", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 394, "r": 219, "f": 5, "c": 36, "b": "The idea is that we shouldn't be clamping experts ids at all, clamping them makes their tokens get prjected as if they were routed to the last expert, instead, we should let the sentinels be: moved to the queue by the sorting (num_experts/sentinel is the biggest value) dropped by the histogram (beca" }, { "n": 45662, "d": "2026-05-13", "t": "Fix EP + FSDP2: experts silently overwritten by rank-0 broadcast", "st": "OPEN", "by": "AmineDiro", "a": 63, "r": 8, "f": 5, "c": 6, "b": "Loading a MoE model with Expert Parallelism (distributed_config=DistributedConfig(enable_expert_parallel=True)) and then calling accelerator.prepare with FSDP2 silently loads wrong of the experts on ranks. The model trains, but on broken weights. Tested on Qwen3-30B-A3B with 128 experts and EP=8. Th" }, { "n": 48208, "d": "2026-08-27", "t": "Fix expert parallelism through Trainer", "st": "OPEN", "by": "qgallouedec", "a": 67, "r": 1, "f": 3, "c": 2, "b": "Stacked on #48200, this PR makes expert parallelism reachable through Trainer, and without the gradient fixes there, the training it unlocks is wrong. Loading a model with DistributedConfig(tp_size=N, enable_expert_parallel=True) and handing it to Trainer fails on main before the first step complete" }, { "n": 48205, "d": "2026-08-27", "t": "Fix expert-parallel training: NaN gradients and missing gradient contributions", "st": "OPEN", "by": "qgallouedec", "a": 10, "r": 0, "f": 2, "c": 3, "b": "Two bugs in expert-parallel training (enable_expert_parallel=True), found while validating FSDP2 × EP (#48204): full fine-tuning of an EP-sharded MoE trains on wrong gradients, then collapses to NaN on step 2. +14 lines total. 1. NaN from uninitialized memory. torch._grouped_mm leaves the sentinel-t" }, { "n": 48201, "d": "2026-08-27", "t": "Skip the expert-parallel sentinel masking when expert parallelism is off", "st": "OPEN", "by": "qgallouedec", "a": 16, "r": 8, "f": 2, "c": 6, "b": "grouped_mm_experts_forward runs the expert-parallel sentinel dance unconditionally: transformers/src/transformers/integrations/moe.py Lines 420 to 421 in d56c55b sentinel_mask = (expert_ids_g >= self.num_experts).unsqueeze(-1) expert_ids_g.clamp_(max=self.num_experts - 1) transformers/src/transforme" } ], "d": "Fixed dense & MoE TP all-reduce and shipped better grouped-GEMM+EP kernels; several open PRs still fix EP-through-Trainer NaN gradients, sentinel masking, and EP+FSDP2 expert overwrite bugs.", "note": null, "links": [], "src": "training.json", "span": [ "2026-02-03", "2026-08-27" ], "roll": { "owners": [ [ "qgallouedec", 3 ], [ "3outeille", 1 ], [ "IlyasMoutawwakil", 1 ], [ "AmineDiro", 1 ] ], "n": 6, "add": 1931, "del": 1032, "files": 94, "commits": 241 } }, { "t": "Pipeline Parallel (PP)", "tag": "native pipeline-parallel inference", "s": "wip", "prs": [ { "n": 44601, "d": "2026-03-12", "t": "[Distributed] Add PP support natively", "st": "DRAFT", "by": "ArthurZucker", "a": 510, "r": 2, "f": 4, "c": 5, "b": "Goal is to be able to run a model with both PP and TP." }, { "n": 47289, "d": "2026-08-25", "t": "Pipeline parallel naive inference", "st": "MERGED", "by": "3outeille", "a": 758, "r": 10, "f": 9, "c": 317, "b": "Summary Adding a very naive pipeline parallel inference engine as a baseline (no specific schedule such as gpipe, 1f1b yet) supports tied/untied weight embeddings To make it work seamless with generate(), we are broadcasting every logits to every ranks. In long term, we want only the last stage to h" } ], "d": "Added naive pipeline-parallel inference support (#47289, merged); a native PP-support draft opened by ArthurZucker (#44601) remains open.", "note": null, "links": [], "src": "training.json", "span": [ "2026-03-11", "2026-08-25" ], "roll": { "owners": [ [ "3outeille", 1 ], [ "ArthurZucker", 1 ] ], "n": 2, "add": 1268, "del": 12, "files": 13, "commits": 322 } }, { "t": "FSDP2 + TP native distributed save/load", "tag": "native FSDP2+TP save/load, no full gather", "s": "done", "prs": [ { "n": 44996, "d": "2026-03-26", "t": " from_pretrained distributed refactor (FSDP2 + TP)", "st": "MERGED", "by": "3outeille", "a": 286, "r": 283, "f": 10, "c": 13, "b": "Introduce DistributedConfig DistributedConfig(tp_size=2, fsdp_size=2) # plans default to \"auto\" replaces passing separate tp_plan, tp_size, fsdp_plan kwargs. Sizes auto-fill (specify one, the other defaults to 1). Plans default to \"auto\" when a size is given. init_device_mesh + distribute_model init" }, { "n": 45409, "d": "2026-04-14", "t": "from_pretrained orchestration + distributed save/load", "st": "MERGED", "by": "3outeille", "a": 310, "r": 49, "f": 5, "c": 12, "b": "Summary Full distributed_config integration in from_pretrained() — mesh creation, apply TP + FSDP, attach model.device_mesh gather_full_state_dict() for streaming DTensor→full tensor saving (rank 0 only) convert_strided_to_shard() / restore_strided_from_shard() for DCP compatibility with _StridedSha" }, { "n": 45028, "d": "2026-05-19", "t": "FSDP + TP & native save/load distributed", "st": "MERGED", "by": "3outeille", "a": 6384, "r": 3015, "f": 287, "c": 133, "b": "Highlights One config, one call. Pass a DistributedConfig to from_pretrained to apply TP, FSDP2, or both — sharding happens during weight load (shard-on-read), not after. model = AutoModelForCausalLM.from_pretrained( name, distributed_config=DistributedConfig( tp_size=2, fsdp_size=2, enable_sequence" }, { "n": 46102, "d": "2026-05-26", "t": "init FSDP through from_pretrained", "st": "MERGED", "by": "3outeille", "a": 20, "r": 118, "f": 3, "c": 9, "b": "Instantiate FSDP through .from_pretrained instead" }, { "n": 46707, "d": "2026-07-09", "t": "Add native FSDP2 module + migration", "st": "MERGED", "by": "3outeille", "a": 300, "r": 80, "f": 7, "c": 53, "b": "#46705 applying FSDP + test_fsdp_mixin.py will be added in PR #46990" }, { "n": 46717, "d": "2026-07-13", "t": "Shard on read Dtensor aware", "st": "MERGED", "by": "3outeille", "a": 845, "r": 47, "f": 4, "c": 96, "b": "? This PR adds shard-on-read support for loading DTensor-sharded model parameters during from_pretrained. Instead of materializing the full checkpoint tensor on every rank and sharding afterward, each rank reads only the slice it needs from the checkpoint, based on the target parameter’s DTensor mes" }, { "n": 47357, "d": "2026-07-23", "t": "Add FSDP CI and end-to-end FSDP tests + save fsdp", "st": "MERGED", "by": "3outeille", "a": 960, "r": 82, "f": 16, "c": 40, "b": "Summary Part of FSDP orchestration stack (2/2). Replaces #46990. #47352 (1/2) → this PR (2/2) Add FSDP DCP (distributed_checkpoint=True) paths in save_pretrained Add FSDP gather into single checkpoint saving Add FSDPTesterMixin and tests/test_fsdp_mixin.py (load/save/generation) Add base_model_fsdp_" } ], "d": "Multi-PR effort (3outeille) giving native FSDP2+TP support with distributed save/load: shard-on-read DTensor-aware loading, a native FSDP2 module, FSDP init through from_pretrained, and an FSDP CI/test suite.", "note": "Earlier attempt #44083 'FSDP2 native support in transformers' was closed unmerged before this line landed.", "links": [], "src": "training.json", "span": [ "2026-03-25", "2026-07-23" ], "roll": { "owners": [ [ "3outeille", 7 ] ], "n": 7, "add": 9105, "del": 3674, "files": 332, "commits": 356 } }, { "t": "FSDP2 + Expert Parallel 2-D mesh umbrella", "tag": "FSDP2 x EP on a (fsdp, tp) mesh", "s": "wip", "prs": [ { "n": 44974, "d": "2026-04-14", "t": "Refactor core_model_loading to support FSDP shard-on-read loading", "st": "OPEN", "by": "3outeille", "a": 428, "r": 82, "f": 5, "c": 3, "b": "TODO: Saving seems to take a bit of time tho. Need investigation Need to check if it works in 1D (FSDP or TP)and 2D (FSDP + TP). Running the script from #44996 (env_pr-44974-fsdp-core-model-loading) ➜ pr-44974-fsdp-core-model-loading git:(pr-44974-fsdp-core-model-loading) ✗ torchrun --nproc_per_node" }, { "n": 48204, "d": "2026-08-25", "t": "[Umbrella] Enable FSDP2 + expert parallelism via a 2-D (fsdp, tp) device mesh", "st": "DRAFT", "by": "qgallouedec", "a": 424, "r": 8, "f": 8, "c": 18, "b": "ImportantDraft to unblock experimentation (fine-tuning large MoEs in TRL without replicating dense weights on every EP rank). The Trainer-side fallbacks are deliberately blunt; we'll migrate to whatever native design transformers lands. Today: tp_size=N, fsdp_size=M with both > 1 raises FSDP+TP is n" }, { "n": 48237, "d": "2026-08-27", "t": "Handle empty FSDP shards in sharded loading: crash and hang fix", "st": "OPEN", "by": "qgallouedec", "a": 14, "r": 0, "f": 2, "c": 5, "b": "Fixes loading models whose FSDP sharding leaves a rank an empty local shard, which today crashes some ranks and silently hangs the rest. Uneven sharding makes this easy to hit: GLM-4.6 has 160 experts; ep=8 leaves 20 per EP rank, and chunking 20 over fsdp=8 gives [3,3,3,3,3,3,2,0] (every fsdp-index-" }, { "n": 48227, "d": "2026-08-27", "t": "Warm the page cache before sharded loading (opt-in via `HF_SHARD_PREFETCH`)", "st": "OPEN", "by": "qgallouedec", "a": 25, "r": 0, "f": 1, "c": 4, "b": "Adds an opt-in page-cache prefetch to sharded model loading: with HF_SHARD_PREFETCH=<threads>, the local ranks split the checkpoint's shard list between them and stream it into the page cache with large sequential reads before loading begins. The load itself then runs at memory speed. Why The loader" }, { "n": 48228, "d": "2026-09-01", "t": "Init the process group with a load-scaled timeout for sharded loading", "st": "MERGED", "by": "qgallouedec", "a": 7, "r": 1, "f": 1, "c": 5, "b": "When transformers initializes torch.distributed itself (the DistributedConfig loading path), the process group gets the default 10-minute NCCL watchdog. Too short for sharded loading of large checkpoints: loading is tens of minutes (GLM-4.6, 665 GiB, 16 ranks: 25–28 min) ranks finish far apart: meas" } ], "d": "Open umbrella (qgallouedec) extending native distributed save/load to a combined (fsdp, tp) device mesh for expert-parallel MoE models: empty-shard handling, page-cache warmup, load-scaled init timeouts, shard-on-read refactor.", "note": null, "links": [], "src": "training.json", "span": [ "2026-04-13", "2026-09-01" ], "roll": { "owners": [ [ "qgallouedec", 4 ], [ "3outeille", 1 ] ], "n": 5, "add": 898, "del": 91, "files": 17, "commits": 35 } }, { "t": "Tensor Parallel native DTensor rewrite", "tag": "TP rewritten on torch DTensor", "s": "done", "prs": [ { "n": 47579, "d": "2026-08-21", "t": "🚨 TP dtensor API inference + training", "st": "MERGED", "by": "3outeille", "a": 1234, "r": 2161, "f": 27, "c": 244, "b": "Summary This PR replaces the legacy Transformers tensor-parallel implementation with a DTensor-native backend while preserving efficient inference for kernels that operate on local tensors + loading time. We outsource all the placement and parameters sharding logic to Dtensor which will enable us to" }, { "n": 48300, "d": "2026-08-26", "t": "Restore BC for the tensor-parallel API", "st": "MERGED", "by": "ArthurZucker", "a": 188, "r": 55, "f": 16, "c": 28, "b": "Restore tp_plan from from_pretrained() and add deprecation cycle to it More guarding regarding qk_rope_head_dim attributes in MLA + expert dim floor-divided by mesh size Doc updates" } ], "d": "Rewrote tensor-parallel inference and training to use torch DTensor natively instead of manual tensor slicing; #47579 was a breaking change (marked with a siren), #48300 restored backward compatibility afterward.", "note": null, "links": [], "src": "training.json", "span": [ "2026-06-17", "2026-08-26" ], "roll": { "owners": [ [ "3outeille", 1 ], [ "ArthurZucker", 1 ] ], "n": 2, "add": 1422, "del": 2216, "files": 43, "commits": 272 } }, { "t": "Independent device mesh per TP replica", "tag": "multiple TP replicas per process", "s": "open", "prs": [ { "n": 48383, "d": "2026-08-28", "t": "Give each tensor parallel model in a process its own device mesh", "st": "OPEN", "by": "qgallouedec", "a": 31, "r": 3, "f": 2, "c": 3, "b": "Two tensor parallel models in the same process silently share a communicator, which deadlocks as soon as they are used at the same time. initialize_tensor_parallelism built its mesh with init_device_mesh(device_type, (tp_size,)). That mesh has no dimension names, and when tp_size == world_size torch" } ], "d": "Open PR giving each tensor-parallel model in a process its own device mesh so several TP replicas can coexist, e.g. for train-while-generate setups.", "note": null, "links": [], "src": "training.json", "span": [ "2026-08-27", "2026-08-28" ], "roll": { "owners": [ [ "qgallouedec", 1 ] ], "n": 1, "add": 31, "del": 3, "files": 2, "commits": 3 } } ] }, { "name": "Quantization & kernels", "hue": 350, "blurb": "Simplify, expand to other libraries, quantization by default.", "items": [ { "t": "Hub kernels / kernelize system", "tag": "hub-kernels-kernelize", "s": "wip", "prs": [ { "n": 39474, "d": "2025-07-22", "t": "Kernels flash attn", "st": "MERGED", "by": "ArthurZucker", "a": 404, "r": 419, "f": 13, "c": 46, "b": "pip install transformers[torch] kernels from transformers import AutoModelForCausalLM, AutoTokenizer import time model = AutoModelForCausalLM.from_pretrained( \"meta-llama/Llama-3.2-3B-Instruct\", device_map=\"auto\", torch_dtype=\"auto\", attn_implementation=\"flash_attention_2\", ).eval() tokenizer = Auto" }, { "n": 45666, "d": "2026-05-26", "t": "Extended n-to-1 kernel fusion via `KernelConfig` ", "st": "OPEN", "by": "michaelbenayoun", "a": 1110, "r": 25, "f": 6, "c": 58, "b": "Extends the KernelConfig API with two orthogonal capabilities: Module fusion: specify how Transformers modules should be fused together before a custom kernel is applied (n-to-1 replacement). Weight transformation in the live model: handle cases where a kernel expects weights in a different layout t" }, { "n": 45974, "d": "2026-05-28", "t": "Enable kernels-community/metal-flash-sdpa on MPS", "st": "MERGED", "by": "ArthurZucker", "a": 30, "r": 5, "f": 2, "c": 6, "b": "Enables kernels-community/metal-flash-sdpa for generate/generate_batch on MPS: synthesize cu_seqlens + .contiguous() in the no-padding branch of _flash_attention_forward, and fix MPS memory accounting in continuous batching. Bench (gsm8k 100 samples, Qwen2.5-0.5B-Instruct, MPS fp16, generate_batch):" }, { "n": 46339, "d": "2026-06-09", "t": "Extended & simplified n-to-1 kernel fusion via KernelConfig", "st": "MERGED", "by": "michaelbenayoun", "a": 320, "r": 25, "f": 4, "c": 72, "b": "Extends the KernelConfig API with two orthogonal capabilities: Module fusion: specify how Transformers modules should be fused together before a custom kernel is applied (n-to-1 replacement). Parameter transformation: handle cases where a kernel expects weights in a different layout than the origina" }, { "n": 46520, "d": "2026-06-18", "t": "Kernelize refactor", "st": "MERGED", "by": "michaelbenayoun", "a": 99, "r": 92, "f": 7, "c": 12, "b": "Small refactor of things related to kernels