v4.45.2 is a patch release. Key changes include:

• Ignore keys on validate_rope #33753 by @zucchini-nlp;
• remove warning v2 #33761 by @itazap.

v4.45.1 is a patch release. Key changes include:

• [MllamaProcessor] Update errors and API with multiple image (#33715) by @ArthurZucker;
• Generate: can_generate() recursive check (#33718) by @gante.

v4.45.0 is a minor release. Key changes include:

• Add MLLama #33703, by @qubvel, @zucchini-nlp, @ArthurZucker;
• SoTA understanding of images of various resolution & ratio: Qwen2-VL achieves state-of-the-art performance on visual understanding benchmarks, including MathVi….

v4.44.2 is a patch release. Key changes include:

• Fix: Jamba cache fails to use torch.nn.module (#32894) Authored by @xgal;
• Fix: No need to dtype A in Jamba (#32924) @xgal.

v4.44.1 is a patch release. Key changes include:

• is_torchdynamo_compiling -- cast a wide exception net (#32476) by @gante;
• Revert "fixes to properly shard FSDP across cpu and meta for cpu_effcient_loading for prequantized 4bit (#32276)" (#32477) by @gante and @matthewdouglas.

v4.44.0 is a minor release. Key changes include:

• 3-5x faster torch.compile forward compilation for autoregressive decoder models #32227* by @fxmarty .;
• Offloaded KV Cache #31325* by @n17s : you just have to set cache_implementation="offloaded" when calling from_pretrained or using this:.

v4.43.4 is a patch release: resize embeds with deepspeed https://github.com/huggingface/transformers/pull/32214

v4.43.3 is a patch release. Key changes include:

• don't log base model architecture in wandb if log model is false #32143;
• [whisper] fix short-form output type #32178, by @sanchit-gandhi which fixes the short audio temperature fallback!.

v4.43.2 is a patch release. Key changes include:

• Fix float8_e4m3fn in modeling_utils;
• Fix resize embedding with Deepspeed.

v4.43.1 is a patch release: fix

v4.43.0 is a minor release. Key changes include:

• Chameleon: add model by @zucchini-nlp in #31534;
• Add ZoeDepth by @NielsRogge in #30136.

v4.42.4 is a patch release. Key changes include:

• [Gemma2] Support FA2 softcapping (#31887) by @ArthurZucker;
• [ConvertSlow] make sure the order is preserved for addedtokens (#31902) by @ArthurZucker.

v4.42.3 is a patch release: gemma capping is a must for big models

v4.42.2 is a patch release. Key changes include:

• Fix Gemma2 4d attention mask (#31674) by @hiyouga;
• don't zero out the attention_mask when using sliding window with flash attention (#31670) by @winglian.

v4.42.1 is a patch release: [hybridcache] fix get_seq_length method

v4.42.0 is a minor release. Key changes include:

• Add gemma 2 by @ArthurZucker in #31659;
• New model support RTDETR by @SangbumChoi in #29077.

v4.41.2 is a patch release. Key changes include:

• Do not trigger autoconversion if local_files_only #31004 from @Wauplin fixes this!;
• Paligemma: Fix devices and dtype assignments (#31008) by @molbap.

v4.41.1 is a patch release. Key changes include:

• https://github.com/huggingface/transformers/commit/a755745546779ae5c42510bc02a859bdac82b3b7 : PaliGemma - fix processor with no input text (https://github.com/…;
• https://github.com/huggingface/transformers/commit/a25f7d3c12975fe21eab437dda7363e9024de7c0 : Paligemma causal attention mask (https://github.com/huggingface/t….

v4.41.0 is a minor release. Key changes include:

• Phi-3 by @gugarosa in https://github.com/huggingface/transformers/pull/30423;
• Add JetMoE model by @yikangshen in https://github.com/huggingface/transformers/pull/30005.

v4.40.2 is a patch release. Key changes include:

• Fix for Neuron;
• Fix copies for DBRX - neuron fix.

v4.40.1 is a patch release: make eostokencriteria compatible with mps #30376

v4.40.0 is a minor release. Key changes include:

• Add Idefics2 by @amyeroberts in #30253;
• Add recurrent gemma by @ArthurZucker in #30143.

v4.39.3 is a patch release. Key changes include:

• Fix BC for AWQ quant #29965;
• generate fix breaking change for patch #29976.

v4.39.2 is a patch release. Key changes include:

• Safe import of LRScheduler #29919;
• [BC] Fix BC for other libraries #29934.

v4.39.1 is a patch release. Key changes include:

• Correct llava mask & fix missing setter for vocab_size #29389;
• [cleanup] vestiges of causal mask #29806.

v4.39.0 is a minor release. Key changes include:

• Strong accuracy on RAG and Tool Use;
• Low latency, and high throughput.

v4.38.2 is a patch release. Key changes include:

• FIX [Gemma] Fix bad rebase with transformers main;
• Improve _update_causal_mask performance.

v4.38.1 is a patch release. Key changes include:

• [Gemma] Fix eager attention #29187 by @sanchit-gandhi;
• attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).

v4.38.0 is a minor release. Key changes include:

• Flash Attention 2;
• bitsandbytes-4bit.

v4.37.2 is a patch release. Key changes include:

• Protecting the imports for SigLIP's tokenizer if sentencepiece isn't installed;
• Fix permissions issue on windows machines when using trainer in multi-node setup.

