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
16

[serge] integration failure triage - 2026-10-06
huggingface/transformers#49379 · opened on Oct 6, 2026
Loading a tokenizer imports torch and scikit-learn when they are installed (about 2 s of a 3 s import)
huggingface/transformers#49378 · opened on Oct 6, 2026
1
set_attn_implementation() does not reach T5 encoder/decoder: model reports SDPA but still runs eager
huggingface/transformers#49372 · opened on Oct 6, 2026
3
Pixtral Flash Attention 2 uses axial RoPE coordinates as packed sequence boundaries (regression in 5.17.0)
huggingface/transformers#49371 · opened on Oct 6, 2026
MetalLinear returns wrong outputs for every batch row after the first (3D inputs)
huggingface/transformers#49337 · opened on Oct 6, 2026
5
Phi3 LongRoPE generation loses the prefix when clearing KV at the short/long crossing
huggingface/transformers#49334 · opened on Oct 6, 2026
2
[PERF] DeepSeek-V3.2 / GLM-5 lightning indexer overallocates memory
huggingface/transformers#49310 · opened on Oct 5, 2026
2
[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
1
WatermarkDetector scores depend on batch order when rows have mixed BOS prefixes
huggingface/transformers#49293 · opened on Oct 4, 2026
1
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
2
New model Aleph-Alpha/Kolibri-1
huggingface/transformers#49281 · opened on Oct 3, 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
Gemma 4 response_template parses reasoning into thinking, which its chat template does not read back
huggingface/transformers#49267 · opened on Oct 2, 2026
4
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
4
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
5
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
8
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

Repository navigation

Issues

Issues

[serge] integration failure triage - 2026-10-06

Loading a tokenizer imports torch and scikit-learn when they are installed (about 2 s of a 3 s import)

set_attn_implementation() does not reach T5 encoder/decoder: model reports SDPA but still runs eager

Pixtral Flash Attention 2 uses axial RoPE coordinates as packed sequence boundaries (regression in 5.17.0)

MetalLinear returns wrong outputs for every batch row after the first (3D inputs)

Phi3 LongRoPE generation loses the prefix when clearing KV at the short/long crossing

[PERF] DeepSeek-V3.2 / GLM-5 lightning indexer overallocates memory

[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

New model Aleph-Alpha/Kolibri-1

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

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

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

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