inside a kernelize function." }, { "n": 46755, "d": "2026-06-19", "t": "[`Kernels`] Trigger proper kernelization on `use_kernels=True`", "st": "MERGED", "by": "vasqu", "a": 14, "r": 12, "f": 3, "c": 2, "b": "As per title, regression in #46520. Partially on me, should've extended the test from the get go When we set use_kernels=True, we call kernelize but ignore any underlying registration of our kernel mappings leading to an empty kernelization (i.e. no op). Imo, this is not intended as we would expect " }, { "n": 46754, "d": "2026-06-24", "t": "Documentation for the kernel API", "st": "MERGED", "by": "michaelbenayoun", "a": 204, "r": 0, "f": 4, "c": 5, "b": "This PR provides the documentation for the big API changes and extension provided by #46339" }, { "n": 47863, "d": "2026-08-17", "t": "Bump default flash-attn2 hub kernel version to v3", "st": "MERGED", "by": "jiqing-feng", "a": 19, "r": 3, "f": 1, "c": 6, "b": "What load_and_register_attn_kernel hardcodes version=1 as the default major version for every attention kernel loaded from the Hub. Pin kernels-community/flash-attn2 to v3 through a small per-repo table instead. Why The v1 branch of kernels-community/flash-attn2 only ships builds up to Torch 2.12, s" }, { "n": 48335, "d": "2026-08-31", "t": "[`Kernels`] Add swiglu and geglu MLP across models", "st": "OPEN", "by": "vasqu", "a": 162, "r": 26, "f": 125, "c": 9, "b": "As per title, adds the tiled mlp kernels for swiglu / geglu across the models cc @kashif @qgallouedec Quoting Quenting on the effects: On Qwen3-8B's MLP shape at 131k tokens, forward+backward peak goes 22.56 GB to 7.97 GB, gradient error 3.5e-03 and flat across sequence lengths. NOTE: There is the p" }, { "n": 48185, "d": "2026-09-01", "t": "Warn once when a hub-kernel function falls back to its reference PyTorch path", "st": "OPEN", "by": "qgallouedec", "a": 13, "r": 0, "f": 1, "c": 11, "b": "use_kernel_func_from_hub_with_fallback resolves, in order, a hub kernel, the original package, and finally the pure-PyTorch reference. The third case is completely silent. For chunk_gated_delta_rule the reference is >20× slower than the kernel, so a user training Qwen3-Next / Qwen3.5 / Qwen3.6-MoE /" } ], "d": "KernelConfig fuses N Hub-kernel modules into one and remaps weight layouts; use_kernels=True now triggers real kernelization; default flash-attn2 hub kernel bumped to v3 (#47863).", "note": null, "links": [], "src": "quant.json", "span": [ "2025-07-17", "2026-09-01" ], "roll": { "owners": [ [ "michaelbenayoun", 4 ], [ "ArthurZucker", 2 ], [ "vasqu", 2 ], [ "jiqing-feng", 1 ] ], "n": 10, "add": 2375, "del": 607, "files": 166, "commits": 227 } }, { "t": "MoE / expert kernels: grouped_mm, DeepGEMM, EP", "tag": "moe-expert-kernels", "s": "done", "prs": [ { "n": 39501, "d": "2025-07-25", "t": "Add ep", "st": "MERGED", "by": "ArthurZucker", "a": 436, "r": 186, "f": 9, "c": 25, "b": "Add support for expert parallel!" }, { "n": 43227, "d": "2026-01-22", "t": "GptOss experts implementation", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 359, "r": 223, "f": 8, "c": 35, "b": "[wip] Fixes #43193 @vasqu @mattteochen Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it" }, { "n": 43435, "d": "2026-02-02", "t": "XPU now supports MoE kernel(MegaBlocks) implementation", "st": "MERGED", "by": "YangKai0616", "a": 87, "r": 36, "f": 3, "c": 9, "b": "XPU now supports the MoE kernel(MegaBlocks) implementation, and includes XPU checks for GPT-OSS UTs compatibility." }, { "n": 43977, "d": "2026-02-16", "t": "stable grouped_mm API", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 48, "r": 23, "f": 5, "c": 9, "b": "Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. Di" }, { "n": 44043, "d": "2026-02-18", "t": "`grouped_mm` fallback", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 170, "r": 162, "f": 31, "c": 8, "b": "With this PR, grouped_mm impl will always be default when the model uses experts impl (even in old torch versions) and will no longer fail if you are torch compiling an fp16/fp32 model, the impl will succeed in all cases, except cuda-graphs (common failure case in all grouped_mm impls). Before submi" }, { "n": 44231, "d": "2026-03-11", "t": "[Performance] FP8 Grouped and Batched Matmuls", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 413, "r": 578, "f": 5, "c": 27, "b": "up to 30x faster than current fp8 experts, the kernels are also tailored for full torch.compile and cuda graphs compatibility. ============================================================ FP8Expert parity: eager / batched_mm / grouped_mm ============================================================ [" }, { "n": 44895, "d": "2026-03-23", "t": "Add static FP8 expert support ", "st": "MERGED", "by": "SunMarc", "a": 57, "r": 10, "f": 1, "c": 2, "b": "This PR enables static FP8 experts. This also works on multi-gpu with device-map. A fix for that was to set was to set torch.cuda.set_device(). Triton's JIT compiler uses he active device context to determine which GPU to compile and launch the kernel on. If tensors are on GPU:1 but the active conte" }, { "n": 45001, "d": "2026-04-10", "t": "Add cuda compatibility check for using `grouped_mm`", "st": "MERGED", "by": "Sai-Suraj-27", "a": 20, "r": 1, "f": 1, "c": 8, "b": "For torch>=2.10.0, the minimum CUDA compute capability requirement for torch.nn.functional.grouped_mm is 8.0. For torch==2.8.0 for torch._grouped_mm(), the minimum CUDA compute capability requirement is 9.0. Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of " }, { "n": 45621, "d": "2026-05-04", "t": "Better Grouped GEMM + EP", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 394, "r": 219, "f": 5, "c": 36, "b": "The idea is that we shouldn't be clamping experts ids at all, clamping them makes their tokens get prjected as if they were routed to the last expert, instead, we should let the sentinels be: moved to the queue by the sorting (num_experts/sentinel is the biggest value) dropped by the histogram (beca" }, { "n": 46743, "d": "2026-06-19", "t": "Fall back to the for-loop grouped_mm on CPU", "st": "MERGED", "by": "Sunt-ing", "a": 10, "r": 7, "f": 1, "c": 3, "b": "The shared MoE experts path in integrations/moe.py runs each expert's matmul through torch._grouped_mm / torch.nn.functional.grouped_mm, and falls back to a per-expert for-loop of torch.mm when grouped matmul is not usable. _can_use_grouped_mm decides which path to take. On CPU it returned True when" }, { "n": 47154, "d": "2026-07-08", "t": "Better deepgemm errors", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 131, "r": 39, "f": 4, "c": 6, "b": "Okay so I fully went down the rabbit hole of DeepGEMM requirements, and what can and can't be done with the Hub kernel build. TLDR; Statically verifying that the environment can compile DeepGEMM kernels before loading them lets us avoid downloading the kernel files entirely (there are too many, and " }, { "n": 47126, "d": "2026-07-09", "t": "[Fix] Make DeepGEMM triton fallback more robust", "st": "MERGED", "by": "remi-or", "a": 152, "r": 34, "f": 4, "c": 5, "b": "This PR makes the triton fallback to DeepGEMM more robust: in case CUDA_HOME is not set, the current behavior is to fail with an AssertionError which is not caught by the try block because it specializes on ImportError. To avoid this, we call get_mk_alignment() before creating the DeepGemm object in" }, { "n": 47313, "d": "2026-07-14", "t": "DeepGEMM shouldn't pad on SM90", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 15, "r": 2, "f": 1, "c": 1, "b": "Fixes # (issue) Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. These often are low-quality, or fix extremely minor issues that occur rarely or never in practice. As a result, we're instituting a rule that fir" }, { "n": 47323, "d": "2026-07-16", "t": "Fix deepgemm on multiple devices", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 112, "r": 58, "f": 4, "c": 10, "b": "Fixes # (issue) Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. These often are low-quality, or fix extremely minor issues that occur rarely or never in practice. As a result, we're instituting a rule that fir" }, { "n": 47334, "d": "2026-07-28", "t": "Kernels and loaders robustification", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 1768, "r": 199, "f": 8, "c": 25, "b": "use a global var to avoid dynamo warnings (originally implemented in an fp8 pr that's still not merged) mock env to test loaders and kernels behaviors, mostly written by the agent with my guidance (compile and env checks) remove populator hop since loader can be the populator in this case (since it " }, { "n": 48361, "d": "2026-08-27", "t": "Add --experts-implementation flag to transformers serve", "st": "OPEN", "by": "NathanHB", "a": 8, "r": 0, "f": 2, "c": 1, "b": "transformers serve had no way to pass experts_implementation to from_pretrained. For FP8 MoE checkpoints whose TP/EP plan requires experts_implementation=\"deepgemm_megamoe\" (the only key in FP8Experts._impl_tp_layer_overrides), this meant update_tp_plan always crashed with AttributeError: 'NoneType'" } ], "d": "FP8 grouped/batched expert matmuls 'up to 30x faster than current fp8 experts' and torch.compile/cuda-graph safe (#44231); grouped_mm API stabilized, DeepGEMM hardened across devices.", "note": null, "links": [], "src": "quant.json", "span": [ "2025-07-17", "2026-08-27" ], "roll": { "owners": [ [ "IlyasMoutawwakil", 9 ], [ "Sai-Suraj-27", 1 ], [ "SunMarc", 1 ], [ "Sunt-ing", 1 ] ], "n": 16, "add": 4180, "del": 1777, "files": 92, "commits": 210 } }, { "t": "Quantization-aware training & FP8 backward", "tag": "qat-fp8-backward-training", "s": "wip", "prs": [ { "n": 40180, "d": "2025-08-14", "t": "Enable native mxfp4 training support for GPT-OSS models", "st": "OPEN", "by": "akacmazz", "a": 921, "r": 28, "f": 4, "c": 1, "b": "Enable native mxfp4 training support for GPT-OSS models Fixes #40170 What does this PR do? This PR adds the foundational infrastructure to enable native mxfp4 training for GPT-OSS models, addressing the issue where users currently must convert 4-bit weights to bfloat16 for training (using 4X more me" }, { "n": 41419, "d": "2025-10-07", "t": "First QAT for Finegrained FP8", "st": "OPEN", "by": "MekkCyber", "a": 121, "r": 5, "f": 6, "c": 4, "b": "Adds QAT support for finegrained FP8, simply calling prepare_model should prepare the model for qat: quantization_config = FineGrainedFP8Config( weight_block_size=(128, 128), ) model = AutoModelForMaskedLM.from_pretrained(model_name, device_map=\"auto\", dtype=torch.bfloat16) quantizer = FineGrainedFP" }, { "n": 41897, "d": "2025-11-04", "t": "[FPQuant] MXFP8 and MXFP4 backwards support", "st": "MERGED", "by": "BlackSamorez", "a": 20, "r": 5, "f": 5, "c": 6, "b": "This PR adds MXFP4 and MXFP8 backwards support in combination with MXFP4 forward, allowing for lightning-fast QAT on Blackwell GPUs. It's blocked by QuTLASS v0.2.0 and fp_quant v0.3.0 release we're hoping to release within a day or two. Before submitting This PR fixes a typo or improves the docs (yo" } ], "d": "FPQuant merged MXFP4/MXFP8 backward pass enabling QAT (#41897); native MXFP4 training for GPT-OSS (#40180) and QAT for finegrained FP8 (#41419) remain open.", "note": null, "links": [], "src": "quant.json", "span": [ "2025-08-14", "2025-11-04" ], "roll": { "owners": [ [ "BlackSamorez", 1 ], [ "akacmazz", 1 ], [ "MekkCyber", 1 ] ], "n": 3, "add": 1062, "del": 38, "files": 15, "commits": 11 } }, { "t": "Quantization method expansion: bnb/GPTQ/AWQ/compressed-tensors/torchao/fbgemm", "tag": "quant-methods-expansion", "s": "wip", "prs": [ { "n": 41897, "d": "2025-11-04", "t": "[FPQuant] MXFP8 and MXFP4 backwards support", "st": "MERGED", "by": "BlackSamorez", "a": 20, "r": 5, "f": 5, "c": 6, "b": "This PR adds MXFP4 and MXFP8 backwards support in combination with MXFP4 forward, allowing for lightning-fast QAT on Blackwell GPUs. It's blocked by QuTLASS v0.2.0 and fp_quant v0.3.0 release we're hoping to release within a day or two. Before submitting This PR fixes a typo or improves the docs (yo" }, { "n": 42256, "d": "2025-11-18", "t": "Integrate Core AutoAWQ Inference Components into Transformers", "st": "OPEN", "by": "fanqiNO1", "a": 0, "r": 0, "f": 0, "c": 1, "b": "Since the AutoAWQ repository has been archived, this PR proposes integrating its essential inference-related components into transformers to ensure continued support and maintenance. Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Di" }, { "n": 42773, "d": "2025-12-11", "t": "[Quantization] FBgemm FP8 for XPU", "st": "MERGED", "by": "MekkCyber", "a": 135, "r": 63, "f": 3, "c": 6, "b": "Fixes some issues with #42639" }, { "n": 45609, "d": "2026-04-23", "t": "make it possible to ser/deser HF MoE models with torchao", "st": "DRAFT", "by": "vkuzo", "a": 20, "r": 17, "f": 1, "c": 1, "b": "Summary: TODO(before review): write me in more detail Test Plan: TODO write me What does this PR do? Fixes # (issue) Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. We are currently bottlenecked by our ability" }, { "n": 47152, "d": "2026-07-22", "t": "[Quantization]: Refactor is_quantization_compressed for format-based detection", "st": "MERGED", "by": "rigen1048", "a": 24, "r": 3, "f": 2, "c": 4, "b": "Problem Models saved with quantization_status: \"frozen\" but in a quantized format (e.g. RedHatAI/Llama-3.2-1B-Instruct-FP8) do not decompress when run_compressed=False. The current check is too strict because it requires status == \"compressed\", causing raw FP8 values to remain in the weights. Soluti" }, { "n": 47797, "d": "2026-08-05", "t": "Migrate torchao integration off deleted torchao.dtypes", "st": "MERGED", "by": "vkuzo", "a": 22, "r": 95, "f": 3, "c": 1, "b": "torchao removed the torchao/dtypes package (AffineQuantizedTensor and the Layout classes). This updates the transformers torchao integration to stop importing from torchao.dtypes so it no longer crashes against current torchao, mirroring the approach in transformers #45321. Changes: integrations/tor" }, { "n": 47579, "d": "2026-08-21", "t": "🚨 TP dtensor API inference + training", "st": "MERGED", "by": "3outeille", "a": 1234, "r": 2161, "f": 27, "c": 244, "b": "Summary This PR replaces the legacy Transformers tensor-parallel implementation with a DTensor-native backend while preserving efficient inference for kernels that operate on local tensors + loading time. We outsource all the placement and parameters sharding logic to Dtensor which will enable us to" }, { "n": 48343, "d": "2026-08-26", "t": "Fix missing FP8 TP layer overrides", "st": "MERGED", "by": "changwangss", "a": 1, "r": 1, "f": 1, "c": 4, "b": "Treats an unmapped FP8 experts implementation as having no TP layer overrides. This prevents update_tp_plan() from calling .get() on None for regular FP8 model configs while preserving grouped-GEMM scale plan handling. Adds a focused regression test for configs without _experts_implementation. Fixes" } ], "d": "FPQuant MXFP4/MXFP8 backward: 'lightning-fast QAT on Blackwell GPUs' (#41897); FBgemm