v4.45.2 is a patch release. Key changes include:

• Ignore keys on validate_rope #33753 by @zucchini-nlp;
• remove warning v2 #33761 by @itazap.

v4.45.1 is a patch release. Key changes include:

• [MllamaProcessor] Update errors and API with multiple image (#33715) by @ArthurZucker;
• Generate: can_generate() recursive check (#33718) by @gante.

v4.45.0 is a minor release. Key changes include:

• Add MLLama #33703, by @qubvel, @zucchini-nlp, @ArthurZucker;
• SoTA understanding of images of various resolution & ratio: Qwen2-VL achieves state-of-the-art performance on visual understanding benchmarks, including MathVi….

v4.44.2 is a patch release. Key changes include:

• Fix: Jamba cache fails to use torch.nn.module (#32894) Authored by @xgal;
• Fix: No need to dtype A in Jamba (#32924) @xgal.

v4.44.1 is a patch release. Key changes include:

• is_torchdynamo_compiling -- cast a wide exception net (#32476) by @gante;
• Revert "fixes to properly shard FSDP across cpu and meta for cpu_effcient_loading for prequantized 4bit (#32276)" (#32477) by @gante and @matthewdouglas.

v4.44.0 is a minor release. Key changes include:

• 3-5x faster torch.compile forward compilation for autoregressive decoder models #32227* by @fxmarty .;
• Offloaded KV Cache #31325* by @n17s : you just have to set cache_implementation="offloaded" when calling from_pretrained or using this:.

v4.43.4 is a patch release: resize embeds with deepspeed https://github.com/huggingface/transformers/pull/32214

v4.43.3 is a patch release. Key changes include:

• don't log base model architecture in wandb if log model is false #32143;
• [whisper] fix short-form output type #32178, by @sanchit-gandhi which fixes the short audio temperature fallback!.

v4.43.2 is a patch release. Key changes include:

• Fix float8_e4m3fn in modeling_utils;
• Fix resize embedding with Deepspeed.

v4.43.1 is a patch release: fix

v4.43.0 is a minor release. Key changes include:

• Chameleon: add model by @zucchini-nlp in #31534;
• Add ZoeDepth by @NielsRogge in #30136.

v4.42.4 is a patch release. Key changes include:

• [Gemma2] Support FA2 softcapping (#31887) by @ArthurZucker;
• [ConvertSlow] make sure the order is preserved for addedtokens (#31902) by @ArthurZucker.

v4.42.3 is a patch release: gemma capping is a must for big models

v4.42.2 is a patch release. Key changes include:

• Fix Gemma2 4d attention mask (#31674) by @hiyouga;
• don't zero out the attention_mask when using sliding window with flash attention (#31670) by @winglian.

v4.42.1 is a patch release: [hybridcache] fix get_seq_length method

v4.42.0 is a minor release. Key changes include:

• Add gemma 2 by @ArthurZucker in #31659;
• New model support RTDETR by @SangbumChoi in #29077.

v4.41.2 is a patch release. Key changes include:

• Do not trigger autoconversion if local_files_only #31004 from @Wauplin fixes this!;
• Paligemma: Fix devices and dtype assignments (#31008) by @molbap.

v4.41.1 is a patch release. Key changes include:

• https://github.com/huggingface/transformers/commit/a755745546779ae5c42510bc02a859bdac82b3b7 : PaliGemma - fix processor with no input text (https://github.com/…;
• https://github.com/huggingface/transformers/commit/a25f7d3c12975fe21eab437dda7363e9024de7c0 : Paligemma causal attention mask (https://github.com/huggingface/t….

v4.41.0 is a minor release. Key changes include:

• Phi-3 by @gugarosa in https://github.com/huggingface/transformers/pull/30423;
• Add JetMoE model by @yikangshen in https://github.com/huggingface/transformers/pull/30005.

v4.40.2 is a patch release. Key changes include:

• Fix for Neuron;
• Fix copies for DBRX - neuron fix.

v4.40.1 is a patch release: make eostokencriteria compatible with mps #30376

v4.40.0 is a minor release. Key changes include:

• Add Idefics2 by @amyeroberts in #30253;
• Add recurrent gemma by @ArthurZucker in #30143.

v4.39.3 is a patch release. Key changes include:

• Fix BC for AWQ quant #29965;
• generate fix breaking change for patch #29976.

v4.39.2 is a patch release. Key changes include:

• Safe import of LRScheduler #29919;
• [BC] Fix BC for other libraries #29934.

v4.39.1 is a patch release. Key changes include:

• Correct llava mask & fix missing setter for vocab_size #29389;
• [cleanup] vestiges of causal mask #29806.

v4.39.0 is a minor release. Key changes include:

• Strong accuracy on RAG and Tool Use;
• Low latency, and high throughput.

v4.38.2 is a patch release. Key changes include:

• FIX [Gemma] Fix bad rebase with transformers main;
• Improve _update_causal_mask performance.

v4.38.1 is a patch release. Key changes include:

• [Gemma] Fix eager attention #29187 by @sanchit-gandhi;
• attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).

v4.38.0 is a minor release. Key changes include:

• Flash Attention 2;
• bitsandbytes-4bit.

v4.37.2 is a patch release. Key changes include:

• Protecting the imports for SigLIP's tokenizer if sentencepiece isn't installed;
• Fix permissions issue on windows machines when using trainer in multi-node setup.