Repository navigation

Issues

New issue

Transformers' version 5 is out!
#43489 · LysandreJik opened on Jan 26, 2026
8
Tell Us: What Would Make Trainer Better?
#43595 · SunMarc opened on Jan 29, 2026
17

[Docs] Add Vietnamese PhoBERT fine-tuning example for low-resource NLP
huggingface/transformers#49967 · opened on Oct 25, 2026
[serge] integration failure triage - 2026-10-24
huggingface/transformers#49961 · opened on Oct 24, 2026
[serve] Streaming error events are not valid JSON when the error message contains quotes
huggingface/transformers#49954 · opened on Oct 24, 2026
1
[serve] max_completion_tokens is ignored on /v1/chat/completions
huggingface/transformers#49952 · opened on Oct 24, 2026
3
SAM3: Sam3ViTLayer.position_ids buffer is cast to bf16 on loading, degrading RoPE in global-attention layers (regression since v5.17.0, PR #48105)
huggingface/transformers#49938 · opened on Oct 23, 2026
2
GraniteSpeech: logits_scaling not applied since v5.10 (logits are 8x too large)
huggingface/transformers#49937 · opened on Oct 23, 2026
3
Speculative sampling accepts zero-probability candidates when the acceptance draw is zero
huggingface/transformers#49936 · opened on Oct 23, 2026
[serve] Add a Warm-Up Step after Loading the Model
huggingface/transformers#49927 · opened on Oct 23, 2026
Qwen3.5: split the Gated DeltaNet layers by heads under tensor parallelism (follow-up to #46846)
huggingface/transformers#49926 · opened on Oct 23, 2026
tp_plan="auto" adds backward hooks that break torch.compile graphs in inference
huggingface/transformers#49925 · opened on Oct 23, 2026
1
[Chat Parsing] Add python-call-list content parser and strip_prefix/strip_suffix text args
huggingface/transformers#49918 · opened on Oct 23, 2026
FPS sampling returns no frames for short, non-empty videos
huggingface/transformers#49907 · opened on Oct 23, 2026
GOT-OCR2 produces incorrect bounding-box coordinates for NumPy images
huggingface/transformers#49902 · opened on Oct 23, 2026
2
default_data_collator incorrectly casts NumPy integer labels to float32
huggingface/transformers#49897 · opened on Oct 23, 2026
1
Qwen4-Exp: eager QSA attention produces NaN logits for left-padded samples
huggingface/transformers#49888 · opened on Oct 23, 2026
1
Add Times Series Forecasting Models
huggingface/transformers#49884 · opened on Oct 22, 2026
Importing Trainer fails under marimo when IPython is not installed
huggingface/transformers#49881 · opened on Oct 22, 2026
1
A generation config loaded from the Hub can't be saved again
huggingface/transformers#49879 · opened on Oct 22, 2026
4
DiffusionGemma: generated text changes with batch size (MoE router projection and self-conditioning matmul)
huggingface/transformers#49876 · opened on Oct 22, 2026
3
FineGrainedFP8 dequantize: weights with a partial 128-block fail to load (DeepSeek-V3/R1) or dequantize silently wrong
huggingface/transformers#49873 · opened on Oct 22, 2026
1
RT-DETR / RT-DETRv2: main loss is computed on the denoising queries too, so real queries train as background
huggingface/transformers#49817 · opened on Oct 21, 2026
1
[CB] generate_batch with flash attention gets slower as the KV cache gets bigger, even when the cache is mostly empty
huggingface/transformers#49793 · opened on Oct 20, 2026
2
ProcessorMixin.apply_chat_template treats template kwargs unused by the template as processor kwargs
huggingface/transformers#49762 · opened on Oct 20, 2026
1
SmolVLMProcessor.apply_chat_template always warns "Kwargs passed to processor.__call__ have to be in processor_kwargs"
huggingface/transformers#49761 · opened on Oct 20, 2026
ProcessorMixin.apply_chat_template drops explicit processor_kwargs when an extra kwarg is passed
huggingface/transformers#49760 · opened on Oct 20, 2026
2

Navigation Menu

Repository navigation

Issues

Issues

[Docs] Add Vietnamese PhoBERT fine-tuning example for low-resource NLP

[serge] integration failure triage - 2026-10-24

[serve] Streaming error events are not valid JSON when the error message contains quotes

[serve] max_completion_tokens is ignored on /v1/chat/completions

SAM3: `Sam3ViTLayer.position_ids` buffer is cast to bf16 on loading, degrading RoPE in global-attention layers (regression since v5.17.0, PR #48105)

GraniteSpeech: `logits_scaling` not applied since v5.10 (logits are 8x too large)

Speculative sampling accepts zero-probability candidates when the acceptance draw is zero

[serve] Add a Warm-Up Step after Loading the Model

Qwen3.5: split the Gated DeltaNet layers by heads under tensor parallelism (follow-up to #46846)

tp_plan="auto" adds backward hooks that break torch.compile graphs in inference

[Chat Parsing] Add python-call-list content parser and strip_prefix/strip_suffix text args

FPS sampling returns no frames for short, non-empty videos

GOT-OCR2 produces incorrect bounding-box coordinates for NumPy images

default_data_collator incorrectly casts NumPy integer labels to float32

Qwen4-Exp: eager QSA attention produces NaN logits for left-padded samples

Add Times Series Forecasting Models

Importing Trainer fails under marimo when IPython is not installed

A generation config loaded from the Hub can't be saved again

DiffusionGemma: generated text changes with batch size (MoE router projection and self-conditioning matmul)

FineGrainedFP8 `dequantize`: weights with a partial 128-block fail to load (DeepSeek-V3/R1) or dequantize silently wrong

RT-DETR / RT-DETRv2: main loss is computed on the denoising queries too, so real queries train as background

[CB] `generate_batch` with flash attention gets slower as the KV cache gets bigger, even when the cache is mostly empty

ProcessorMixin.apply_chat_template treats template kwargs unused by the template as processor kwargs

SmolVLMProcessor.apply_chat_template always warns "Kwargs passed to processor.call have to be in processor_kwargs"

ProcessorMixin.apply_chat_template drops explicit processor_kwargs when an extra kwarg is passed