FP8 on XPU; torchao off deleted dtypes; AutoAWQ folded in-tree after upstream archival (#42256, open).", "note": null, "links": [], "src": "quant.json", "span": [ "2025-10-27", "2026-08-26" ], "roll": { "owners": [ [ "vkuzo", 2 ], [ "BlackSamorez", 1 ], [ "MekkCyber", 1 ], [ "rigen1048", 1 ] ], "n": 8, "add": 1456, "del": 2345, "files": 42, "commits": 267 } }, { "t": "MacOS/Apple: GGUF, Metal/MPS kernels, MLX", "tag": "macos-metal-gguf-mlx", "s": "open", "prs": [ { "n": 44348, "d": "2026-02-28", "t": "Enable MetalConfig to load pre-quantized MLX models from HuggingFace Hub", "st": "OPEN", "by": "n0kovo", "a": 341, "r": 32, "f": 4, "c": 4, "b": "Summary Most quantized models for Apple Silicon on the Hub are in MLX format. The MetalConfig quantization backend supports on-the-fly quantization of standard checkpoints but cannot load pre-quantized MLX models. This PR fixes the five issues blocking that: quantizers/auto.py: Detect MLX affine qua" }, { "n": 45975, "d": "2026-05-14", "t": "GGUF: optional Metal dequant fast path via kernels-community", "st": "OPEN", "by": "ArthurZucker", "a": 127, "r": 6, "f": 3, "c": 3, "b": "Summary Adds an optional Metal dequant fast path in GGUFDequantize. When the kernels package is installed and kernels-community/gguf-dequant is reachable, MPS GGUF loads (Q4_0 / Q8_0 / Q4_K / Q5_K / Q6_K / IQ4_NL / IQ4_XS) route through one Metal compute kernel per tensor — 2.3–7.5× faster than the " }, { "n": 45974, "d": "2026-05-28", "t": "Enable kernels-community/metal-flash-sdpa on MPS", "st": "MERGED", "by": "ArthurZucker", "a": 30, "r": 5, "f": 2, "c": 6, "b": "Enables kernels-community/metal-flash-sdpa for generate/generate_batch on MPS: synthesize cu_seqlens + .contiguous() in the no-padding branch of _flash_attention_forward, and fix MPS memory accounting in continuous batching. Bench (gsm8k 100 samples, Qwen2.5-0.5B-Instruct, MPS fp16, generate_batch):" }, { "n": 46357, "d": "2026-06-02", "t": "[docs] gguf mps", "st": "OPEN", "by": "stevhliu", "a": 71, "r": 0, "f": 1, "c": 2, "b": "adds docs for: #45975, dequant fast path through a Metal kernel #45977, native quantized and on-the-fly quantization inference on a Metal kernel. i don't think the docs here overlap with the existing docs in your PR :)" }, { "n": 46349, "d": "2026-06-18", "t": "[docs] metal flash attention", "st": "MERGED", "by": "stevhliu", "a": 15, "r": 0, "f": 1, "c": 1, "b": "adds docs for #45974 (install kernels to automatically use the metal flash kernel on MPS with transformers serve)" }, { "n": 46792, "d": "2026-06-26", "t": "Enable CB paged decode fast path for MPS hub kernels + sensible block_size default", "st": "OPEN", "by": "ArthurZucker", "a": 52, "r": 7, "f": 4, "c": 3, "b": "Summary Three changes to make continuous batching (CB) work well on Apple Silicon (MPS) with the metal-flash-sdpa kernel: 1. Enable paged decode fast path for hub kernels (initialization.py) ensure_decode_fast_path_is_available previously only allowed the paged decode fast path (block tables + flash" }, { "n": 46740, "d": "2026-07-24", "t": "Allow metal-flash-sdpa for OpenAIPrivacyFilter on MPS", "st": "MERGED", "by": "ArthurZucker", "a": 26, "r": 3, "f": 4, "c": 4, "b": "Adds kernels-community/metal-flash-sdpa to OpenAIPrivacyFilter._compatible_flash_implementations, enabling the sliding-window + attention-sink flash path on MPS instead of the eager fallback. Requires metal-flash-sdpa PR #5 (window + s_aux). Validated: cos 0.997, argmax 100%." }, { "n": 47779, "d": "2026-08-31", "t": "Refactor GGUF to speed-up inference", "st": "OPEN", "by": "SunMarc", "a": 2435, "r": 137, "f": 35, "c": 43, "b": "This PR updates the GGUF in transformers in order to take advantages of some features that landed for the past few months and are now stable: conversion using WeightConverter, instead of our custom tensor processor. Support for loading quantized models. Before we were dequantizing by default. Add su" } ], "d": "GGUF Metal dequant '2.3-7.5x faster than the chained PyTorch path' on M3 Max (#45975); WeightConverter GGUF reload hits 76.33 vs 29.23 tok/s BF16 (#47779); MPS serve latency 38-57s to ~4.3s (#46792).", "note": "PRs 45975, 47779, 46792, 44348, 46357 all verified OPEN as of 2026-09-01; none merged yet.", "links": [], "src": "quant.json", "span": [ "2026-02-28", "2026-08-31" ], "roll": { "owners": [ [ "ArthurZucker", 4 ], [ "stevhliu", 2 ], [ "SunMarc", 1 ], [ "n0kovo", 1 ] ], "n": 8, "add": 3097, "del": 190, "files": 54, "commits": 66 } }, { "t": "Quantization on WeightConverter / dynamic weight loader", "tag": "weightconverter-quant-loader", "s": "wip", "prs": [ { "n": 44794, "d": "2026-05-14", "t": "Refacto GGUF weight conversion", "st": "OPEN", "by": "ArthurZucker", "a": 1741, "r": 610, "f": 13, "c": 16, "b": "Refactor GGUF loading to use our weight-conversion pipeline + swap the slow NumPy dequant for a pure-PyTorch port that also runs on MPS / CUDA (vendored from diffusers / existing places) 1. Use the weight-conversion framework _GGUF_ARCH_CONVERTERS is now a flat per-model_type list of WeightRenaming " }, { "n": 46339, "d": "2026-06-09", "t": "Extended & simplified n-to-1 kernel fusion via KernelConfig", "st": "MERGED", "by": "michaelbenayoun", "a": 320, "r": 25, "f": 4, "c": 72, "b": "Extends the KernelConfig API with two orthogonal capabilities: Module fusion: specify how Transformers modules should be fused together before a custom kernel is applied (n-to-1 replacement). Parameter transformation: handle cases where a kernel expects weights in a different layout than the origina" }, { "n": 47779, "d": "2026-08-31", "t": "Refactor GGUF to speed-up inference", "st": "OPEN", "by": "SunMarc", "a": 2435, "r": 137, "f": 35, "c": 43, "b": "This PR updates the GGUF in transformers in order to take advantages of some features that landed for the past few months and are now stable: conversion using WeightConverter, instead of our custom tensor processor. Support for loading quantized models. Before we were dequantizing by default. Add su" } ], "d": "GGUF now converts via WeightConverter instead of a custom tensor processor, unlocking quantized (non-dequantized) loads; KernelConfig reuses WeightRenaming to remap kernel-specific weight layouts.", "note": null, "links": [], "src": "quant.json", "span": [ "2026-04-10", "2026-08-31" ], "roll": { "owners": [ [ "SunMarc", 1 ], [ "ArthurZucker", 1 ], [ "michaelbenayoun", 1 ] ], "n": 3, "add": 4496, "del": 772, "files": 52, "commits": 131 } }, { "t": "transformers serve: quantization & expert routing", "tag": "serve-quant-experts-routing", "s": "open", "prs": [ { "n": 46792, "d": "2026-06-26", "t": "Enable CB paged decode fast path for MPS hub kernels + sensible block_size default", "st": "OPEN", "by": "ArthurZucker", "a": 52, "r": 7, "f": 4, "c": 3, "b": "Summary Three changes to make continuous batching (CB) work well on Apple Silicon (MPS) with the metal-flash-sdpa kernel: 1. Enable paged decode fast path for hub kernels (initialization.py) ensure_decode_fast_path_is_available previously only allowed the paged decode fast path (block tables + flash" }, { "n": 48361, "d": "2026-08-27", "t": "Add --experts-implementation flag to transformers serve", "st": "OPEN", "by": "NathanHB", "a": 8, "r": 0, "f": 2, "c": 1, "b": "transformers serve had no way to pass experts_implementation to from_pretrained. For FP8 MoE checkpoints whose TP/EP plan requires experts_implementation=\"deepgemm_megamoe\" (the only key in FP8Experts._impl_tp_layer_overrides), this meant update_tp_plan always crashed with AttributeError: 'NoneType'" } ], "d": "serve.py gains an MPS-aware block_size default fixing a large MPS slowdown (#46792) and a new --experts-implementation flag to pick grouped/batched/kernel MoE backends at serve time (#48361).", "note": null, "links": [], "src": "quant.json", "span": [ "2026-06-21", "2026-08-27" ], "roll": { "owners": [ [ "ArthurZucker", 1 ], [ "NathanHB", 1 ] ], "n": 2, "add": 60, "del": 7, "files": 6, "commits": 4 } } ] }, { "name": "Export & on-device", "hue": 190, "blurb": "Get a transformers model out of Python and onto a device.", "items": [ { "t": "ONNX export deprecation and fixes", "tag": "Legacy onnx export retired", "s": "done", "prs": [ { "n": 41214, "d": "2025-10-01", "t": "[V5] Remove deprecated transformers.onnx", "st": "MERGED", "by": "cyyever", "a": 1, "r": 1306, "f": 9, "c": 4, "b": "Remove the source files, doc and the dependencies." }, { "n": 41700, "d": "2025-10-21", "t": "[v5] Remove deprecated tranformers.onnx", "st": "MERGED", "by": "echarlaix", "a": 93, "r": 5309, "f": 91, "c": 8, "b": "remove deprecated transformers.onnx for v5 remove all XxxOnnxConfig from concerned models config files Also following #41214" }, { "n": 48345, "d": "2026-08-26", "t": "[qwen4_exp] disable torch/onnx export tests due to data-dependent control flow", "st": "MERGED", "by": "ydshieh", "a": 2, "r": 0, "f": 1, "c": 1, "b": "Summary Sets test_torch_exportable = False in both Qwen4ExpTextModelTest and Qwen4ExpCompositeModelTest QSA index selection has data-dependent control flow ((u0//2) > 0) that cannot be guarded during torch.export, causing GuardOnDataDependentSymNode errors Fixes: https://github.com/huggingface/trans" }, { "n": 48191, "d": "2026-08-27", "t": "[ONNX] Skip affected models on torch 2.13 (two dynamo regressions)", "st": "MERGED", "by": "ydshieh", "a": 102, "r": 0, "f": 1, "c": 8, "b": "Summary torch 2.13.0 introduced two regressions in dynamo ONNX export that break 45 tests across 15 model families. This PR skips all affected model classes when torch >= 2.13 using the existing EXPORT_SKIPS infrastructure. Root causes (both filed upstream): pytorch/pytorch#194381 — aten.sub type-pr" } ], "d": "Deprecated transformers.onnx module removed for v5 in favor of optimum/exporters; ongoing skips added for torch 2.13 dynamo regressions and data-dependent-control-flow models (e.g. qwen4_exp).", "note": null, "links": [], "src": "infra.json", "span": [ "2025-09-30", "2026-08-27" ], "roll": { "owners": [ [ "ydshieh", 2 ], [ "echarlaix", 1 ], [ "cyyever", 1 ] ], "n": 4, "add": 198, "del": 6615, "files": 102, "commits": 21 } }, { "t": "torch.export compatibility hardening", "tag": "Models trace cleanly via export", "s": "done", "prs": [ { "n": 42317, "d": "2026-01-22", "t": "Make Transformers more torch-exportable and dynamo-friendly", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 1011, "r": 896, "f": 183, "c": 70, "b": "First proposals include: check_with(error_type, cond, lambda: msg) instead of if cond: raise error_type(msg), which also works with torch.export/torch.compile to hint to the compiler that the condition is expected to be true at export/compile time. vectorization of some loops / comprehension lists i" }, { "n": 43801, "d": "2026-02-12", "t": "Fix lifted_tensor in Gemma3n export which dynamo can't reason about", "st": "MERGED", "by": "robell", "a": 8, "r": 4, "f": 2, "c": 3, "b": "When using the torch.export path with this model we hit an issue on correction_coefs being modified with unknown side-effects the dynamo can't catch. So it lifts this tensor unnecessarily, and fails assuming mutation is needed RuntimeError: Constant model.model.language_model.layers.0.lifted_tensor_" }, { "n": 44266, "d": "2026-02-25", "t": "fix(utils): Make torch_compilable_check compatible with torch.export strict mode", "st": "MERGED", "by": "harshaljanjani", "a": 27, "r": 1, "f": 2, "c": 1, "b": "The following issue was identified and fixed in this PR: → Reasoning: The impact of this fix goes beyond Mask2Former and DeformableDetr and should fix any model that uses torch_compilable_check. Most users run models in eager where the fn works just fine; only torch._check_with only breaks during ex" }, { "n": 46738, "d": "2026-06-30", "t": "🚨 Modeling changes for export, compile, and hybrid-attention standardization", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 2353, "r": 3720, "f": 152, "c": 48, "b": "Modeling changes for export, compile, and hybrid-attention standardization A broad sweep of modeling-side changes that make a large set of models cleanly exportable (ONNX / torch.export / ExecuTorch), fullgraph-compilable, and consistent in how they declare their layer types and build masks/caches. " } ], "d": "Structural modeling changes (attention/cache/get_decoder standardization) made more architectures torch.export/dynamo-friendly, fixing strict-mode checks and lifted-tensor tracing (e.g. Gemma3n).", "note": null, "links": [], "src": "infra.json", "span": [ "2025-11-21", "2026-06-30" ], "roll": { "owners": [ [ "IlyasMoutawwakil", 2 ], [ "harshaljanjani", 1 ], [ "robell", 1 ] ], "n": 4, "add": 3399, "del": 4621, "files": 339, "commits": 122 } }, { "t": "transformers.exporters unified module", "tag": "One exporter API: PT/ONNX/ET", "s": "wip", "prs": [ { "n": 41992, "d": "2026-07-02", "t": "[PoC] HF exporters", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 5468, "r": 264, "f": 89, "c": 406, "b": "HfExporters: Native, Unified export for PyTorch / ONNX / ExecuTorch Edit: pieces of this PR were merged separately (#42697, #42317), so this one is now focused on the HfExporters machinery itself. What this adds A native, in-Transformers export pipeline — one base class (HfExporter), three subclasse" }, { "n": 47161, "d": "2026-07-08", "t": "Add exporters min version requirements and test skip", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 42, "r": 8, "f": 4, "c": 4, "b": "Fixes the failing CI using torch 2.8 Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. We are currently bottlenecked by our ability to review and respond to them. As a result, we ask that new users do not submit" }, { "n": 47374, "d": "2026-07-29", "t": "[docs] Exporters", "st": "MERGED", "by": "stevhliu", "a": 266, "r": 210, "f": 4, "c": 5, "b": "Refactors the exporters docs a bit by user intent (using exporters vs extending exporters) and make it shorter and more scannable. Split the doc by audience so the user guide is focused on using an exporter and a separate guide for contributors who want to patch/fix. Replace dense stages for every e" }, { "n": 47711, "d": "2026-08-03", "t": "Fix ImportError in transformers.exporters on torch < 2.8", "st": "MERGED", "by": "Neal006", "a": 1, "r": 1, "f": 1, "c": 2, "b": "Fixes #47710. import transformers.exporters raises ImportError on torch 2.5, 2.6 and 2.7, all inside the declared torch>=2.5 support range. exporter_executorch.py imports guard_or_true at module scope, and that symbol was added in torch 2.8 where it replaced definitely_true. is_torch_available() che" }, { "n": 47747, "d": "2026-08-25", "t": "Support quantization in Dynamo, ONNX and ET exporters", "st": "OPEN", "by": "IlyasMoutawwakil", "a": 1028, "r": 181, "f": 12, "c": 14, "b": "Fixes # (issue) Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. These often are low-quality, or fix extremely minor issues that occur rarely or never in practice. As a result, we're instituting a rule that fir" } ], "d": "New transformers.exporters module (PoC merged #41992) unifies export to PyTorch/ONNX/ExecuTorch; min-version gating and docs landed, quantization-aware export still open (#47747).", "note": null, "links": [], "src": "infra.json", "span": [ "2026-03-19", "2026-08-25" ], "roll": { "owners": [ [ "IlyasMoutawwakil", 3 ], [ "Neal006", 1 ], [ "stevhliu", 1 ] ], "n": 5, "add": 6805, "del": 664, "files": 110, "commits": 431 } }, { "t": "ExecuTorch pulled from CI, then reworked", "tag": "ExecuTorch pulled from CI, redone", "s": "wip", "prs": [ { "n": 47196, "d": "2026-07-09", "t": "Remove executorch from all-latest-gpu image + add torch smoke test", "st": "MERGED", "by": "tarekziade", "a": 16, "r": 5, "f": 1, "c": 1, "b": "executorch is only used for CPU tests, and it was breaking the GPU image because it pulled torch 2.13 this patch removes it but : keeps onnxruntime (gpu flavor) do a smoke test to verify it imports correctly" }, { "n": 47201, "d": "2026-07-09", "t": "Make executorch exporter tests always use xnnpack backend", "st": "MERGED", "by": "tarekziade", "a": 19, "r": 6, "f": 3, "c": 19, "b": "The ExecuTorch export tests target CPU (via xnnpack on CPU), and its deps now live in the CPU torch-light CI image. Mark test_executorch_export with a new require_non_gpu decorator so they run on the CPU image, not the GPU nightly. Lower DynamoExporter.min_versions torch 2.11.0 -> 2.10.0 so the test" }, { "n": 47204, "d": "2026-07-09", "t": "torch-light: add executorch (+onnx) for CPU exporter tests", "st": "OPEN", "by": "tarekziade", "a": 8, "r": 0, "f": 1, "c": 1, "b": "Install onnxscript + onnxruntime and executorch==1.2.0 in the CPU test image so the CPU-only exporter tests (tests/exporters, executorch tests marked require_non_gpu) can run here. executorch 1.2.0 matches transformers' exporter API (targets 1.3.1) and requires torch>=2.11.0, satisfied by the torch<" }, { "n": 47242, "d": "2026-07-10", "t": "Remove Executorch from CI until latest version is supported and fully tested on CI env", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 0, "r": 6, "f": 1, "c": 1, "b": "Fixes # (issue) Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. These often are low-quality, or fix extremely minor issues that occur rarely or never in practice. As a result, we're instituting a rule that fir" }, { "n": 47243, "d": "2026-08-03", "t": "Executorch exporter fixes", "st": "MERGED", "by": "IlyasMoutawwakil", "a": 1894, "r": 314, "f": 28, "c": 41, "b": "Been hunting them since we introduced their skips in the original HF exporters PR. Also adds static cache export support, for in-place cache modification with cache position. Also adds multi-token decoding support, so that the decode graph can do prefill (empty cache), single token decode and multi-" }, { "n": 47900, "d": "2026-08-21", "t": "Fix DynamicCache reconstruction during ExecuTorch export", "st": "MERGED", "by": "eladsegal", "a": 97, "r": 25, "f": 3, "c": 19, "b": "Fixes incorrect cache-update graphs when exporting models with DynamicCache to ExecuTorch. torch.export uses the registered pytree functions to flatten and reconstruct DynamicCache during export. The current pytree registration keeps populated key/value tensors, but loses each layer's type and windo" } ], "d": "ExecuTorch was removed from CI (#47242, #47196) as untested; exporter bugs (xnnpack backend, DynamicCache tracing) were fixed, and a lighter CPU-only ET+ONNX test image is being reintroduced (#47204).", "note": null, "links": [], "src": "infra.json", "span": [ "2026-07-09", "2026-08-21" ], "roll": { "owners": [ [ "tarekziade", 3 ], [ "IlyasMoutawwakil", 2 ], [ "eladsegal", 1 ] ], "n": 6, "add": 2034, "del": 356, "files": 37, "commits": 82 } } ] }, { "name": "Tests, CI & infra", "hue": 150, "blurb": "Weekly releases mean the tests have to stay green.", "items": [ { "t": "AMD MI300/MI325 runner migration", "tag": "AMD MI300/MI325 CI runners", "s": "done", "prs": [ { "n": 38132, "d": "2025-05-26", "t": "Add AMD MI300 CI caller leveraging self-hosted runner scale set workflow in hf-workflows", "st": "MERGED", "by": "jitesh-gupta", "a": 59, "r": 0, "f": 1, "c": 10, "b": "Added mi300 caller workflow to leverage self-hosted runner scale set in hf-worfklows Copied the exiting workflows and changed 'runner' parameter to the new 'runner_scale_set' parameter Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). " }, { "n": 39651, "d": "2025-07-28", "t": "Add self-hosted runner scale set workflow for mi325 CI", "st": "MERGED", "by": "jitesh-gupta", "a": 63, "r": 0, "f": 1, "c": 1, "b": "Fixes # (issue) Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. Di" }, { "n": 41193, "d": "2025-09-29", "t": "CI Runners - move amd runners mi355 and 325 to runner group", "st": "MERGED", "by": "glegendre01", "a": 10, "r": 10, "f": 2, "c": 6, "b": "Move CI runners 325/355 from scale set targetting to runner group" }, { "n": 47259, "d": "2026-07-11", "t": "Switch AMD daily CI to mi300 runners", "st": "MERGED", "by": "Abdennacer-Badaoui", "a": 16, "r": 16, "f": 5, "c": 1, "b": "MI325 hardware is being retired; migrate the AMD daily CI to MI300. Updates runner groups to hfc-amd-mi300-{1,2}gpu and labels to amd-mi300-ci-{1,2}gpu, and renames the mi325 caller workflows to mi300." }, { "n": 47307, "d": "2026-07-14", "t": "Update AMD CI runner groups to amd-mi300", "st": "MERGED", "by": "Abdennacer-Badaoui", "a": 5, "r": 5, "f": 2, "c": 1, "b": "Switch the AMD daily CI runner groups from hfc-amd-mi300 to amd-mi300-{1,2}gpu." } ], "d": "AMD daily/benchmark CI moved off older AMD hardware onto MI300 and MI325 self-hosted runner scale sets defined in the shared hf-workflows repo.", "note": null, "links": [], "src": "infra.json", "span": [ "2025-05-08", "2026-07-14" ], "roll": { "owners": [ [ "jitesh-gupta", 2 ], [ "Abdennacer-Badaoui", 2 ], [ "glegendre01", 1 ] ], "n": 5, "add": 153, "del": 31, "files": 11, "commits": 19 } }, { "t": "Benchmark v2 CI + Trackio reporting", "tag": "Benchmark v2 + Trackio CI", "s": "done", "prs": [ { "n": 38814, "d": "2025-07-22", "t": "🎯 Trackio integration", "st": "MERGED", "by": "qgallouedec", "a": 150, "r": 9, "f": 7, "c": 13, "b": "from trl import SFTTrainer, SFTConfig from datasets import load_dataset trainer = SFTTrainer( model=\"Qwen/Qwen2.5-0.5B\", args=SFTConfig(report_to=\"trackio\"), train_dataset=load_dataset(\"trl-lib/Capybara\", split=\"train\"), ) trainer.train() Before submitting This PR fixes a typo or improves the docs (" }, { "n": 40716, "d": "2025-09-19", "t": "Benchmarking v2 GH workflows", "st": "MERGED", "by": "ahadnagy", "a": 311, "r": 10, "f": 7, "c": 33, "b": "This PR is a follow-up to #40486 that adds the workflows and HF Datasets upload to start collecting data. Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discusse" }, { "n": 41408, "d": "2025-10-14", "t": "Benchmark overhaul", "st": "MERGED", "by": "remi-or", "a": 1007, "r": 1864, "f": 13, "c": 25, "b": "This PR overhauls the benchmarking suite that is included in transformers. The benchmarking suite is now based around three main components: BenchmarkingConfig is a dataclass-like object which contains everything needed to reproduce a benchmark on the same machine: input length, generation length, w" }, { "n": 41672, "d": "2025-10-20", "t": "feat: add benchmark v2 ci with results pushed to dataset", "st": "MERGED", "by": "McPatate", "a": 136, "r": 48, "f": 6, "c": 4, "b": "Reusing the old benchmark.yml workflow for the benchmark v2, we'll eventually rename and clean up everything once we're confident it works nicely. Benchmark results will be pushed to a dataset, for now located here. I chose JSONL for simplicity of parsing in the frontend space, and file name have th" }, { "n": 41916, "d": "2025-11-07", "t": "feat(ci): add continuous batching to benchmarks", "st": "MERGED", "by": "McPatate", "a": 190, "r": 107, "f": 10, "c": 15, "b": "" }, { "n": 43153, "d": "2026-01-13", "t": "Make benchmark-v2 to be device agnostic, to support more torch built-in devices like xpu", "st": "MERGED", "by": "yao-matrix", "a": 59, "r": 8, "f": 3, "c": 4, "b": "@remi-or , pls help review, thx very much." }, { "n": 43750, "d": "2026-03-19", "t": "enable tp for benchmark", "st": "MERGED", "by": "sywangyi", "a": 40, "r": 5, "f": 3, "c": 4, "b": "enable tp in benchmark_v2, to ensure large model could run." }, { "n": 45329, "d": "2026-04-13", "t": "Update `trackio` integration to use Buckets and \"freeze\" Space after training", "st": "MERGED", "by": "abidlabs", "a": 152, "r": 57, "f": 2, "c": 48, "b": "Updates TrackioCallback and TrainingArguments for the latest version of Trackio using HF Buckets as the backend, and control over creating a static Space for the Trackio dashboard during or at the end of training. These are now the TrainingArguments relevant for trackio: trackio_space_id — Only affe" } ], "d": "Benchmarking rebuilt as a device-agnostic benchmark_v2 GitHub Actions workflow pushing results to a HF dataset; Trackio integrated for training-metric dashboards, deployed to HF Spaces by default.", "note": null, "links": [], "src": "infra.json", "span": [ "2025-06-13", "2026-04-13" ], "roll": { "owners": [ [ "McPatate", 2 ], [ "ahadnagy", 1 ], [ "remi-or", 1 ], [ "yao-matrix", 1 ] ], "n": 8, "add": 2045, "del": 2108, "files": 51, "commits": 146 } }, { "t": "Public CI dashboard", "tag": "daily CI, AMD + Nvidia", "s": "done", "prs": [], "d": "Daily CI status for important models across AMD MI300 and Nvidia A10, published as a Space.", "links": [ [ "CI dashboard Space", "https://huggingface.co/spaces/transformers-community/transformers-ci-dashboard" ] ], "note": null, "src": "curated", "span": [ "2025-08-05", "2026-07-12" ], "roll": { "owners": [], "n": 0, "add": 0, "del": 0, "files": 0, "commits": 0 } }, { "t": "PR-triggered GPU slow-test runs", "tag": "run-slow v2 GPU tests on PRs", "s": "wip", "prs": [ { "n": 41914, "d": "2025-11-01", "t": "Run slow v2", "st": "MERGED", "by": "ydshieh", "a": 355, "r": 339, "f": 12, "c": 5, "b": "Run slow v2! Send feedback as a comment. Include verified failed tests specific to a PR. Discussed offline in the office. merge now 🔥" }, { "n": 47521, "d": "2026-07-24", "t": "Run only @slow tests in PR comment CI", "st": "MERGED", "by": "ydshieh", "a": 8, "r": 0, "f": 3, "c": 4, "b": "Add pytest.mark.slow to the slow() decorator, register it in conftest.py, and pass pytest_marker: slow to self-scheduled.yml from self-comment-ci.yml so that run-slow PR comment triggers only run @slow-decorated tests." }, { "n": 48190, "d": "2026-08-21", "t": "Post two CI badges on a PR: CPU PR CI and GPU run-slow", "st": "MERGED", "by": "tarekziade", "a": 41, "r": 3, "f": 2, "c": 1, "b": "A PR accumulates two independent CI streams -- regular PR CI on CPU, and the GPU runs a maintainer asks for with a run-slow: <models> comment. A single badge had to pick one, and picked whichever ran last, so a PR could carry a red badge sourced from a run-slow run beside a dashboard link explaining" }, { "n": 48065, "d": "2026-08-25", "t": "[WIP] Add PR comment CI for AMD (MI300)", "st": "OPEN", "by": "ydshieh", "a": 458, "r": 0, "f": 2, "c": 28, "b": "Summary Adds self-comment-ci-amd.yml — AMD equivalent of the Nvidia PR comment CI workflow Triggered by run-slow comments (same as Nvidia), runs on hfc-amd-mi300 runners Calls check_failed_tests_amd.yml (in hf-workflows) to bisect new failures and report back as a PR comment Commit status context: p" } ], "d": "run-slow v2 lets maintainers trigger @slow GPU tests via PR comment; PRs now show both CPU-CI and GPU run-slow dashboard badges (#48190); AMD MI300 PR-comment CI is still WIP (#48065).", "note": null, "links": [], "src": "infra.json", "span": [ "2025-10-31", "2026-08-25" ], "roll": { "owners": [ [ "ydshieh", 3 ], [ "tarekziade", 1 ] ], "n": 4, "add": 862, "del": 342, "files": 19, "commits": 38 } }, { "t": "CircleCI to GitHub Actions migration", "tag": "CircleCI -> GitHub Actions", "s": "done", "prs": [ { "n": 45180, "d": "2026-04-02", "t": "🔒 Pin GitHub Actions to commit SHAs", "st": "MERGED", "by": "paulinebm", "a": 48, "r": 48, "f": 18, "c": 18, "b": "🔒 Pin GitHub Actions to commit SHAs This PR pins all GitHub Actions to their exact commit SHA instead of mutable tags or branch names. Why? Pinning to a SHA prevents supply chain attacks where a tag (e.g. v4) could be moved to point to malicious code. Changes Workflow Action Avant Après SHA add-mode" }, { "n": 45476, "d": "2026-05-04", "t": "Github Actions PR CI (caller)", "st": "MERGED", "by": "ydshieh", "a": 9, "r": 0, "f": 1, "c": 1, "b": "Call CI workflow" }, { "n": 45955, "d": "2026-05-14", "t": "chore(ci): pin all GitHub Actions and reusable workflows by SHA", "st": "MERGED", "by": "XciD", "a": 146, "r": 146, "f": 38, "c": 1, "b": "Summary Replace mutable tag/branch references with commit SHAs across .github/workflows/. Mitigates the tj-actions/changed-files class of supply-chain attack: if a third-party action repo is compromised (account takeover, force-push to tag), the workflow keeps running the audited commit. Ran pinact " }, { "n": 46157, "d": "2026-05-26", "t": "chore: enable Dependabot weekly GitHub Actions bumps", "st": "MERGED", "by": "hf-dependantbot-rollout", "a": 11, "r": 0, "f": 1, "c": 1, "b": "Summary Adds .github/dependabot.yml so this repo's pinned GitHub Action SHAs get bumped automatically once a week. All action updates are grouped into one weekly PR (not one PR per action) to keep the noise down, and Dependabot waits 7 days after a release before opening the bump (cooldown). The 7-d" }, { "n": 46721, "d": "2026-06-17", "t": "ci: disable CircleCI by replacing config with no-op", "st": "MERGED", "by": "ydshieh", "a": 241, "r": 228, "f": 3, "c": 3, "b": "Replaces the full CircleCI config with a minimal no-op job that immediately passes, and disables the CircleCI failure summary workflow. Why not just delete the config? Removing config.yml causes CircleCI to post a failing \"no configuration found\" check. Replacing it with a no-op keeps the check gree" }, { "n": 48056, "d": "2026-08-18", "t": "[CircleCI] Enable CI for private forks, no-op for public repo", "st": "MERGED", "by": "ydshieh", "a": 201, "r": 14, "f": 2, "c": 5, "b": "Summary Public repo (huggingface/transformers): genuine noop workflow — CircleCI does nothing (GitHub Actions handles CI here) Private fork (e.g. new model additions): full CI runs automatically, gated by [run-circleci] in the commit message to control costs Add peft_integration CircleCI job to matc" } ], "d": "GH Actions PR-CI caller replaced CircleCI for the public repo (config swapped for a no-op); CircleCI kept alive only for private-fork model additions, gated by [run-circleci]. Actions pinned to SHAs.", "note": null, "links": [], "src": "infra.json", "span": [ "2026-04-02", "2026-08-18" ], "roll": { "owners": [ [ "ydshieh", 3 ], [ "paulinebm", 1 ], [ "XciD", 1 ], [ "hf-dependantbot-rollout", 1 ] ], "n": 6, "add": 656, "del": 436, "files": 63, "commits": 29 } }, { "t": "Grafana CI dashboard on every PR", "tag": "Per-PR Grafana CI dashboard link", "s": "done", "prs": [ { "n": 