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
16

WatermarkDetector counts repeated token windows when deduplication is enabled
huggingface/transformers#49299 · opened on Oct 4, 2026
[FSDP2] cpu_ram_efficient_loading materializes full model on CPU for every non-rank-0, defeating rank-0 broadcast (host OOM)
huggingface/transformers#49297 · opened on Oct 4, 2026
WatermarkDetector scores depend on batch order when rows have mixed BOS prefixes
huggingface/transformers#49293 · opened on Oct 4, 2026
MiniMax lightning attention: decay recurrence runs in bf16, so slow-decay heads never decay and decode drifts with length
huggingface/transformers#49289 · opened on Oct 4, 2026
1
model.to(bf16)/half() and Trainer bf16_full_eval cast RoPE inv_freq buffers, unlike from_pretrained(dtype=...): large angle error at long context
huggingface/transformers#49288 · opened on Oct 4, 2026
1
[serge] integration failure triage - 2026-10-03
huggingface/transformers#49286 · opened on Oct 3, 2026
New model Aleph-Alpha/Kolibri-1
huggingface/transformers#49281 · opened on Oct 3, 2026
Multi-image generation reorders encoded image features during beam and sampled batch expansion
huggingface/transformers#49274 · opened on Oct 2, 2026
pipeline(..., model_kwargs={"code_revision": ...}) raises TypeError, and top-level code_revision does not reach the model load
huggingface/transformers#49272 · opened on Oct 2, 2026
1
Add MLX KV-cache compression integration with VeloxQuant-MLX
huggingface/transformers#49268 · opened on Oct 2, 2026
Gemma 4 response_template parses reasoning into thinking, which its chat template does not read back
huggingface/transformers#49267 · opened on Oct 2, 2026
1
generate() with a static cache: host-side inputs miss tensors nested in dicts, and TP undersizes the early cache
huggingface/transformers#49261 · opened on Oct 2, 2026
ImageTextToTextPipeline warns about generation arguments it injects itself (generation_config + return_dict_in_generate, and max_new_tokens vs max_length)
huggingface/transformers#49253 · opened on Oct 2, 2026
2
AutoTokenizer with tokenizer_class=LlamaTokenizerFast silently builds a sentencepiece pipeline for a byte-level BPE tokenizer.json, dropping all CJK input
huggingface/transformers#49252 · opened on Oct 2, 2026
4
ExponentialDecayLengthPenalty creates NaN EOS scores when combined with minimum-length constraints
huggingface/transformers#49238 · opened on Oct 1, 2026
4
Trainer counts fp16 steps skipped by GradScaler as completed update steps (global_step, max_steps, on_optimizer_step)
huggingface/transformers#49237 · opened on Oct 1, 2026
7
Encoder repetition penalty uses incorrect source rows after beam or sampled-return expansion
huggingface/transformers#49233 · opened on Oct 1, 2026
1
CSM codebook embedding tying is ignored when tie_word_embeddings=False
huggingface/transformers#49196 · opened on Sep 30, 2026
SAM3 Video reconditioning causes mask to grow onto the edges and make the bounding boxes balloon
huggingface/transformers#49195 · opened on Sep 30, 2026
1
[eager MOE] Selected expert IDs should probably not be treated as 0-d GPU tensors
huggingface/transformers#49190 · opened on Sep 29, 2026
8
[BLT, Byte-Latent Transformer] Entropy patcher is missing BLT's 512-token sliding-window attention — patch rate collapses on inputs longer than 512 bytes
huggingface/transformers#49185 · opened on Sep 29, 2026
3
tiny-aya models are loaded as CohereTokenizer in v5, and digits are split differently from tokenizer.json
huggingface/transformers#49176 · opened on Sep 29, 2026
Rename gradient_checkpointing to activation_checkpointing in the next major
huggingface/transformers#49166 · opened on Sep 28, 2026
3
SigLIP2's tokenizer performs no case-normalization, unlike SigLIP v1's SiglipTokenizer
huggingface/transformers#49139 · opened on Sep 27, 2026
10
set_attn_implementation fails to propagate attention backend to nested sub-configs
huggingface/transformers#49137 · opened on Sep 27, 2026

Navigation Menu

Issues

Issues

WatermarkDetector counts repeated token windows when deduplication is enabled

[FSDP2] cpu_ram_efficient_loading materializes full model on CPU for every non-rank-0, defeating rank-0 broadcast (host OOM)

WatermarkDetector scores depend on batch order when rows have mixed BOS prefixes

MiniMax lightning attention: decay recurrence runs in bf16, so slow-decay heads never decay and decode drifts with length

model.to(bf16)/half() and Trainer bf16_full_eval cast RoPE inv_freq buffers, unlike from_pretrained(dtype=...): large angle error at long context

[serge] integration failure triage - 2026-10-03

New model Aleph-Alpha/Kolibri-1

Multi-image generation reorders encoded image features during beam and sampled batch expansion

pipeline(..., model_kwargs={"code_revision": ...}) raises TypeError, and top-level code_revision does not reach the model load

Add MLX KV-cache compression integration with VeloxQuant-MLX

Gemma 4 `response_template` parses reasoning into `thinking`, which its chat template does not read back

`generate()` with a static cache: host-side inputs miss tensors nested in dicts, and TP undersizes the early cache

ImageTextToTextPipeline warns about generation arguments it injects itself (generation_config + return_dict_in_generate, and max_new_tokens vs max_length)

AutoTokenizer with tokenizer_class=LlamaTokenizerFast silently builds a sentencepiece pipeline for a byte-level BPE tokenizer.json, dropping all CJK input

ExponentialDecayLengthPenalty creates NaN EOS scores when combined with minimum-length constraints

Trainer counts fp16 steps skipped by GradScaler as completed update steps (`global_step`, `max_steps`, `on_optimizer_step`)

Encoder repetition penalty uses incorrect source rows after beam or sampled-return expansion

CSM codebook embedding tying is ignored when `tie_word_embeddings=False`

SAM3 Video reconditioning causes mask to grow onto the edges and make the bounding boxes balloon

[eager MOE] Selected expert IDs should probably not be treated as 0-d GPU tensors

[BLT, Byte-Latent Transformer] Entropy patcher is missing BLT's 512-token sliding-window attention — patch rate collapses on inputs longer than 512 bytes

tiny-aya models are loaded as CohereTokenizer in v5, and digits are split differently from tokenizer.json

Rename `gradient_checkpointing` to `activation_checkpointing` in the next major

SigLIP2's tokenizer performs no case-normalization, unlike SigLIP v1's SiglipTokenizer

set_attn_implementation fails to propagate attention backend to nested sub-configs