46410, "d": "2026-06-04", "t": "Add workflow to post CI Grafana dashboard link to PR", "st": "MERGED", "by": "ydshieh", "a": 39, "r": 0, "f": 1, "c": 1, "b": "Summary Adds pr-ci-post-dashboard-link.yml workflow that triggers via workflow_run on completion of \"PR CI\" Uses actions/github-script to find the PR by head SHA and post a comment with the Grafana observability dashboard link Works for both same-repo and fork PRs since workflow_run always runs in t" }, { "n": 46412, "d": "2026-06-04", "t": "Improve CI dashboard comment: rename and deduplicate", "st": "MERGED", "by": "ydshieh", "a": 18, "r": 1, "f": 1, "c": 1, "b": "Summary Renames comment from **CI Observability Dashboard:** to **CI Dashboard:** Before posting, deletes any existing dashboard comment on the PR (handles both old and new text) to avoid duplicates on CI reruns" }, { "n": 46415, "d": "2026-06-04", "t": "ci: mention code quality failure in CI dashboard comment", "st": "MERGED", "by": "ydshieh", "a": 16, "r": 1, "f": 1, "c": 1, "b": "Summary When check_code_quality fails, the CI dashboard comment posted to the PR now includes a warning explaining that all test jobs were skipped The Grafana link is always posted regardless, so partial OTEL data is still accessible How it works After the \"PR CI\" workflow completes, listJobsForWork" }, { "n": 46499, "d": "2026-06-08", "t": "Log some information in `.github/workflows/pr-ci-post-dashboard-link.yml`", "st": "MERGED", "by": "ydshieh", "a": 17, "r": 0, "f": 1, "c": 1, "b": "So we can see exactly which workflow \"PR CI\" run triggered a run of \"Post CI Dashboard Link\". This enables to track back if a comment is posted with strange information. For example, this comment #46469 (comment) it is from this run: https://github.com/huggingface/transformers/actions/runs/271360354" }, { "n": 46774, "d": "2026-06-28", "t": "Insert a Grafana badge in the PR", "st": "MERGED", "by": "tarekziade", "a": 714, "r": 73, "f": 3, "c": 4, "b": "Moved the script to Python, in utils, and insert a badge in the PR" }, { "n": 47668, "d": "2026-07-31", "t": "Fix missing github_utils.py download in PR CI dashboard workflow", "st": "MERGED", "by": "ydshieh", "a": 4, "r": 0, "f": 1, "c": 1, "b": "Summary update_pr_ci_dashboard_recap.py imports github_request from github_utils, but the workflow only downloaded the main script, causing ModuleNotFoundError: No module named 'github_utils' Now both update_pr_ci_dashboard_recap.py and github_utils.py are fetched from the same SCRIPT_REF (the workf" }, { "n": 48104, "d": "2026-08-19", "t": "chore: update pr-ci-post-dashboard-link.yml", "st": "OPEN", "by": "hf-security-analysis", "a": 9, "r": 1, "f": 1, "c": 1, "b": "Update .github/workflows/pr-ci-post-dashboard-link.yml workflow configuration. cc @ydshieh Closes huggingface/tracking-issues#908" } ], "d": "Workflow posts a Grafana pytest-observability dashboard badge/link as a PR comment, later deduped and extended to flag code-quality failures; #48104 is ongoing maintenance.", "note": null, "links": [], "src": "infra.json", "span": [ "2026-06-04", "2026-08-19" ], "roll": { "owners": [ [ "ydshieh", 5 ], [ "tarekziade", 1 ], [ "hf-security-analysis", 1 ] ], "n": 7, "add": 817, "del": 76, "files": 9, "commits": 10 } }, { "t": "Dedicated transformers-ci system", "tag": "Split CI into transformers-ci repo", "s": "wip", "prs": [ { "n": 46811, "d": "2026-06-26", "t": "call transformers-ci in a nightly run", "st": "MERGED", "by": "tarekziade", "a": 41, "r": 0, "f": 1, "c": 3, "b": "Adds a nighly job that computes integration tests errors and call Serge to generate fixes in PRs" }, { "n": 47040, "d": "2026-07-03", "t": "Update workflow callers to use `transformers-ci`", "st": "MERGED", "by": "ydshieh", "a": 2, "r": 2, "f": 2, "c": 1, "b": "See https://github.com/huggingface/transformers-test-ci/pull/3 As discussed offline, this transformers-ci would be our lovely home for CI." }, { "n": 47202, "d": "2026-07-09", "t": "Drop CPU/light docker image builder (moved to transformers-ci)", "st": "OPEN", "by": "tarekziade", "a": 0, "r": 229, "f": 7, "c": 1, "b": "Remove build-ci-docker-images.yml and the six light dockerfiles it built (quality, consistency, custom-tokenizers, torch-light, exotic-models, examples-torch). These CPU/light CI images are now built and pushed from the transformers-ci repo, which owns the PR test workflow; keeping a duplicate build" }, { "n": 47360, "d": "2026-07-27", "t": "CI: use transformers-ci daily workflow with OTEL", "st": "MERGED", "by": "tarekziade", "a": 29, "r": 942, "f": 12, "c": 12, "b": "Switches the daily workflow to using transformers-ci workflow see https://github.com/huggingface/transformers-ci/blob/main/.github/workflows/daily-ci_reusable.yml This workflow is also OTEL-ed and results are landing in https://transformers-ci.lor-e.huggingface.cool/d/pytest-observability-daily/dail" }, { "n": 47884, "d": "2026-08-10", "t": "Add `CI_CPU_MEMORY_LIMIT_GB` to check_failed_tests workflow", "st": "MERGED", "by": "ydshieh", "a": 4, "r": 0, "f": 1, "c": 2, "b": "Summary The check_failed_tests workflow re-runs failing tests on the same K8S instance-sharing runners as the daily CI model job, where each runner sees the full machine CPU RAM (~750 GB) instead of its allocated share Without CI_CPU_MEMORY_LIMIT_GB, device_map=\"auto\" in the re-run produces a differ" }, { "n": 48001, "d": "2026-08-17", "t": "Let the GPU verify caller turn on the memory probe", "st": "MERGED", "by": "tarekziade", "a": 6, "r": 0, "f": 1, "c": 1, "b": "transformers-ci#75 added a memory_probe input to the reusable serge-verify-slow workflow: it installs the resource plugin, sets PYTEST_RESOURCE_METRICS_FILE / PYTEST_RESOURCE_GC_PROBE and uploads a per-test JSONL saying how much device memory each test leaves behind and how much of that survives a g" }, { "n": 48085, "d": "2026-08-21", "t": "Assign a reviewer even when a codeowner has left, and route models by modality", "st": "MERGED", "by": "tarekziade", "a": 494, "r": 463, "f": 11, "c": 14, "b": "#48070 went out with zero reviewers. The workflow ran and reported success: Top owners [('gante', 4)] Failed to request review for ['gante']: Reviews may only be requested from collaborators. The .github/scripts/codeowners_for_review_action was outdated and missed a lot of new models. Also, an audit" } ], "d": "CI logic split into a separate transformers-ci system now owning PR CI, the GPU verify caller, a daily workflow with OTEL, and reviewer routing; docker image builder migration still open (#47202).", "note": "Founding 'split out transformers-ci' commit (e6c1043, 2026-04-21) has no associated PR number; appears to be a direct maintainer push.", "links": [], "src": "infra.json", "span": [ "2026-06-26", "2026-08-21" ], "roll": { "owners": [ [ "tarekziade", 5 ], [ "ydshieh", 2 ] ], "n": 7, "add": 576, "del": 1636, "files": 35, "commits": 34 } }, { "t": "Serge automated test-fix bot", "tag": "Serge auto-fixes failing tests", "s": "wip", "prs": [ { "n": 47381, "d": "2026-07-17", "t": "CI: add serge verify (GPU) caller workflow", "st": "MERGED", "by": "tarekziade", "a": 79, "r": 0, "f": 1, "c": 2, "b": "What A workflow_dispatch caller that delegates to the reusable serge-verify-slow workflow in huggingface/transformers-ci (PR #30 there). It runs a serge failure group's targeted @slow test node-ids on a GPU twice — the pre-patch baseline (must be red) and serge's patched candidate (want green), opti" }, { "n": 47492, "d": "2026-07-23", "t": "CI: add reproduce mode to serge verify caller", "st": "MERGED", "by": "tarekziade", "a": 13, "r": 3, "f": 1, "c": 1, "b": "Pass a mode input (verify|reproduce) through to the transformers-ci reusable serge-verify-slow workflow. reproduce runs the targeted tests only on the baseline tree so serge can confirm a failure is real BEFORE spending LLM budget investigating it; verify (default) is unchanged. commit_sha is now op" }, { "n": 48414, "d": "2026-08-29", "t": "[serge] Fix 2 integration tests for model `cwm` failing with `import_or_config` (other (2))", "st": "OPEN", "by": "sergereview", "a": 2, "r": 0, "f": 2, "c": 1, "b": "Original CI failure Failure group: 2 integration tests for model cwmfailing withimport_or_config (other (2)) tests/models/cwm/test_modeling_cwm.py::CwmIntegrationTest::test_cwm_integration [multi-gpu] (import_or_config, seen 7/7) CI traceback — tests/models/cwm/test_modeling_cwm.py::CwmIntegrationTe" }, { "n": 48437, "d": "2026-08-31", "t": "[serge] Fix 2 integration tests for model `big_bird` failing with `other` (other (2))", "st": "OPEN", "by": "sergereview", "a": 3, "r": 3, "f": 1, "c": 1, "b": "Original CI failure Failure group: 2 integration tests for model big_birdfailing withother (other (2)) tests/models/big_bird/test_modeling_big_bird.py::BigBirdModelIntegrationTest::test_fill_mask [multi-gpu] (other, seen 6/7) CI traceback — tests/models/big_bird/test_modeling_big_bird.py::BigBirdMod" } ], "d": "A 'serge' bot triages GPU-verify integration-test failures per model and opens auto-generated fix PRs (dozens open, e.g. #48414, #48437); GPU verify caller + reproduce mode added to CI.", "note": null, "links": [], "src": "infra.json", "span": [ "2026-07-17", "2026-08-31" ], "roll": { "owners": [ [ "tarekziade", 2 ], [ "sergereview", 2 ] ], "n": 4, "add": 97, "del": 6, "files": 5, "commits": 5 } }, { "t": "mlinter automated PR review bot", "tag": "mlinter review bot in CI", "s": "done", "prs": [ { "n": 47800, "d": "2026-08-05", "t": "Add post-mlinter-review job to post-dashboard-link workflow", "st": "MERGED", "by": "ydshieh", "a": 21, "r": 0, "f": 1, "c": 3, "b": "Summary Adds a post-mlinter-review job to pr-ci-post-dashboard-link.yml that downloads the mlinter-findings artifact from the triggering PR CI run and prints its content First step toward posting mlinter violations as inline PR review comments Note This needs to be merged to main to take effect (wor" }, { "n": 47918, "d": "2026-08-18", "t": "Moving mlinter to 0.1.4", "st": "MERGED", "by": "tarekziade", "a": 2249, "r": 131, "f": 54, "c": 13, "b": "Bumps transformers-mlinter to 0.1.4 to pick up the new rules (TRF020 to TRF057), and resolves everything those rules flag. The base pinned transformers-mlinter==0.1.2, whose utils/rules.toml carried only TRF001–TRF019. This bump adds 37 new rules in the TRF020–TRF057 range, and every finding they re" }, { "n": 48088, "d": "2026-08-19", "t": "Fix mlinter artifact path", "st": "MERGED", "by": "ydshieh", "a": 2, "r": 1, "f": 1, "c": 5, "b": "Summary Necessary follow-up after #47949 (Potential fix for code scanning alert no. 267: Artifact poisoning). That PR downloads the mlinter findings artifact into ${{ runner.temp }}/mlinter-findings (outside the workspace, for security), but post_mlinter_review.py was still looking for the files in " }, { "n": 48106, "d": "2026-08-19", "t": "Accept artifact dir as argument in post_mlinter_review.py", "st": "MERGED", "by": "ydshieh", "a": 3, "r": 2, "f": 1, "c": 1, "b": "Summary Follow-up to #48088 which changed the workflow to pass $MLINTER_FINDINGS_DIR to post_mlinter_review.py, but the script was still hardcoding Path(\"mlinter-findings.json\") in the CWD and ignoring the argument — causing Artifact files missing; skipping. Accept sys.argv[1] as the artifact direct" }, { "n": 48107, "d": "2026-08-19", "t": "Delete old mlinter review comments before posting new ones", "st": "MERGED", "by": "ydshieh", "a": 28, "r": 6, "f": 1, "c": 3, "b": "Summary Prevents accumulation of duplicate inline comments on the mlinter verification PR (seen in #47774). When findings change between CI runs (e.g. a new violation is added), the bot posts a new review containing all current findings as inline comments — including ones already commented in previo" }, { "n": 48117, "d": "2026-08-19", "t": "Enable mlinter findings artifact for inline PR reviews", "st": "MERGED", "by": "ydshieh", "a": 1, "r": 1, "f": 1, "c": 1, "b": "Summary Last piece of the mlinter inline review pipeline, verified in #48109. Adds --output-json mlinter-findings.json to the mlinter invocation so the check_code_quality job uploads a mlinter-findings artifact. The post-mlinter-review job (already on main) then downloads it and posts inline review " }, { "n": 48267, "d": "2026-08-25", "t": "ignore mlinter ci file", "st": "MERGED", "by": "tarekziade", "a": 1, "r": 0, "f": 1, "c": 1, "b": "mlinter-findings.json is created by the CI to provide developers feedback, we should ignore it in Git" }, { "n": 48259, "d": "2026-08-28", "t": "Bump transformers-mlinter to 0.1.5 and clear the new findings", "st": "MERGED", "by": "tarekziade", "a": 872, "r": 536, "f": 23, "c": 13, "b": "setup.py points transformers-mlinter at main ahead of the 0.1.5 release. The new rules surface 30 findings; this branch takes that to 0. make typing and make check-repo pass. ⚠️ Before merge: pin setup.py back to ==0.1.5 once released (make fix-repo propagates to dependency_versions_table.py). Findi" } ], "d": "transformers-mlinter static-analysis bot posts inline review comments via a post-mlinter-review CI job; iterated through Aug 2026 on artifact handling, comment dedup, and version bumps (0.1.4, 0.1.5).", "note": null, "links": [], "src": "infra.json", "span": [ "2026-08-05", "2026-08-28" ], "roll": { "owners": [ [ "ydshieh", 5 ], [ "tarekziade", 3 ] ], "n": 8, "add": 3177, "del": 677, "files": 83, "commits": 40 } }, { "t": "Daily/nightly CI scope reduction", "tag": "Nightly CI trimmed down", "s": "done", "prs": [ { "n": 48292, "d": "2026-08-25", "t": "Disable daily nightly CI", "st": "MERGED", "by": "remi-or", "a": 6, "r": 6, "f": 1, "c": 1, "b": "As per title: since we don't use the results of this CI in a daily way, disable its daily run." }, { "n": 48338, "d": "2026-08-26", "t": "[debug] Trace previous CI run selection in get_previous_daily_ci.py", "st": "MERGED", "by": "ydshieh", "a": 36, "r": 2, "f": 2, "c": 3, "b": "Summary Adds debug print statements to trace why an unexpectedly old run (Aug 3) was selected as the previous CI baseline instead of the most recent daily run. Context On the Aug 26 CI run, the Slack notification reported 246 new failed tests compared to run 30781855254 — which is the Aug 3 run, ~3 " } ], "d": "Daily/nightly full-suite CI run disabled ('we don't use the results of this CI in a daily way'), shifting emphasis to per-PR CI via transformers-ci; run-selection logic needed a follow-up debug fix.", "note": null, "links": [], "src": "infra.json", "span": [ "2026-08-25", "2026-08-26" ], "roll": { "owners": [ [ "remi-or", 1 ], [ "ydshieh", 1 ] ], "n": 2, "add": 42, "del": 8, "files": 3, "commits": 4 } } ] }, { "name": "Multimodality", "hue": 220, "blurb": "Backend-agnostic processors; serve every modality.", "items": [ { "t": "Video/audio decoding consolidates on torchcodec", "tag": "torchcodec as decode backend", "s": "wip", "prs": [ { "n": 35206, "d": "2025-05-12", "t": "🔴 Video processors as a separate class", "st": "MERGED", "by": "zucchini-nlp", "a": 5419, "r": 2005, "f": 83, "c": 103, "b": "This is quite a breaking PR, because from now on some processor will not have a separate self.video_processor and save/load priority is given to new file names image/video_preprocessor_config.json. Save/load perserved BC and can load from any file name Reasons for PR: Currently the main issue is to " }, { "n": 47935, "d": "2026-08-14", "t": "[docs] torchcodec + vision processor features", "st": "MERGED", "by": "stevhliu", "a": 54, "r": 5, "f": 3, "c": 3, "b": "Adds docs for some vision related things in #47850 and #47573 Remove mention of torchvision from the docs. Document dynamic resolution and frame sampling" }, { "n": 47963, "d": "2026-08-18", "t": "[DONT MERGE] Image decoding moves to torchcodec", "st": "OPEN", "by": "zucchini-nlp", "a": 22, "r": 1, "f": 2, "c": 3, "b": "Torch team is migrating from torchvision to torchcodec Do not merge yet, we don't know yet at which version torchvision is deprecated/removed. They released torchcodec.decoder.decode_image just yesterday - https://github.com/meta-pytorch/torchcodec/releases/tag/v0.16.0 Keeping the PR so I don't forg" }, { "n": 47629, "d": "2026-08-20", "t": "[docs] Audio backends", "st": "OPEN", "by": "stevhliu", "a": 36, "r": 0, "f": 2, "c": 2, "b": "Adds a section for audio inputs and the different decoding backends for chat templates" }, { "n": 48158, "d": "2026-08-24", "t": "Deprecate backend=\"torchaudio\" for `load_audio`", "st": "OPEN", "by": "eustlb", "a": 19, "r": 35, "f": 2, "c": 2, "b": "backend=\"torchaudio\" was introduced to use torchaudio load + resampling for bit exact correspondance with sglang during inkling integration. Nevertheless, torchaudio.load is deprecated upstream and has delegated to torchcodec since torchaudio 2.9, so backend=\"torchaudio\" no longer provides a distinc" } ], "d": "Building on video processors as their own class, decoding moves onto torchcodec: torchaudio load backend deprecated, docs added on backend choice, image decoding via torchcodec being explored.", "note": null, "links": [], "src": "multimodal.json", "span": [ "2024-10-30", "2026-08-24" ], "roll": { "owners": [ [ "zucchini-nlp", 2 ], [ "stevhliu", 2 ], [ "eustlb", 1 ] ], "n": 5, "add": 5550, "del": 2046, "files": 92, "commits": 113 } }, { "t": "Image processors: unified backend architecture", "tag": "unified image processor backends", "s": "done", "prs": [ { "n": 43514, "d": "2026-03-19", "t": "🚨🚨 Refactor Image Processors to support different backends", "st": "MERGED", "by": "yonigozlan", "a": 33586, "r": 57296, "f": 675, "c": 59, "b": "Image Processor Backend Refactor Summary Replaces the dual-file BaseImageProcessor (slow/PIL) + BaseImageProcessorFast (fast/torchvision) design with a unified backend architecture. The image_processing_utils_fast module is removed; all logic lives in image_processing_utils and image_processing_back" }, { "n": 45033, "d": "2026-03-26", "t": "Add BC for `_further_process_kwargs`", "st": "MERGED", "by": "hmellor", "a": 3, "r": 0, "f": 1, "c": 1, "b": "In #43514, BaseImageProcessorFast became BaseImageProcessor and _further_process_kwargs was renamed to _standardize_kwargs This PR adds some BC for the old name of this method." } ], "d": "Replaced the dual BaseImageProcessor (PIL) / BaseImageProcessorFast (torchvision) classes with one shared pipeline dispatching to pluggable TorchvisionBackend/PilBackend implementations.", "note": null, "links": [], "src": "multimodal.json", "span": [ "2026-01-27", "2026-03-26" ], "roll": { "owners": [ [ "yonigozlan", 1 ], [ "hmellor", 1 ] ], "n": 2, "add": 33589, "del": 57296, "files": 676, "commits": 60 } }, { "t": "One encode path for multimodal generate", "tag": "single multimodal encode path", "s": "wip", "prs": [ { "n": 45493, "d": "2026-05-20", "t": ":rotating_light: Modularize `ProcessorMixin` into smaller components", "st": "MERGED", "by": "zucchini-nlp", "a": 1746, "r": 3021, "f": 58, "c": 46, "b": "Modularizes ProcessorMixin to make it easier for new processors to override smaller fn rather than the whole __call__. Splits __call__ into smaller functions such as validation, input preparation, replacing multimodal placeholders, and a few properties for common special tokens In simple cases like " }, { "n": 46405, "d": "2026-07-22", "t": "Consistent output shape from `get_image_features`", "st": "MERGED", "by": "zucchini-nlp", "a": 293, "r": 52, "f": 21, "c": 11, "b": "as per title, branches off from #45783 The pooled image (OR video, NOT audio yet) feature output will now always satisfy three cond, where image features can be a list or a 3D tensor: len(image_features) = len(input_images) # NOTE: or num_videos, not number of total video frames image_features[0].nd" }, { "n": 45783, "d": "2026-08-31", "t": "[generation] Encode multimodal data only once", "st": "OPEN", "by": "zucchini-nlp", "a": 2996, "r": 4493, "f": 120, "c": 72, "b": "As per title, moves multimodal encoding out of the loop so it happens only once and on un-expanded pixels. For example, in assisted generation we used to encode twice (once in assistant and once in target model), or in beam search we used to expand all pixel by x num_beams times. This is very ineffi" } ], "d": "Modularized ProcessorMixin, moved multimodal encoding out of the generate loop to run once on un-expanded pixels instead of per-beam, and standardized get_image_features output shape.", "note": null, "links": [], "src": "multimodal.json", "span": [ "2026-03-10", "2026-08-31" ], "roll": { "owners": [ [ "zucchini-nlp", 3 ] ], "n": 3, "add": 5035, "del": 7566, "files": 199, "commits": 129 } }, { "t": "Audio processors: AudioProcessor replaces SequenceFeatureExtractor", "tag": "torch-native AudioProcessor migration", "s": "wip", "prs": [ { "n": 48110, "d": "2026-08-20", "t": "[AudioProcessor] add PreprocessingMixin", "st": "DRAFT", "by": "eustlb", "a": 826, "r": 1016, "f": 5, "c": 3, "b": "Introduces PreprocessingMixin that factorizes pre-processor's (image and the coming audio in sub-stacked-prs) logic" }, { "n": 48111, "d": "2026-08-20", "t": "[AudioProcessor] add BaseAudioProcessor and SpectrogramConfig", "st": "DRAFT", "by": "eustlb", "a": 2244, "r": 1057, "f": 10, "c": 3, "b": "Adds SpectrogramConfig (with StftConfig / MelScaleConfig) to describe an audio frontend declaratively instead of as a call into spectrogram() with a dozen positional arguments, and the modality layer on top of it: AudioProcessingMixin and BaseAudioProcessor, which own padding, truncation, masking an" }, { "n": 48112, "d": "2026-08-20", "t": "[AudioProcessor] add numpy and torch backends", "st": "DRAFT", "by": "eustlb", "a": 607, "r": 0, "f": 2, "c": 1, "b": "Adds NumpyAudioBackend and TorchAudioBackend, the two concrete implementations of BaseAudioProcessor's array primitives — bit-exact with librosa/kaldi and with torchaudio respectively. Every model processor picks a backend and supplies a SpectrogramConfig rather than reimplementing STFT/mel/log fram" }, { "n": 48113, "d": "2026-08-20", "t": "[AudioProcessor] add AutoAudioProcessor", "st": "DRAFT", "by": "eustlb", "a": 386, "r": 228, "f": 5, "c": 1, "b": "Adds AutoAudioProcessor, successor to AutoFeatureExtractor which is now deprecated and delegates to it. Since a model can ship a torch and a numpy processor, the auto mapping becomes backend-keyed ({\"torch\": ..., \"numpy\": ...}), resolved by from_pretrained(..., backend=...) and falling back with a w" }, { "n": 48114, "d": "2026-08-20", "t": "[AudioProcessor] migrate all audio models", "st": "DRAFT", "by": "eustlb", "a": 5344, "r": 14191, "f": 186, "c": 1, "b": "Migrates every in-library audio frontend off SequenceFeatureExtractor and onto the AudioProcessor classes added earlier in this stack. Each of the 30 models gains an XxxAudioProcessor — plus an XxxAudioProcessorNumpy sibling wherever the frontend is shared between backends — that declares its fronte" }, { "n": 48115, "d": "2026-08-20", "t": "[AudioProcessor] deprecate SequenceFeatureExtractor", "st": "DRAFT", "by": "eustlb", "a": 258, "r": 2566, "f": 6, "c": 1, "b": "Now that every audio model is an AudioProcessor, retires the legacy layer: SequenceFeatureExtractor becomes a thin shim over BaseAudioProcessor, FeatureExtractionMixin warns on instantiation, and audio_utils drops mel_filter_bank, spectrogram, spectrogram_batch, power_to_db_batch and amplitude_to_db" }, { "n": 44394, "d": "2026-08-28", "t": "🚨🚧 FeatureExtractor → AudioProcessor", "st": "DRAFT", "by": "eustlb", "a": 8244, "r": 14769, "f": 189, "c": 89, "b": "should cover #45549 #45134" } ], "d": "Stacked PR series adds PreprocessingMixin, BaseAudioProcessor/SpectrogramConfig and numpy+torch backends, then migrates ~30 audio models off SequenceFeatureExtractor (kept as deprecated alias).", "note": null, "links": [], "src": "multimodal.json", "span": [ "2026-05-12", "2026-08-28" ], "roll": { "owners": [ [ "eustlb", 7 ] ], "n": 7, "add": 17909, "del": 33827, "files": 403, "commits": 99 } }, { "t": "Consolidating VLM/processor test mixins", "tag": "shared per-modality test mixin", "s": "wip", "prs": [ { "n": 47005, "d": "2026-07-03", "t": "Add tiny_model_id support to ProcessorTesterMixin for memory-sensitive tests", "st": "MERGED", "by": "ydshieh", "a": 82, "r": 30, "f": 3, "c": 4, "b": "Summary Adds tiny_model_id attribute to ProcessorTesterMixin allowing test classes to specify a lightweight Hub repo (e.g. containing only a tiny tokenizer) for use in most tests, while keeping model_id as the full Hub repo used only when explicitly requested via get_processor(use_full=True) / get_c" }, { "n": 47922, "d": "2026-08-31", "t": "Processing tests [part 2]", "st": "OPEN", "by": "zucchini-nlp", "a": 1426, "r": 5349, "f": 108, "c": 39, "b": "Added a new testcase for return_replacement_offsets which will be used by vllm and finally become the only preferred way over mm_token_type_ids as it enables processing cache for simple models Deleted a bunch of overriden test_chat_template for videos by allowing each child to define only input samp" } ], "d": "Replaces per-model overridden processor/chat-template tests with tests parameterized per modality plus return_replacement_offsets coverage; lets heavy tests use tiny Hub checkpoints via the mixin.", "note": null, "links": [], "src": "multimodal.json", "span": [ "2026-07-01", "2026-08-31" ], "roll": { "owners": [ [ "zucchini-nlp", 1 ], [ "ydshieh", 1 ] ], "n": 2, "add": 1508, "del": 5379, "files": 111, "commits": 43 } }, { "t": "Standardizing vision/mRoPE across VLMs", "tag": "standard vision RoPE contract", "s": "wip", "prs": [ { "n": 47974, "d": "2026-08-28", "t": "Unified multimodal mixin (mrope)", "st": "OPEN", "by": "IlyasMoutawwakil", "a": 3705, "r": 5664, "f": 58, "c": 27, "b": "Fixes # (issue) Code Agent Policy The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by code agents. These often are low-quality, or fix extremely minor issues that occur rarely or never in practice. As a result, we're instituting a rule that fir" }, { "n": 48105, "d": "2026-08-28", "t": "Vision (2d/3d) rotary embeddings ", "st": "OPEN", "by": "zucchini-nlp", "a": 2747, "r": 1985, "f": 107, "c": 73, "b": "We need a standard way to compute rope freq in vision models which is usually only 2D or 3D. Atm we either interleave the grids inside rope module's forward or inside each attention layer. I think doing one interleave/permute or whatever needed in a rope module is easier and we keep the rest of mode" } ], "d": "Proposes one recomposition_2d/3d hook for vision rotary embeddings plus a unified multimodal mRoPE mixin, replacing per-model interleave/permute logic duplicated across attention layers.", "note": null, "links": [], "src": "multimodal.json", "span": [ "2026-07-24", "2026-08-28" ], "roll": { "owners": [ [ "IlyasMoutawwakil", 1 ], [ "zucchini-nlp", 1 ] ], "n": 2, "add": 6452, "del": 7649, "files": 165, "commits": 100 } } ] } ];
// Generated by build2.py — do not hand-edit.
// Spans are REAL: first commit of the earliest PR -> last commit / merge of the latest.
// Source board: https://miro.com/app/board/uXjVL_vDn1k=/
const AXIS = ["2023-12-14", "2026-09-01"];
const RELEASES = [{"n": "v4.42.0", "d": "2024-06-27", "big": false}, {"n": "v4.43.0", "d": "2024-07-23", "big": false}, {"n": "v4.44.0", "d": "2024-08-06", "big": false}, {"n": "v4.45.0", "d": "2024-09-25", "big": false}, {"n": "v4.46.0", "d": "2024-10-24", "big": false}, {"n": "v4.47.0", "d": "2024-12-05", "big": false}, {"n": "v4.48.0", "d": "2025-01-10", "big": false}, {"n": "v4.49.0", "d": "2025-02-17", "big": false}, {"n": "v4.50.0", "d": "2025-03-21", "big": false}, {"n": "v4.51.0", "d": "2025-04-05", "big": false}, {"n": "v4.52.0", "d": "2025-05-20", "big": false}, {"n": "v4.53.0", "d": "2025-06-26", "big": false}, {"n": "v4.54.0", "d": "2025-07-25", "big": false}, {"n": "v4.55.0", "d": "2025-08-05", "big": false}, {"n": "v4.56.0", "d": "2025-08-29", "big": false}, {"n": "v4.57.0", "d": "2025-10-03", "big": false}, {"n": "v5.0.0rc0", "d": "2025-12-01", "big": true}, {"n": "v5.0.0rc1", "d": "2025-12-11", "big": false}, {"n": "v5.0.0rc2", "d": "2026-01-07", "big": false}, {"n": "v5.0.0rc3", "d": "2026-01-14", "big": false}, {"n": "v5.0.0", "d": "2026-01-26", "big": true}, {"n": "v5.1.0", "d": "2026-02-05", "big": false}, {"n": "v5.2.0", "d": "2026-02-16", "big": false}, {"n": "v5.3.0", "d": "2026-03-04", "big": false}, {"n": "v5.4.0", "d": "2026-03-27", "big": false}, {"n": "v5.5.0", "d": "2026-04-02", "big": false}, {"n": "v5.6.0", "d": "2026-04-22", "big": false}, {"n": "v5.7.0", "d": "2026-04-28", "big": false}, {"n": "v5.8.0", "d": "2026-05-05", "big": false}, {"n": "v5.9.0", "d": "2026-05-20", "big": false}, {"n": "v5.10.0", "d": "2026-06-03", "big": false}, {"n": "v5.11.0", "d": "2026-06-10", "big": false}, {"n": "v5.12.0", "d": "2026-06-12", "big": false}, {"n": "v5.13.0", "d": "2026-07-03", "big": false}, {"n": "v5.14.0", "d": "2026-07-15", "big": false}, {"n": "v5.15.0", "d": "2026-08-10", "big": false}, {"n": "v5.16.0", "d": "2026-08-26", "big": false}];
const MODELS = [{"m": "rt_detr", "d": "2024-06-22", "pr": 29077, "by": "Sangbum Daniel Choi", "t": "New model support RTDETR"}, {"m": "instructblipvideo", "d": "2024-06-25", "pr": 30182, "by": "Raushan Turganbay", "t": "Add video modality for InstrucBLIP"}, {"m": "llava_next_video", "d": "2024-06-26", "pr": 31252, "by": "Raushan Turganbay", "t": "Add LLaVa NeXT Video"}, {"m": "gemma2", "d": "2024-06-27", "pr": 31659, "by": "Arthur", "t": "Add gemma 2"}, {"m": "zoedepth", "d": "2024-07-08", "pr": 30136, "by": "NielsRogge", "t": "Add ZoeDepth"}, {"m": "hiera", "d": "2024-07-12", "pr": 30356, "by": "Naman Garg", "t": "Adding hiera"}, {"m": "chameleon", "d": "2024-07-17", "pr": 31534, "by": "Raushan Turganbay", "t": "Chameleon: add model"}, {"m": "nemotron", "d": "2024-08-06", "pr": 31699, "by": "Ao Tang", "t": "Add Nemotron HF Support"}, {"m": "mamba2", "d": "2024-08-06", "pr": 32080, "by": "Pablo Montalvo", "t": "Add codestral mamba2"}, {"m": "qwen2_audio", "d": "2024-08-08", "pr": 32137, "by": "Yunfei Chu", "t": "Add Qwen2-Audio"}, {"m": "falcon_mamba", "d": "2024-08-12", "pr": 32615, "by": "Younes Belkada", "t": "Add new model"}, {"m": "dac", "d": "2024-08-19", "pr": 31494, "by": "Kamil Akesbi", "t": "Add Descript-Audio-Codec model"}, {"m": "qwen2_vl", "d": "2024-08-26", "pr": 32318, "by": "Shijie", "t": "support qwen2-vl"}, {"m": "granite", "d": "2024-08-27", "pr": 31502, "by": "Mayank Mishra", "t": "Granite language models"}, {"m": "olmoe", "d": "2024-09-03", "pr": 32406, "by": "Niklas Muennighoff", "t": "Add OLMoE"}, {"m": "llava_onevision", "d": "2024-09-05", "pr": 32673, "by": "Raushan Turganbay", "t": "Llava Onevision: add model"}, {"m": "pixtral", "d": "2024-09-14", "pr": 33449, "by": "Arthur", "t": "Add support for Pixtral"}, {"m": "mimi", "d": "2024-09-18", "pr": 33565, "by": "Yoach Lacombe", "t": "Codec integration"}, {"m": "granitemoe", "d": "2024-09-20", "pr": 33207, "by": "Mayank Mishra", "t": "Granitemoe"}, {"m": "omdet_turbo", "d": "2024-09-25", "pr": 31843, "by": "Yoni Gozlan", "t": "Add OmDet-Turbo"}, {"m": "mllama", "d": "2024-09-25", "pr": 33703, "by": "Arthur", "t": "Add MLLama"}, {"m": "idefics3", "d": "2024-09-25", "pr": 32473, "by": "Andrés Marafioti", "t": "Add Idefics 3!"}, {"m": "phimoe", "d": "2024-10-04", "pr": 33363, "by": "Amit Garg", "t": "PhiMoE"}, {"m": "zamba", "d": "2024-10-04", "pr": 30950, "by": "pglorio", "t": "Add Zamba"}, {"m": "myt5", "d": "2024-10-06", "pr": 31286, "by": "TomLim", "t": "[WIP] Add Tokenizer for MyT5 Model"}, {"m": "moshi", "d": "2024-10-16", "pr": 33624, "by": "Yoach Lacombe", "t": "Moshi integration"}, {"m": "glm", "d": "2024-10-18", "pr": 33823, "by": "Cyril Vallez", "t": "add Glm"}, {"m": "olmo_1124", "d": "2024-11-18", "pr": 34551, "by": "Shane A", "t": "Add OLMo November 2024"}, {"m": "ijepa", "d": "2024-12-05", "pr": 33125, "by": "João Marcelo", "t": "Add I-JEPA"}, {"m": "aria", "d": "2024-12-06", "pr": 34157, "by": "Aymeric Roucher", "t": "Add Aria"}, {"m": "timm_wrapper", "d": "2024-12-11", "pr": 34564, "by": "Pavel Iakubovskii", "t": "Add TimmWrapper"}, {"m": "cohere2", "d": "2024-12-13", "pr": 35224, "by": "alexrs-cohere", "t": "Add Cohere2 model"}, {"m": "colpali", "d": "2024-12-17", "pr": 33736, "by": "Tony Wu", "t": "Add ColPali to 🤗 transformers"}, {"m": "bamba", "d": "2024-12-19", "pr": 34982, "by": "Yu Chin Fabian Lim", "t": "Add the Bamba Model"}, {"m": "modernbert", "d": "2024-12-19", "pr": 35158, "by": "Benjamin Warner", "t": "Add ModernBERT to Transformers"}, {"m": "dinov2_with_registers", "d": "2024-12-24", "pr": 35348, "by": "NielsRogge", "t": "Add DINOv2 with registers"}, {"m": "diffllama", "d": "2025-01-07", "pr": 34083, "by": "松本和真", "t": "Add diffllama"}, {"m": "textnet", "d": "2025-01-08", "pr": 34979, "by": "Jade Choghari", "t": "Add TextNet"}, {"m": "vitpose", "d": "2025-01-08", "pr": 30530, "by": "NielsRogge", "t": "Add ViTPose"}, {"m": "vitpose_backbone", "d": "2025-01-08", "pr": 30530, "by": "NielsRogge", "t": "Add ViTPose"}, {"m": "moonshine", "d": "2025-01-10", "pr": 34784, "by": "eustlb", "t": "Add Moonshine"}, {"m": "emu3", "d": "2025-01-10", "pr": 33770, "by": "Raushan Turganbay", "t": "[WIP] Emu3: add model"}, {"m": "helium", "d": "2025-01-13", "pr": 35669, "by": "Arthur", "t": "Add-helium"}, {"m": "superglue", "d": "2025-01-20", "pr": 29886, "by": "StevenBucaille", "t": "Add SuperGlue model"}, {"m": "qwen2_5_vl", "d": "2025-01-23", "pr": 35569, "by": "ShuaiBai623", "t": "add qwen2.5vl"}, {"m": "zamba2", "d": "2025-01-27", "pr": 34517, "by": "pglorio", "t": "Add Zamba2"}, {"m": "got_ocr2", "d": "2025-01-31", "pr": 34721, "by": "Yoni Gozlan", "t": "Add GOT-OCR 2.0 to Transformers"}, {"m": "dab_detr", "d": "2025-02-04", "pr": 30803, "by": "David", "t": "Add DAB-DETR for object detection"}, {"m": "rt_detr_v2", "d": "2025-02-06", "pr": 34773, "by": "Jade Choghari", "t": "Adding RT-DETRv2 for object detection"}, {"m": "depth_pro", "d": "2025-02-10", "pr": 34583, "by": "Armaghan Shakir", "t": "Add Apple's Depth-Pro for depth estimation"}, {"m": "granitemoeshared", "d": "2025-02-14", "pr": 35894, "by": "Mayank Mishra", "t": "add shared experts for upcoming Granite 4.0 language models"}, {"m": "smolvlm", "d": "2025-02-20", "pr": 36126, "by": "Orr Zohar", "t": "SmolVLM2"}, {"m": "siglip2", "d": "2025-02-21", "pr": 36323, "by": "Pavel Iakubovskii", "t": "Add SigLIP 2"}, {"m": "aya_vision", "d": "2025-03-04", "pr": 36521, "by": "Arthur", "t": "Add aya"}, {"m": "gemma3", "d": "2025-03-12", "pr": 36658, "by": "Ryan Mullins", "t": "Gemma3"}, {"m": "mistral3", "d": "2025-03-18", "pr": 36790, "by": "Cyril Vallez", "t": "Add Mistral3"}, {"m": "shieldgemma2", "d": "2025-03-20", "pr": 36678, "by": "Ryan Mullins", "t": "Shieldgemma2"}, {"m": "prompt_depth_anything", "d": "2025-03-21", "pr": 35401, "by": "Haotong LIN", "t": "Add Prompt Depth Anything Model"}, {"m": "phi4_multimodal", "d": "2025-03-25", "pr": 36939, "by": "Cyril Vallez", "t": "Add Phi4 multimodal"}, {"m": "deepseek_v3", "d": "2025-03-28", "pr": 35926, "by": "Minho Ryu", "t": "[WIP] add deepseek-v3"}, {"m": "qwen3", "d": "2025-03-31", "pr": 36878, "by": "Bo Zheng", "t": "Adding Qwen3 and Qwen3MoE"}, {"m": "qwen3_moe", "d": "2025-03-31", "pr": 36878, "by": "Bo Zheng", "t": "Adding Qwen3 and Qwen3MoE"}, {"m": "llama4", "d": "2025-04-05", "pr": 37307, "by": "Arthur", "t": "Add llama4"}, {"m": "glm4", "d": "2025-04-09", "pr": 37388, "by": "Arthur", "t": "Add glm4"}, {"m": "granite_speech", "d": "2025-04-11", "pr": 36801, "by": "Alex Brooks", "t": "Add Granite Speech Support"}, {"m": "qwen2_5_omni", "d": "2025-04-14", "pr": 36752, "by": "BakerBunker", "t": "Add Qwen2.5-Omni"}, {"m": "mlcd", "d": "2025-04-15", "pr": 36182, "by": "Huajie Tan", "t": "Add MLCD model"}, {"m": "timesfm", "d": "2025-04-16", "pr": 34082, "by": "Jinan Zhou", "t": "Add TimesFM Time Series Forecasting Model"}, {"m": "janus", "d": "2025-04-17", "pr": 36053, "by": "Yaswanth Gali", "t": "Add Janus model"}, {"m": "internvl", "d": "2025-04-18", "pr": 35968, "by": "Yoni Gozlan", "t": "Add InternVL (2.5 MPO)"}, {"m": "bitnet", "d": "2025-04-28", "pr": 37742, "by": "Mohamed Mekkouri", "t": "Add Bitnet model"}, {"m": "sam_hq", "d": "2025-04-28", "pr": 35147, "by": "sushmanth reddy", "t": "Samhq model addition"}, {"m": "d_fine", "d": "2025-04-29", "pr": 36261, "by": "Vladislav Bronzov", "t": "Add D-FINE Model into Transformers"}, {"m": "hgnet_v2", "d": "2025-04-29", "pr": 36261, "by": "Vladislav Bronzov", "t": "Add D-FINE Model into Transformers"}, {"m": "granitemoehybrid", "d": "2025-05-05", "pr": 37658, "by": "Sukriti Sharma", "t": "Add GraniteMoeHybrid support for 4.0"}, {"m": "csm", "d": "2025-05-07", "pr": 36719, "by": "eustlb", "t": "Add CSM model"}, {"m": "falcon_h1", "d": "2025-05-21", "pr": 38249, "by": "Younes Belkada", "t": "[MODEL] Add Falcon H1"}, {"m": "colqwen2", "d": "2025-06-02", "pr": 35778, "by": "Tony Wu", "t": "Add ColQwen2 to 🤗 transformers"}, {"m": "minimax", "d": "2025-06-04", "pr": 35831, "by": "Armaghan Shakir", "t": "Add support for MiniMax's MiniMax-Text-01"}, {"m": "vjepa2", "d": "2025-06-11", "pr": 38746, "by": "Pavel Iakubovskii", "t": "Add V-JEPA 2"}, {"m": "lightglue", "d": "2025-06-17", "pr": 31718, "by": "StevenBucaille", "t": "Add LightGlue model"}, {"m": "arcee", "d": "2025-06-24", "pr": 38621, "by": "Crystalcareai", "t": "Add Arcee model support"}, {"m": "stt", "d": "2025-06-24", "pr": 38909, "by": "eustlb", "t": "Add kyutai stt"}, {"m": "glm4v", "d": "2025-06-25", "pr": 38431, "by": "Yuxuan Zhang", "t": "GLM-4.1V Model support"}, {"m": "t5gemma", "d": "2025-06-25", "pr": 38332, "by": "Biao Zhang", "t": "Encoder-Decoder Gemma"}, {"m": "dots1", "d": "2025-06-25", "pr": 38143, "by": "redmoe-moutain", "t": "[Model] add dots1"}, {"m": "smollm3", "d": "2025-06-25", "pr": 38755, "by": "Anton Lozhkov", "t": "Add SmolLM3"}, {"m": "dia", "d": "2025-06-26", "pr": 38405, "by": "Jaeyong Sung", "t": "Add Dia model"}, {"m": "gemma3n", "d": "2025-06-26", "pr": 39059, "by": "Ryan Mullins", "t": "Gemma 3n"}, {"m": "eomt", "d": "2025-06-27", "pr": 37610, "by": "Yaswanth Gali", "t": "✨ Add EoMT Model || 🚨 Fix Mask2Former loss calculation"}, {"m": "doge", "d": "2025-07-08", "pr": 35891, "by": "Jingze Shi", "t": "Add Doge model"}, {"m": "aimv2", "d": "2025-07-08", "pr": 36625, "by": "Yaswanth Gali", "t": "Add Aimv2 model"}, {"m": "deepseek_v2", "d": "2025-07-09", "pr": 36400, "by": "Vladislav Bronzov", "t": "Add DeepSeek V2 Model into Transformers"}, {"m": "lfm2", "d": "2025-07-10", "pr": 39340, "by": "Paul Pak", "t": "LFM2"}, {"m": "perception_lm", "d": "2025-07-11", "pr": 37878, "by": "Shuming Hu", "t": "PerceptionLM"}, {"m": "modernbert_decoder", "d": "2025-07-15", "pr": 38967, "by": "Orion Weller", "t": "Add ModernBERT Decoder Models - ModernBERT, but trained with CLM!"}, {"m": "voxtral", "d": "2025-07-18", "pr": 39429, "by": "eustlb", "t": "Add voxtral"}, {"m": "glm4_moe", "d": "2025-07-21", "pr": 39393, "by": "Yuxuan Zhang", "t": "GLM-4 Update"}, {"m": "ernie4_5", "d": "2025-07-21", "pr": 39228, "by": "Anton Vlasjuk", "t": "[`Ernie 4.5`] Add ernie text models"}, {"m": "ernie4_5_moe", "d": "2025-07-21", "pr": 39228, "by": "Anton Vlasjuk", "t": "[`Ernie 4.5`] Add ernie text models"}, {"m": "efficientloftr", "d": "2025-07-22", "pr": 36355, "by": "StevenBucaille", "t": "Add EfficientLoFTR model"}, {"m": "deepseek_vl", "d": "2025-07-25", "pr": 36248, "by": "Armaghan Shakir", "t": "Add support for DeepseekAI's DeepseekVL"}, {"m": "deepseek_vl_hybrid", "d": "2025-07-25", "pr": 36248, "by": "Armaghan Shakir", "t": "Add support for DeepseekAI's DeepseekVL"}, {"m": "xlstm", "d": "2025-07-25", "pr": 39665, "by": "Cyril Vallez", "t": "Add xlstm model"}, {"m": "evolla", "d": "2025-07-26", "pr": 36232, "by": "Xibin Bayes Zhou", "t": "Add evolla rebase main"}, {"m": "exaone4", "d": "2025-07-26", "pr": 39129, "by": "lgai-exaone", "t": "Add EXAONE 4.0 model"}, {"m": "cohere2_vision", "d": "2025-07-31", "pr": 39810, "by": "Raushan Turganbay", "t": "[Model] Cohere2 Vision"}, {"m": "mm_grounding_dino", "d": "2025-08-01", "pr": 37925, "by": "rziga", "t": "Add MM Grounding DINO"}, {"m": "gpt_oss", "d": "2025-08-05", "pr": 39923, "by": "Arthur", "t": "Add GPT OSS model from OpenAI"}, {"m": "glm4v_moe", "d": "2025-08-08", "pr": 39805, "by": "Yuxuan Zhang", "t": "GLM-4.5V Model Support"}, {"m": "sam2", "d": "2025-08-14", "pr": 32317, "by": "Sangbum Daniel Choi", "t": "Add Segment Anything 2 (SAM2)"}, {"m": "sam2_video", "d": "2025-08-14", "pr": 32317, "by": "Sangbum Daniel Choi", "t": "Add Segment Anything 2 (SAM2)"}, {"m": "dinov3_convnext", "d": "2025-08-14", "pr": 40167, "by": "Pavel Iakubovskii", "t": "Fix docs typo"}, {"m": "dinov3_vit", "d": "2025-08-14", "pr": 40167, "by": "Pavel Iakubovskii", "t": "Fix docs typo"}, {"m": "xcodec", "d": "2025-08-15", "pr": 38248, "by": "Manal ML", "t": "Add X-Codec model"}, {"m": "ovis2", "d": "2025-08-18", "pr": 37088, "by": "Eon Kim", "t": "Add Ovis2 model and processor implementation"}, {"m": "kosmos2_5", "d": "2025-08-19", "pr": 31711, "by": "tic-top", "t": "Add Kosmos-2.5"}, {"m": "metaclip_2", "d": "2025-08-20", "pr": 39826, "by": "NielsRogge", "t": "Add MetaCLIP 2"}, {"m": "florence2", "d": "2025-08-20", "pr": 38188, "by": "Duc-Viet Hoang", "t": "Add support for Florence-2"}, {"m": "hunyuan_v1_dense", "d": "2025-08-22", "pr": 39606, "by": "yjc9696", "t": "HunYuan opensource"}, {"m": "hunyuan_v1_moe", "d": "2025-08-22", "pr": 39606, "by": "yjc9696", "t": "HunYuan opensource"}, {"m": "seed_oss", "d": "2025-08-22", "pr": 40272, "by": "Fazzie", "t": "Addiing ByteDance Seed Seed-OSS"}, {"m": "apertus", "d": "2025-08-28", "pr": 39381, "by": "EduardDurech", "t": "Add Apertus"}, {"m": "qwen3_next", "d": "2025-09-10", "pr": 40771, "by": "Bo Zheng", "t": "Adding Support for Qwen3-Next"}, {"m": "ministral", "d": "2025-09-11", "pr": 40247, "by": "Manuel de Prada Corral", "t": "⚠️ 🔴 Add ministral model"}, {"m": "vaultgemma", "d": "2025-09-12", "pr": 40851, "by": "Ryan Mullins", "t": "add: differential privacy research model"}, {"m": "qwen3_vl", "d": "2025-09-15", "pr": 40795, "by": "JJJYmmm", "t": "Adding Support for Qwen3-VL Series"}, {"m": "qwen3_vl_moe", "d": "2025-09-15", "pr": 40795, "by": "JJJYmmm", "t": "Adding Support for Qwen3-VL Series"}, {"m": "olmo3", "d": "2025-09-16", "pr": 40778, "by": "Shane A", "t": "Add Olmo3 model"}, {"m": "longcat_flash", "d": "2025-09-17", "pr": 40730, "by": "Pablo Montalvo", "t": "Add LongCat-Flash"}, {"m": "flex_olmo", "d": "2025-09-18", "pr": 40921, "by": "Shane A", "t": "Add FlexOlmo model"}, {"m": "lfm2_vl", "d": "2025-09-18", "pr": 40624, "by": "Raushan Turganbay", "t": "Add new model LFM2-VL"}, {"m": "blt", "d": "2025-09-19", "pr": 38579, "by": "Ita Zaporozhets", "t": "blt wip"}, {"m": "qwen3_omni_moe", "d": "2025-09-21", "pr": 41025, "by": "BakerBunker", "t": "Adding support for Qwen3Omni"}, {"m": "parakeet", "d": "2025-09-25", "pr": 39062, "by": "Nithin Rao", "t": "Add Parakeet"}, {"m": "edgetam", "d": "2025-09-29", "pr": 39800, "by": "Yoni Gozlan", "t": "Add EdgeTAM"}, {"m": "edgetam_video", "d": "2025-09-29", "pr": 39800, "by": "Yoni Gozlan", "t": "Add EdgeTAM"}, {"m": "lfm2_moe", "d": "2025-10-07", "pr": 41401, "by": "Paul Pak", "t": "[Model] Lfm2Moe"}, {"m": "cwm", "d": "2025-10-09", "pr": 41199, "by": "Jacob Kahn", "t": "Add Code World Model (CWM)"}, {"m": "video_llama_3", "d": "2025-10-13", "pr": 40499, "by": "Kehan Li", "t": "[model] Add VideoLLaMA3 implementation"}, {"m": "audioflamingo3", "d": "2025-11-12", "pr": 40290, "by": "Lasha Koroshinadze", "t": "[models] Add AudioFlamingo3 integration"}, {"m": "glm46v", "d": "2025-11-15", "pr": 42122, "by": "Yuxuan Zhang", "t": "GLM-V update with new processor"}, {"m": "sam3", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "sam3_tracker", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "sam3_tracker_video", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "sam3_video", "d": "2025-11-19", "pr": 42285, "by": "Yoni Gozlan", "t": "Add SAM3 to 🤗 Transformers"}, {"m": "nanochat", "d": "2025-11-27", "pr": 41634, "by": "burtenshaw", "t": "[MODEL] Nanochat implementation"}, {"m": "afmoe", "d": "2025-11-29", "pr": 42168, "by": "Raghav Ravishankar", "t": "Add afmoe model"}, {"m": "t5gemma2", "d": "2025-12-01", "pr": 41834, "by": "Biao Zhang", "t": "T5gemma2"}, {"m": "ministral3", "d": "2025-12-01", "pr": 42498, "by": "Patrick von Platen", "t": "[Ministral 3] Add ministral 3"}, {"m": "fast_vlm", "d": "2025-12-02", "pr": 41112, "by": "Kamila Luczaj (Woo-chai)", "t": "Add FastVLM"}, {"m": "lasr", "d": "2025-12-05", "pr": 42648, "by": "eustlb", "t": "Lasr model"}, {"m": "paddleocr_vl", "d": "2025-12-11", "pr": 42178, "by": "zhang-prog", "t": "[Model] Add PaddleOCR-VL Model Support"}, {"m": "pe_audio", "d": "2025-12-16", "pr": 42905, "by": "eustlb", "t": " Sam: Perception Encoder Audiovisual"}, {"m": "pe_audio_video", "d": "2025-12-16", "pr": 42905, "by": "eustlb", "t": " Sam: Perception Encoder Audiovisual"}, {"m": "pe_video", "d": "2025-12-16", "pr": 42905, "by": "eustlb", "t": " Sam: Perception Encoder Audiovisual"}, {"m": "jais2", "d": "2025-12-17", "pr": 42684, "by": "sarathc-cerebras", "t": "adds jais2 model support"}, {"m": "pixio", "d": "2025-12-17", "pr": 42795, "by": "Lihe Yang", "t": "Add Pixio pre-trained models"}, {"m": "ernie4_5_vl_moe", "d": "2025-12-19", "pr": 39585, "by": "Anton Vlasjuk", "t": "[`Ernie 4.5`] Ernie VL models"}, {"m": "glmasr", "d": "2025-12-24", "pr": 42875, "by": "Yuxuan Zhang", "t": "GLM-ASR Support"}, {"m": "minimax_m2", "d": "2026-01-09", "pr": 42028, "by": "Roger Young", "t": "Add support for MiniMax-M2"}, {"m": "lw_detr", "d": "2026-01-12", "pr": 40991, "by": "StevenBucaille", "t": "Add LWDetr model"}, {"m": "glm4_moe_lite", "d": "2026-01-13", "pr": 43031, "by": "Yuxuan Zhang", "t": "[GLM-4.7] GLM-Lite Supoort"}, {"m": "glm_image", "d": "2026-01-13", "pr": 43100, "by": "Yuxuan Zhang", "t": "[GLM-Image] AR Model Support for GLM-Image"}, {"m": "lighton_ocr", "d": "2026-01-14", "pr": 41621, "by": "Aubertin Baptiste", "t": "Add LightOnOCR model implementation"}, {"m": "solar_open", "d": "2026-01-22", "pr": 43244, "by": "Inseo Song", "t": "Add Solar-Open Model"}, {"m": "glm_ocr", "d": "2026-01-27", "pr": 43391, "by": "Yuxuan Zhang", "t": "[GLM-OCR] GLM-OCR Support"}, {"m": "youtu", "d": "2026-01-28", "pr": 43166, "by": "Lu junru", "t": "Add Youtu-LLM model"}, {"m": "pp_doclayout_v3", "d": "2026-01-29", "pr": 43098, "by": "zhang-prog", "t": "[Model] Add PP-DocLayoutV3 Model Support"}, {"m": "eomt_dinov3", "d": "2026-02-02", "pr": 41212, "by": "NielsRogge", "t": "Add EoMT with DINOv3 backbone"}, {"m": "exaone_moe", "d": "2026-02-04", "pr": 43080, "by": "Junwon Hwang", "t": "Add EXAONE-MoE implementations"}, {"m": "moonshine_streaming", "d": "2026-02-04", "pr": 43702, "by": "eustlb", "t": "Add moonshine streaming"}, {"m": "vibevoice_acoustic_tokenizer", "d": "2026-02-06", "pr": 43400, "by": "Eric Bezzam", "t": "Add VibeVoice Acoustic Tokenizer"}, {"m": "qwen3_5", "d": "2026-02-09", "pr": 43830, "by": "Bo Zheng", "t": "Adding Support for Qwen3.5"}, {"m": "qwen3_5_moe", "d": "2026-02-09", "pr": 43830, "by": "Bo Zheng", "t": "Adding Support for Qwen3.5"}, {"m": "glm_moe_dsa", "d": "2026-02-09", "pr": 43858, "by": "Cyril Vallez", "t": "Add GlmMoeDsa"}, {"m": "voxtral_realtime", "d": "2026-02-16", "pr": 43769, "by": "eustlb", "t": "Add Voxtral Realtime"}, {"m": "higgs_audio_v2", "d": "2026-02-19", "pr": 40294, "by": "Shuai Zheng", "t": "Add Higgs Audio V2 Model"}, {"m": "higgs_audio_v2_tokenizer", "d": "2026-02-19", "pr": 40294, "by": "Shuai Zheng", "t": "Add Higgs Audio V2 Model"}, {"m": "colmodernvbert", "d": "2026-02-23", "pr": 42504, "by": "Paul Teiletche", "t": "Add ModernVBERT models"}, {"m": "modernvbert", "d": "2026-02-23", "pr": 42504, "by": "Paul Teiletche", "t": "Add ModernVBERT models"}, {"m": "olmo_hybrid", "d": "2026-02-26", "pr": 43358, "by": "Yanhong Li", "t": "Add OLMo Hybrid model"}, {"m": "pp_doclayout_v2", "d": "2026-02-27", "pr": 43018, "by": "zhang-prog", "t": "[Model] Add PP-DocLayoutV2 Model Support"}, {"m": "timesfm2_5", "d": "2026-02-27", "pr": 41763, "by": "Kashif Rasul", "t": "Timesfm 2.5"}, {"m": "vibevoice_asr", "d": "2026-03-02", "pr": 43625, "by": "Eric Bezzam", "t": "Add VibeVoice ASR"}, {"m": "nemotron_h", "d": "2026-03-03", "pr": 44390, "by": "Li Ding", "t": "add support for nemotron_3"}, {"m": "eurobert", "d": "2026-03-04", "pr": 39455, "by": "Arthur", "t": "Add eurobert"}, {"m": "chmv2", "d": "2026-03-11", "pr": 44595, "by": "Yoni Gozlan", "t": "Add CHMv2"}, {"m": "pp_ocrv5_server_det", "d": "2026-03-13", "pr": 43274, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_server_det Model Support"}, {"m": "pp_lcnet", "d": "2026-03-13", "pr": 43247, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_mobile_det Model Support"}, {"m": "pp_lcnet_v3", "d": "2026-03-13", "pr": 43247, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_mobile_det Model Support"}, {"m": "pp_ocrv5_mobile_det", "d": "2026-03-13", "pr": 43247, "by": "XingweiDeng", "t": "[Model] Add PP-OCRV5_mobile_det Model Support"}, {"m": "pi0", "d": "2026-03-16", "pr": 44160, "by": "Pablo Montalvo", "t": "Add model lerobot PI0 to transformers"}, {"m": "mistral4", "d": "2026-03-16", "pr": 44760, "by": "Julien Denize", "t": "Add Mistral 4"}, {"m": "jina_embeddings_v3", "d": "2026-03-19", "pr": 44251, "by": "Sai-Suraj-27", "t": "Add `Jina-Embeddings-V3` Model"}, {"m": "pp_ocrv5_mobile_rec", "d": "2026-03-19", "pr": 44808, "by": "zhang-prog", "t": "[Model] Add PP-OCRv5_server_rec and PP-OCRv5_mobile_rec models Support"}, {"m": "pp_ocrv5_server_rec", "d": "2026-03-19", "pr": 44808, "by": "zhang-prog", "t": "[Model] Add PP-OCRv5_server_rec and PP-OCRv5_mobile_rec models Support"}, {"m": "pp_chart2table", "d": "2026-03-20", "pr": 43767, "by": "XingweiDeng", "t": "[Model] Add PP-Chart2Table Model Support"}, {"m": "slanext", "d": "2026-03-21", "pr": 43707, "by": "Liu Jiaxuan", "t": "[Model] Add SLANeXt Model Support"}, {"m": "uvdoc", "d": "2026-03-21", "pr": 43385, "by": "XingweiDeng", "t": "[Model] Add UVDoc Model Support"}, {"m": "videomt", "d": "2026-03-25", "pr": 44285, "by": "NielsRogge", "t": "Add VidEoMT"}, {"m": "cohere_asr", "d": "2026-03-26", "pr": 45023, "by": "eustlb", "t": "Add cohere asr"}, {"m": "musicflamingo", "d": "2026-03-30", "pr": 43538, "by": "Lasha Koroshinadze", "t": "Add Music Flamingo"}, {"m": "nomic_bert", "d": "2026-04-02", "pr": 43067, "by": "Sonny Cooper", "t": "Internalise the NomicBERT model"}, {"m": "gemma4", "d": "2026-04-02", "pr": 45192, "by": "Ryan Mullins", "t": "casually dropping the most capable open weights on the planet"}, {"m": "sam3_lite_text", "d": "2026-04-13", "pr": 44320, "by": "NielsRogge", "t": "Add SAM3-LiteText"}, {"m": "qianfan_ocr", "d": "2026-04-17", "pr": 45280, "by": "bairongz", "t": "add Qianfan-OCR model definition"}, {"m": "slanet", "d": "2026-04-22", "pr": 45532, "by": "zhang-prog", "t": "[Model] Add SLANet Model Support"}, {"m": "hy_v3", "d": "2026-04-22", "pr": 45574, "by": "Anton Vlasjuk", "t": "Fix typos"}, {"m": "openai_privacy_filter", "d": "2026-04-22", "pr": 45580, "by": "Anton Vlasjuk", "t": "[`Privacy Filter`] Add model"}, {"m": "deimv2", "d": "2026-04-27", "pr": 44339, "by": "Harshal Janjani", "t": "model: Add DEIMv2 to Transformers"}, {"m": "laguna", "d": "2026-04-28", "pr": 45673, "by": "Joe Rowell", "t": "Laguna XS.2 implementation"}, {"m": "minicpmv4_6", "d": "2026-04-28", "pr": 45689, "by": "Anton Vlasjuk", "t": "fixing more typos"}, {"m": "granite_speech_plus", "d": "2026-04-29", "pr": 45695, "by": "Zvi Kons", "t": "Support for a new Granite-Speech-Plus model"}, {"m": "pp_formulanet", "d": "2026-04-30", "pr": 45626, "by": "zhang-prog", "t": "[Model] Add PP-FormulaNet Model Support"}, {"m": "deepseek_v4", "d": "2026-05-02", "pr": 45643, "by": "Arthur", "t": "Add DeepSeek V4"}, {"m": "exaone4_5", "d": "2026-05-04", "pr": 45471, "by": "Junwon Hwang", "t": "Add EXAONE 4.5 implementations"}, {"m": "gemma4_assistant", "d": "2026-05-05", "pr": 45788, "by": "Sindhu Raghuram Panyam", "t": "First model"}, {"m": "granite4_vision", "d": "2026-05-05", "pr": 45597, "by": "artem-spector", "t": "Add Granite 4.1 Vision (granite4_vision)"}, {"m": "rf_detr", "d": "2026-05-07", "pr": 36895, "by": "StevenBucaille", "t": "Add RF-DETR"}, {"m": "hyperclovax", "d": "2026-05-08", "pr": 44956, "by": "bigshanedogg", "t": "Add HyperCLOVAX SEED Think 14B"}, {"m": "hrm_text", "d": "2026-05-18", "pr": 46025, "by": "yifei wu", "t": "Add hrm text"}, {"m": "pp_lcnet_v4", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_medium_det", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_small_det", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_small_rec", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "pp_ocrv6_tiny_rec", "d": "2026-05-19", "pr": 45838, "by": "zhang-prog", "t": "[Model] Add PP-OCRv6 Models Support"}, {"m": "cohere2_moe", "d": "2026-05-20", "pr": 46115, "by": "Cyril Vallez", "t": "Add new cohere2_moe model"}, {"m": "glmga", "d": "2026-05-26", "pr": 46184, "by": "Yuxuan Zhang", "t": "[GLM-4.6V] Update with GLM-GA Processor"}, {"m": "mellum", "d": "2026-05-28", "pr": 46112, "by": "Madeesh Kannan", "t": "feat: Add support for JetBrains' `Mellum` v2 code generation model"}, {"m": "deepseek_ocr2", "d": "2026-06-02", "pr": 45075, "by": "Eon Kim", "t": "Add Deepseek-OCR-2 model"}, {"m": "sapiens2", "d": "2026-06-03", "pr": 45919, "by": "guarin", "t": "Add Sapiens2 Model"}, {"m": "gemma4_unified", "d": "2026-06-03", "pr": 46385, "by": "Douglas Reid", "t": "who needs encoders?"}, {"m": "gemma4_unified_assistant", "d": "2026-06-03", "pr": 46385, "by": "Douglas Reid", "t": "who needs encoders?"}, {"m": "cosmos3_omni", "d": "2026-06-05", "pr": 46146, "by": "MaciejBalaNV", "t": "Added cosmos3 model"}, {"m": "diffusion_gemma", "d": "2026-06-10", "pr": 46540, "by": "Joao Gante", "t": "GPU go brr"}, {"m": "deepseek_v32", "d": "2026-06-11", "pr": 41251, "by": "Arthur", "t": "Add deepseek 3.2 exp"}, {"m": "minimax_m3_vl", "d": "2026-06-12", "pr": 46600, "by": "Arthur", "t": "Add minimax m3vl"}, {"m": "nemotron_asr_streaming", "d": "2026-06-19", "pr": 46332, "by": "eustlb", "t": "Add Nemotron ASR Streaming"}, {"m": "videoprism", "d": "2026-06-20", "pr": 39895, "by": "MAHIR DAIYAN", "t": "Add Videoprism"}, {"m": "minicpm3", "d": "2026-06-23", "pr": 41116, "by": "Minho Ryu", "t": "Add MiniCPM3"}, {"m": "xcodec2", "d": "2026-06-25", "pr": 44178, "by": "Eric Bezzam", "t": "Add xcodec2 model"}, {"m": "qwen3_asr", "d": "2026-06-26", "pr": 43838, "by": "Muhammed", "t": "Qwen3 ASR and Forced Aligner"}, {"m": "nemotron3_5_asr", "d": "2026-06-27", "pr": 46565, "by": "eustlb", "t": "Add Nemotron 3.5 ASR Streaming"}, {"m": "radio", "d": "2026-07-01", "pr": 46425, "by": "meatybobby", "t": "Add support for RADIO models"}, {"m": "zaya", "d": "2026-07-01", "pr": 45862, "by": "JJJYmmm", "t": "[new model] Add Zyphra/ZAYA1-8B"}, {"m": "mimo_v2_flash", "d": "2026-07-01", "pr": 45144, "by": "casinca", "t": "Add Xiaomi MiMo-V2"}, {"m": "hunyuan_vl", "d": "2026-07-03", "pr": 46417, "by": "Mi Jiazhi", "t": "Add HunYuan VL model"}, {"m": "kimi_k25", "d": "2026-07-03", "pr": 45630, "by": "Raushan Turganbay", "t": "Add new model: Kimi2-6"}, {"m": "tipsv2", "d": "2026-07-07", "pr": 46347, "by": "Zixuan Zhu", "t": "Add TIPSv2"}, {"m": "tipsv2_dpt", "d": "2026-07-07", "pr": 46347, "by": "Zixuan Zhu", "t": "Add TIPSv2"}, {"m": "inkling", "d": "2026-07-15", "pr": 47347, "by": "Pablo Montalvo", "t": "Add Inkling model"}, {"m": "cosmos3_edge", "d": "2026-07-16", "pr": 47181, "by": "Atharva Joshi", "t": "Add Cosmos3 Edge model support"}, {"m": "axk1", "d": "2026-07-24", "pr": 46867, "by": "Minsang Kim", "t": "add_axk1"}, {"m": "axk2", "d": "2026-07-25", "pr": 47528, "by": "Anton Vlasjuk", "t": "Add AXK2 from SKT"}, {"m": "granite_swa", "d": "2026-07-29", "pr": 47179, "by": "Davis Wertheimer", "t": "Add Granite-swa and Granitemoe-swa model support"}, {"m": "granitemoe_swa", "d": "2026-07-29", "pr": 47179, "by": "Davis Wertheimer", "t": "Add Granite-swa and Granitemoe-swa model support"}, {"m": "muse_glimmer", "d": "2026-08-10", "pr": 47867, "by": "Cyril Vallez", "t": "Add MuseGlimmer"}, {"m": "muse_glimmer_assistant", "d": "2026-08-10", "pr": 47867, "by": "Cyril Vallez", "t": "Add MuseGlimmer"}, {"m": "cohere_compass", "d": "2026-08-10", "pr": 47878, "by": "calpt", "t": "Add CohereCompass modeling"}, {"m": "step3p7", "d": "2026-08-16", "pr": 46658, "by": "Ita Zaporozhets", "t": "[new model] step 3.7"}, {"m": "esmc", "d": "2026-08-20", "pr": 46419, "by": "Matt", "t": "Port ESMC and ESMFold2 to Transformers"}, {"m": "esmfold2", "d": "2026-08-20", "pr": 46419, "by": "Matt", "t": "Port ESMC and ESMFold2 to Transformers"}, {"m": "granite_speech5", "d": "2026-08-25", "pr": 48288, "by": "eustlb", "t": "gs"}, {"m": "qwen4_exp", "d": "2026-08-26", "pr": 48337, "by": "Cyril Vallez", "t": "Add Qwen4Exp model"}, {"m": "glm5_next", "d": "2026-08-26", "pr": 48342, "by": "Shijin Zhang", "t": "[Glm 5.3 Flash] GLM 5.3 Flash Support"}, {"m": "vibevoice", "d": "2026-08-27", "pr": 40546, "by": "Zhiliang Peng", "t": "Implement VibeVoice"}];
const MODEL_ROW = {"n": 267, "span": ["2024-06-22", "2026-08-27"], "owners": [["eustlb", 14], ["Arthur", 12], ["zhang-prog", 12], ["Yoni Gozlan", 10]], "authors": 126};
const TRACKS = [
{
"name": "Deprecation → v5",
"hue": 258,
"blurb": "Cut the library to one framework, then ship the major version.",
"items": [
{
"t": "Remove TensorFlow, Flax & JAX support",
"tag": "3 frameworks -> 1",
"s": "done",
"prs": [
{
"n": 38758,
"d": "2025-06-11",
"t": "Deprecate TF + JAX",
"st": "MERGED",
"by": "Rocketknight1",
"a": 27,
"r": 50030,
"f": 121,
"c": 3,
"b": "The time has finally come 🔫 🥃"
},
{
"n": 38760,
"d": "2025-06-11",
"t": "Prepare for TF+Jax deprecation",
"st": "MERGED",
"by": "Rocketknight1",
"a": 4,
"r": 37,
"f": 2,
"c": 2,
"b": "This PR removes a lot of TF + Jax testing in repo-consistency and .circleci. In a follow-up PR, I'll delete a lot of the test files, but this needs to be in a separate PR because repo-consistency checks run against the target branch, not the PR branch, and so they will fail if we remove other tests "
},
{
"n": 40372,
"d": "2025-08-25",
"t": "[docs] flax/jax purge",
"st": "MERGED",
"by": "gante",
"a": 16,
"r": 183,
"f": 7,
"c": 2,
"b": "This should be the last PR removing JAX/FLAX from our docs 🤗 After this PR, the only JAX/FLAX mentions in our docs are either historical (e.g. \"(written in JAX)\" as a note regarding the model origin) or useful for model addition information (e.g. tips to debug models originally written in jax). Meth"
},
{
"n": 40654,
"d": "2025-09-03",
"t": "Remove TF/Flax examples",
"st": "MERGED",
"by": "Rocketknight1",
"a": 1,
"r": 21961,
"f": 74,
"c": 3,
"b": "As the title says, the frameworks have been deprecated, so we should probably drop the examples as well! One thing to note: check_copies.py had a check_full_copies function that checked when files should be exact copies of each other. This was only used for TF/Flax examples. I've removed the FULL_CO"
},
{
"n": 40760,
"d": "2025-09-18",
"t": "🚨🚨🚨 Fully remove Tensorflow and Jax support library-wide",
"st": "MERGED",
"by": "Cyrilvallez",
"a": 3792,
"r": 181269,
"f": 854,
"c": 35,
"b": "Apart from obvious tf/jax support, I believe the following should be the only potential breaking changes to torch-only code: pipelines do not take framework argument anymore onnx config methods do not take framework argument anymore It may break current torch code if users do framework=\"pt\" explicit"
},
{
"n": 41058,
"d": "2025-09-23",
"t": "Remove mention of TensorFlow/Flax/JAX from English documentation",
"st": "MERGED",
"by": "cyyever",
"a": 3,
"r": 4,
"f": 2,
"c": 1,
"b": "There is a remaining one. Before submitting This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). Did you read the contributor guideline, Pull Request section? Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's th"
}
],
"d": "Deprecated then fully removed TF/Flax/JAX code paths, examples, and docs library-wide, collapsing transformers to a single PyTorch backend.",
"note": null,
"links": [],
"src": "speed.json",
"span": [
"2025-06-11",
"2025-09-23"